\documentclass{report}
\usepackage[section]{placeins}
\usepackage[top=1in,left=1.5in,right=1in,bottom=1in]{geometry}
\usepackage{siunitx}
\usepackage{multicol}
\setlength{\columnsep}{1cm}
\usepackage[acronym]{glossaries}
\usepackage[T1]{fontenc}
\usepackage{enumitem}
\usepackage{titlesec}
\usepackage{titlecaps}
\usepackage{upgreek}
\usepackage{graphicx}
\usepackage{subcaption}
\usepackage{nth}
\usepackage{hyperref} % must be before cleveref
\usepackage[capitalize]{cleveref}
\usepackage[version=4]{mhchem}
\usepackage{pgfgantt}
\usepackage{setspace}
\usepackage{listings}
\usepackage{tocloft}
\usepackage{epigraph} 

\hypersetup{
  colorlinks=true,
  linkcolor=black,
  filecolor=black,
  citecolor=black,      
  urlcolor=black,
}

\newcommand{\dmspaper}{Dwarshuis et al. Functionalized microcarriers improve T
  cell manufacturing by facilitating migratory memory T cell production and
  increasing CD4/CD8 ratio.~2019.~biorxiv.~https://doi.org/10.1101/646760}

\newcommand{\modelpaper}{Odeh-Couvertier et al. Predicting T Cell Quality During
  Manufacturing Through an Artificial Intelligence-based Integrative Multi-Omics
  Analytical Platform.~2019.~biorxiv.~https://doi.org/10.1101/2021.05.05.442854}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% my attempt to make MATLAB code look pretty

\definecolor{dkgreen}{rgb}{0,0.6,0}
\definecolor{gray}{rgb}{0.5,0.5,0.5}
\definecolor{mauve}{rgb}{0.58,0,0.82}

\lstset{frame=tb,
  language=Matlab,
  aboveskip=3mm,
  belowskip=3mm,
  showstringspaces=false,
  columns=flexible,
  basicstyle={\small\ttfamily},
  numbers=none,
  numberstyle=\tiny\color{gray},
  keywordstyle=\color{blue},
  commentstyle=\color{dkgreen},
  stringstyle=\color{mauve},
  breaklines=true,
  breakatwhitespace=true,
  tabsize=3
}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% benevolently force figures stay in their own subsection
%
% NOTE the placeins package only has a 'section' option which puts
% floatbarriers after every \section call; this does the same for \subsection

\makeatletter
\AtBeginDocument{%
  \expandafter\renewcommand\expandafter\subsection\expandafter
    {\expandafter\@fb@secFB\subsection}%
  \newcommand\@fb@subsecFB{\FloatBarrier
    \gdef\@fb@afterHHook{\@fb@topbarrier \gdef\@fb@afterHHook{}}}%
  \g@addto@macro\@afterheading{\@fb@afterHHook}%
  \gdef\@fb@afterHHook{}%
}
\makeatother

% ...also center them
\makeatletter
\g@addto@macro\@floatboxreset\centering
\makeatother

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% header configuration
%
% NOTE glossary can't apparently be used in section header (even thought it
% would be nice)

\doublespacing{}

\titleformat{\chapter}[block]{\filcenter\bfseries\Large}
{\MakeUppercase{\chaptertitlename} \thechapter: }{0pt}{\uppercase}
\titleformat{\section}[block]{\bfseries\large}{}{0pt}{\titlecap}
\titleformat{\subsection}[block]{\itshape\large}{}{0pt}{\titlecap}
\titleformat{\subsubsection}[runin]{\bfseries}{}{0pt}{\titlecap}

\setlist[description]{font=$\bullet$~\textbf\normalfont}

\renewcommand*{\contentsname}{TABLE OF CONTENTS}
\renewcommand{\listfigurename}{LIST OF FIGURES}
\renewcommand{\listtablename}{LIST OF TABLES}

\renewcommand{\cfttoctitlefont}{\hspace*{\fill}\Large\bfseries}
\renewcommand{\cftaftertoctitle}{\hspace*{\fill}}
\renewcommand{\cftlottitlefont}{\hspace*{\fill}\Large\bfseries}
\renewcommand{\cftafterlottitle}{\hspace*{\fill}}
\renewcommand{\cftloftitlefont}{\hspace*{\fill}\Large\bfseries}
\renewcommand{\cftafterloftitle}{\hspace*{\fill}}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% acronyms for the lazy
%
% adding as many as possible has the added benefit of making the thesis longer
% and making me sound more sophisticated

% the many flavors of T cells
\newcommand{\tcellacronym}[4]{
  \newacronym[shortplural={T\textsubscript{#2}#4
    cells}]{#1}{T\textsubscript{#2}#4}{#3 T cell}
}

\renewcommand{\glossarysection}[2][]{} % remove glossary title
\makeglossaries

\tcellacronym{tn}{n}{naive}{}
\tcellacronym{tcm}{cm}{central memory}{}
\tcellacronym{tscm}{scm}{stem-memory}{}
\tcellacronym{tem}{em}{effector-memory}{}
\tcellacronym{teff}{eff}{effector}{}
\tcellacronym{treg}{reg}{regulatory}{}
\tcellacronym{th}{h}{helper}{}
\tcellacronym{tc}{c}{cytotoxic}{}
\tcellacronym{th1}{h}{type 1 helper}{1}
\tcellacronym{th2}{h}{type 2 helper}{2}
\tcellacronym{th17}{h}{IL-17 helper}{17}

\newacronym{bcaa}{BCAA}{branched-chain amino acid}
\newacronym{til}{TIL}{tumor infiltrating lymphocyte}
\newacronym{tcr}{TCR}{T cell receptor}
\newacronym{act}{ACT}{adoptive cell therapies}
\newacronym{qc}{QC}{quality control}
\newacronym{car}{CAR}{chimeric antigen receptor}
\newacronym[longplural={monoclonal antibodies}]{mab}{mAb}{monoclonal antibody}
\newacronym{ecm}{ECM}{extracellular matrix}
\newacronym{cqa}{CQA}{critical quality attribute}
\newacronym{cpp}{CPP}{critical process parameter}
\newacronym{dms}{DMS}{degradable microscaffold}
\newacronym{doe}{DOE}{design of experiments}
\newacronym{adoe}{ADOE}{adaptive design of experiments}
\newacronym{gmp}{GMP}{Good Manufacturing Practices}
\newacronym{cho}{CHO}{Chinese hamster ovary}
\newacronym{all}{ALL}{acute lymphoblastic leukemia}
\newacronym{cll}{CLL}{chronic lymphoblastic leukemia}
\newacronym{pdms}{PDMS}{polydimethylsiloxane}
\newacronym{dc}{DC}{dendritic cell}
\newacronym{il}{IL}{interleukin}
\newacronym{il2}{IL2}{interleukin 2}
\newacronym{il15}{IL15}{interleukin 15}
\newacronym{il15r}{IL15R}{interleukin 15 receptor}
\newacronym{rhil2}{rhIL2}{recombinant human interleukin 2}
\newacronym{apc}{APC}{antigen presenting cell}
\newacronym{mhc}{MHC}{major histocompatibility complex}
\newacronym{elisa}{ELISA}{enzyme-linked immunosorbent assay}
\newacronym{nmr}{NMR}{nuclear magnetic resonance}
\newacronym{haba}{HABA}{4-hydroxyazobenene-2-carboxylic-acid}
\newacronym{pbs}{PBS}{phosphate buffered saline}
\newacronym{bca}{BCA}{bicinchoninic acid assay}
\newacronym{bsa}{BSA}{bovine serum albumin}
\newacronym{hsa}{HSA}{human serum albumin}
\newacronym{stp}{STP}{streptavidin}
\newacronym{stppe}{STP-PE}{streptavidin-phycoerythrin}
\newacronym{snb}{SNB}{sulfo-nhs-biotin}
\newacronym{cug}{CuG}{Cultispher G}
\newacronym{cus}{CuS}{Cultispher S}
\newacronym{pbmc}{PBMC}{peripheral blood mononuclear cells}
\newacronym{macs}{MACS}{magnetic activated cell sorting}
\newacronym{aopi}{AO/PI}{acridine orange/propidium iodide}
\newacronym{igg}{IgG}{immunoglobulin G}
\newacronym{pe}{PE}{phycoerythrin}
\newacronym{fitc}{FITC}{Fluorescein}
\newacronym{fitcbt}{FITC-BT}{Fluorescein-biotin}
\newacronym{ptnl}{PTN-L}{Protein L}
\newacronym{af647}{AF647}{Alexa Fluor 647}
\newacronym{anova}{ANOVA}{analysis of variance}
\newacronym{crispr}{CRISPR}{clustered regularly interspaced short
  palindromic repeats}
\newacronym{mtt}{MTT}{3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide}
\newacronym{bmi}{BMI}{body mass index}
\newacronym{a2b1}{A2B1}{integrin $\upalpha$1$\upbeta$1}
\newacronym{a2b2}{A2B2}{integrin $\upalpha$1$\upbeta$2}
\newacronym{nsg}{NSG}{NOD scid gamma}
\newacronym{colb}{COL-B}{collagenase B}
\newacronym{cold}{COL-D}{collagenase D}
\newacronym{tsne}{tSNE}{t-stochastic neighbor embedding}
\newacronym{umap}{UMAP}{uniform manifold approximation and projection}
\newacronym{anv}{AXV}{Annexin-V}
\newacronym{pi}{PI}{propidium iodide}
\newacronym{rt}{RT}{room temperature}
\newacronym{cas37}{Cas3/7}{Caspase-3/7}
\newacronym{bcl2}{BCL-2}{B cell lymphoma 2}
\newacronym{tmb}{TMB}{3,3',5,5'-Tetramethylbenzidine}
\newacronym{gvhd}{GVHD}{graft-vs-host disease}
\newacronym{bcma}{BCMA}{B-cell maturation antigen}
\newacronym{di}{DI}{deionized}
\newacronym{moi}{MOI}{multiplicity of infection}
\newacronym{ifng}{IFN$\upgamma$}{interferon-$\upgamma$}
\newacronym{tnfa}{TNF$\upalpha$}{tumor necrosis factor-$\upalpha$}
\newacronym{sql}{SQL}{structured query language}
\newacronym{fcs}{FCS}{flow cytometry standard}
\newacronym{ivis}{IVIS}{in vivo imaging system}
\newacronym{iacuc}{IACUC}{institutional animal care and use committee}
\newacronym{hbss}{HBSS}{Hank's buffered saline solution}
\newacronym{leaf}{LEAF}{low endotoxin, azide-free}
\newacronym{cytof}{CyTOF}{cytometry time-of-flight}
\newacronym{spade}{SPADE}{spanning-tree progression analysis of
  density-normalized events}
\newacronym{ml}{ML}{machine learning}
\newacronym{rf}{RF}{random forest}
\newacronym{sr}{SR}{symbolic regression}
\newacronym{gbm}{GBM}{gradient boosted trees}
\newacronym{cif}{CIF}{conditional inference forests}
\newacronym{lasso}{LASSO}{least absolute shrinkage and selection operator}
\newacronym{svm}{SVM}{support vector machines}
\newacronym{plsr}{PLSR}{partial least squares regression}
\newacronym{mse}{MSE}{mean squared error}
\newacronym{loocv}{LOO-CV}{leave-one-out cross validation}
\newacronym{hsqc}{HSQC}{heteronuclear single quantum coherence}
\newacronym{hla}{HLA}{human leukocyte antigen}
\newacronym{zfn}{ZFN}{zinc-finger nuclease}
\newacronym{talen}{TALEN}{transcription activator-like effector nuclease}
\newacronym{qbd}{QbD}{quality-by-design}
\newacronym{aws}{AWS}{Amazon Web Services}
\newacronym{qpcr}{qPCR}{quantitative polymerase chain reaction}
\newacronym{cstr}{CSTR}{continuously stirred tank bioreactor}
\newacronym{esc}{ESC}{embryonic stem cell}
\newacronym{msc}{MSC}{mesenchymal stromal cells}
\newacronym{scfv}{scFv}{single-chain fragment variable}
\newacronym{hepes}{HEPES}{4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid}
\newacronym{nhs}{NHS}{N-hydroxysulfosuccinimide}
\newacronym{tocsy}{TOCSY}{total correlation spectroscopy}
\newacronym{hplc}{HPLC}{high-performance liquid chromatography}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% SI units for uber nerds

% NOTE the \SI macro is depreciated but the arch repo (!!!) hasn't been updated
% with the latest package yet (texlive-science)

\sisetup{per-mode=symbol,list-units=single}
\DeclareSIUnit\IU{IU}
\DeclareSIUnit\rpm{RPM}
\DeclareSIUnit\carrier{carrier}
\DeclareSIUnit\dms{DMS}
\DeclareSIUnit\cell{cells}
\DeclareSIUnit\ab{mAb}
\DeclareSIUnit\normal{N}
\DeclareSIUnit\molar{M}
\DeclareSIUnit\mM{\milli\molar}
\DeclareSIUnit\uM{\micro\molar}
\DeclareSIUnit\gforce{\times{} g}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% commands for lazy farts like me

% gatech format conformity
\newcommand{\mytitle}{
  \Large{
    \textbf{
      Optimizing T Cell Manufacturing and Quality Using Functionalized
      Degradable Microscaffolds
    }
  }
}

\newcommand{\mycommitteemember}[3]{
  \begin{flushleft}
    \noindent
    #1 \\
    #2 \\
    \textit{#3}
  \end{flushleft}
}

% a BME's best friend
\newcommand{\invivo}{\textit{in vivo}}
\newcommand{\invitro}{\textit{in vitro}}
\newcommand{\exvivo}{\textit{ex vivo}}
\newcommand{\Invivo}{\textit{In vivo}}
\newcommand{\Invitro}{\textit{In vitro}}
\newcommand{\Exvivo}{\textit{Ex vivo}}

% various CD-whatever crap
\newcommand{\cd}[1]{CD{#1}}
\newcommand{\anti}[1]{anti-{#1}}
\newcommand{\antih}[1]{anti-human {#1}}
\newcommand{\antim}[1]{anti-mouse {#1}}
\newcommand{\acd}[1]{\anti{\cd{#1}}}
\newcommand{\ahcd}[1]{\antih{\cd{#1}}}
\newcommand{\amcd}[1]{\antim{\cd{#1}}}
\newcommand{\pos}[1]{#1+}
\newcommand{\cdp}[1]{\pos{\cd{#1}}}
\newcommand{\pcd}[1]{\cdp{#1}~\si{\percent}}
\newcommand{\cdn}[1]{\cd{#1}-}
\newcommand{\ptmem}{\cdp{62L}\pos{CCR7}}
\newcommand{\ptmemp}{\ptmem{}~\si{\percent}}
\newcommand{\pth}{\cdp{4}}
\newcommand{\pthp}{\pth{}~\si{\percent}}
\newcommand{\ptk}{\cdp{8}}
\newcommand{\ptmemh}{\pth\ptmem}
\newcommand{\ptmemk}{\ptk\ptmem}
\newcommand{\dpthp}{$\Updelta$\pthp{}}
\newcommand{\ptcar}{\gls{car}+}
\newcommand{\ptcarp}{\ptcar~\si{\percent}}

% so I don't need to worry about abbreviating all the different interleukins
\newcommand{\il}[1]{\gls{il}-#1}

% DOE stuff I don't feel like typing ad-nauseam
\newcommand{\pilII}{\gls{il2} concentration}
\newcommand{\pdms}{\gls{dms} concentration}
\newcommand{\pmab}{functional \gls{mab} surface density}
\newcommand{\rmemh}{total \ptmemh{} cells}
\newcommand{\rmemk}{total \ptmemk{} cells}
\newcommand{\rratio}{CD4/CD8 ratio}

% vendor and product stuff I don't feel like typing
\newcommand{\catnum}[2]{(#1, #2)}
\newcommand{\product}[3]{#1 \catnum{#2}{#3}}
\newcommand{\thermo}{Thermo Fisher}
\newcommand{\gehc}{GE Healthcare}
\newcommand{\sigald}{Sigma Aldrich}
\newcommand{\miltenyi}{Miltenyi Biotech}
\newcommand{\bl}{Biolegend}
\newcommand{\bd}{Becton Dickenson}

% the obligatory misc category
\newcommand{\inlinecode}{\texttt}
\newcommand{\subcap}[2]{\subref{#1}) #2}
\newcommand{\sigkey}{Significance test key: *p<0.1; **p < 0.05; ***p<0.01}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% ditto for environments

\newenvironment{mytitlepage}{
  \begin{singlespace}
    \begin{center}
    }
    {
    \end{center}
  \end{singlespace}
}

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% begin document (proceed with caution)

\begin{document}

\begin{titlepage}
  \begin{mytitlepage}
    \mytitle{}

    \vfill

    \Large{
      A Dissertation \\
      Presented to \\
      The Academic Faculty \\

      \vspace{1.5em}

      by

      \vspace{1.5em}

      Nathan John Dwarshuis, B.S. \\

      \vfill

      In Partial Fulfillment \\
      of the Requirements for the Degree \\
      Doctor of Philosophy in Biomedical Engineering in the \\
      Wallace H. Coulter Department of Biomedical Engineering

      \vfill

      Georgia Institute of Technology and Emory University \\
      August 2021

      \vfill

      COPYRIGHT \copyright{} BY NATHAN J. DWARSHUIS
    }
  \end{mytitlepage}
\end{titlepage}

\onecolumn \pagenumbering{roman}
\clearpage
\begin{mytitlepage}
  \mytitle{}
\end{mytitlepage}

\vfill

\large{
  \noindent
  Committee Members

  \begin{multicols}{2}
    \begin{singlespace}

      \mycommitteemember{Dr.\ Krishnendu\ Roy\ (Advisor)}
      {Department of Biomedical Engineering}
      {Georgia Institute of Technology and Emory University}

      \vspace{1.5em}

      \mycommitteemember{Dr.\ Madhav\ Dhodapkar}
      {Department of Hematology and Medical Oncology}
      {Emory University}

      \vspace{1.5em}

      \mycommitteemember{Dr.\ Melissa\ Kemp}
      {Department of Biomedical Engineering}
      {Georgia Institute of Technology and Emory University}

      \columnbreak{}
      \null{}
      \vfill

      \mycommitteemember{Dr.\ Wilbur\ Lam}
      {Department of Biomedical Engineering}
      {Georgia Institute of Technology and Emory University}

      \vspace{1.5em}

      \mycommitteemember{Dr.\ Sakis\ Mantalaris}
      {Department of Biomedical Engineering}
      {Georgia Institute of Technology and Emory University}

    \end{singlespace}
  \end{multicols}

  \vspace{1.5em}

  \hfill Date Approved:
}

\clearpage

\vspace*{\fill}

\begin{quote}
You cannot answer a question that you cannot ask, and you cannot ask a
question for which you have no words.

\hfill \textit{-- Judea Pearl}
\end{quote}

\vspace{1in}
\vspace*{\fill}

\clearpage

\chapter*{acknowledgements}
\addcontentsline{toc}{chapter}{Acknowledgments} 

There are many people without which this work would not have been possible.
Firstly, I would like to thank my advisor, Krish Roy, for his mentoring and
guidance as well as his support in my exploring my own ideas throughout this
project. In addition to the Roy lab as a whole, I should specifically recognize
Dr's Ranjna Madan-Lala and Kyung-Ho Roh for writing the NSF EAGER grant that
initially funded this work and providing the foundation of ideas, Dr Hannah
Wilson Song for assisting me with much of the cell-based work, Dr Pallab Pradhan
for assisting with the animal studies, and Miguel Armenta Ochoa and Ritika Jain
for their assistance as well as their ideas for new directions of this work as
they continue it beyond my tenure. I would also like to thank the undergraduates
and high school students I had the pleasure of mentoring, Anokhi Patel, Kate
Richardson, Zahra Mousavi Karimi, Sambhav Jain, Lauren Bailey.

Beyond the Roy lab, I should thank the Cell Manufacturing Technologies (CMaT)
family at large, especially our collaborators Art Edison, Max Colonna, Wandaliz
Torres-Garcia, Valerie Odeh-Couvertier, Theresa Kotanchek, and Bruce Levine.
Additionally, I would like to thank the staff and faculty who fundamentally
supported this work, including Andrea Soyland, Punya Mardhanan, Carol Mills,
Carla Zachery, Adrienne Williams,  Sommer Durham, Laxmi Krishnan, Andrew Shaw,
Aaron Lifland, and Paramita Chatterjee.

Finally, I would like to thank my friends and family for their unconditional
support throughout this process, as well as Arch Enemy and Megadeth
for obvious reasons :)

\clearpage

\tableofcontents

\clearpage

\listoffigures
\addcontentsline{toc}{chapter}{LIST OF FIGURES} 

\clearpage

\listoftables
\addcontentsline{toc}{chapter}{LIST OF TABLES} 

\clearpage

\chapter*{LIST OF SYMBOLS AND ABBREVIATIONS}
\addcontentsline{toc}{chapter}{LIST OF SYMBOLS AND ABBREVIATIONS}

\printglossary[type=\acronymtype]

\clearpage
\pagenumbering{arabic}

\clearpage

\chapter*{summary}
\addcontentsline{toc}{chapter}{SUMMARY} 

\Gls{act} using \gls{car} T cells have shown promise in treating cancer, but
manufacturing large numbers of high quality cells remains challenging. Currently
approved T cell expansion technologies involve \acd{3} and \acd{28} \glspl{mab},
usually mounted on magnetic beads. This method fails to recapitulate many key
signals found \invivo{} and is also heavily licensed by a few companies,
limiting its long-term usefulness to manufactures and clinicians. Furthermore,
highly potent anti-tumor T cells are generally less-differentiated subtypes such
as \acrlongpl{tcm} and \acrlongpl{tscm}. Despite this understanding, little has
been done to optimize T cell expansion for generating these subtypes, including
measurement and feedback control strategies that are necessary for any modern
manufacturing process.

The goal of this dissertation was to develop a microcarrier-based \gls{dms} T
cell expansion system and determine biologically-meaningful \glspl{cqa} and
\glspl{cpp} that could be used to optimize for highly-potent T cells. In
\cref{aim1}, we develop and characterized the \gls{dms} system, including
quality control steps. We also demonstrate the feasiblity of expanding
high-quality T cells. In \cref{aim2a,aim2b}, we use \gls{doe} methodology to
optimize the \gls{dms} platform, and develop a computational pipeline to
identify and model the effect of measurable \glspl{cqa}, and \glspl{cpp} on the
final product. In \cref{aim3}, we demonstrate the effectiveness of the \gls{dms}
platform \invivo{}. This thesis lays the groundwork for a novel T cell expansion
method which can be utilized at scale for a clinical trial and beyond.

\clearpage

\chapter{INTRODUCTION}

\section*{overview}

T cell-based immunotherapies have received great interest from clinicians and
industry due to their potential to treat, and often cure, cancer and other
diseases\cite{Fesnak2016,Rosenberg2015}. In 2017, Novartis and Kite Pharma
received FDA approval for \textit{Kymriah} and \textit{Yescarta} respectively,
two genetically-modified \gls{car} T cell therapies against B cell malignancies.
Despite these successes, \gls{car} T cell therapies are constrained by an
expensive and difficult-to-scale manufacturing process with little control on
cell quality and phenotype\cite{Roddie2019, Dwarshuis2017}. State-of-the-art T
cell manufacturing techniques focus on \acd{3} and \acd{28} activation and
expansion, typically presented on superparamagnetic, iron-based microbeads
(Invitrogen Dynabead, Miltenyi MACS beads), on nanobeads (Miltenyi TransACT), or
in soluble tetramers (Expamer)\cite{Roddie2019,Dwarshuis2017,Wang2016,
  Piscopo2017, Bashour2015}. These strategies overlook many of the signaling
components present in the secondary lymphoid organs where T cells expand
\invivo{}. Typically, T cells are activated under close cell-cell contact, which
allows for efficient autocrine/paracrine signaling via growth-stimulating
cytokines such as \gls{il2}. Additionally, the lymphoid tissues are comprised of
\gls{ecm} components such as collagen and stromal cells, which provide signals
to upregulate proliferation, cytokine production, and pro-survival
pathways\cite{Gendron2003, Ohtani2008, Boisvert2007, Ben-Horin2004}.

A variety of solutions have been proposed to make the T cell expansion process
more physiological. Including feeder cell cultures\cite{Forget2014} and
biomaterials-based methods such as lipid-coated microrods or 3D scaffold
gels\cite{Cheung2018,Delalat2017,meyer15_immun,Lambert2017} that attempt to
recapitulate the cellular membrane, large interfacial contact area,
3D-structure, or soft surfaces T cells normally experience \invivo{}. While
these have been shown to activation and expand T cells, they either are not
scalable (in the case of feeder cells) or still lack many of the signals and
cues T cells experience as the expand. Additionally, none have been shown to
preferentially expand highly-potent T cell necessary for anti-cancer therapies.
Such high potency cells including subtypes with low differentiation state such
as \gls{tscm} and \gls{tcm} cells or CD4 cells, all of which have been shown to
be necessary for durable responses\cite{Xu2014, Fraietta2018, Gattinoni2011,
  Gattinoni2012,Wang2018, Yang2017}. Methods to increase memory and CD4 T cells
in the final product are needed. Furthermore, \gls{qbd} principles such as
discovering and validating novel \glspl{cqa} and \glspl{cpp} in the space of T
cell manufacturing are required to reproducibly manufacture these subtypes and
ensure low-cost and safe products with maximal effectiveness in the clinic

This dissertation describes a novel \acrlong{dms}-based method derived from
porous microcarriers functionalized with \acd{3} and \acd{28} \glspl{mab} for
use in T cell expansion cultures. Microcarriers have historically been used
throughout the bioprocess industry for adherent cultures such as \gls{cho} cells
but not with suspension cells such as T cells\cite{Heathman2015, Sart2011}. The
microcarriers chosen to make the \gls{dms} in this work have a microporous
structure that allows T cells to grow inside and along the surface, providing
ample cell-cell contact for enhanced autocrine and paracrine signaling.
Furthermore, the 3D surface of the carriers provides a larger contact area for T
cells to interact with the \glspl{mab} relative to beads; this may better
emulate the large contact surface area that occurs between T cells and
\glspl{dc}.

\section*{hypothesis}

The hypothesis of this dissertation was that using \glspl{dms} created from
off-the-shelf microcarriers and coated with activating \glspl{mab} would lead to
higher quantity and quality T cells as compared to state-of-the-art bead-based
expansion. We also hypothesized that T cells have measurable biological
signatures that are predictive of downstream outcomes and phenotypes. The
objective of this dissertation was to develop this platform, test its
effectiveness both \invitro{} and \invivo{}, and develop computational pipelines
to discover novel \glspl{cpp} and \glspl{cqa} that can be translated to a
manufacturing environment and a clinical trial setting.

\section*{specific aims}

The specific aims of this dissertation are outlined in
\cref{fig:graphical_overview}.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics[width=\textwidth]{../figures/overview.png}

  \endgroup
  \caption[Project Overview]{High-level overview.}
  \label{fig:graphical_overview}
\end{figure*}

\subsection*{aim 1: develop and optimize a novel T cell expansion process that
  mimics key components of the lymph nodes}

In this first aim, we demonstrated the process for manufacturing \glspl{dms},
including quality control steps that are necessary for translation of this
platform into a scalable manufacturing setting. We also demonstrate that the
\gls{dms} platform leads to higher overall expansion of T cells and higher
overall fractions of potent memory and CD4+ subtypes desired for T cell
therapies. Finally, we demonstrate \invitro{} that the \gls{dms} platform can be
used to generate functional \gls{car} T cells targeted toward CD19.

\subsection*{aim 2: develop methods to control and predict T cell quality}

For this second aim, we investigated methods to identify and control \glspl{cqa}
and \glspl{cpp} for manufacturing T cells using the \gls{dms} platform. This was
accomplished through two sub-aims:

\begin{itemize}
  \item[A --] Develop computational methods to control and predict T cell
    expansion and quality
  \item[B --] Perturb \gls{dms} expansion to identify additional mechanistic
    controls for expansion and quality
\end{itemize}

\subsection*{aim 3: confirm potency of T cells from novel T cell expansion
  process using \invivo{} xenograft mouse model}

In this final aim, we demonstrate the effectiveness of \gls{dms}-expanded T
cells compared to state-of-the-art beads using \invivo{} mouse models for
\gls{all}.

\section*{outline}

In \cref{background}, we provide additional background on the current state of T
cell manufacturing and how the work in this dissertation moves the field
forward. In \cref{aim1,aim2a,aim2b,aim3} we present the work pertaining to Aims
1, 2a, 2b, and 3 respectively. Finally, in \cref{conclusions} we present our
conclusions as well as provide insights for how this work can be extended in the
future.

\chapter{BACKGROUND AND INNOVATION}\label{background}

\section{Background}

\subsection{Quality by Design in Cell Manufacturing}

The challenges for the cell manufacturing field are significant. Unlike other
industries which manufacture inanimate products such as automobiles and
semiconductors, the cell manufacturing industry needs to contend with the fact
that cells are living entities which can change with every process
manipulation\cite{Kirouac2008, Little2006, Pirnay2012, Rousseau2013}. This is
further compounded by the lack of standardization and limited regulation.

In order to overcome these barriers, adopting a systemic approach to cell
manufacturing using \acrlong{qbd} principles will be extremely
important\cite{Kirouac2008}. In \gls{qbd}, the objective is to reproducibly
manufacturing products which minimizes risk for downstream stakeholders (in this
case, the patient). Broadly, this entails determining \acrlongpl{cqa} and
\acrlongpl{cpp} and incorporating them into models which can explain and predict
the cell manufacturing process.

\Glspl{cqa} are measurable properties of the product that can be used to define
its functionality and hence quality. \glspl{cqa} are important for defining the
characteristics of a `good' product (release criteria) but also for ensuring
that a process is on track to making such a product (process control). In the
space of cell manufacturing, examples of \glspl{cqa} include markers on the
surface of cells and readouts from functional assays such as killing assays. In
general, these are poorly understood if they exist at all.

\glspl{cpp} are parameters which may be tuned and varied to control the outcome
of process and the quality of the final product. In cell manufacturing, these
are poorly understood. Examples in the cell manufacturing space include the type
of media used and the amount of \il{2} added. Once \glspl{cpp} are known, they
can be optimized to ensure that costs are minimized and potency of the cellular
product is maximized.

The topic of discovering novel \glspl{cpp} and \glspl{cqa} in the context of
this work are discussed further in \cref{sec:background_doe} and
\cref{sec:background_quality}/\cref{sec:background_cqa} respectively.

\subsection{T Cells for Immunotherapies}

A variety of T cell therapies have been utilized with varying degrees of
success, and we describe a few of the most prominent below. We should note that
while this work focuses on the application of \gls{car} T cell therapies, in
theory the technology developed in this dissertation could theoretically apply
to any T cell-based therapy with little to no modification.

One of the first successful T cell-based immunotherapies against cancer is
\glspl{til}\cite{Rosenberg2015}. This method works by taking tumor specimens
from a patient, allowing the tumor-reactive lymphocytes to expand \exvivo{}, and
then administered back to the patient along with a high dose of
\il{2}\cite{Rosenberg1988}. In particular, \gls{til} therapy has shown robust
results in treating melanoma\cite{Rosenberg2011}, although \glspl{til} have been
found in other solid tumors such as gastointestinal, cervical, lung, and
ovarian\cite{Rosenberg2015, Wang2014, Foppen2015, Solinas2017, June2007,
  Santoiemma2015}, and their presence is generally associate with favorable
outcomes\cite{Clark1989}. \glspl{til} are heterogeneous cell mixtures and
generally are comprised of CD3 T cells and $\upgamma\updelta$ T
cells\cite{Nishimura1999, Cordova2012}. To date, there are over 250 open
clinical trials using \glspl{til}.

Besides \glspl{til}, the other broad class of T cell immunotherapies that has
achieved great success in treating cancer in recent decades are gene-modified T
cells. Rather than expand T cells that are present natively (as is the case with
\gls{til} therapy), gene-modified T cell therapies entail extracting T cells
from either the cancer patient (autologous) or a healthy donor (allogeneic) and
reprogramming them genetically to target a tumor antigen (see
\cref{sec:background_source} for an overview of how T cells can be sourced).
This approach offers much more flexibility, as the degree of reprogramming is
only limited by the scale and possibilities of gene-editing technology, which
has rapidly accelerated in recent decades\cite{Rosenberg2015}.

T cells with transduced \glspl{tcr} were first designed to overcome the
limitations of \glspl{til}\cite{Rosenberg2015, Wang2014}. In this case, T cells
are transduced \exvivo{} with a lentiviral vector to express a \gls{tcr}
targeting a tumor antigen. T cells transduced with \glspl{tcr} have shown robust
results in melanoma patients\cite{Robbins2011}, synovial
sarcoma\cite{Morgan2006}, and others\cite{Ikeda2016}. To date, there are over
200 clinical trials using T cells with transduced \glspl{tcr}.

While transduced \glspl{tcr} offer some flexibility in retargeting T cells
toward relevant tumor antigens, they are still limited in that they can only
target antigens that are presented via \gls{mhc} complexes. \acrlong{car} T
cells overcome this limitation by using linking a \gls{tcr}-independent antigen
recognition domain with the stimulatory and costimulatory machinery of a T cell
\gls{car} T cells were first demonstrated in 1989, where the authors swapped the
antigen-recognition domains of a native \gls{tcr} with a that of a foreign
\gls{tcr}\cite{Gross1989}. Since then, this method has progressed to using an
\gls{scfv} with a CD3$\upzeta$ stimulatory domain along with the CD28, OX-40, or
4-1BB domains for costimulation. Since these can all be expressed with one
protein sequence, \gls{car} T cells are relatively simple to produce and require
only a single genetic transduction step (usually a lentiviral vector) to
reprogram a batch T cells \exvivo{} toward the desired antigen. \gls{car} T
cells have primarily found success in against CD19- and CD20-expressing tumors
such as \gls{all} and \gls{cll} (eg B-cell malignancies)\cite{Kalos2011,
  Brentjens2011, Kochenderfer2010, Maude2014, Till2012, Till2008}.

Out of all the T cell therapies discussed thus far, \gls{car} T cells have
experienced the most commercial success and excitement. In 2017, Novartis and
Kite Pharma acquired FDA approval for \textit{Kymriah} and \textit{Yescarta}
respectively, both of which are \gls{car} T cell therapies against B-cell
malignancies. \gls{car} T cells are under further exploration for use in many
other tumors, including multiple myeloma, mesothelioma, pancreatic cancer,
glioblastoma, neuroblastoma, and prostate cancer, breast cancer, non-small-cell
lung cancer, and others\cite{Rosenberg2015, Wang2014, Fesnak2016, Guo2016}. To
date, there are almost 1000 clinical trials using \gls{car} T cells.

\subsection{Scaling T Cell Expansion}

In order to scale T cell therapies to meet clinical demands, automation and
bioreactors will be necessary. To this end, there are several choices that have
found success in the clinic.

The WAVE bioreactor (GE Healthcare) is the choice of expansion for many clinical
applications\cite{Brentjens2011, Hollyman2009, Brentjens2013}. It is part of a
broader class of bioreactors that consist of rocking platforms that agitate a
bag filled with media and cells. Importantly, it has built-in sensors for
measuring media flow rate, \ce{CO2}, \ce{O2}, pH, and nutrient consumption which
enables automation. Unfortunately, in some settings this is not considered
scalable as only one bag per bioreactor is allowed at once\cite{Roddie2019}. The
other disadvantage with the WAVE system is that it keeps cells far apart by
design, which could have negative impact on cross-talk, activation, and
growth\cite{Somerville2012}.

% BACKGROUND find clinical trials (if any) that use this
Alternatively, the CliniMACS Prodigy (Miltenyi) is an all-in-one system that is
a fully closed system that removes the need for expensive cleanrooms and
associated personnel\cite{Kaiser2015, Bunos2015}. It contains modules to perform
transduction, expansion, and washing. This setup also implies that fewer
mistakes and handling errors will be made, since many of the steps are internal
to the machine. Initial investigations have shown that it can yield T cells
doses required for clinical use\cite{Zhu2018}. At the time of writing, several
clinical trial are underway which use the CliniMACS, although mostly for
stem-cell based cell treatments.

Finally, another option that has been investigated for T cell expansion is the
Grex bioreactor (Wilson Wolf). This is effectively a tall tissue-culture plate
with a porous membrane at the bottom, which allows gas exchange to the active
cell culture at the bottom of the plate while permitting large volumes of media
to be loaded on top without suffocating the cells. While this is quite similar
to plates and flasks normally used for small-scale research, the important
difference is that its larger size requires fewer interactions and keeps the
cells at a higher nutrient concentration for longer periods of time. However, it
is still a an open system and requires manual (by default) interaction from an
operator to load, feed, and harvest the cell product. Grex bioreactors have been
using to grow \glspl{til}\cite{Jin2012} and virus-specific T
cells\cite{Gerdemann2011}.

Much work is still required in the space of bioreactor design for T cell
manufacturing, but novel T cell expansion technologies such as that described in
this work need to consider how it may be used at scale in such a system.

\subsection{Cell Sources in T Cell Manufacturing}\label{sec:background_source}

T cells for cell manufacturing can be obtained broadly via two paradigms:
autologous and allogeneic. The former involves obtaining T cells from a patient
and giving them back to the same patient after \exvivo{} expansion and genetic
modification. The latter involves taking T cells from a healthy donor, expanding
them and manipulating them as desired, storing them long term, and then giving
them to multiple patients. There are advantages and disadvantages to both, and
in some cases such as \gls{til} therapy, the only option is to use autologous
therapy.

Autologous T cells by default are much safer. By definition, they will have no
cross-reactivity with the patient and thus \gls{gvhd} is not a
concern\cite{Decker2012}. However, there are numerous disadvantages. Autologous
therapies are over 20 times more costly as the process needs to be repeated for
every patient\cite{Harrison2019}. Compounding this, many cell products are
manufactured at a centralized location, so patient T cells need to be shipped
twice on dry ice from the hospital and back. This adds days to the process,
which is critical for patients with fast moving diseases. Manufacturing could be
done on-site in a decentralized manner, but this requires more equipment and
personnel overall. Using cells from a diseased patient has many drawbacks in
itself. Cancer patients (especially those with chronic illnesses) often have
exhausted T cells which expand far less readily and are consequently less
potent\cite{Wherry2015, Ando2019, Zheng2017}. Additionally, they may have high
frequencies of \glspl{treg} which inhibitory\cite{Decker2012}. Removing these
cells as well as purifying \glspl{th1} may enhance the potency of the final
product\cite{Goldstein2012, Drela2004, Rankin2003, Luheshi2013, Grotz2015};
however, this would make the overall process more expensive as an additional
separation step would be required.

Allogeneic T cell therapies overcome nearly all of these disadvantages. Donor
\glspl{pbmc} are easy to obtain, they can be processed in centralized locations,
they can be stored easily under liquid nitrogen, and donors could be screened to
find those with optimal anti-tumor cells. The key is overcoming \gls{gvhd}.
Obviously this could be done the same way as done for transplants where patients
find a `match' for their \gls{hla} type, but this severally limits options. This
can be overcome by using advanced gene-editing tools which can both add and
delete genetic material (eg \glspl{zfn}, \glspl{talen}, or \gls{crispr}) to
remove the native \gls{tcr} which would prevent the donor T cells from attacking
host tissue\cite{Liu2019, Wiebking2020, Provasi2012, Berdien2014, Themeli2015}.
This obviously complicates the process, as additional edits besides the
insertion of the \gls{car} would be required, and these technologies are not yet
very efficient. To date there are about 10 open clinical trials utilizing
allogeneic T cell therapies edited with \gls{crispr} to reduce the likelihood of
\gls{gvhd}.

\subsection{Overview of T Cell Quality}\label{sec:background_quality}

T cells are highly heterogeneous and can exist in a variety of states and
subtypes, many of which can be measured (at least indirectly) though biomarkers
such as cell surface proteins. Identifying and understanding these biomarkers
are the basis for \glspl{cqa} which can be used to for process control, release
criteria, and initial cell source screening.

One of the most important dimensions of T cell quality is that of
differentiation. T cells begin their life in circulation (eg after they exit the
thymus) as \glspl{tn}. When they become activated in the secondary lymph node
organs, they differentiate from \gls{tn} to \glspl{tscm}, \glspl{tcm},
\glspl{tem}, and finally \glspl{teff}\cite{Gattinoni2012}. Subtypes earlier in
this process are generally called `memory' or `memory-like' cells (eg \gls{tscm}
and \gls{tcm}), and have been shown to have increased potency toward a variety
of tumors, presumably due to their higher capacity for self-renewal and
replication, enhanced migratory capacity, and/or increased engraftment
potential\cite{Xu2014, Gattinoni2012, Fraietta2018, Gattinoni2011, Turtle2009}.
The capacity for self-renewal is especially important for T cells therapies, as
evidenced by the fact that \gls{til} therapies with longer telomeres tend to
create more durable responses\cite{Donia2012}. Additionally, clonal diversity
decreases following the infusion of \gls{car} T cell therapies, which
demonstrates that only a few clones are self-renewing and therefore responsible
for the overall response\cite{Sheih2020}. Memory T cells can be quantified
easily using surface markers such as CD62L, CCR7, CD27, CD45RA, and CD45RO.
Furthermore, memory markers are inversely related to exhaustion markers which
are negatively associated with clinical outcomes\cite{Lee2013}. These cells in
particular are seen in patients with chronic immune activation such as patients
with chronic cancers.

In addition to memory, the other major axis by which T cells may be classified
is the CD4/CD8 ratio. \Glspl{th} are CD4+ are responsible for secreting
cytokines which coordinate the immune response while CD8+ \glspl{tc} are
responsible for killing tumors or infected cells using specialized lytic
enzymes. Since \glspl{tc} actually perform the killing function, it seems
intuitive that \glspl{tc} would be most important for anti-tumor
immunotherapies. However, in mouse models with glioblastoma, survival was
negatively impacted when \glspl{th} were removed\cite{Wang2018}. Furthermore,
\glspl{th} have been shown to have cytotoxic properties on their own and also
show resistance to exhaustion compared to \glspl{tc}\cite{Yang2017}. While T
cell products with a defined ratio of CD4 and CD8 T cells have been utilized,
they are more expensive than products with undefined ratios as the T cells need
to be sorted and recombined, adding additional complexity\cite{Turtle2016}.

While less of a focus in this dissertation, other quality markers exists to
assess the overall killing potential and safety of the T cell product. Numerous
methods exists to detect the killing capacity of \gls{car} T cells, many of
which involve either measuring the lysis of a target cell using a dye or a
radioactive tracer, by measuring the degranulation of the T cells themselves, or
by measuring a cytokine that is secreted upon T cell activation and killing such
as \gls{ifng}. Furthermore, the viability of T cells may be assessed using a
number of methods, including exclusion dyes such as \gls{aopi} or a functional
assay to detect metabolism. Finally, for the purposes of safety, T cell products
using retro- or lentiviral vectors as their means of gene-editing must be tested
for replication competent vectors\cite{Wang2013} and for contamination via
bacteria or other pathogens.

\subsection{T cell Activation Methods}\label{sec:background_activation}

In order for T cells to be expanded \exvivo{} they must be activated with a
stimulatory signal (Signal 1) and a costimulatory signal (Signal 2). \Invivo{},
Signal 1 is administered via the \gls{tcr} and the CD3 receptor when \glspl{apc}
present a peptide via \gls{mhc} that the T cell in question is able to
recognize. Signal 2 is administered via CD80 and CD86 which are also present on
\glspl{apc} and is necessary to prevent the T cell from becoming anergic. While
these two signal are the bare minimum to trigger a T cell to expand, there are
many other potential signals present. T cells have many other costimulatory
receptors such as OX40, 4-1BB and ICOS which are costimulatory along with CD28,
and \glspl{apc} have corresponding ligands for these depending on the nature of
the pathogen they have detected\cite{Azuma2019}. Furthermore, T cells exist in
high cell density within the secondary lymphoid organs, which allows efficient
cytokine cross-talk in an autocrine and paracrine manner. These cytokines are
responsible for triggering proliferation (in the case of \il{2}) and subset
differentiation (in the case of many others)\cite{Luckheeram2012}. By tuning the
signal strength and duration of Signal 1, Signal 2, the various costimulatory
signals, and the cytokine milieu, a variety of T cell phenotypes can be
actualized.

\Invitro{}, T cells can be activated in a number of ways but the simplest and
most common is to use \glspl{mab} that cross-link the CD3 and CD28 receptors,
which supply Signal 1 and Signal 2 without the need for antigen (which also
means all T cells are activated and not just a few specific clones). Additional
signals may be supplied in the form of cytokines (eg \il{2}, \il{7}, or \il{15})
or feeder cells\cite{Forget2014}.

As this is a critical unit operation in the manufacturing of T cell therapies, a
number of commercial technologies exist to activate T cells\cite{Wang2016,
  Piscopo2017, Roddie2019, Bashour2015}. The simplest is to use \acd{3} and
\acd{28} \glspl{mab} bound to a 2D surface such as a plate, and this can be
accomplished in a \gls{gmp} manner (at least from a reagents perspective) as
soluble \gls{gmp}-grade \glspl{mab} are commericially available. A similar but
distinct method along these lines is to use multivalent activators such as
ImmunoCult (StemCell Technologies) or Expamer (Juno Therapeutics) which may have
increased cross-linking capacity compared to traditional \glspl{mab}. Beyond
soluble protein, \glspl{mab} against CD3 and CD28 can be mounted on magnetic
microbeads (\SIrange{3}{5}{\um} in diameter) such as DynaBeads (Invitrogen) and
MACSbeads (\miltenyi{}), which are easier to separate using magnetic washing
plates. Magnetic nanobeads such as TransAct (\miltenyi{}) work by a similar
principle except they can be removed via centrifugation rather than a magnetic
washing plate. Cloudz (RnD Systems) are another bead-based T cell expansion
which mounts \acd{3} and \acd{28} \glspl{mab} on alginate microspheres, which
are not only easily degradable but are also softer, which can have a positive
impact on T cell activation and phenotype\cite{Lambert2017, OConnor2012}.

A problem with all of these commercial solutions is that they only focus on
Signal 1 and Signal 2 and ignore the many other physiological cues present in
the secondary lymphoid organs. A variety of novel T cell activation and
expansion solutions have been proposed to overcome this. One strategy is to use
modified feeder cell cultures to provide activation signals similar to those of
\glspl{dc}\cite{Forget2014}. While this has the theoretical capacity to mimic
several key components of the lymph node, it is hard to reproduce on a large
scale due to the complexity and inherent variability of using cell lines in a
fully \gls{gmp}-compliant manner. Others have proposed biomaterials-based
solutions to circumvent this problem, including lipid-coated
microrods\cite{Cheung2018}, 3D-scaffolds via either Matrigel\cite{Rio2018} or
3d-printed lattices\cite{Delalat2017}, ellipsoid beads\cite{meyer15_immun}, and
\gls{mab}-conjugated \gls{pdms} beads\cite{Lambert2017} that respectively
recapitulate the cellular membrane, large interfacial contact area,
3D-structure, or soft surfaces T cells normally experience \textit{in vivo}.
None have been demonstrated to demonstrably expand high quality T cells as
outlined in \cref{sec:background_quality}.

\subsection{Microcarriers in Bioprocessing}

In this work, we explored microcarriers as the basis for an alternative to the
methods described in \cref{sec:background_activation}.

Microcarriers have historically been used to grow a number of adherent cell
types for a variety of applications. They were introduced in 1967 as a means to
grow adherent cells `in suspension', effectively turning a 2D flask system into
a 3D culture system\cite{WEZEL1967}. Microcarriers are generally spherical and
are several hundred \si{\um} in diameter, which means they collectively have a
much higher surface area than a traditional flask when matched for volume.
Consequently, this means that microcarrier-based cultures can operate with much
lower footprints than flask-like systems. Microcarriers also allow cell culture
to operate more like a traditional chemical engineering process, wherein a
\gls{cstr} may be employed to enhance oxygen transfer, maintain pH,
and continuously supply nutrients\cite{Derakhti2019}.

A variety of microcarriers have been designed, primarily differing in their
choice of material and macroporous structure. Key concerns driving these
choiceshave been cell attachment at the beginning of culture and cell detachment
at the harvesting step\cite{Derakhti2019}. Many microcarriers simply use
polystyrene (the material used for tissue culture flasks and dishes in general)
with no modification (SoloHill Plastic, Nunc Biosilon), with a cationic surface
charge (SoloHill Hillex) or coated with an \gls{ecm} protein such as collagen
(SoloHill Fact III). Other base materials have been used such as dextran (GE
Healthcare Cytodex), cellulose (GE Healthcare Cytopore), and glass (\sigald{}
G2767), all with none or similar surface modifications. Additionally, some
microcarriers such as \gls{cus} and \gls{cug} are composed entirely out of
protein (in these cases, porcine collagen) which also allows the microcarriers
to be enzymatically degraded. In the case of non-protein materials, cells may
still be detached using enzymes but these require similar methods to those
currently used in flasks such as trypsin which target the cellular \gls{ecm}
directly. Since trypsin and related enzymes tends to be harsh on cells, an
advantage of using entirely protein-based microcarriers is that they can be
degraded using a much safer enzyme such as collagenase, at the cost of being
more expensive and also being harder to make
\gls{gmp}-compliant\cite{Derakhti2019}. Going one step further, some
microcarriers are composed of a naturally degrading scaffold such as alginate,
which do not need an enzyme for degradation. Finally, microcarriers can differ
in their overall structure. \gls{cug} and \gls{cus} microcarriers as well as the
Cytopore microcarriers are macroporous, meaning they have a porous network in
which cells can attach throughout their interior. This drastically increases the
effective surface area and consequently the number of cells which may be grown
per unit volume. Other microcarriers are microporous (eg only to small
molecules) or not porous at all (eg polystyrene); in either case the cells can
only grow on the surface.

Microcarriers in general have seen the most use in growing \gls{cho} cells and
hybridomas in the case of protein manufacturing (eg \gls{igg}
production)\cite{Xiao1999, Kim2011} as well as \glspl{esc} and \glspl{msc} more
recently in the case of cell manufacturing\cite{Heathman2015, Sart2011,
  Chen2013, Schop2010, Rafiq2016}. Interestingly, some groups have even explored
using biodegradable microcarriers \invivo{} as a delivery vehicle for stem cell
therapies in the context of regenerative medicine\cite{Zhang2016, Saltz2016,
  Park2013, Malda2006}. However, the characteristic shared by all the cell types
in this application is the fact that they are adherent. In this work, we explore
the use of microcarrier for T cells, which are naturally non-adherent.

The microcarriers used in this work were \gls{cus} and \gls{cug} (mostly the
former) which are both composed of cross-linked gelatin and have a macroporous
morphology. Their protein-based composition makes functionalization easy; the
surface is rich in lysine residues which can be easily bonded with a
base-reactive linker such as \gls{snb}. These specific carriers have been used
in the past for pancreatic islet cells\cite{Guerra2001},
\glspl{esc}\cite{Fernandes2007, Storm_2010}, and \glspl{msc}\cite{Eibes2010}.
Furthermore, they are readily available in over 30 countries and are used in an
FDA fast-track-approved combination retinal pigment epithelial cell product
(Spheramine, Titan Pharmaceuticals)\cite{purcellmain}. This regulatory history
will aid in clinical translation.

\subsection{Integrins and T Cell Signaling}

Because the microcarriers used in this work are derived from collagen, one key
question is how these collagen domains may interact with the T cells during
culture. This question is further explored in \cref{aim2b}.

T cells naturally expand in the lymph nodes which have an \gls{ecm} composed of
collagen\cite{Dustin2001, Ebnet1996, Ohtani2008}. Despite this, T cells don't
interact with collagen fibers in the lymph node as the collagen fibers are
sheathed with stromal fibroblasts\cite{Dustin2001, Ebnet1996}. However, the
\gls{ecm} of peripheral tissues is dense with exposed collagen fibers are
available to interact with T cells. Furthermore, T cells have been shown
\invitro{} to crawl along collagen fibers in the presence of \glspl{apc},
facilitating short encounters with \glspl{apc}\cite{Gunzer2000}. While this may
not be ideal \invivo{} due to the lack of cumulative signal received by
\glspl{apc}\cite{Dustin2001}, it may be advantageous to include collagen domains
\invitro{} as the mode of activation is not specific to any given clone.

The major surface receptors for collagen are \gls{a2b1} and
\gls{a2b2}\cite{Dustin2001, Hemler1990}. These receptors are not expressed on
naive T cells and thus presence and stimulation of collagen alone is not
sufficient to cause activation or expansion of T cells\cite{Hemler1990}. These
receptors have been shown to lead to a number of functions that may be useful in
the context of T cell expansion. First, they have been shown to act in a
costimulatory manner which leads to increased proliferation\cite{Rao2000}.
Furthermore, \gls{a2b1} and \gls{a2b2} have been shown to protect Jurkat cells
against Fas-mediated apoptosis in the presence of collagen I\cite{Aoudjit2000,
  Gendron2003}. Finally, these receptors have been shown to increase \gls{ifng}
production \invitro{} when T cells derived from human \glspl{pbmc} are
stimulated in the presence of collagen I\cite{Boisvert2007}.

\subsection{The Role of IL15 in Memory T Cell Proliferation}

\il{15} is a cytokine that is involved with the proliferation and homeostasis of
memory T cells. Its role in the work of this dissertation is the subject of
further exploration in \cref{aim2b}.

Functionally, mice lacking the gene for either \il{15}\cite{Kennedy2000} or its
high affinity receptor \il{15R$\upalpha$}\cite{Lodolce1998} are generally
healthy but show a deficit in memory CD8 T cells, thus underscoring its
importance in manufacturing high-quality memory T cells for immunotherapies. T
cells themselves express \il{15} and all of its receptor
components\cite{MirandaCarus2005}. Additionally, blocking \il{15} itself or
\il{15R$\upalpha$} \invitro{} has been shown to inhibit homeostatic
proliferation of resting human T cells\cite{MirandaCarus2005}.

% ACRO fix the il2R and IL15R stuff
\il{15} has been puzzling historically because it shares almost the same pathway
as \il{2} yet has different functions\cite{Stonier2010, Osinalde2014, Giri1994,
  Giri1995}. In particular, both cytokines share the common $\upgamma$ subchain
(CD132) as well as the \il{2} $\upbeta$ receptor (CD122). The main difference in
the heterodimeric receptors for \il{2} and \il{15} is the \il{2} $\upalpha$
receptor (CD25) and the \il{15} $\upalpha$ chain respectively, both of which
have high affinity for their respective ligands. The \il{2R$\upalpha$} chain
itself does not have any signaling capacity, and therefore all the signaling
resulting from \il{2} is mediated thought the $\upbeta$ and $\upgamma$ chains,
namely via JAK1 and JAK3 leading to STAT5 activation consequently T cell
activation. \il{15R$\upalpha$} itself has some innate signaling capacity, but
this is poorly characterized in lymphocytes\cite{Stonier2010}. Thus there is a
significant overlap between the functions of \il{2} and \il{15} due to their
receptors sharing the $\upbeta$ and $\upgamma$ chains in their heterodimeric
receptors, and perhaps the main driver of their differential functions it the
half life of each respective receptor\cite{Osinalde2014}.

Where \il{15} is unique is that many (or possibly most) of its functions derive
from being membrane-bound to its receptor\cite{Stonier2010}. Particularly,
\il{15R$\upalpha$} binds to soluble \il{15} which produces a complex that can
transmit signals to close neighboring cells (so called \textit{trans}
presentation). This has been demonstrated in adoptive cell models, where T cells
lacking \il{15R$\upalpha$} were able to generate memory T cells and proliferate
only when other cells were present which expressed
\il{15R$\upalpha$}\cite{Burkett2003, Schluns2004}. The implication of this
mechanism is that cells expression \il{15R$\upalpha$} either need to express
\il{15} themselves or be near other cells expressing \il{15}, and other cells in
proximity require the $\upbeta$ and $\upgamma$ chains to receive the signal. In
addition to \textit{trans} presentation, \il{15} may also work in a \textit{cis}
manner, where \il{15R$\upalpha$}/\il{15} complexes may bind to the $\upbeta$ and
$\upgamma$ chains on the same cell, assuming all receptors are expressed and
soluble \il{15} is available\cite{Olsen2007}. Finally, \il{15R$\upalpha$} itself can exist in
a soluble form, which can bind to \il{15} and signal to cells which are not
adjacent to the source independent of the \textit{cis/trans} mechanisms already
described\cite{Budagian2004}.

\subsection{Overview of Design of Experiments}\label{sec:background_doe}

The \gls{dms} system described in this dissertation has a number of parameters
that can be optimized and controlled (eg \glspl{cpp}). A \gls{doe} is an ideal
framework to test multiple parameters simultaneously and determine which are
relevant \glspl{cpp}.

The goal of \gls{doe} is to answer a data-driven question with the least number
of resources\cite{Wu2009}. It was developed in many non-biological industries
throughout the \nth{20} century such as the automotive and semiconductor
industries where engineers needed to minimize downtime and resource consumption
on full-scale production lines.

At its core, a \gls{doe} is simply a matrix of conditions to test where each row
(usually called a `run') corresponds to one experimental unit for which the
conditions are applied, and each column represents a parameter of concern to be
tested. The values in each cell represent the level at which each parameter is
to be tested. When the experiment is performed using this matrix of conditions,
the results are be summarized into one or more `responses' that correspond to
each run. These responses are then be modeled (usually using linear regression)
to determine the statistic relationship (also called an `effect') between each
parameter and the response(s).

Collectively, the space spanned by all parameters at their feasible ranges is
commonly referred to as the `design space', and generally the goal of a
\gls{doe} is to explore this design space using using the least number of runs
possible. While there are many types of \glspl{doe} depending on the nature
of the parameters and the goal of the experimenter, they all share common
principles:

\begin{description}
\item [randomization --] The order in which the runs are performed should
  ideally be as random as possible. This is to mitigate against any confounding
  factors that may be present which depend on the order or position of the runs.
  For an example in context, the evaporation rate of media in a tissue culture
  plate will be much faster at the perimeter of the plate vs the center. While
  randomization does not eliminate this error, it will ensure the error is
  `spread' evenly across all runs in an unbiased manner.
\item [replication --] Since the analysis of a \gls{doe} is inherently
  statistical, replicates should be used to ensure that the underlying
  distribution of errors can be estimated. While this is not strictly necessary
  to conclude results using a \gls{doe}, failure to use replications requires
  strong assumptions about the model structure (particularly in the case of
  high-complexity models which could easily fit the data perfectly) and also
  precludes the use of statistical tests such as the lack-of-fit test which can
  be useful in rejecting or accepting a particular analysis. Note that the
  subject of replication is within but not the same as power analysis, which
  concerns the number of runs required to estimate a certain effect size.
\item [orthogonality --] Orthogonality refers to the independence of each
  parameter in the design matrix. In other words, the levels tested in any given
  parameter add mutually-exclusive information about the response(s). Again,
  while not strictly necessary, orthogonality drastically simplifies the
  analysis of the experiment by allowing each parameter to be treated
  separately. In cases where orthogonality is impossible (which is often true in
  experiments with many categorical variables) strategies exist to maximize
  orthogonality.
\item [blocking --] In the case where the experiment must be non-randomly spread
  over multiple groups, runs are assigned to `blocks' which are not necessarily
  relevant to the goals of the experiment but nonetheless could affect the
  response. A key assumption that is (usually) made in the case of blocking is
  that there is no interaction between the blocking variable and any of the
  experimental parameters. For example, in T cell expansion, if media lot were a
  blocking variable and expansion method were a parameter, we would by default
  assume that the effect of the expansion method does not depend on the media
  lot (even if the media lot itself might change the mean of the response).
\end{description}

\Glspl{doe} served three purposes in this dissertation. First, we used them as
screening tools for potential \glspl{cpp}, which allowed us to test many input
parameters and filter out the few that likely have the greatest effect on the
response. Second, they were used to make a robust response surface model to
predict optimums using relatively few resources, especially compared to full
factorial or one-factor-at-a-time approaches. Third, we used \glspl{doe} to
discover novel effects and interactions that generated hypotheses that could
influence the directions for future work. To this end, the types of \glspl{doe}
we generally used were fractional factorial designs with three levels, which
enable the estimation of both main effects and second order quadratic effects.

While there are advantages of using \glspl{doe}, it is important to recognize
that they are not necessary or recommended for all experimental aims. In
particular, \glspl{doe} excel when multiple factors (possible with multiple
levels) need to be investigated at once and with a known degree of power. This
is especially important when interaction is expected or needs to be
investigated. However, it could be the case that one already has data on many of
the factors of concern. If one only cares about main effects, performing a
\gls{doe} (particularly a lower-powered screening experiment such as a
resolution III design) with these factors and a few others may not be
productive, and one is better off performed a few extra pilot experiments before
doing a more complex design such as a central composite if desired. Furthermore,
it should be noted that while the goal of a \gls{doe} is to minimize resources,
the size necessary to justify a \gls{doe} may not be worth the experimental
return. For biological work (or any domain with little automation), performing a
randomized experiment with 20 to 30 runs is not trivial from a logistical
perspective, especially when considering the number of manual manipulations and
the chance of human error.

Despite these caveats, many of the principles used for a \gls{doe} are important
in general for experimentation. The most obvious is randomization, which is
often not employed (and also not explicitly mentioned in papers) even though it
is empirically obvious that well plates have different evaporation rates
depending on well position. Assuming the experiment is manual, the largest
reason to avoid randomization is that the experimentalist has no pattern to
follow when administering treatment (such as ``add X to row 1 in well plate''),
thus cognitive burden and the likelihood of mistakes increases. While
\glspl{doe} are usually bigger with more parameters, the one-factor-at-a-time
experiment usually performed in biological disciplines is much smaller and only
has a few parameters, thus these concerns are minimal. There is no reason to
avoid randomization in these cases, as the added cognitive cost is far offset by
the guarantee of eliminated bias due to run position.

\subsection{Identification and Standardization of CPPs and
  CQAs}\label{sec:background_cqa}

% BACKGROUND at least attempt to show that there is alot of work in the space
% identifying signaling networks

In the context of T cell manufacturing, ideally we would have a set of
non-destructive biomarkers that could both identify functional T cells and
predict when a process is on track to deliver such functional T cells. T cells
secrete numerous cytokines and metabolites in the media, which may reflect the
internal state accurately and thus serve as a potential set of \glspl{cqa}.

The complexity of these pathways dictates that we take a big-data approach to
this problem. To this end, there are several pertinent multi-omic (or simply
`omic') techniques that can be used to collect such datasets, which can then be
mined, modeled, and correlated to relevent responses (such as an endpoint
quantification of memory T cells) to identify pertinent \glspl{cqa}.

An overview of the techniques used in this work are:

\begin{description}
\item[Luminex --] This is a multiplexed bead-based assay similar to \gls{elisa} that can measure
  many bulk (not single cell) cytokine concentrations simultaneously
  in a media sample. This is a destructive assay but does not require cells to
  obtain a measurement.
\item[\gls{nmr} --] It is well known that T cells of different
  lineages have different metabolic profiles; for instance memory T
  cells have larger aerobic capacity and fatty acid
  oxidation\cite{Buck2016, van_der_Windt_2012}. \gls{nmr} is a technique that
  can non-destructively quantify small molecules in a media sample, and thus is
  an attractive method that could be used for inline, real-time monitoring.
\item[Flow and Mass Cytometry --] Flow cytometry using fluorophores has been
  used extensively for immune cell analysis, but has a practical limit of
  approximately 18 colors\cite{Spitzer2016}. Mass cytometry is analogous to
  traditional flow cytometry except that it uses heavy-metal \gls{mab}
  conjugates, which has a practical limit of over 50 markers. While mass
  cytometry is less practical than simple flow cytometers such as the BD Accuri,
  we may find that only a few markers are required to accurately predict
  performance, and thus this could easily translate to industry using relatively
  cost-effective equipment. Both of these destructively analyze the cells
  themselves, but they have the advantage in that they are measuring a direct
  property of the cells and not a secreted product.
\end{description}

% BACKGROUND what about ssRNAseq?

Upon collecting these omic datasets, determining the \glspl{cqa} becomes a
computational problem. Predictions of the final product using data collected
earlier in time can be made using any number of supervised learning techniques
(linear and non-linear regression in all its forms) which in turn can be used to
develop process control models. Unsupervised learning and dimensionality
reduction techniques such as \gls{tsne}, \gls{umap}, and
\gls{spade}\cite{Qiu2011, Qiu2017} can be performed to delineate clusters of
interesting cell types and the markers that define them.

Ultimately, identifying \glspl{cqa} will likely be an iterative process, wherein
putative \glspl{cqa} will be identified, the corresponding \glspl{cpp} will be
set and optimized to maximize products with these \glspl{cpp}, and then
additional data will be collected in the clinic as the product is tested on
various patients with different indications. Additional \glspl{cqa} may be
identified which better predict specific clinical outcomes, which can be fed
back into the process model and optimized again.

\section{Innovation}

Several aspects of the \gls{dms} platform described in this dissertation are
novel considering the state-of-the-art technology for T cell manufacturing:

\begin{itemize}
\item \Glspl{dms} offers a compelling alternative to state-of-the-art magnetic
  bead technologies (e.g. DynaBeads, MACS-Beads), which is noteworthy because
  the licenses for these techniques is controlled by only a few companies
  (Invitrogen and Miltenyi respectively). Because of this, bead-based expansion
  is more expensive to implement and therefore hinders companies from entering
  the rapidly growing T cell manufacturing arena. Providing an alternative will
  provide more options for manufacturers, leading to increased competition,
  lower costs, and higher innovation in the T cell manufacturing space.
\item This is the first technology for T cell immunotherapies that selectively
  expands memory T cell populations with greater efficiency relative to
  bead-based expansion Others have demonstrated methods that can achieve greater
  expansion of T cells, but not necessarily specific populations that are known
  to be potent.
\item We used \glspl{doe} to discover and validate novel \glspl{cpp}, which is a
  strategy commonly used in non-biological industries but has yet to gain wide
  usage in the development of cell therapies where research often employs a
  one-factor-at-a-time approach. We believe this method is highly relevant to
  the development of cell therapies, not only for process optimization but also
  hypotheses generation. Furthermore, it is a perfectly natural strategy to use
  even at small scale, where the cost of reagents, cells, and materials often
  precludes large sample sizes.
\item The \gls{dms} system is be compatible with static bioreactors such as the
  G-Rex which has been adopted throughout the cell therapy industry. Thus this
  technology can be easily incorporated into existing cell therapy process that
  are performed at scale.
\item We analyzed our system using a multiomics approach, which will enable the
  discovery of novel biomarkers to be used as \glspl{cqa}. While this approach
  has been applied to T cells previously, it has not been done in the context of
  a large \gls{doe}-based model. This approach is aware of the whole design
  space, and thus enables greater understanding of process parameters and their
  effect on cell phenotype.
\end{itemize}

\chapter{AIM 1}\label{aim1}

\section{Introduction}

The first aim was to develop a microcarrier system that mimics several key
aspects of the \invivo{} lymph node microenvironment. We compared compare this
system to state-of-the-art T cell activation technologies for both expansion
potential and memory cell formation. The governing hypothesis was that
microcarriers functionalized with \acd{3} and \acd{28} \glspl{mab} will
provide superior expansion and memory phenotype compared to state-of-the-art
bead-based T cell expansion technology\footnote{adapted from \dmspaper{}}.

\section{Methods}

\subsection{DMS Functionalization}\label{sec:dms_fab}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/dms_flowchart.png}

  \endgroup
  \caption[\gls{dms} Flowchart]{Overview of \gls{dms} manufacturing process.}
  \label{fig:dms_flowchart}
\end{figure*}

\product{\gls{cus}}{\gehc}{DG-2001-OO} or \product{\gls{cug}}{\gehc}{DG-0001-OO}
were suspended at \SI{20}{\mg\per\ml} in 1X \gls{pbs} and autoclaved. All
subsequent steps were done aseptically, and all reactions were carried out at
\SI{20}{\mg\per\ml} carriers at room temperature and agitated using an orbital
shaker with a \SI{3}{\mm} orbit diameter. After autoclaving, the microcarriers
were washed using sterile \gls{pbs} three times in a 10:1 volume ratio.
\product{\Gls{snb}}{\thermo}{21217} was dissolved at approximately \SI{10}{\uM}
in sterile ultrapure water, and the true concentration was then determined using
the \gls{haba} assay (see below). \SI{5}{\ul\of{\ab}\per\mL} \gls{pbs} was added
to carrier suspension and allowed to react for \SI{60}{\minute} at
\SI{700}{\rpm} of agitation. After the reaction, the amount of biotin remaining
in solution was quantified using the \gls{haba} assay (see below). The carriers
were then washed three times, which entailed adding sterile \gls{pbs} in a 10:1
volumetric ratio, agitating at \SI{900}{\rpm} for \SI{10}{\minute}, adding up to
a 15:1 volumetric ratio (relative to reaction volume) of sterile \gls{pbs},
centrifuging at \SI{1000}{\gforce} for \SI{1}{\minute}, and removing all liquid
back down to the reaction volume.

To coat with \gls{stp}, \SI{40}{\ug\per\mL} \product{\gls{stp}}{Jackson
  Immunoresearch}{016-000-114} was added and allowed to react for
\SI{60}{\minute} at \SI{700}{RPM} of agitation. After the reaction, supernatant
was taken for the \product{\gls{bca} assay}{\thermo}{23225}, and the carriers
were washed analogously to the previous wash step to remove the biotin, except
two washes were done and the agitation time was \SI{30}{\minute}. Biotinylated
\glspl{mab} against human CD3 \catnum{\bl}{317320} and CD28 \catnum{\bl}{302904}
were combined in a 1:1 mass ratio and added to the carriers at
\SI{0.2}{\ug\of{\ab}\per\mg\of{\dms}}. Along with the \glspl{mab}, sterile
\product{\gls{bsa}}{\sigald}{A9576} was added to a final concentration of
\SI{2}{\percent} in order to prevent non-specific binding of the antibodies to
the reaction tubes. \glspl{mab} were allowed to bind to the carriers for
\SI{60}{\minute} with \SI{700}{\rpm} agitation. After binding, supernatants were
sampled to quantify remaining \gls{mab} concentration using an
\product{\anti{\gls{igg}} \gls{elisa} kit}{Abcam}{157719}. Fully functionalized
\glspl{dms} were washed in sterile \gls{pbs} analogous to the previous washing
step to remove excess \gls{stp}. They were washed once again in the cell culture
media to be used for the T cell expansion.

\begin{table}[!h] \centering
  \caption{Properties of the microcarriers used}
  \label{tab:carrier_props}
  \input{../tables/carrier_properties.tex}
\end{table}

The concentration of the final \gls{dms} suspension was found by taking a
\SI{50}{\uL} sample, plating in a well, and imaging the entire well. The image
was then manually counted to obtain a concentration. Surface area for
\si{\ab\per\um\squared} was calculated using the properties for \gls{cus} and
\gls{cug} as given by the manufacturer \cref{tab:carrier_props}.

\subsection{DMS Quality Control Assays}

Biotin was quantified using the \product{\gls{haba} assay}{\sigald}{H2153-1VL}.
In the case of quantifying \gls{snb} prior to adding it to the microcarriers,
the sample volume was quenched in a 1:1 volumetric ratio with \SI{1}{\molar}
NaOH and allowed to react for \SI{1}{\minute} in order to prevent the reactive
ester linkages from binding to the avidin proteins in the \gls{haba}/avidin
premix. All quantifications of \gls{haba} were performed on an Eppendorf D30
Spectrophotometer using \product{\SI{70}{\ul} cuvettes}{BrandTech}{759200}. The
extinction coefficient at \SI{500}{\nm} for \gls{haba}/avidin was assumed to be
\SI{34000}{\per\cm\per\molar}.

The \gls{stp} binding to the microcarriers was quantified indirectly using a
\product{\gls{bca} kit}{\thermo}{23227} according to the manufacturer’s
instructions, with the exception that the standard curve was made with known
concentrations of purified \gls{stp} instead of \gls{bsa}. Absorbance at
\SI{592}{\nm} was quantified using a BioTek plate reader.

The \gls{mab} binding to the microcarriers was quantified indirectly using an
\gls{elisa} assay per the manufacturer’s instructions, with the exception that
the same \glspl{mab} used to coat the carriers were used as the standard for the
\gls{elisa} standard curve.

Open biotin binding sites on the \glspl{dms} after \gls{stp} coating was
quantified indirectly using \product{\gls{fitcbt}}{\thermo}{B10570}.
Briefly, \SI{400}{\pmol\per\ml} \gls{fitcbt} were added to \gls{stp}-coated
carriers and allowed to react for \SI{20}{\minute} at room temperature under
constant agitation. The supernatant was quantified against a standard curve of
\gls{fitcbt} using a BioTek plate reader.

\Gls{stp} binding was verified after the \gls{stp}-binding step visually by
adding \gls{fitcbt} to the \gls{stp}-coated \glspl{dms}, resuspending in
\SI{1}{\percent} agarose gel, and imaging on a \product{lightsheet
  microscope}{Zeiss}{Z.1}. Overall \gls{mab} binding was verified visually
by first staining with \product{\anti{\gls{igg}}-\gls{fitc}}{\bl}{406001},
incubating for \SI{30}{\minute}, washing with \gls{pbs}, and imaging on a
confocal microscope.

\subsection{T Cell Culture}\label{sec:tcellculture}

Cryopreserved primary human T cells were either obtained as sorted
\product{\cdp{3} T cells}{Astarte Biotech}{1017} or isolated from
\product{\glspl{pbmc}}{Zenbio}{SER-PBMC} using a negative selection
\product{\cdp{3} \gls{macs} kit}{\miltenyi}{130-096-535}. T cells were activated
using \glspl{dms} or \product{\SI{3.5}{\um} CD3/CD28 magnetic
  beads}{\miltenyi}{130-091-441}. In the case of beads, T cells were activated
at the manufacturer recommended cell:bead ratio of 2:1. In the case of
\glspl{dms}, cells were activated using \SI{2500}{\dms\per\cm\squared} unless
otherwise noted. Initial cell density was \SIrange{2e6}{2.5e6}{\cell\per\ml} to
in a 96 well plate with \SI{300}{\ul} volume. Serum-free media was either
\product{OpTmizer}{\thermo}{A1048501} or
\product{TexMACS}{\miltenyi}{170-076-307} supplemented with
\SIrange{100}{400}{\IU\per\ml} \product{\gls{rhil2}}{Peprotech}{200-02} unless
otherwise noted. Cell cultures were expanded for \SI{14}{\day} as counted from
the time of initial seeding and activation. Cell counts and viability were
assessed using \product{trypan blue}{\thermo}{T10282} or
\product{\gls{aopi}}{Nexcelom Bioscience}{CS2-0106-5} and a \product{Countess
  Automated Cell Counter}{Thermo Fisher}{Countess 3 FL}. Media was added to
cultures every \SIrange{2}{3}{\day} depending on media color or a
\SI{300}{\mg\per\deci\liter} minimum glucose threshold. Media glucose was
measured using a \product{GlucCell glucose meter}{Chemglass}{CLS-1322-02}.

Cells on the \glspl{dms} were visualized by adding \SI{0.5}{\ul}
\product{\gls{stppe}}{\bl}{405204} and \SI{2}{ul}
\product{\acd{45}-\gls{af647}}{\bl}{368538}, incubating for \SI{1}{\hour}, and
imaging on a spinning disk confocal microscope.

In the case of Grex bioreactors, we either used a \product{24 well plate}{Wilson
  Wolf}{P/N 80192M} or a \product{6 well plate}{Wilson Wolf}{P/N 80240M}.

\subsection{Quantifying Cells on DMS Interior}

To visualize T cells on the interior of the \glspl{dms}, we stained them with
\gls{mtt}. \glspl{dms} with attached and loosely attached cells were sampled as
desired and filtered through a \SI{40}{\um} cell strainer. While still in the
cell strainer, \glspl{dms} were washed twice with \gls{pbs} and then dried by
pulling liquid through the bottom of the cell strainer via a micropipette and
dabbing with a KimWipe. \glspl{dms} were transferred to a 24 well plate with
\SI{400}{\ul} media. \SI{40}{\ul} \gls{mtt} was added to each well and allowed
to incubate for \SI{3}{\hour}, after which \glspl{dms} with cell were visualized
via a brightfield microscope.

To quantify cells on the interior of \glspl{dms}, cells and \glspl{dms} were
isolated analogously to those for the \gls{mtt} stain up until the drying step.
Cells were then transferred to a tube containing \SI{400}{\ul} at
\SI{5}{\mg\per\ml} dispase solution. \glspl{dms} were incubated and rotated for
\SI{45}{\minute} at \SI{37}{\degreeCelsius}, after which cells were counted as
already described in \cref{sec:tcellculture}.

\subsection{Quantification of Apoptosis Using Annexin-V}

Apoptosis was quantified using \gls{anv} according to the manufacturer's
instructions. Briefly, cells were transferred to flow tubes and washed twice
with \gls{pbs} by adding \SI{3}{\ml} to each tube, centrifuging for
\SI{400}{\gforce}, and aspirating the liquid down to \SI{200}{\ul}. The cells
were analogously washed a third time with staining buffer (\SI{10}{\mM}
\gls{hepes}, \SI{140}{\mM} NaCl, \SI{2.5}{\mM} \ce{CaCl2}) and aspirated down to
a final volume of \SI{100}{\ul}. Cells were stained in this volume with
\SI{1}{\ul} \product{\gls{anv}-\gls{fitc}}{\bl}{640906} and \SI{5}{\ul}
\product{\gls{pi}}{\thermo}{P3566} and incubated for \SI{15}{\minute} at
\gls{rt} in the dark. After incubation, \SI{400}{\ul} staining buffer was added
to each tube. Each tube was then analyzed via flow cytometry.

\subsection{Quantification of Caspase-3/7}

\Gls{cas37} was quantified using \product{CellEvent dye}{\thermo}{C10723}
according the manufacturer's instructions. Briefly, a 2X (\SI{8}{\mM}) working
solution of CellEvent dye was added to \SI{100}{\ul} cell suspension (at least
\num{5e4} cells) and incubated at \SI{37}{\degreeCelsius} for \SI{30}{\minute}.
After incubation, cells were immediately analyzed via flow cytometry.

\subsection{Quantification of BCL-2}

\Gls{bcl2} was quantified using an \product{Human Total Bcl-2 DuoSet \gls{elisa}
  kit}{Rnd Systems}{DYC827B-2} according to the manufacturer's instructions and
supplemented with \product{\gls{tmb} substrate
  solution}{eBioscience}{00-4201-56}, \product{5X diluent buffer}{\bl}{421203},
and \SI{2}{\normal} \ce{H2SO4} stop solution made in house. Briefly, cells were
lysed using \product{10X lysis buffer}{Cell Signaling}{9803S}, and the lysate
was quantified for protein using a \product{\gls{bca} assay}{\thermo}{23225} as
directed. Standardized lysates were measured using the \gls{elisa} kit as
directed.

\subsection{Chemotaxis Assay}

Migratory function was assayed using a transwell chemotaxis assay as previously
described\cite{Hromas1997}. Briefly, \SI{3e5}{\cell} were loaded into a
\product{transwell plate with \SI{5}{\um} pore size}{Corning}{3421} with the
basolateral chamber loaded with \SI{600}{\ul} media and 0, 250, or
\SI{1000}{\ng\per\mL} \product{CCL21}{Peprotech}{250-13}. The plate was
incubated for \SI{4}{\hour} after loading, and the basolateral chamber of each
transwell was quantified for total cells using \product{countbright
  beads}{\thermo}{C36950}. The final readout was normalized using the
\SI{0}{\ng\per\mL} concentration as background.

\subsection{Degranulation Assay}

Cytotoxicity of expanded \gls{car} T cells was assessed using a degranulation
assay as previously described\cite{Schmoldt1975}. Briefly, \num{3e5} T cells
were incubated with \num{1.5e5} target cells consisting of either \product{K562
  wild type cells}{ATCC}{CCL-243} or CD19- expressing K562 cells transformed
with \gls{crispr} (kindly provided by Dr.\ Yvonne Chen, UCLA)\cite{Zah2016}.
Cells were seeded in a flat bottom 96 well plate with \SI{1}{\ug\per\ml}
\product{\acd{49d}}{eBioscience}{16-0499-81}, \SI{2}{\micro\molar}
\product{monensin}{eBioscience}{ 00-4505-51}, and \SI{1}{\ug\per\ml}
\product{\acd{28}}{eBioscience}{302914} (all functional grade \glspl{mab}) with
\SI{250}{\ul} total volume. After \SI{4}{\hour} incubation at
\SI{37}{\degreeCelsius}, cells were stained for CD3, CD4, and CD107a and
analyzed on a \bd{} LSR Fortessa. Readout was calculated as the percent
\cdp{107a} cells of the total \cdp{8} fraction.

\subsection{CAR Expression}

\gls{car} expression of the \anti{CD19} \gls{car} was quantified as previously
described\cite{Zheng2012}. Briefly, cells were washed once and stained with
\product{biotinylated \gls{ptnl}}{\thermo}{29997}. After a subsequent wash,
cells were stained with \product{\gls{pe}-\gls{stp}}{\bl}{405204}, washed again,
and analyzed on a \bd{} Accuri. Readout was percent \gls{pe}+ cells as compared
to secondary controls (\gls{pe}-\gls{stp} with no \gls{ptnl}).

\gls{car} expression of the \anti{\gls{bcma}} \gls{car} was quantified using a
\product{\gls{fitc}-labeled \gls{bcma} protein}{Acro}{Bca-hf254}. \SI{100}{\ng}
was added to tubes analogously to \gls{ptnl} and incubated for \SI{45}{\minute}
prior to analyzing on a \bd{} Accuri

\subsection{CAR Plasmid and Lentiviral Transduction}

The anti-CD19-CD8-CD137-CD3$\upzeta$ \gls{car} with the EF1$\upalpha$
promotor\cite{Milone2009} was synthesized (Aldevron) and subcloned into a
\product{FUGW}{Addgene}{14883} kindly provided by the Emory Viral Vector Core.
Lentiviral vectors were synthesized by the Emory Viral Vector Core or the
Cincinnati Children's Hospital Medical Center Viral Vector Core. RNA titer was
calculated using a \product{Lenti-X \gls{qpcr} titer kit}{Takara}{631235}. To
transduce primary human T cells, \product{retronectin}{Takara}{T100A} was coated
onto non-TC treated 96 well plates and used to immobilize lentiviral vector
particles according to the manufacturer's instructions. Briefly, retronectin
solution was adsorbed overnight at \SI{4}{\degreeCelsius} and blocked the next
day using \gls{bsa}. Prior to transduction, lentiviral supernatant was
spinoculated at \SI{2000}{\gforce} for \SI{2}{\hour} at \SI{4}{\degreeCelsius}.
T cells were activated in 96 well plates using beads or \glspl{dms} for
\SI{24}{\hour}, and then cells and beads/\glspl{dms} were transferred onto
lentiviral vector coated plates and incubated for another \SI{24}{\hour}. Cells
and beads/\glspl{dms} were removed from the retronectin plates using vigorous
pipetting and transferred to another 96 well plate wherein expansion continued.

% METHOD fill in missing product numbers
\gls{bcma} \gls{car} lentiviral vector was synthesized in house as
follows\footnote{lentiviral synthesis was performed by Ritika Jain in our
  laboratory and included here for reference}. \SI{10}{\ng} of
\anti{\gls{bcma}}-CD8-CD137-CD3$\upzeta$ plasmid (generously provided by Jim
Kochenderfer at the NIH)\cite{Lam2020} was added to \SI{50}{\ul}
\product{DH5$\upalpha$ cells}{\thermo}{18265017} and incubated for
\SI{30}{\minute} on ice. The cell mixture was then heat-shocked at
\SI{42}{\degreeCelsius} for \SI{20}{\minute} before being placed on ice for
another \SI{2}{\minute}. \SI{950}{\ul} luria broth was added to the cells which
were then centrifuged for \SI{1}{\hour} at \SI{225}{\rpm}. \SI{20}{\ul} of the
cell mixture was then spread over selection plates and incubated overnight at
\SI{37}{\degreeCelsius}. Colonies were selected the following day and incubated
in luria broth with \product{ampicillin}{\sigald{}}{A9518-5G} at
\SI{37}{\degreeCelsius} for \SIrange{12}{16}{\hour} prior to using the
\product{miniprep kit}{Qiagen}{27104} as per the manufacturer's instructions to
isolate the plasmid DNA. Transfer plasmid along with
\product{pMDLg/pRRE}{Addgene}{12251}, \product{pRSV-Rev}{Addgene}{12253}, and
\product{pMD2.G}{Addgene}{12259} (generously provided by the Sloan lab at Emory
University) in \product{Opti-Mem}{\thermo}{31-985-070} with
\product{lipfectamine 2000}{\thermo}{11668019} were added dropwise to HEK 293T
cells and incubated for \SI{6}{\hour}, after which all media was replaced with
fresh fresh media. After \SI{24}{\hour} and \SI{48}{\hour}, supernatent was
collected, pooled, and concentrated using a \product{Lenti-X
  concentrator}{Takara}{631231} prior to storing at \SI{-80}{\degreeCelsius}.

\subsection{Sulfo-NHS-Biotin Hydrolysis Quantification}

The equation for hydrolysis of \gls{snb} to biotin and \gls{nhs} is given by
\cref{chem:snb_hydrolysis}.

\begin{equation}
  \label{chem:snb_hydrolysis}
  \ce{NHS-CO-Biotin + OH- -> NHS- + Biotin-COOH}
\end{equation}

Measuring the hydrolysis of \gls{snb} was performed spectroscopically. \gls{snb}
was added to either \gls{di} water or \gls{pbs} in a UV-transparent 96 well
plate. Kinetic analysis using a BioTek plate reader began immediately after
prep, and readings at \SI{260}{\nm} were taken every minute for \SI{2}{\hour}.

\subsection{Reaction Kinetics Quantification}

The diffusion of \gls{stp} into biotin-coated microcarriers was determined
experimentally. \SI{40}{\ug\per\ml} \gls{stp} was added to multiple batches of
biotin-coated microcarriers, and supernatents were taken at fixed intervals and
quantified for \gls{stp} protein using the \gls{bca} assay.

The geometric diffusivity of the microcarriers was determined using a
pseudo-steady-state model. Each microcarrier was assumed to be a porous sphere
with a fixed number of uniformly distributed `\gls{stp} binding sites' equal to
the number of \gls{stp} molecules experimentally determined to bind to the
microcarriers. Because the reaction rate between biotin and \gls{stp} is so fast
(it is the strongest non-covalent bond in known existence), we assumed that the
interface of free biotin shrunk as a function of \gls{stp} diffusing to the
unbound biotin interface until the center of the microcarriers was reached. We
also assumed that the pores in the microcarriers were large enough that the
interactions between the \gls{stp} and surfaces would be small, thus the
geometric diffusivity could be represented as a fraction of the diffusion
coefficient of \gls{stp} in water. This model was given by
\cref{eqn:stp_diffusion_1,eqn:stp_diffusion_2}:

% TODO actually derive these equations, eg state the initial conditions and
% governing equation
\begin{equation}
  \label{eqn:stp_diffusion_1}
  \frac{dr}{dt} = \frac{-D_{app}C_b}{Br(1-r/R)}
\end{equation}

\begin{equation}
  \label{eqn:stp_diffusion_2}
  \frac{dC_b}{dt} = \frac{-4 \pi n D_{app} C_b}{V(1/r-1/R)}
\end{equation}

\noindent where
\begin{itemize}[label={}]
\item $D_{app}$ is the apparent diffusion rate of species $X$ which is equal to
  $D\beta$
\item $D$ the diffusion rate of species $X$ in water at room temperature
  (where $X$ is \gls{stp} in this example and \glspl{mab} later in this section)
\item $\beta$ a fractional parameter representing the tortuousity and void
  fraction of the microcarriers (here called the `geometric diffusivity')
\item $r$ is the interfatial radius of the unbound binding sites for species $X$
  within a microcarrier
\item $t$ is the reaction time
\item $C_b$ is the concentration of species $X$ in the bulk solution
\item $V$ is the volume of the bulk medium
\item $R$ is the average radius of the microcarriers
\item $n$ is the number of microcarriers in the reaction volume
\end{itemize}

The diffusion rate of \gls{stp} was assumed to be
\SI{6.2e-7}{\cm\squared\per\second}\cite{Kamholz2001}. Since all but $\beta$ was
known, the experimental data was fit using these equations using
\inlinecode{ode45} in MATLAB and least squares as the fitting error. These
equations were then used analogously to describe the reaction profile of
\glspl{mab} assuming a diffusion rate of
\SI{4.8e-7}{\cm\squared\per\second}\cite{Sherwood1992}.

To model the washing of the microcarriers, they once again were assumed to be
porous spheres filled with whatever amount of reagent was left unbound from the
previous step (which was assumed to be equal to concentration in the
supernatent). The diffusion out of the microcarriers is given by the following
partial differential equation and boundary conditions:

\begin{equation}
  \label{eqn:stp_washing}
  \frac{\partial C_i}{\partial t} = \frac{1}{r^2}\frac{\partial}{\partial
    r}\left(r^2 D_{app} \frac{\partial C_i}{\partial r}\right)
\end{equation}

\begin{equation}
  \label{eqn:stp_washing_left_bc}
  C_i(r, 0) = C_{i,0}
\end{equation}

\begin{equation}
  \label{eqn:stp_washing_left_bc}
  N_i(0, t) = 0
\end{equation}

\begin{equation}
  \label{eqn:stp_washing_right_bc}
  C_i(R, t) = (C_{b,0}+C_{b,\infty}) / 2
\end{equation}

\noindent where (in addition to the variables given already for
\cref{eqn:stp_diffusion_1,eqn:stp_diffusion_2})
\begin{itemize}[label={}]
\item $N_i$ is the radial flux of species $X$ inside the microcarriers
\item $C_i$ is the concentration of species $X$ inside the microcarriers
\item $C_{i,0}$ is the initial concentration of species $X$ inside
  the microcarriers (which is assumed to be the concentration in the bulk before
  the wash volume is added)
\item $C_{b,0}$ is the initial bulk concentration of species $X$ outside the
  microcarriers after the initial wash volume has been added
\item $C_{b,\infty}$ is the final bulk concentration of species $X$ outside the
  microcarriers
\end{itemize}

Note that in order to avoid solving a moving boundary value problem, the
concentration at the boundary of the microcarriers was fixed at the average of
the final and initial concentration expected to be observed in bulk. This should
be a reasonable assumption given that the volume inside the microcarriers is
tiny compared to the amount of volume added in the wash, thus the boundary
concentration should change little.

The same diffusion coefficients were used in determining the kinetics of the
washing steps, and \SI{5.0e-6}{\cm\squared\per\second}\cite{Niether2020} was
used as the diffusion coefficient for free biotin (which should be the only
reactive species left in solution after all the \gls{snb} has hydrolyzed).

All diffusion coefficients were taken to be valid at \gls{rt} and in \gls{di}
water, which is a safe assumption given that our reaction medium was 1X
\gls{pbs}.

See \cref{sec:appendix_binding} and \cref{sec:appendix_washing} for the MATLAB
code (and output in the case of the washing steps) used.

\subsection{Luminex Analysis}\label{sec:luminex_analysis}

Luminex was performed using a \product{ProcartaPlex kit}{\thermo}{custom} for
the markers outlined in \cref{tab:luminex_panel} with modifications (note that
some markers were run in separate panels to allow for proper dilutions).
Briefly, media supernatents from cells were sampled as desired and immediately
placed in a \SI{-80}{\degreeCelsius} freezer until use. Before use, samples were
thawed at \gls{rt} and vortexed to ensure homogeneity. To run the plate,
\SI{25}{\ul} of magnetic beads were added to the plate and washed 3X using
\SI{300}{\ul} of wash buffer. \SI{25}{\ul} of samples or standard were added to
the plate and incubated for \SI{120}{\minute} at \SI{850}{\rpm} at \gls{rt}
before washing analogously 3X with wash buffer. \SI{12.5}{\ul} detection
\glspl{mab} and \SI{25}{\ul} \gls{stppe} were sequentially added, incubated for
\SI{30}{\minute} and vortexed, and washed analogously to the sample step.
Finally, samples were resuspended in \SI{120}{\ul} reading buffer and analyzed
via a BioRad Bioplex 200 plate reader. An 8 point log\textsubscript{2} standard
curve was used, and all samples were run with single replicates.

Luminex data was preprocessed using R for inclusion in downstream analysis as
follows. Any cytokine level that was over-range (`OOR >' in output spreadsheet)
was set to the maximum value of the standard curve for that cytokine. Any value
that was under-range (`OOR <' in output spreadsheet) was set to zero. All values
that were extrapolated from the standard curve were left unchanged.

\begin{table}[!h] \centering
  \caption{Luminex Panel}
  \label{tab:luminex_panel}
  \input{../tables/luminex_panel.tex}
\end{table}

\subsection{Data Aggregation and Meta-Analysis}

In order to perform meta-analysis on all experimental data generate using the
\gls{dms} system, we developed a program to curate and aggregate the raw
datafiles into a \gls{sql} database (\cref{sec:appendix_meta}).

The data files to be aggregated included Microsoft Excel files which held
timeseries measurements for cell cultures (eg cell counts, viability, glucose,
\gls{il2} added, media added, and media removed), \gls{fcs} files for cellular
phenotypes, and FlowJo files which held gating parameters and statistics based
on the \gls{fcs} files. Additional information which was held in electronic lab
notebooks (eg OneNote files) was not easily parsable, and thus this data was
summarized in YAML files. The data included in these YAML files included reagent
characteristics (vendor, catalog number, lot number, manufacturing date), cell
donor characteristics (age, \gls{bmi}, phenotype, demographic, gender), and all
experimental parameters such as the number of bead or \gls{dms} added.

To aggregate the data in a database, we wrote a program using Python, R, and
Docker which retrieved the data from its source location and inserted the data
in a Postgres database (specifically Aurora implementation hosted on \gls{aws}).
This program included checks to ensure the integrity of source data (for
example, flagging entries which had a reagent whose manufacturing date was after
the date the experiment started, which signifies a human input error).


\subsection{Statistical Analysis}\label{sec:statistics}

For 1-way \gls{anova} analysis with Tukey multiple comparisons test,
significance was assessed using the \inlinecode{stat\_compare\_means} function
with the \inlinecode{t.test} method from the \inlinecode{ggpubr} library in R.
For 2-way \gls{anova} analysis, the significance of main and interaction effects
was determined using the car library in R.

For least-squares linear regression, statistical significance was evaluated the
\inlinecode{lm} function in R. All results with categorical variables are
reported relative to baseline reference. Each linear regression was assessed for
validity using residual plots (to assess constant variance and independence
assumptions), QQplots and Shapiro-Wilk normality test (to assess normality
assumptions), Box-Cox plots (to assess need for power transformations), and
lack-of-fit tests where replicates were present (to assess model fit in the
context of pure error). Significance was evaluated at $\upalpha$ = 0.05.

\subsection{Flow Cytometry}\label{sec:flow_cytometry}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/gating_strategy.png}

  \endgroup
  \caption[Gating Strategy]
  {Gating strategy for quantifying \ptmemp{}, \pthp{}, and \ptcarp{}.}
  \label{fig:gating_strategy}
\end{figure*}

\begin{table}[!h] \centering
  \caption{\glspl{mab} used for flow cytometry}
  \label{tab:flow_mabs}
  \input{../tables/flow_mabs.tex}
\end{table}

All \glspl{mab} used for flow cytometry are shown in \cref{tab:flow_mabs}. Other
reagents for specialized assays such as degranulation are described in their
respective sections. Cells were gated according to \cref{fig:gating_strategy}.

\section{Results}

\subsection{DMSs Can be Fabricated in a Controlled Manner}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/dms_coating.png}
  \phantomsubcaption\label{fig:cug_vs_cus}
  \phantomsubcaption\label{fig:biotin_coating}
  \phantomsubcaption\label{fig:stp_coating}
  \phantomsubcaption\label{fig:mab_coating}
  \phantomsubcaption\label{fig:stp_carrier_fitc}
  \phantomsubcaption\label{fig:mab_carrier_fitc}

  \endgroup
  \caption[\gls{dms} Coating]
  {\gls{dms} functionalization results.
    \subcap{fig:cug_vs_cus}{Bound \gls{stp} surface
      density on either \gls{cus} or \gls{cug} microcarriers. Surface density
      was estimated using the properties in~\cref{tab:carrier_props}}.
    Total binding curve of \subcap{fig:biotin_coating}{biotin},
    \subcap{fig:stp_coating}{\gls{stp}}, and
    \subcap{fig:mab_coating}{\glspl{mab}} as a function of biotin added for
    batches manufactured on different dates.
    \subcap{fig:stp_carrier_fitc}{\gls{stp}-coated or uncoated \glspl{dms}
      treated with \gls{fitcbt} and imaged using a lightsheet microscope.}
    \subcap{fig:mab_carrier_fitc}{\gls{mab}-coated or \gls{stp}-coated
      \glspl{dms} treated with \anti{\gls{igg}} \glspl{mab} and imaged using a
      lightsheet microscope.}
    }
  \label{fig:dms_coating}
\end{figure*}

Two types of gelatin-based microcariers, \gls{cus} and \gls{cug}, were
covalently conjugated with varying concentration of \gls{snb} and then coated
with \gls{stp} and \glspl{mab} to make \glspl{dms}. Aside from slight
differences in swelling ratio and crosslinking chemistry\cite{purcellmain} the
properties of \gls{cus} and \gls{cug} were the same (\cref{tab:carrier_props}).
We chose to continue with the \gls{cus}-based \glspl{dms}, which showed higher
overall \gls{stp} binding compared to \gls{cug}-based \glspl{dms}
(\cref{fig:cug_vs_cus}). We showed that by varying the concentration of
\gls{snb}, we were able to precisely control the amount of attached biotin
(\cref{fig:biotin_coating}), mass of attached \gls{stp}
(\cref{fig:stp_coating}), and mass of attached \glspl{mab}
(\cref{fig:mab_coating}). Furthermore, we showed that the microcarriers were
evenly coated with \gls{stp} on the surface and throughout the interior as
evidenced by the presence of biotin-binding sites occupied with \gls{fitcbt} on
the microcarrier surfaces after the \gls{stp}-coating step
(\cref{fig:stp_carrier_fitc}). Finally, we confirmed that biotinylated
\glspl{mab} were bound to the \glspl{dms} by staining either \gls{stp}- or
\gls{stp}/\gls{mab}-coated carriers with \antim{\gls{igg}-\gls{fitc}} and
imaging on a confocal microscope (\cref{fig:mab_carrier_fitc}). Taking this
together, we noted that the maximal \gls{mab} binding capacity occurred near
\SI{50}{\nmol} biotin input (which corresponded to
\SI{2.5}{\nmol\per\mg\of{\dms}}) thus we used this in subsequent processes.

We then asked how sensitive the \gls{dms} manufacturing process was to a variety
of variables. In particular, we focused on the biotin-binding step, since it
appeared that the \gls{mab} binding was quadratically related to biotin binding
(\cref{fig:mab_coating}) and thus controlling the biotin binding step would be
critical to controlling the amount and \glspl{mab} and thus the amount of signal
the T cells receive downstream.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/dms_qc.png}
  \phantomsubcaption\label{fig:dms_qc_doe}
  \phantomsubcaption\label{fig:dms_qc_ph}
  \phantomsubcaption\label{fig:dms_qc_washes}
  \phantomsubcaption\label{fig:dms_snb_decay_curves}

  \endgroup
  \caption[\gls{dms} Process Parameters]
  {Investigation of influential parameters for the \gls{dms} process.
    \subcap{fig:dms_qc_doe}{\gls{doe} investigating the effect of initial mass
      of microcarriers, reaction temperature, and biotin concentration on
      biotin attachment.}
    \subcap{fig:dms_qc_ph}{Effect of reaction ph on biotin attachment.}
    \subcap{fig:dms_qc_washes}{effect of post-autoclave washing of the
      microcarriers on biotin attachment.}
    \subcap{fig:dms_snb_decay_curves}{Hydrolysis curves of \gls{snb} in
      \gls{pbs} or \gls{di} water.}
    All statistical tests where p-values are noted are given by two-tailed t
    tests.
  }
  \label{fig:dms_qc}
\end{figure*}

To answer this question, we first performed a \gls{doe} to understand the effect
of reaction parameters on biotin binding. The parameters included in this
\gls{doe} were temperature, microcarrier mass, and \gls{snb} input mass. These
were parameters that we specifically controlled but hypothesized might have some
sensitivity on the final biotin mass attachment rate depending on their noise
and uncertainty. In particular, temperature was `controlled' only by allowing
the microcarrier suspension to come to \gls{rt} after autoclaving. After
performing a full factorial \gls{doe} with three center points as the target
reaction conditions, we found that the final biotin binding mass is only highly
dependent on biotin input concentration (\cref{fig:dms_qc_doe}). Overall,
temperature had no effect and carrier mass had no effect at higher temperatures,
but carrier mass had a slightly positive effect when temperature was low. This
could be because lower temperature might make spontaneous decay of \gls{snb}
occur slower, which would give \gls{snb} molecule more opportunity to diffuse
into the microcarriers and react with amine groups to form attachments. It seems
that concentration only has a linear effect and doesn't interact with any of the
other variables, which is not surprisingly given the behavior observed in
(\cref{fig:biotin_coating})

We also observed that the reaction pH does not influence the amount of biotin
attached (\cref{fig:dms_qc_ph}), which indicates that while higher pH might
increase the number of deprotonated amines on the surface of the microcarrier,
it also increases the number of \ce{OH-} groups which can spontaneously
hydrolyze the \gls{snb} in solution (\cref{chem:snb_hydrolysis}).

Furthermore, we observed that washing the microcarriers after autoclaving
increases the biotin binding rate (\cref{fig:dms_qc_washes}). While we did not
explore this further, one possible explanation for this behavior is that the
microcarriers have some loose protein in the form of powder or soluble peptides
that competes for \gls{snb} binding against the surface of the microcarriers,
and when measuring the supernatent using the \gls{haba} assay, these soluble or
lightly-suspended peptides/protein fragments are also measured and therefore
inflate the readout.

Lastly, we asked what the effect on reaction pH had on spontaneous degradation
of \gls{snb} while in solution (\cref{chem:snb_hydrolysis}). If the \gls{snb}
significantly degrades within minutes of preparation, then it is important to
carefully control the timing between \gls{snb} solution preparation and addition
to the microcarriers. We found that in the presence of \gls{di} water, \gls{snb}
is extremely stable (\cref{fig:dms_snb_decay_curves}) where it decays rapidly in
the presence of \gls{pbs} buffered to pH of 7.1. In fact, the \gls{di} water
curve actually decreases slightly, possibly due to \gls{snb} absorbing to the
plate surface. \gls{snb} is known to hydrolyze in the presence of \ce{OH-}, but
the lack of hydrolysis in \gls{di} water can be explained by the fact that
biotin itself is acidic, and thus the reaction is self-inhibitory in an
unbuffered and neutral pH system. Because we dissolve our \gls{snb} in \gls{di}
water prior to adding it to the microcarrier suspension (which itself is in
\gls{pbs}) this result indicated that hydrolysis is not of concern when adding
\gls{snb} within minutes.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/dms_timing.png}
  \phantomsubcaption\label{fig:dms_biotin_rxn_mass}
  \phantomsubcaption\label{fig:dms_biotin_rxn_frac}
  \phantomsubcaption\label{fig:dms_stp_per_time}
  \phantomsubcaption\label{fig:dms_mab_per_time}
  \phantomsubcaption\label{fig:dms_biotin_washed}

  \endgroup
  \caption[\gls{dms} Reaction kinetics]
  {Reaction kinetics for microcarrier functionalization.
    \subcap{fig:dms_biotin_rxn_mass}{Biotin mass bound per time}
    \subcap{fig:dms_biotin_rxn_frac}{Fraction of input biotin bound per time}
    \subcap{fig:dms_stp_per_time}{\Gls{stp} bound per time. Each dot is an
      experimental run and the line is the fitted model.}
    \subcap{fig:dms_mab_per_time}{Simulated \glspl{mab} bound per time.}
    \subcap{fig:dms_biotin_washed}{Biotin quantification via the \gls{haba}
      assay after washing.}
  }
  \label{fig:dms_kinetics}
\end{figure*}

\subsection{Reaction Kinetics for Coating the DMSs}

We investigated the reaction kinetics of all three coating steps (accompanying
MATLAB codes are provided in \cref{sec:appendix_binding}). To quantify the
reaction kinetics of the biotin binding step, we reacted multiple batches of
\SI{20}{\mg\per\ml} microcarriers in \gls{pbs} at \gls{rt} with \gls{snb} in
parallel and sacrificially analyzed each at varying timepoints using the
\gls{haba} assay. This was performed at two different concentrations. We
observed that for either concentration, the reaction was over in
\SIrange{20}{30}{\minute} (\cref{fig:dms_biotin_rxn_mass}). Furthermore, when
put in terms of fraction of input \gls{snb}, we observed that the curves are
almost identical (\cref{fig:dms_biotin_rxn_frac}). Given this, the reaction step
for biotin attached can be set to \SI{30}{\minute}\footnote{we actually used
  \SI{60}{\minute} as outlined in methods, which shouldn't make any difference
  except save for being excessive according to this result}.

Next, we quantified the amount of \gls{stp} reacted with the surface of the
biotin-coated microcarriers. Different batches of biotin-coated \glspl{dms} were
coated with \SI{40}{\ug\per\ml} \gls{stp} and sampled at intermediate timepoints
using the \gls{bca} assay to indirectly quantify the amount of attached
\gls{stp} mass. We found this reaction took approximately \SI{30}{\minute}
(\cref{fig:dms_stp_per_time}). Assuming a quasi-steady-state paradigm, we used
this experimental binding data to compute the geometric diffusivity of the
microcarriers and fit a continuous model for the \gls{stp} binding reaction. We
computed the number of `binding sites' using the maximum mass observed to bind
to the \gls{dms}, which should describe the upper-bound for reaction time
(\cref{fig:stp_coating}). Using the diffusion rate of the \gls{stp}
(\SI{6.2e-7}{\cm\squared\per\second}), we then calculated the geometric
diffusivity of the microcarriers to be 0.190 (see
\cref{eqn:stp_diffusion_1,eqn:stp_diffusion_2}).

Using this geometric diffusivity and the known diffusion coefficient of a
\gls{mab} protein in water, we calculated the binding of \glspl{mab} per time
onto the microcarriers (this obviously assumes that the effectively diffusivity
is independent of the protein used, which should be reasonable given that the
pores of the microcarriers are huge compared to the proteins, and we don't
expect any significant reaction between the protein and the microcarrier surface
save for the \gls{stp}-biotin binding reaction). Once again, we used the maximum
number of \glspl{mab} observed to determine the number of `binding sites' for
\glspl{mab} on the microcarriers, which should correspond to the upper-bound for
the reaction time (\cref{fig:mab_coating}). According to this model, the
\gls{mab} binding reaction should be complete within \SI{75}{\minute} under the
conditions used for our protocol (\cref{fig:dms_mab_per_time})\footnote{We
  actually used \SI{60}{\minute} as describe in the method section as this model
  was not updated with new parameters until recently; however, we should point
  out that even at \SI{60}{\minute} the reaction appears to be
  >\SI{95}{\percent} complete}.

Finally, we calculated the number of wash steps needed to remove the reagents
between each step, including the time for each wash which required the geometric
diffusivity of the microcarriers as calculated above. This is important, as
failing to wash out residual free \gls{snb} (for example) could occupy binding
sites on the \gls{stp} molecules, lowering the effective binding capacity of the
\gls{mab} downstream. Each wash was a 1:15 dilution (\SI{1}{\ml} reaction volume
in a \SI{15}{\ml} conical tube), and in the case of \gls{snb} we wished to wash
out enough biotin such that less than \SI{1}{\percent} of the binding sites in
\gls{stp} would be occupied. Given this dilution factor, a maximum of
\SI{20}{\nmol} of biotin remaining \cref{fig:biotin_coating} \SI{2.9}{\nmol}
biotin binding sites on \SI{40}{\ug} \gls{stp} (assuming 4 binding sites per
\gls{stp} protein), this turned out to be 3 washes. By similar logic, using 2
washes after the \gls{stp} binding step will ensure that the number of free
\gls{stp} binding sites is less than 20X the number of \gls{mab} molecules
added\footnote{This step may benefit from an additional wash, as the number of
  washes used here was develop when \SI{40}{\ug} rather than \SI{4}{\ug}
  \gls{mab} was used to coat the \gls{dms}, yielding a much wider margin.
  However, it is also not clear to what extent this matters, as the \gls{mab}
  have multiple biotin molecules per \gls{mab} protein, and thus one \gls{mab}
  would require binding to several \gls{stp} molecules to be prevented from
  binding at all.}

To determine the length of time required for each wash, we again assumed the
microcarriers to be porous spheres, this time with an initial concentration of
\gls{snb}, \gls{stp}, or \glspl{mab} equal to the final concentration of the
bulk concentration of the previous binding step, and calculated the amount of
time it would take for the concentration profile inside the microcarriers to
equilibrate to the bulk in the wash step. Using this model, we found that the
wash times for \gls{snb}, \gls{stp}, and \glspl{mab} was \SI{3}{\minute},
\SI{15}{\minute}, and \SI{17}{\minute} respectively. We verified that the
\gls{snb} was totally undetectable after washing (\cref{fig:dms_biotin_washed}).
The other two species need to be verified in a similar manner; however, we
should not that the washing time for both the \gls{stp} and \gls{mab} coating
steps were \SI{30}{\minute}, which is a significant margin of safety (albeit
one that could be optimized).

MATLAB code and output for all the wash step calculations are given in
\cref{sec:appendix_washing}.

\subsection{DMSs can efficiently expand T cells compared to beads}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/cells_on_dms.png}
  \phantomsubcaption\label{fig:dms_cells_phase}
  \phantomsubcaption\label{fig:dms_cells_fluor}

  \endgroup
  \caption[T cells growing on \glspl{dms}]
  {Cells grow in tight clusters in and around functionalized \gls{dms}.
    \subcap{fig:dms_cells_phase}{Phase-contrast image of T cells growing on
      \glspl{dms}}
    \subcap{fig:dms_cells_fluor}{Confocal images of T cells in varying z-planes
      growing on \glspl{dms} on day 9. \Glspl{dms} were stained using
      \gls{stppe} (red) and T cells were stained using \acd{45}-\gls{af647}.}
    Images are from day 7 of culture.
  }
  \label{fig:dms_cells}
\end{figure*}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/dms_expansion.png}
  \phantomsubcaption\label{fig:dms_expansion_bead}
  \phantomsubcaption\label{fig:dms_expansion_isotype}

  \endgroup
  \caption[\glspl{dms} selectively expand T cells]
  {T cells are selectively expanded on \gls{dms}.
    \subcap{fig:dms_expansion_bead}{T cells expanded with either \glspl{dms} or
      bead for 12 days. Significance was assessed using a two-tailed
      heteroschodastic T test.}
    \subcap{fig:dms_expansion_isotype}{T cells grown on \glspl{dms} coated with
      either activating \glspl{mab} or \gls{igg} isotype control \glspl{mab}.}
  }
  \label{fig:dms_expansion}
\end{figure*}

We next sought to determine how our \glspl{dms} could expand T cells compared to
state-of-the-art methods used in industry. All bead expansions were performed as
per the manufacturer’s protocol, with the exception that the starting cell
densities were matched between the beads and carriers to
~\SI{2.5e6}{\cell\per\ml}. Throughout the culture we observed that T cells in
\gls{dms} culture grew in tight clumps on the surface of the \glspl{dms} as well
as inside the pores of the \glspl{dms}
(\cref{fig:dms_cells_phase,fig:dms_cells_fluor}). Furthermore, we observed that
the \glspl{dms} conferred greater expansion compared to traditional beads, and
significantly greater expansion after \SI{12}{\day} of culture
(\cref{fig:dms_expansion_bead}). We also observed no T cell expansion using
\glspl{dms} coated with an isotype control mAb compared to \glspl{dms} coated
with \acd{3}/\acd{28} \glspl{mab} (\cref{fig:dms_expansion_isotype}), confirming
specificity of the expansion method. Given that \il{2} does not lead to
expansion on its own, we know that the expansion of the T cells shown here is
due to the \acd{3} and \acd{28} \glspl{mab}\cite{Waysbort2013}.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/apoptosis.png}
  \phantomsubcaption\label{fig:apoptosis_annV}
  \phantomsubcaption\label{fig:apoptosis_cas}
  \phantomsubcaption\label{fig:apoptosis_bcl2}

  \endgroup
  \caption[Apoptosis Quantification for \glspl{dms}]
  {\glspl{dms} produce cells with lower apoptosis marker expression on average
    compared to bead.
    \subcap{fig:apoptosis_annV}{Quantification of apoptosis and necrosis by
      \gls{anv} and \gls{pi}.}
    \subcap{fig:apoptosis_cas}{Quantification of Caspase-3/7 expression using
      CellEvent dye.}
    \subcap{fig:apoptosis_bcl2}{Quantification of BCL-2 expression using
      \gls{elisa}. All statistical tests shown are two-tailed homoschodastic
      t-tests. All cells were harvested at day 8.}
  }
  \label{fig:dms_apoptosis}
\end{figure*}

Given that the \gls{dms} system seemed to expand T cells more effectively, we
asked if this difference was due to a reduction in apoptosis or an increase in
proliferation rate (or both). We assessed the apoptotic state of T cells grown
using either bead or \gls{dms} harvested on day 8 using \gls{pi} and \gls{anv}.
\gls{anv} is a marker which stains phospholipid phosphatidylserine, which is
usually present only on the cytoplasmic surface of the cell membrane, but flips
to the outside when the cell becomes apoptotic. \gls{pi} stains the nucleus of
the cell, but only penetrates necrotic cells which have a perforated cell
membrane. When staining for these two markers and assessing via flow cytometry,
we observe that the \gls{dms}-expanded T cells have lower frequencies of
apoptotic and necrotic cells (\cref{fig:apoptosis_annV}). Furthermore, we
stained our cultures with CellEvent dye, which is an indicator of \gls{cas37},
which is activated in apoptotic cells. In line with the \gls{pi}/\gls{anv}
results, we observed that the \gls{dms} T cells had lower frequency of
\gls{cas37} expression, indicating less apoptosis for our method
(\cref{fig:apoptosis_cas}). Finally, we lysed our cells and stained for
\gls{bcl2}, which is also upregulated in apoptosis. In this case, some (but not
all) of the bead-expanded cultures showed higher \gls{bcl2} expression, which
could indicate more apoptosis in those groups (\cref{fig:apoptosis_bcl2}). None
of the \gls{dms} cultures showed similar heightened expression. Taken together,
these data suggest that the \gls{dms} platform at least in part achieves higher
expansion by lowering apoptosis of the cells in culture.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/dms_inside.png}
  \phantomsubcaption\label{fig:dms_inside_bf}
  \phantomsubcaption\label{fig:dms_inside_regression}

  \endgroup
  \caption[A subset of T cells grow in interior of \glspl{dms}]
  {A percentage of T cells grow in the interior of \glspl{dms}.
    \subcap{fig:dms_inside_bf}{T cells stained dark with \gls{mtt} after
      growing on either coated or uncoated \glspl{dms} for 15 days visualized
      with brightfield microscope.}
    \subcap{fig:dms_inside_regression}{Linear regression performed on T cell
      percentages harvested on the interior of the \glspl{dms} vs the initial
      starting cell density.}
  }
  \label{fig:dms_inside}
\end{figure*}

\begin{table}[!h] \centering
  \caption{Regression for fraction of cells in \gls{dms} at day 14}
  \label{tab:inside_regression}
  \input{../tables/inside_fraction_regression.tex}
\end{table}

% RESULT state the CI of what are inside the carriers
We also asked how many cells were inside the \glspl{dms} vs. free-floating in
suspension and/or loosely attached to the surface. We qualitatively verified the
presence of cells inside the \glspl{dms} using a \gls{mtt} stain to opaquely
mark cells and enable visualization on a brightfield microscope
(\cref{fig:dms_inside_bf}). After seeding \glspl{dms} at different densities and
expanding for \SI{14}{\day}, we filtered the \glspl{dms} out of the cell
suspension and digested them using dispase to free any cells attached on the
inner surface. We observed that approximately \SI{15}{\percent} of the total
cells after \SI{14}{\day} were on the interior surface of the \glspl{dms}
(\cref{fig:dms_inside_regression,tab:inside_regression}). Performing linear
regression on this data revealed that the percentage of T cells inside the
\glspl{dms} does not depend on the initial starting cell density (at least when
harvested after \SI{14}{\day}) (\cref{tab:inside_regression}).

\subsection{DMSs lead to greater expansion and memory and CD4+ phenotypes}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/dms_vs_bead_expansion.png}
  \phantomsubcaption\label{fig:dms_exp_fold_change}
  \phantomsubcaption\label{fig:dms_exp_mem}
  \phantomsubcaption\label{fig:dms_exp_cd4}
  \phantomsubcaption\label{fig:dms_exp_mem4}
  \phantomsubcaption\label{fig:dms_exp_mem8}

  \endgroup
  \caption[\gls{dms} vs bead expansion]
  {\gls{dms} lead to superior expansion of T cells compared to beads across
    multiple donors.
    \subcap{fig:dms_exp_fold_change}{Longitudinal fold change of T cells grown
      using either \glspl{dms} or beads. Significance was evaulated using t
      tests at each timepoint}
    Fold change of subpopulations expanded using either \gls{dms} or beads at
    day 14, including 
    \subcap{fig:dms_exp_mem}{\ptmem{} cells},
    \subcap{fig:dms_exp_cd4}{\pth{} cells},
    \subcap{fig:dms_exp_mem4}{\ptmemh{} cells}, and
    \subcap{fig:dms_exp_mem8}{\ptmemk{} cells}. \sigkey{}
  }
  \label{fig:dms_exp}
\end{figure*}

After observing differences in expansion, we further hypothesized that the
\gls{dms} cultures could lead to a different T cell phenotype. In particular, we
were interested in the formation of \glspl{tn}, \gls{tscm}, and \glspl{tcm} as
these represent a subset with higher capacity to replicate and therefore
improved clinical prognosis\cite{Gattinoni2011, Wang2018}. We measured the
frequency of these subtypes by staining for CCR7 and CD62L. Using three donor
lots, we noted again \glspl{dms} produced more T cells over a \SI{14}{\day}
expansion than beads, with significant differences in number appearing as early
after \SI{5}{\day} (\cref{fig:dms_exp_fold_change}). Furthermore, we noted that
\glspl{dms} produced more memory/naïve cells after \SI{14}{\day} when compared
to beads for all donors (\cref{fig:dms_exp_mem,fig:dms_exp_cd4}) showing that
the \gls{dms} platform is able to selectively expand potent, early
differentiation T cells.

Of additional interest was the preservation of the CD4+ compartment. In healthy
donor samples (such as those used here), the typical CD4:CD8 ratio is 2:1. We
noted that \glspl{dms} produced more CD4+ T cells than bead cultures as well as
naïve/memory, showing that the \gls{dms} platform can selectively expand CD4 T
cells to a greater degree than beads \cref{fig:dms_exp_cd4}. The trends held
true when observing the CD4+ and CD8+ fractions of the naïve/memory subset
(\ptmem{}) (\cref{fig:dms_exp_mem4,fig:dms_exp_mem8}).

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/dms_phenotypes.png}
  \phantomsubcaption\label{fig:dms_phenotype_mem}
  \phantomsubcaption\label{fig:dms_phenotype_cd4}

  \endgroup
  \caption[Representative flow plots of \ptmem{} and \pth{} T cells]
  {Representative flow plots of \ptmem{} and \pth{} T cells at day 14 of
    expansion using either beads or \glspl{dms}. For three representative donor
    samples, phenotypes are shown for \subcap{fig:dms_phenotype_mem}{\ptmem{}}
    and \subcap{fig:dms_phenotype_cd4}{\pth}. Each population was also gated on
    \cdp{3} T cells.
  }
  \label{fig:dms_phenotype}
\end{figure*}

We also observed that, at least with the donors and conditions tested in these
experiments\footnote{these results were not always consistent, see the
  metaanalysis at the end of this aim for an in-depth quantification of this
  observation} that the fraction of \ptmem{} and \pth{} T cells was higher in
the \gls{dms} groups compared to the bead groups
(\cref{fig:dms_phenotype})\footnote{these where not the same donors as used for
  \cref{fig:dms_exp}}. This result was seen for multiple donors. We should note
that in the case of \pthp{}, the donors we used had an initial \pthp{} that was
much higher (healthy donors generally have a CD4:CD8 ratio of 2:1), so the
proper interpretation of this is that the \pthp{} decreases less over the
culture period with the \gls{dms} platform as opposed to the beads (or
alternatively, the \gls{dms} has less preferential expansion for CD8 T cells).
We cannot say the same about the \ptmemp{} since we did not have the initial
data for this phenotype; however (although it should be the vast majority of
cells given that cryopreserved T cells from a healthy donor should generally be
composed of circulated memory and naive T cells). Taken together, these data
indicate the \gls{dms} platform has the capacity to expand higher numbers and
percentages of highly potent \ptmem{} and \pth{} T cells compared to
state-of-the-art bead technology.

\subsection*{DMSs can be used to produce functional CAR T cells}

After optimizing for naïve/memory and CD4 yield, we sought to determine if the
\glspl{dms} were compatible with lentiviral transduction protocols used to
generate \gls{car} T cells\cite{Tumaini2013, Lamers2014}. We added a
\SI{24}{\hour} transduction step on day 1 of the \SI{14}{\day} expansion to
insert an anti-CD19 \gls{car}\cite{Milone2009} and subsequently measured the
surface expression of the \gls{car} on day 14
(\cref{fig:car_cd19_flow,fig:car_cd19_endpoint}). We noted
that there was robust \gls{car} expression in over \SI{25}{\percent} of expanded
T cells, and there was no observable difference in \gls{car} expression between
beads and \glspl{dms}.

We also verified the functionality of expanded \gls{car} T cells using a
degranulation assay\cite{Zheng2012}. Briefly, T cells were cocultured with
target cells (either wild-type K562 or CD19-expressing K562 cells) for
\SI{4}{\hour}, after which the culture was analyzed via flow cytometry for the
appearance of CD107a on CD8+ T cells. CD107a is found on the inner-surface of
cytotoxic granules and will emerge on the surface after cytotoxic T cells are
activated and degranulate. Indeed, we observed degranulation in T cells expanded
with both beads and \glspl{dms}, although not to an observably different degree
(\cref{fig:car_degran_flow,fig:car_degran_endpoint}).
Taken together, these results indicated that the \glspl{dms} provide similar
transduction efficiency compared to beads.

We also verified that expanded T cells were migratory using a chemotaxis assay
for CCL21; since \glspl{dms} produced a larger percentage of naïve and memory T
cells (which have CCR7, the receptor for CCL21) we would expect higher migration
in \gls{dms}-expanded cells vs.\ their bead counterparts. Indeed, we noted a
significantly higher migration percentage for T cells grown using \glspl{dms}
versus beads (\cref{fig:car_degran_migration}). Interestingly, there also
appeared to be a decrease in CCL21 migration between transduced and untransduced
T cells expanded using beads, but this interaction effect was only weakly
significant (p = 0.068). No such effect was seen for \gls{dms}-expanded T cells,
showing that migration was likely independent of \gls{car} transduction.

% FIGURE break this up to give the text more flexibility
\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/car_cd19.png}
  \phantomsubcaption\label{fig:car_cd19_flow}
  \phantomsubcaption\label{fig:car_cd19_endpoint}

  \endgroup
  \caption[\glspl{dms} lead to efficient CD19 transduction]
  {\glspl{dms} lead to efficient CD19 transduction.
    \subcap{fig:car_cd19_flow}{Representative flow cytometry plot for
      transduced or untransduced T cells stained with \gls{ptnl}.}
    \subcap{fig:car_cd19_endpoint}{Endpoint plots with \gls{anova} for
      transduced or untransduced T cells stained with \gls{ptnl}.}
    All data is from T cells expanded for \SI{14}{\day}.
  }
  \label{fig:car_production}
\end{figure*}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/car_degranulation.png}
  \phantomsubcaption\label{fig:car_degran_flow}
  \phantomsubcaption\label{fig:car_degran_endpoint}
  \phantomsubcaption\label{fig:car_degran_migration}

  \endgroup
  \caption[\glspl{dms} produce functional \gls{car} T cells]
  {\glspl{dms} produce functional \gls{car} T cells.
    \subcap{fig:car_degran_flow}{Representative flow plot for
      degenerating T cells.}
    \subcap{fig:car_degran_endpoint}{Endpoint plots for transduced or
    untransduced T cells stained with \cd{107a} for the degranulation assay.}
    \subcap{fig:car_degran_migration}{Endpoint plot for transmigration assay
    with \gls{anova}.} All data is from T cells expanded for \SI{14}{\day}.
  }
  \label{fig:car_production}
\end{figure*}

In addition to CD19 \gls{car} T cells, we also demonstrated that the \gls{dms}
platform can be used to expand \gls{car} T cells against \gls{bcma}. Analogously
to the case with CD19, \gls{dms} and bead produced similar fractions of \ptcar{}
cells (albeit in this case at day 7 and with an undefined \gls{moi})
(\cref{fig:car_bcma_percent}). Also consistent with CD19 and non-\gls{car} data,
we also found that the number of \ptcar{} T cells was greater for \gls{dms} than
for bead (\cref{fig:car_bcma_total}).

% FIGURE the right half if bigger than the left half
\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/car_bcma.png}
  \phantomsubcaption\label{fig:car_bcma_percent}
  \phantomsubcaption\label{fig:car_bcma_total}

  \endgroup
  \caption[BMCA Transduction Results]
  {\glspl{dms} produce larger numbers of \gls{bcma} \gls{car} T cells compared
    to beads.
    \subcap{fig:car_bcma_percent}{\ptcarp{} at day 14.}
    \subcap{fig:car_bcma_total}{Total number of \ptcarp{} cells at day 14.}
  }
  \label{fig:car_bcma}
\end{figure*}

\subsection{DMSs efficiently expand T cells in Grex bioreactors}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/grex_results.png}
  \phantomsubcaption\label{fig:grex_results_fc}
  \phantomsubcaption\label{fig:grex_results_viability}
  \phantomsubcaption\label{fig:grex_mem}
  \phantomsubcaption\label{fig:grex_cd4}

  \endgroup
  \caption[Grex bioreactor results]
  {\glspl{dms} expand T cells robustly in Grex bioreactors.
    \subcap{fig:grex_results_fc}{Fold change of T cells over time.}
    \subcap{fig:grex_results_viability}{Viability of T cells over time.}
    \subcap{fig:grex_mem}{\ptmemp{}} and 
    \subcap{fig:grex_cd4}{\pthp{}} of T cells after \SI{14}{\day}
    of expansion. Significance tests were performed using the Wilcoxon
   non-parametric test.
  }
  \label{fig:grex_results}
\end{figure*}

We also asked if the \gls{dms} platform could expand T cells in a static
bioreactor such a Grex. We incubated T cells in a Grex analogously to that for
plates and found that T cells in Grex bioreactors expanded as efficiently as
bead over \SI{14}{\day} and had similar viability
(\cref{fig:grex_results_fc,fig:grex_results_viability}). Furthermore, consistent
with past results, \glspl{dms}-expanded T cells had higher \pthp{} compared to
beads and higher \ptmemp{} compared to beads (\cref{fig:grex_mem,fig:grex_cd4}).
Overall the \ptmemp{} was much lower than that seen from cultures grown in
tissue-treated plates (\cref{fig:dms_phenotype_mem}). 

These discrepancies might be explained in light of our other data as follows.
The Grex bioreactor has higher media capacity relative to its surface area, and
we did not move the T cells to a larger bioreactor as they grew in contrast with
our plate cultures. This means that the cells had higher growth area
constraints, which may have nullified any advantage to the expansion that we
seen elsewhere (\cref{fig:dms_exp_fold_change}). Furthermore, the higher growth
area could mean higher signaling and higher differentiation rate to
\glspl{teff}, which was why the \ptmemp{} was so low compared to other data
(\cref{fig:dms_phenotype_mem}).

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/grex_luminex.png}

  \endgroup
  \caption[Grex luminex results]
  {\gls{dms} lead to higher cytokine production in Grex bioreactors.}
  \label{fig:grex_luminex}
\end{figure*}

We also quantified the cytokines released during the Grex expansion using
Luminex. We noted that in nearly all cases, the \gls{dms}-expanded T cells
released higher concentrations of cytokines compared to beads
(\cref{fig:grex_luminex}). This included higher concentrations of
pro-inflammatory cytokines such as GM-CSF, \gls{ifng}, and \gls{tnfa}. This
demonstrates that \gls{dms} could lead to more robust activation and fitness.

Taken together, these data suggest that \gls{dms} also lead to robust expansion
in Grex bioreactors, although more optimization may be necessary to maximize the
media feed rate and growth area to get comparable results to those seen in
tissue-culture plates.

\subsection{DMSs do not leave antibodies attached to cell product}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/nonstick.png}

  \endgroup
  \caption[\glspl{mab} do not detach from \glspl{dms}]
  {\glspl{mab} do not detach from microcarriers onto T cells in a detectable
    manner. Plots are representative manufacturing runs harvest after
    \SI{14}{\day} of expansion and stained with \anti{\gls{igg}}.
  }
  \label{fig:nonstick}
\end{figure*}

% DISCUSSION alude to this figure
We asked if \glspl{mab} from the \glspl{dms} detached from the \gls{dms} surface
and could be detected on the final T cell product. This test is important for
clinical translation as any residual \glspl{mab} on T cells injected into the
patient could elicit an undesirable \antim{\gls{igg}} immune response. We did
not detect the presence of either \ahcd{3} or \ahcd{28} \glspl{mab} (both of
which were \gls{igg}) on the final T cell product after \SI{14}{\day} of
expansion (\cref{fig:nonstick}).

\subsection{DMSs consistently outperform bead-based expansion compared to
beads in a variety of conditions}

In order to establish the robustness of our method, we combined all experiments
performed in our lab using beads or \glspl{dms} and combined them into one
dataset. Since each experiment was performed using slightly different process
conditions, we hypothesized that performing causal inference on such a dataset
would not only indicate if the \glspl{dms} indeed led to better results under a
variety of conditions, but would also indicate other process parameters that
influence the outcome. The dataset was curated by compiling all experiments and
filtering those that ended at day 14 and including flow cytometry results for
the \ptmem{} and \pth{} populations. We further filtered our data to only
include those experiments where the surface density of the CD3 and CD28
\gls{mab} were held constant (since some of our experiments varied these on the
\glspl{dms}). This ultimately resulted in a dataset with 177 runs spanning 16
experiments between early 2017 and early 2021.

Since the aim of the analysis was to perform causal inference, we determined 6
possible treatment variables which we controlled when designing the experiments
included in this dataset. Obviously the principle treatment parameter was
‘activation method’ which represented the effect of activating T cells with
either beads or our \gls{dms} method. We also included ‘bioreactor’ which was a
categorical for growing the T cells in a Grex bioreactor vs polystyrene plates,
‘feed criteria’ which represented the criteria used to feed the cells (using
media color or a glucose meter), ‘IL2 Feed Conc’ as a continuous parameter for
the concentration of IL2 added each feed cycle, and ‘CD19-CAR Transduced’
representing if the cells were lentivirally transduced or not. Unfortunately,
many of these parameters correlated with each other highly despite the large
size of our dataset, so the only two parameters for which causal relationships
could be evaluated were ‘activation method’ and ‘bioreactor’. We should also
note that these were not the only set of theoretical treatment parameters that
we could have used. For example, media feed rate is an important process
parameter, but in our experiments this was dependent on the feeding criteria and
the growth rate of the cells, which in turn is determined by activation method.
Therefore, ‘media feed rate’ (or similar) is a ‘post-treatment parameter’ and
would have violated the backdoor criteria and severely biased our estimates of
the treatment parameters themselves.

In addition to these treatment parameters, we also included covariates to
improve the precision of our model. Among these were donor parameters including
age, \gls{bmi}, demographic, and gender, as well as the initial viability and
CD4/CD8 ratio of the cryopreserved cell lots used in the experiments
(\cref{tab:meta_donors}). We also included the age of key reagents such as IL2,
media, and the anti-aggregate media used to thaw the T cells prior to
activation. Each experiment was performed by one of three operators, so this was
included as a three-level categorical parameter. Lastly, some of our experiments
were sampled longitudinally, so we included a boolean categorical to represented
this modification as removing conditioned media as the cell are expanding could
disrupt signaling pathways.

\begin{table}[!h] \centering
  \caption{Causal Inference on treatment variables only}
  \label{tab:ci_treat}
  \input{../tables/causal_inference_treat.tex}
\end{table}

\begin{table}[!h] \centering
  \caption{Causal Inference on treatment variables and control variables}
  \label{tab:ci_controlled}
  \input{../tables/causal_inference_control.tex}
\end{table}

\begin{table}[!h] \centering
  \caption{Causal Inference on treatment variables and control variables (single
  donor)}
  \label{tab:ci_single}
  \input{../tables/causal_inference_single.tex}
\end{table}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/metaanalysis_effects.png}
  \phantomsubcaption\label{fig:metaanalysis_fx_exp}
  \phantomsubcaption\label{fig:metaanalysis_fx_mem}
  \phantomsubcaption\label{fig:metaanalysis_fx_cd4}

  \endgroup
  \caption[Meta-analysis effect sizes]
  {\glspl{dms} exhibit superior performance compared to beads controlling for
    many experimental and process conditions. Effect sizes for
    \subcap{fig:metaanalysis_fx_exp}{fold change},
    \subcap{fig:metaanalysis_fx_mem}{\ptmemp{}}, and
    \subcap{fig:metaanalysis_fx_cd4}{\dpthp{}}. The dotted line represents
    the mean of the bead population. The red and blue dots represent the effect
    size of using \gls{dms} instead of beads only considering treatment
    variables (\cref{tab:ci_treat}) or treatment and control variables
    (\cref{tab:ci_controlled}) respectively.
  }
  \label{fig:metaanalysis_fx}
\end{figure*}

We first asked what the effect of each of our treatment parameters was on the
responses of interest, which were fold change of the cells, the \ptmemp{}, and
\dpthp{} (the shift in \pthp{} at day 14 compared to the initial \pthp{}). We
performed a linear regression using activation method and bioreactor as
predictors (the only treatments that were shown to be balanced)
(\cref{tab:ci_treat}). Note that fold change was log transformed to reflect the
exponential nature of T cell growth. We observe that the treatments are
significant in all cases except for the \dpthp{}; however, we also observe that
relatively little of the variability is explained by these simple models ($R^2$
between 0.17 and 0.44).

We then included all covariates and unbalanced treatment parameters and
performed linear regression again
(\cref{tab:ci_controlled,fig:metaanalysis_fx}). We observe that after
controlling for additional noise, the models explained much more variability
($R^2$ between 0.76 and 0.87). Furthermore, the coefficient for activation
method in the case of fold change changed very little but still remained quite
high (note the log-transformation) with \SI{131}{\percent} increase in fold
change compared to beads. Furthermore, the coefficient for \ptmemp{} dropped to
a \SI{3.5}{\percent} increase and almost became non-significant at $\upalpha$ =
0.05, and the \dpthp{} response increased to a \SI{7.4}{\percent} increase and
became highly significant. Looking at the bioreactor treatment, we see that
using the bioreactor in the case of fold change and \ptmemp{} is actually
harmful to the response, while at the same time it seems to increase the
\dpthp{} response. We should note that this parameter merely represents whether
or not the choice was made experimentally to use a bioreactor or not; it does
not indicate why the bioreactor helped or hurt a certain response. For example,
using a Grex entails changing the cell surface and feeding strategy for the T
cells, and any one of these ‘mediating variables’ might actually be the cause of
the responses.

Finally, we stratified on the most common donor (vendor ID 338 from Astarte
Biotech) as this was responsible for almost half the data (80 runs) and repeated
the regression (\Cref{tab:ci_single}). Note that in this case, we did not
include any of the donor-dependent variables as well as any of the variables
that were the same value for these 80 runs. In this analysis, fold change and
\dpthp{} remained high (but slightly lowered from the full analysis) and
\ptmemp{} was non-significant. Given this, it appears that high \ptmemp{} may
have been due to other donors besides this one, and that high fold change and
\dpthp{} may have been driven by this single donor but more extreme in other
donors.

\section{Discussion}

% DISCUSSION this is fluffy
We have developed a T cell expansion shows superior expansion with higher number
of naïve/memory and CD4+ T cells compared to state-of-the-art microbead
technology (\cref{fig:dms_exp}). Other groups have used biomaterials approaches
to mimic the \invivo{} microenvironment\cite{Cheung2018, Rio2018, Delalat2017,
  Lambert2017, Matic2013}; however, to our knowledge this is the first system
that specifically drives naïve/memory and CD4+ T cell formation in a scalable,
potentially bioreactor-compatible manufacturing process. 

Memory and naïve T cells have been shown to be important clinically. Compared to
\glspl{teff}, they have a higher proliferative capacity and are able to engraft
for months; thus they are able to provide long-term immunity with smaller
doses\cite{Gattinoni2012, Joshi2008}. Indeed, less differentiated T cells have
led to greater survival both in mouse tumor models and human
patients\cite{Fraietta2018, Adachi2018, Rosenberg2011}. Furthermore, clinical
response rates have been positively correlated with T cell expansion, implying
that highly-proliferative naïve and memory T cells are a significant
contributor\cite{Xu2014, Besser2010}. Circulating memory T cells have also been
found in complete responders who received CAR T cell therapy\cite{Kalos2011}.

Similarly, CD4 T cells have been shown to play an important role in CAR T cell
immunotherapy. It has been shown that CAR T doses with only CD4 or a mix of CD4
and CD8 T cells confer greater tumor cytotoxicity than only CD8 T
cells\cite{Wang2018, Sommermeyer2015}. There are several possible reasons for
these observations. First, CD4 T cells secrete proinflammatory cytokines upon
stimulation which may have a synergistic effect on CD8 T cells. Second, CD4 T
cells may be less prone to exhaustion and may more readily adopt a memory
phenotype compared to CD8 T cells\cite{Wang2018}. Third, CD8 T cells may be more
susceptible than CD4 T cells to dual stimulation via the \gls{car} and
endogenous \gls{tcr} , which could lead to overstimulation, exhaustion, and
apoptosis\cite{Yang2017}. Despite evidence for the importance of CD4 T cells,
more work is required to determine the precise ratios of CD4 and CD8 T cell
subsets to be included in CAR T cell therapy given a disease state.

% DISCUSSION this mentions the DOE which is in the next aim
When analyzing all our experiments comprehensively using causal inference, we
found that all three of our responses were significantly increased when
controlling for covariates (\cref{fig:metaanalysis_fx,tab:ci_controlled}). By
extension, this implies that not only will \glspl{dms} lead to higher fold
change overall, but also much higher fold change in absolute numbers of memory
and CD4+ T cells. Furthermore, we found that using a Grex bioreactor is
detrimental to fold change and memory percent while helping CD4+. Since there
are multiple consequences to using a Grex compared to tissue-treated plates, we
can only speculate as to why this might be the case. Firstly, when using a Grex
we did not expand the surface area on which the cells were growing in a
comparable way to that of polystyrene plates. One possible explanation is that
the T cells spent longer times in highly activating conditions (since the beads
and DMSs would have been at higher per-area concentrations in the Grex vs
polystyrene plates) which has been shown to skew toward \gls{teff}
populations\cite{Lozza2008}. Furthermore, the simple fact that the T cells spent
more time at high surface densities could simply mean that the T cells didn’t
expands as much due to spacial constraints. This would all be despite the fact
that Grex bioreactors are designed to lead to better T cell expansion due to
their gas-permeable membranes and higher media-loading capacities. If anything,
our data suggests we were using the bioreactor sub-optimally, and the
hypothesized causes for why our T cells did not expand could be verified with
additional experiments varying the starting cell density and/or using larger
bioreactors.

A key question in the space of cell manufacturing is that of donor variability.
To state this precisely, this is a second order interaction effect that
represents the change in effect of treatment (eg bead vs \gls{dms}) given the
donor. While our meta-analysis was relatively large compared to many published
experiments usually seen for technologies at this developmental stage, we have a
limited ability in answering this question. We can control for donor as a
covariate, and indeed our models show that many of the donor characteristics are
strongly associated with each response on average, but these are first order
effects and represent the association of age, gender, demographic, etc given
everything else in the model is held constant. Second order interactions require
that our treatments be relatively balanced and random across each donor, which
is a dubious assumption for our dataset. However, this can easily be solved by
performing more experiments with these restrictions in mind, which will be a
subject of future work.

Furthermore, this dataset offers an interesting insight toward novel hypothesis
that might be further investigated. One limitation of our dataset is that we
were unable to investigate the effects of time using a method such as
autoregression, and instead relied on aggregate measures such as the total
amount of a reagent added over the course of the expansion. Further studies
should be performed to investigate the temporal relationship between phenotype,
cytokine concentrations, feed rates, and other measurements which may perturb
cell cultures, as this will be the foundation of modern process control
necessary to have a fully-automated manufacturing system.

% It is important to note that all T cell cultures in this study were performed up
% to 14 days. Others have demonstrated that potent memory T cells may be obtained
% simply by culturing T cells as little as 5 days using traditional
% beads\cite{Ghassemi2018}. It is unknown if the naïve/memory phenotype of our DMS
% system could be further improved by reducing the culture time, but we can
% hypothesize that similar results would be observed given the lower number of
% doublings in a 5 day culture. We should also note that we investigated one
% subtype (\ptmem{}) in this study. Future work will focus on other memory
% subtypes such as tissue resident memory and stem memory T cells, as well as the
% impact of using the DMS system on the generation of these subtypes.

% DISCUSSION this sounds sketchy
% Another advantage is that the DMS system appears to induce a faster growth rate
% of T cells given the same IL2 concentration compared to beads (Supplemental
% Figure 8) along with retaining naïve and memory phenotype. This has benefits in
% multiple contexts. Firstly, some patients have small starting T cell populations
% (such as infants or those who are severely lymphodepleted), and thus require
% more population doublings to reach a usable dose. Our data suggests the time to
% reach this dose would be reduced, easing scheduling a reducing cost. Secondly,
% the allogeneic T cell model would greatly benefit from a system that could
% create large numbers of T cells with naïve and memory phenotype. In contrast to
% the autologous model which is currently used for Kymriah and Yescarta,
% allogeneic T cell therapy would reduce cost by spreading manufacturing expenses
% across many doses for multiple patients\cite{Harrison2019}. Since it is
% economically advantageous to grow as many T cells as possible in one batch in
% the allogeneic model (reduced start up and harvesting costs, fewer required cell
% donations), the DMSs offer an advantage over current technology.

The \gls{dms} system could be used as a drop in replacement for beads in many of
current allogeneic therapies. Indeed, given its higher potential for expansion
(\cref{fig:dms_exp,tab:ci_controlled}), it may work in cases where the beads
fail (although this would need to be tested by gathering data with many
unhealthy donors). However, in the autologous setting patients only need a fixed
dose, and thus any expansion beyond the indicated dose would be wasted. Given
this, it will be interesting to apply this technology in an allogeneic paradigm
where this increased expansion potential would be well utilized.

Finally, we should note that while we demonstrated a method providing superior
performance compared to bead-based expansion, the cell manufacturing field would
tremendously benefit from simply having an alternative to state-of-the-art bead
based expansion. The patents for bead-based expansion are owned by few companies
and licensed accordingly; having an alternative would provide more competition
in the market, reducing costs and improving access for academic researchers and
manufacturing companies.

\chapter{AIM 2A}\label{aim2a}

\section{Introduction}

The purpose of this sub-aim was to develop computational methods to identify
novel \glspl{cqa} and \glspl{cpp} that could be used for release criteria,
process control, and process optimization for the \gls{dms} platform. We
hypothesized that T cells grown using the \gls{dms} system would produce
detectable biological signatures in the media supernatent which corresponded to
clinically relevent responses such as fold expansion or phenotype. We tested
this hypothesis by activating T cells under a variety of conditions using a
\gls{doe}, sampling the media at intermediate timepoints, and creating models to
predict the outcome of the cultures. We should stress that the specific
\glspl{cpp} and \glspl{cqa} determined by this aim are not necessarily
universal, as this was not performed with equipment that would normally be used
at scale. However, the process outlined here is one that can easily be adaptable
to any system, and the specific findings themselves offer interesting insights
that warrant further study\footnote{adapted from \modelpaper{}}.

\section{Methods}

\subsection{Study Design}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/modeling_overview.png}
  \phantomsubcaption\label{fig:mod_overview_flow}
  \phantomsubcaption\label{fig:mod_overview_doe}

  \endgroup
  \caption[Modeling Overview]
  {Overview of modeling experiments.
    \subcap{fig:mod_overview_flow}{Relationship
      between \gls{doe} experiments and AI driven prediction. \glspl{doe} will
      be used to determine optimal process input conditions, and longitudinal
      multiomics data will be used to fit predictive models. Together, these
      will reveal predictive species that may be used for \glspl{cqa} and
      \glspl{cpp}.}
    \subcap{fig:mod_overview_doe}{Overview of the two \gls{doe} experiments; the
      initial \gls{doe} is given by the blue points and the augmented \gls{doe}
      is given by the blue points.}
  }
  \label{fig:mod_overview}
\end{figure*}

The overall workflow of this aim is shown in \cref{fig:mod_overview_flow}.
Experimental conditions within the design space were explored using a \gls{doe},
and longitudinal samples were collected for each condition as the cultures
progressed. Data from inputs and/or longitudinal samples were used to predict
the endpoint response. The fusion of cytokine and \gls{nmr} profiles from media
to model these responses included 30 cytokines from a custom Thermo Fisher
ProcartaPlex Luminex kit and 20 \gls{nmr} features. These 20 spectral features
from \gls{nmr} media analysis were selected out of approximately 250 peaks
through the implementation of a variance-based feature selection approach and
some manual inspection steps.

The first \gls{doe} resulted in a randomized 18-run I-optimal custom design
where each \gls{dms} parameter was evaluated at three levels: \pilII{} (10, 20,
and 30 U/uL), \pdms{} (500, 1500, 2500 \si{\dms\per\ul}), and \pmab{} (60, 80,
100 \si{\percent}). These 18 runs consisted of 14 unique parameter combinations
where 4 of them were replicated twice to assess prediction error. To further
optimize the initial region explored, an \gls{adoe} was designed with 10 unique
parameter combinations, two of these replicated twice for a total of 12
additional samples (\cref{fig:mod_overview_doe}). Process parameters for the
\gls{adoe} were evaluated at multiple levels: \pilII{} (30, 35, and 40
\si{\IU\per\ml}), \pdms{} (500, 1000, 1500, 2000, 2500, 3000, 3500
\si{\dms\per\ml}), and \pmab{} (\SI{100}{\percent}) (\cref{fig:mod_overview}).

\subsection{DMS fabrication}

\glspl{dms} were fabricated as described in \cref{sec:dms_fab} with the
following modifications in order to obtain a variable functional \gls{mab}
surface density. During the \gls{mab} coating step, the \acd{3}/\acd{28}
\gls{mab} mixture was further combined with a biotinylated isotype control to
reduce the overall fraction of targeted \glspl{mab} (for example the
\SI{60}{\percent} \gls{mab} surface density corresponded to 3 mass parts
\acd{3}, 3 mass parts \acd{28}, and 4 mass parts isotype control).

\subsection{T Cell Culture}

T cell culture was performed as described in \cref{sec:tcellculture} with the
following modifications. At days 4, 6, 8, and 11, \SI{100}{\ul} media were
collected for the Luminex assay and \gls{nmr} analysis. The volume of removed
media was equivalently replaced during the media feeding step, which took place
immediately after sample collection. Additionally, the same media feeding
schedule was followed for the \gls{doe} and \gls{adoe} to improve consistency,
and the same donor lot was used for both experiments. All cell counts were
performed using \gls{aopi}.

\subsection{Flow Cytometry}

Flow cytometry was performed analogously to \cref{sec:flow_cytometry}.

\subsection{Cytokine Quantification}

Cytokines were quantified via Luminex as described in
\cref{sec:luminex_analysis}.

\subsection{NMR Metabolomics}

Prior to analysis, samples were centrifuged at \SI{2990}{\gforce} for
\SI{20}{\minute} at \SI{4}{\degreeCelsius} to clear any debris\footnote{all
  \gls{nmr} analysis was done by our collaborators Max Colonna and Art Edison at
  the University of Georgia; methods included here for reference}. \SI{5}{\ul} of
100/3 \si{\mM} DSS-D6 in deuterium oxide (Cambridge Isotope Laboratories) were
added to \SI{1.7}{\mm} \gls{nmr} tubes (Bruker BioSpin), followed by
\SI{45}{\ul} of media from each sample that was added and mixed, for a final
volume of \SI{50}{\ul} in each tube. Samples were prepared on ice and in
predetermined, randomized order. The remaining volume from each sample in the
rack (approx. \SI{4}{\ul}) was combined to create an internal pool. This
material was used for internal controls within each rack as well as metabolite
annotation.

\gls{nmr} spectra were collected on a Bruker Avance III HD spectrometer at
\SI{600}{\MHz} using a \SI{5}{\mm} TXI cryogenic probe and TopSpin software
(Bruker BioSpin). One-dimensional spectra were collected on all samples using
the noesypr1d pulse sequence under automation using ICON NMR software.
Two-dimensional \gls{hsqc} and \gls{tocsy} spectra were collected on internal
pooled control samples for metabolite annotation.

One-dimensional spectra were manually phased and baseline corrected in TopSpin.
Two-dimensional spectra were processed in NMRpipe\cite{Delaglio1995}. One
dimensional spectra were referenced, water/end regions removed, and normalized
with the PQN algorithm\cite{Dieterle2006} using an in-house MATLAB (The
MathWorks, Inc.) toolbox.

To reduce the total number of spectral features from approximately 250 peaks and
enrich for those that would be most useful for statistical modeling, a
variance-based feature selection was performed within MATLAB. For each digitized
point on the spectrum, the variance was calculated across all experimental
samples and plotted. Clearly-resolved features corresponding to peaks in the
variance spectrum were manually binned and integrated to obtain quantitative
feature intensities across all samples. In addition to highly variable features,
several other clearly resolved and easily identifiable features were selected
(glucose, \gls{bcaa} region, etc). Some features were later discovered to belong
to the same metabolite but were included in further analysis.

Two-dimensional spectra collected on pooled samples were uploaded to COLMARm web
server, where \gls{hsqc} peaks were automatically matched to database peaks.
\gls{hsqc} matches were manually reviewed with additional 2D and proton spectra
to confirm the match. Annotations were assigned a confidence score based upon
the levels of spectral data supporting the match as previously
described\cite{Dashti2017}. Annotated metabolites were matched to previously
selected features used for statistical analysis.

Several low abundance features selected for analysis did not have database
matches and were not annotated. Statistical total correlation spectroscopy41
suggested that some of these unknown features belonged to the same molecules
(not shown). Additional multidimensional \gls{nmr} experiments will be required
to determine their identity.

\subsection{Machine Learning and Statistical Analysis}

Linear regression analysis of the \glspl{doe} was performed as described in
\cref{sec:statistics}.

Seven \gls{ml} techniques were implemented to predict three responses related to
the memory phenotype of the cultured T cells under different process
conditions (\rmemh{}, \rmemk{}, and \rratio{}). The \gls{ml} methods
executed were \gls{rf}, \gls{gbm}, \gls{cif}, \gls{lasso}, \gls{plsr},
\gls{svm}, and DataModeler’s \gls{sr}\footnote{\gls{sr} was performed by Theresa
  Kotanchek at Evolved Analytics, \gls{rf}, \gls{gbm}, \gls{cif}, \gls{plsr},
  \gls{svm} were performed by Valerie Odeh-Couvertier at UPRM. Methods included
  here for reference}. Primarily, \gls{sr} models were used to optimize process
parameter values based on \ptmem{} phenotype and to extract early predictive
variable combinations from the multi-omics experiments. Furthermore, all
regression methods were executed, and the high-performing models were used to
perform a consensus analysis of the important variables to extract potential
critical quality attributes and critical process parameters predictive of T cell
potency, safety, and consistency at the early stages of the manufacturing
process.

\gls{sr} was done using Evolved Analytics’ DataModeler software (Evolved
Analytics LLC, Midland, MI). DataModeler utilizes genetic programming to evolve
symbolic regression models (both linear and non-linear) rewarding simplicity and
accuracy. Using the selection criteria of highest accuracy
($R^2$>\SI{90}{\percent}) and lowest complexity, the top-performing models were
identified. Driving variables, variable combinations, and model dimensionality
tables were generated. The top-performing variable combinations were used to
generate model ensembles. In this analysis, DataModeler’s
\inlinecode{SymbolicRegression} function was used to develop explicit algebraic
(linear and nonlinear) models. The fittest models were analyzed to identify the
dominant variables using the \inlinecode{VariablePresence} function, the
dominant variable combinations using the \inlinecode{VariableCombinations}
function, and the model dimensionality (number of unique variables) using the
\inlinecode{ModelDimensionality} function. \inlinecode{CreateModelEnsemble} was
used to define trustable model ensembles using selected variable combinations
and these were summarized (model expressions, model phenotype, model tree plot,
ensemble quality, model quality, variable presence map, \gls{anova} tables,
model prediction plot, exportable model forms) using the
\inlinecode{ModelSummaryTable} function. Ensemble prediction and residual
performance were assessed via the \inlinecode{EnsemblePredictionPlot} and
\inlinecode{EnsembleResidualPlot} subroutines respectively. Model maxima
(\inlinecode{ModelMaximum} function) and model minima (\inlinecode{ModelMinimum}
function) were calculated and displayed using the
\inlinecode{ResponsePlotExplorer} function. Trade-off between multiple
responses was explored using \inlinecode{MultiTargetResponseExplorer} and
\inlinecode{ResponseComparisonExplorer} with additional insights derived from
\inlinecode{ResponseContourPlotExplorer}. Graphics and tables were generated by
DataModeler. These model ensembles were used to identify predicted response
values, potential optima in the responses, and regions of parameter values where
the predictions diverge the most.

Non-parametric tree-based ensembles were done through the
\inlinecode{randomForest}, inlinecode{gbm}, and \inlinecode{cforest} regression
functions in R, for \gls{rf}, \gls{gbm}, and \gls{cif} models, respectively.
Both \gls{rf} and \gls{cif} construct multiple decision trees in parallel, by
randomly choosing a subset of features at each decision tree split, in the
training stage. \gls{rf} individual decision trees are split using the Gini
Index, while conditional inference forest uses a statistical significance test
procedure to select the variables at each split, reducing correlation bias. In
contrast, \gls{gbm} construct regression trees in series through an iterative
procedure that adapts over the training set. This model learns from the mistakes
of previous regression trees in an iterative fashion to correct errors
(\gls{mse}) from its precursors’ trees. Prediction performance was evaluated
using \gls{loocv} and permutation-based variable importance scores assessing
percent increase of \gls{mse}, relative influence based on the increase of
prediction error, coefficient values for \gls{rf}, \gls{gbm}, and \gls{cif},
respectively. \gls{plsr} was executed using the \inlinecode{plsr} function from
the \inlinecode{pls} package in R while \gls{lasso} regression was performed
using the \inlinecode{cv.glmnet} R package, both using \gls{loocv}. Finally, the
\inlinecode{kernlab} R package was used to construct the \gls{svm} models.

Parameter tuning was done for all models in a grid search manner using the train
function from the \inlinecode{caret} R package using \gls{loocv} as the
optimization criteria. Specifically, the number of features randomly sampled as
candidates at each split (\inlinecode{mtry}) and the number of trees to grow
(\inlinecode{ntree}) were tuned parameters for random forest and conditional
inference forest. In particular, minimum sum of weights in a node to be
considered for splitting and the minimum sum of weights in a terminal node were
manually tuned for building the \gls{cif} models. Moreover, \gls{gbm} parameters
such as the number of trees to grow, maximum depth of each tree, learning rate,
and the minimal number of observations at the terminal node, were tuned for
optimum \gls{loocv} performance as well. For \gls{plsr}, the optimal number of
components to be used in the model was assessed based on the standard error of
the cross-validation residuals using the function \inlinecode{selectNcomp} from
the \inlinecode{pls} package. Moreover, \gls{lasso} regression was performed
using the \inlinecode{cv.glmnet} package with $\upalpha$ = 1. The best
$\uplambda$ for each response was chosen using the minimum error criteria.
Lastly, a fixed linear kernel (\inlinecode{svmLinear}) was used to build
the \gls{svm} regression models evaluating the cost parameter value with best
\gls{loocv}. Prediction performance was measured for all models using the final
model with \gls{loocv} tuned parameters.

\subsection{Consensus Analysis}

Consensus analysis of the relevant variables extracted from each machine
learning model was done to identify consistent predictive features of quality at
the early stages of manufacturing. First importance scores for all features were
measured across all \gls{ml} models using \inlinecode{varImp} with
\inlinecode{caret} R package except for scores for \gls{svm} which
\inlinecode{rminer} R package was used. These importance scores were percent
increase in \gls{mse}, relative importance through average increase in
prediction error when a given predictor is permuted, permuted coefficients
values, absolute coefficient values, weighted sum of absolute coefficients
values, and relative importance from sensitivity analysis determined for
\gls{rf}, \gls{gbm}, \gls{cif}, \gls{lasso}, \gls{plsr}, and \gls{svm},
respectively. Using these scores, key predictive variables were selected if
their importance scores were within the \nth{80} percentile ranking for the
following \gls{ml} methods: \gls{rf}, \gls{gbm}, \gls{cif}, \gls{lasso},
\gls{plsr}, \gls{svm} while for \gls{sr} variables present in >\SI{30}{\percent}
of the top-performing \gls{sr} models from DataModeler ($R^2\ge$
\SI{90}{\percent}, Complexity $\ge$ 100) were chosen to investigate consensus
except for \gls{nmr} media models at day 4 which considered a combination of the
top-performing results of models excluding lactate ppms, and included those
variables which were in >\SI{40}{\percent} of the best performing models. Only
variables with those high percentile scoring values were evaluated in terms of
their logical relation (intersection across \gls{ml} models) and depicted using
a Venn diagram from the \inlinecode{venn} R package.

\section{Results}

\subsection{T Cells Can be Grown on DMSs with Lower IL2 Concentrations}

Prior to the main experiments in this aim, we performed a preliminary experiment
to assess the effect of lowering the \gls{il2} concentration on the T cells
grown with either bead or \gls{dms}. One of the hypotheses for the \gls{dms}
system was that the higher cell density would enable more efficient cross-talk
between T cells. Since \gls{il2} is secreted by activated T cells themselves,
T cells in the \gls{dms} system may need less or no \gls{il2} if this hypothesis
were true.

% FIGURE this plots proportions look dumb
\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/il2_modulation.png}
  \phantomsubcaption\label{fig:il2_mod_timecourse}
  \phantomsubcaption\label{fig:il2_mod_total}
  \phantomsubcaption\label{fig:il2_mod_mem}
  \phantomsubcaption\label{fig:il2_mod_flow}

  \endgroup
  \caption[T cells grown at varying IL2 concentrations]
  {\glspl{dms} grow T cells effectively at lower IL2 concentrations.
    \subcap{fig:il2_mod_timecourse}{Longitudinal cell counts of T cells grown
      with either bead or \glspl{dms} using varying IL2 concentrations}
    Day 14 counts of either \subcap{fig:il2_mod_total}{total cells} or
    \subcap{fig:il2_mod_mem}{\ptmem{} cells} plotted against \gls{il2}
    concentration.
    \subcap{fig:il2_mod_flow}{Flow cytometry plots of the \ptmem{} gated
      populations at day 14 of culture for each \gls{il2} concentration.}
  }
  \label{fig:il2_mod}
\end{figure*}

We varied the concentration of \gls{il2} from \SIrange{0}{100}{\IU\per\ml} and
expanded T cells as described in \cref{sec:tcellculture}. T cells grown with
either method expanded robustly as \gls{il2} concentration was increased
(\cref{fig:il2_mod_timecourse}). Surprisingly, neither the bead or the \gls{dms}
group expanded at all with \SI{0}{\IU\per\ml} \gls{il2}. When examining the
endpoint fold change after \SI{14}{\day}, we observe that the difference between
the bead and \gls{dms} appears to be greater at lower \gls{il2} concentrations
(\cref{fig:il2_mod_total}).
% This is further supported by fitting a non-linear
% least squares equation to the data following a hyperbolic curve (which should be
% a plausible model given that this curve describes receptor-ligand kinetics,
% which we can assume \gls{il2} to follow).
Furthermore, the same trend can be
seen when only examining the \ptmem{} cell expansion at day 14
(\cref{fig:il2_mod_mem}). In this case, the \ptmemp{} of the T cells seemed to
be relatively close at higher \gls{il2} concentrations, but separated further at
lower concentrations (\cref{fig:il2_mod_flow})

Taken together, these data do not support the hypothesis that the \gls{dms}
system does not need \gls{il2} at all; however, it appears to have a modest
advantage at lower \gls{il2} concentrations compared to beads. For this reason,
we decided to investigate the lower range of \gls{il2} concentrations starting
at \SI{10}{\IU\per\ml} throughout the remainder of this aim.

\subsection{DOE Shows Optimal Conditions for Expanded Potent T Cells}

% TABLE not all of these were actually used, explain why by either adding columns
% or marking with an asterisk
\begin{table}[!h] \centering
  \caption{DOE Runs}
  \label{tab:doe_runs}
  \input{../tables/doe_runs.tex}
\end{table}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/doe_responses_first.png}
  \phantomsubcaption\label{fig:doe_response_first_mem}
  \phantomsubcaption\label{fig:doe_response_first_cd4}

  \endgroup
  \caption[Response plots for first DOE]
  {Response plots from the first \gls{doe} experiment for
    \subcap{fig:doe_response_first_mem}{\ptmemp{}} and
    \subcap{fig:doe_response_first_cd4}{\pthp{}}. Each point is one run.
  }
  \label{fig:doe_response_first}
\end{figure*}

% RESULT maybe add regression tables to this, although it doesn't really matter
% since we end up doing regression on the full thing later anyways.
We conducted two consecutive \glspl{doe} to optimize the \pth{} and \ptmem{}
responses for the \gls{dms} system. In the first \gls{doe} we, tested \pilII{} in
the range of \SIrange{10}{30}{\IU\per\ml}, \pdms{} in the range of
\SIrange{500}{2500}{\dms\per\ml}, and \pmab{} in the range of
\SIrange{60}{100}{\percent}. When looking at the total \ptmemp{} output, we
observed that \pilII{} showed a positive linear trend with the \pdms{} and
\pmab{} showing possible second-order effects with maximums and minimums at the
intermediate level (\cref{fig:doe_response_first_mem}). In the case of \pth{},
we observed that all parameters seemed to have a positive linear response, with
\pilII{} and \pdms{} showing slight second order effects that suggest a maximum
might exist at a higher value for each.

After performing the first \gls{doe} we augmented the original design matrix
with an \gls{adoe} which was built with three goals in mind. Firstly we wished
to validate the first \gls{doe} by assessing the strength and responses of each
effect. Secondly, we wished to improve our confidence in regions that showed
high complexity, such as the peak in the \gls{dms} concentration for the total
\ptmem{} cell response. Thirdly, we wished to explore additional ranges of each
response. Since \pilII{} and \pdms{} appeared to continue positively influence
multiple responses beyond our tested range, we were curious if there was an
optimum at some higher setting of either of these values. For this reason, we
increased the \pilII{} to include \SI{40}{\IU\per\ml} and the \pdms{} to
\SI{3500}{\dms\per\ml}. Note that it was impossible to go beyond
\SI{100}{\percent} for the \pmab{}, so runs were positioned for this parameter
with validation and confidence improvements in mind. The runs for each \gls{doe}
were shown in \cref{tab:doe_runs}\footnote{Not all runs in this table were used.
It was determined later that the total \glspl{mab} surface density may not be
consistent across each batch of \gls{dms} used, primarily due to the fact that a
subset were created at different scale and with a different operator. To remove
this bias in our data, these runs were not used.}.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/doe_responses.png}
  \phantomsubcaption\label{fig:doe_responses_mem}
  \phantomsubcaption\label{fig:doe_responses_cd4}
  \phantomsubcaption\label{fig:doe_responses_mem4}
  \phantomsubcaption\label{fig:doe_responses_ratio}

  \endgroup
  \caption[T cell optimization through Design of Experiments]
  {\gls{doe} methodology reveals optimal conditions for expanding T cell
    subsets. Responses vs IL2 concentration, \gls{dms} concentration, and
    functional \gls{mab} percentage are shown for 
    \subcap{fig:doe_responses_mem}{total \ptmem{} T cells},
    \subcap{fig:doe_responses_cd4}{total \pth{} T cells},
    \subcap{fig:doe_responses_mem4}{total \ptmemh{} T cells}, and 
    \subcap{fig:doe_responses_ratio}{ratio of CD4 and CD8 T cells in the
      \ptmem{} compartment}. Each point represents one run.
  }
  \label{fig:doe_responses}
\end{figure*}

\begin{table}[!h] \centering
  \caption{Total CD62L+CCR7+ T cell response (first order regression)}
  \label{tab:doe_mem1.tex}
  \input{../tables/doe_mem1.tex}
\end{table}

\begin{table}[!h] \centering
  \caption{Total CD62L+CCR7+ T cell response (third order regression)}
  \label{tab:doe_mem2.tex}
  \input{../tables/doe_mem2.tex}
\end{table}

\begin{table}[!h] \centering
  \caption{Total CD4+ T cell response}
  \label{tab:doe_cd4.tex}
  \input{../tables/doe_cd4.tex}
\end{table}

\begin{table}[!h] \centering
  \caption{Linear regression for total \ptmemh{} cells}
  \label{tab:doe_mem4.tex}
  \input{../tables/doe_mem4.tex}
\end{table}

\begin{table}[!h] \centering
  \caption{Linear regression for CD4:CD8 ratio in the \ptmem{} compartment}
  \label{tab:doe_ratio.tex}
  \input{../tables/doe_ratio.tex}
\end{table}

The response plots from both \glspl{doe} are shown in \cref{fig:doe_responses}
for total \ptmem{} cells, total \pth{} cells, total \ptmemh{} cells, and CD4:CD8
ratio in the \ptmem{} compartment. In general, the responses for the first and
second \gls{doe} seemed to overlap, although not perfectly. Interestingly, only
the \ptmem{} response seemed to have anything more complex than a linear
relationship, particularly in the case of \pilII{} and \pdms{}, which showed
intermediate optimums (\cref{fig:doe_responses_mem}). In the case of \pilII{},
it was not clear if this optimum was simply due to a batch effect of being from
the first or second \gls{doe}. The optimum for \pdms{} appeared in the same
location albeit more pronounced in the second \gls{doe} so, giving more
confidence to the location of this second order feature. The remainder of the
responses showed mostly linear relationships in all parameter cases
(\cref{fig:doe_responses_cd4,fig:doe_responses_mem4,fig:doe_responses_ratio}).

% RESULT it seems arbitrary that I went straight to a third order model, the real
% reason is because it seemed weird that a second order model didn't find
% anything to be significant
We performed linear regression on the three input parameters as well as a binary
parameter representing if a given run came from the first or second \gls{doe}
(called `dataset'). Starting with the total \ptmem{} cells response, we fit a
first order regression model using these four parameters
(\cref{tab:doe_mem1.tex}). While \pilII{} was found to be a significant
predictor, the model fit was extremely poor ($R^2$ of 0.331). This was not
surprising given the apparent complexity of this response
(\cref{fig:doe_responses_mem}). To obtain a better fit, we added second and
third degree terms (\cref{tab:doe_mem2.tex}). Note that the dataset parameter
was not included in the second order interaction as this was treated as a
blocking variable, which are typically not assumed to have interaction effects.
Also note that the response was log-transformed, which yielded a better fit. In
this model many more parameters emerged as being significant, including the
quadratic terms for \pdms{} and \pilII{}, in agreement with what can be
qualitatively observed in the response plot (\cref{fig:doe_responses_mem}).
Furthermore, the dataset parameter was weakly significant, indicating a possible
batch effect between the \glspl{doe}. We should also note that despite many
parameters being significant, this model was still only mediocre in describing
this response; the $R^2$ was 0.741 but the $R_{adj}^2$ was 0.583, indicating
that our data might be underpowered for a model this complex. Further
experiments beyond what was performed here may be needed to fully describe this
response.

% TABLE combine these tables into one
We performed linear regression on the other three responses, all of which
performed much better than the \ptmem{} response as expected given the much
lower apparent complexity in the response plots
(\cref{fig:doe_responses_cd4,fig:doe_responses_mem4,fig:doe_responses_ratio}).
All these models appeared to fit will, with $R^2$ and $R_{adj}^2$ upward of
0.8. In all but the CD4:CD8 \ptmem{} ratio, the dataset parameter emerged as
significant, indicating a batch effect between the \glspl{doe}. All other
parameters except \pilII{} in the case of CD4:CD8 \ptmem{} ratio were
significant predictors.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/doe_sr_contour.png}
  \phantomsubcaption\label{fig:doe_sr_contour_mem4}
  \phantomsubcaption\label{fig:doe_sr_contour_ratio}

  \endgroup
  \caption[Contour plots for DOE responses]
  {Symbolic regression and contour plots reveal optimal conditions for
    \subcap{fig:doe_sr_contour_mem4}{\ptmemh{} cells} and 
    \subcap{fig:doe_sr_contour_ratio}{CD4:CD8 ratio in the \ptmem{}
      compartment}.
  }
  \label{fig:doe_sr_contour}
\end{figure*}

We then visualized the total \ptmemh{} cells and \rratio{} using the response
explorer in DataModeler to create contour plots around the maximum responses.
For both, it appeared that maximizing all three input parameters resulted in the
maximum value for either response (\cref{fig:doe_sr_contour}). While not all
combinations at and around this optimum were tested, the model nonetheless
showed that there were no other optimal values or regions elsewhere in the
model.

\subsection{Modeling With Artificial Intelligence Methods Reveals Potential
  CQAs}

Due to the heterogeneity of the multivariate data collected and knowing that no
single model is perfect for all applications, we implemented an agnostic
modeling approach to better understand these \ptmem{} responses. To achieve
this, a consensus analysis using seven \gls{ml} techniques, \gls{rf}, \gls{gbm},
\gls{cif}, \gls{lasso}, \gls{plsr}, \gls{svm}, and DataModeler’s \gls{sr}, was
implemented to molecularly characterize \ptmem{} cells and to extract predictive
features of quality early in their expansion process.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/doe_luminex.png}

  \endgroup
  \caption[Cytokine release profile of T cells from DOE]
  {T cells show robust and varying cytokine responses over time}
  \label{fig:doe_luminex}
\end{figure*}

We collected secretome data via luminex for days 4, 6, 8, 11, and 14.
Plotting the concentrations of these cytokines showed a large variation over all
runs and between different timepoints, demonstrated that these could potentially
be used to differentiate between different process conditions qualitatively
simply based on variance (\cref{fig:doe_luminex}). These were also much higher
in most cases that a set of bead based runs which were run in parallel, in
agreement with the luminex data obtained previously in the Grex system (these
data were collected in plates) (\cref{fig:grex_luminex}).

% TABLE this table looks like crap, break it up into smaller tables
\begin{table}[!h] \centering
  \caption[Results for data-driven modeling]
  {Results for data-driven modeling using process parameters (PP) with
    only \gls{nmr} on day 4 (N4), only \gls{nmr} on day 6 (N6), only secretome
    on day 6 (S6), or various combindation of each for all seven \gls{ml}
    techniques}
  \label{tab:mod_results}
  \input{../tables/model_results.tex}
\end{table}

\gls{sr} models achieved the highest predictive performance
($R^2$>\SI{93}{\percent}) when using multi-omics predictors for all endpoint
responses (\cref{tab:mod_results}). \gls{sr} achieved $R^2$>\SI{98}{\percent}
while \gls{gbm} ensembles showed \gls{loocv} $R^2$ > \SI{95}{\percent} for
\rmemh{} and \rmemk{} responses. Similarly, \gls{lasso}, \gls{plsr}, and
\gls{svm} methods showed consistently high \gls{loocv}, (\SI{92.9}{\percent},
\SI{99.7}{\percent}, and \SI{90.5}{\percent} respectively), to predict the
\rratio{}. Yet, about \SI{10}{\percent} reduction in \gls{loocv},
\SIrange{72.5}{81.7}{\percent}, was observed for \rmemh{} with these three
methods. Lastly, \gls{sr} and \gls{plsr} achieved $R^2$>\SI{90}{\percent} while
other \gls{ml} methods exhibited exceedingly variable \gls{loocv}
(\SI{0.3}{\percent} for \gls{rf} to \SI{51.5}{\percent} for \gls{lasso}) for
\rmemk{}.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/sr_omics.png}

  \endgroup
  \caption[Symbolic Regression Cytokine Dependencies]
  {Multi-omics culturing media prediction profiles at day 6 using symbolic
    regression.}
  \label{fig:sr_omics}
\end{figure*}

The top-performing technique, \gls{sr}, showed that the median aggregated
predictions for \rmemh{} \rmemk{} increases when IL2 concentration, IL15, and
IL2R increase while IL17a decreases in conjunction with other features. These
patterns combined with low values of \pdms{} and GM-CSF uniquely characterized
maximum \rmemk{}. Meanwhile, higher glycine but lower IL13 in combination with
others showed maximum \rmemh{} predictions (\cref{fig:sr_omics}).

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/modeling_flower.png}
  \phantomsubcaption\label{fig:mod_flower_48r}
  \phantomsubcaption\label{fig:mod_flower_cd4}

  \endgroup
  \caption[Data-Driven \gls{cqa} identification]
  {Data-driven modeling using techniques with regularization reveals species
    predictive species which are candidates for \glspl{cqa}. Flower plots are
    shown for \subcap{fig:mod_flower_48r}{CD4:CD8 ratio} and
    \subcap{fig:mod_flower_cd4}{total \ptmemh{} cells}. The left and right
    columns includes models that were trained only on the secretome and
    metabolome respectively. Each flower on each plot represents one model,
    moving toward the center indicates higher agreement between models.}
  \label{fig:mod_flower}
\end{figure*}

Selecting \gls{cpp} and \glspl{cqa} candidates consistently for T cell memory is
desired. Here, \gls{tnfa} was found in consensus across all seven \gls{ml}
methods for predicting \rratio{} when considering features with the highest
importance scores across models (\cref{fig:mod_flower_48r}). Other features,
IL2R, IL4, IL17a, and \pdms{}, were commonly selected in $\ge$ 5 \gls{ml}
methods (\cref{fig:mod_flower_48r}). When restricting the models only to include
metabolome, formate emerged as the dominant predictor shared across all seven
models.

% Moreover, IL13 and IL15 were found predictive in combination
% with these using \gls{sr} (Supp.Table.S4).

When performing similar analysis on \rmemh{}, we observe that no species for
either the secretome or metabolome was agreed upon by all seven models
(\cref{fig:mod_flower_cd4}). We also observed that these models did not fit as
well as they did for \rratio{} (\cref{tab:mod_results}). For the secretome, the
species that were agreed upon by $\ge$ 5 models were IL4, IL17a, and IL2R. For
the metabolome, formate once again was agreed upon by $\ge$ 5 models as well as
lactate.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/nmr_cors.png}
  \phantomsubcaption\label{fig:nmr_cors_lactate}
  \phantomsubcaption\label{fig:nmr_cors_formate}
  \phantomsubcaption\label{fig:nmr_cors_glucose}
  \phantomsubcaption\label{fig:nmr_cors_matrix}

  \endgroup
  \caption[NMR Day 4 correlations]
  {\gls{nmr} features at day 4 are strongly correlated with each other and the
    response variables. Highly correlated relationships are shown for
    \subcap{fig:nmr_cors_lactate}{lactate},
    \subcap{fig:nmr_cors_formate}{formate}, and
    \subcap{fig:nmr_cors_glucose}{glucose}. Blue and blue connections indicate
    positive and negative correlations respectively. The threshold for
    visualizing connections in all cases was 0.8.
    \subcap{fig:nmr_cors_matrix}{The correlation matrix for all predictive
      features and the total \ptmemh{} response.}
  }
  \label{fig:nmr_cors}
\end{figure*}

We also investigated the \gls{nmr} features extracted from day of expansion to
assess if there was any predictive power for \ptmemh{}; in general these models
had almost as good of fit despite being 2 days earlier in the process
(\cref{fig:nmr_cors}). Lactate and formate were observed to correlate with each
other, and both correlated with \rmemh{}. Furthermore, lactate was observed to
positively correlate with \pdms{} and negatively correlate with glucose
(\cref{fig:nmr_cors_lactate}). Formate also had the same correlation patterns
(\cref{fig:nmr_cors_formate}). Glucose was only negatively correlated with
formate and lactate (\cref{fig:nmr_cors_glucose}). Together, these data suggest
that lactate, formate, \pdms{}, and \rmemh{} are fundamentally linked.

\section{Discussion}

\gls{cpp} modeling and understanding are critical to new product development and
in cell therapy development, it can have life-saving implications. The
challenges for effective modeling grow with the increasing complexity of
processes due to high dimensionality, and the potential for process interactions
and nonlinear relationships. Another critical challenge is the limited amount of
available data, mostly small \gls{doe} datasets. \gls{sr} has the necessary
capabilities to resolve the issues of process effects modeling and has been
applied across multiple industries\cite{Kordona}. \gls{sr} discovers
mathematical expressions that fit a given sample and differs from conventional
regression techniques in that a model structure is not defined \textit{a
priori}\cite{Koza1992}. Hence, a key advantage of this methodology is that
transparent, human-interpretable models can be generated from small and large
datasets with no prior assumptions\cite{Kotancheka}.

Since the model search process lets the data determine the model, diverse and
competitive model structures are typically discovered. An ensemble of diverse
models can be formed where its constituent models will tend to agree when
constrained by observed data yet diverge in new regions. Collecting data in
these regions helps to ensure that the target system is accurately modeled, and
its optimum is accurately located\cite{Kotancheka}. Exploiting these features
allows adaptive data collection and interactive modeling. Consequently, this
\gls{adoe} approach is useful in a variety of scenarios, including maximizing
model validity for model-based decision making, optimizing processing parameters
to maximize target yields, and developing emulators for online optimization and
human understanding\cite{Kotancheka}.

An in-depth characterization of potential \gls{dms} based T cell \glspl{cqa}
includes a list of cytokine and \gls{nmr} features from media samples that are
crucial in many aspects of T cell fate decisions and effector functions of
immune cells. Cytokine features were observed to slightly improve prediction and
dominated the ranking of important features and variable combinations when
modeling together with \gls{nmr} media analysis and process parameters
(\cref{fig:mod_flower}).

Predictive cytokine features such as \gls{tnfa}, IL2R, IL4, IL17a, IL13, and
IL15 were biologically assessed in terms of their known functions and activities
associated with T cells. T helper cells secrete more cytokines than T cytotoxic
cells, as per their main functions, and activated T cells secrete more cytokines
than resting T cells. It is possible that some cytokines simply reflect the
\rratio{} and the activation degree by proxy proliferation. However, the exact
ratio of expected cytokine abundance is less clear and depends on the subtypes
present, and thus examination of each relevant cytokine is needed.

IL2R is secreted by activated T cells and binds to IL2, acting as a sink to
dampen its effect on T cells\cite{Witkowska2005}. Since IL2R was much greater
than IL2 in solution, this might reduce the overall effect of IL2, which could
be further investigated by blocking IL2R with an antibody. In T cells, TNF can
increase IL2R, proliferation, and cytokine production\cite{Mehta2018}. It may
also induce apoptosis depending on concentration and alter the CD4+ to CD8+
ratio\cite{Vudattu2005}. Given that TNF has both a soluble and membrane-bound
form, this may either increase or decrease CD4+ ratio and/or memory T cells
depending on the ratio of the membrane to soluble TNF\cite{Mehta2018}. Since
only soluble TNF was measured, membrane TNF is needed to understand its impact
on both CD4+ ratio and memory T cells. Furthermore, IL13 is known to be critical
for \gls{th2} response and therefore could be secreted if there are significant
\glspl{th2} already present in the starting population\cite{Wong2011}. This
cytokine has limited signaling in T cells and is thought to be more of an
effector than a differentiation cytokine\cite{Junttila2018}. It might be
emerging as relevant due to an initially large number of \glspl{th2} or because
\glspl{th2} were preferentially expanded; indeed, IL4, also found important, is
the conical cytokine that induces \gls{th2} differentiation
(\cref{fig:mod_flower}). The role of these cytokines could be investigated by
quantifying \glspl{th1}, \glspl{th2}, or \glspl{th17} both in the starting
population and longitudinally. Similar to IL13, IL17 is an effector cytokine
produced by \glspl{th17}\cite{Amatya2017} thus may reflect the number of
\glspl{th17} in the population. GM-CSF has been linked with activated T cells,
specifically \glspl{th17}, but it is not clear if this cytokine is inducing
differential expansion of CD8+ T cells or if it is simply a covariate with
another cytokine inducing this expansion\cite{Becher2016}. Finally, IL15 has
been shown to be essential for memory signaling and effective in skewing
\gls{car} T cells toward \glspl{tscm} when using membrane-bound IL15Ra and
IL15R\cite{Hurton2016}. Its high predictive behavior goes with its ability to
induce large numbers of memory T cells by functioning in an autocrine/paracrine
manner and could be explored by blocking either the cytokine or its receptor.

Moreover, many predictive metabolites found here are consistent with metabolic
activity associated with T cell activation and differentiation, yet it is not
clear how the various combinations of metabolites relate with each other in a
heterogeneous cell population. Formate and lactate were found to be highly
predictive and observed to positively correlate with higher values of total live
\rmemh{} cells (~\cref{fig:nmr_cors}). Formate is a byproduct of the
one-carbon cycle implicated in promoting T cell activation\cite{RonHarel2016}.
Importantly, this cycle occurs between the cytosol and mitochondria of cells and
formate excreted\cite{Pietzke2020}. Mitochondrial biogenesis and function are
shown necessary for memory cell persistence\cite{van_der_Windt_2012,
  Vardhana2020}. Therefore, increased formate in media could be an indicator of
one-carbon metabolism and mitochondrial activity in the culture.

In addition to formate, lactate was found as a putative \gls{cqa} of \ptmem{}
cells. Lactate is the end-product of aerobic glycolysis, characteristic of
highly proliferating cells and activated T cells\cite{Lunt2011, Chang2013}.
Glucose import and glycolytic genes are immediately upregulated in response to T
cell stimulation, and thus generation of lactate. At earlier time-points, this
abundance suggests a more robust induction of glycolysis and higher overall T
cell proliferation. Interestingly, our models indicate that higher lactate
predicts higher CD4+, both in total and in proportion to CD8+, seemingly
contrary to previous studies showing that CD8+ T cells rely more on glycolysis
for proliferation following activation\cite{Cao2014}. It may be that glycolytic
cells dominate in the culture at the early time points used for prediction, and
higher lactate reflects more cells.

Metabolites that consistently decreased over time are consistent with the
primary carbon source (glucose) and essential amino acids (\gls{bcaa},
histidine) that must be continually consumed by proliferating cells. Moreover,
the inclusion of glutamine in our predictive models also suggests the importance
of other carbon sources for certain T cell subpopulations. Glutamine can be used
for oxidative energy metabolism in T cells without the need for
glycolysis\cite{Cao2014}. Overall, these results are consistent with existing
literature that show different T cell subtypes require different relative levels
of glycolytic and oxidative energy metabolism to sustain the biosynthetic and
signaling needs of their respective phenotypes\cite{Almeida2016,Wang_2012}. It
is worth noting that the trends of metabolite abundance here are potentially
confounded by the partial replacement of media that occurred periodically during
expansion, thus likely diluting some metabolic byproducts (such as formate,
lactate) and elevating depleted precursors (such as glucose and amino acids).
More definitive conclusions of metabolic activity across the expanding cell
population can be addressed by a closed system, ideally with on-line process
sensors and controls for formate, lactate, along with ethanol and glucose.

Practically, knowledge of how cytokines and/or metabolites are related to
outcome can be utilized for process control, which involves measuring the
current state of the culture, comparing it to a desired state, and intervening
if it is outside an acceptable range. In the case of lactate and formate, a
benchtop \gls{nmr} can be utilized to sample the media in real time during
culture. This \gls{nmr} can be tuned to automatically quantify the presence of
lactate and formate. Formate is part of the one-carbon pathway, and thus culture
fate may be controlled by altering the inputs to this pathway (glycine, serine,
choline) and/or adding folic acid inhibitors\cite{Ducker2017}. Since lactate is
a direct byproduct of glycolysis, this may be controlled by altering the
concentration of glucose in solution. Each of these control schemes would need
further study to assess if they have enough precision and temporal resolution to
reasonably ensure product quality. In the case of cytokines, there is currently
no analogue to a benchtop \gls{nmr}; however, research is underway to develop
protein-specific sensors using aptamers\cite{Parolo2020}. Even without these
developments, one could still use \gls{elisa} or Luminex to assess protein
levels in a semi-automated manner, but the disadvantage is that these assays are
temporally discrete and impose a significant time lag before the intervention
can be performed.

\chapter{AIM 2B}\label{aim2b}

\section{Introduction}

The purpose of this sub-aim was to further explore the \gls{dms} platform,
specifically for mechanisms and pathways that could be the basis for additional
\glspl{cpp} that could be optimized to yield higher quantity and quality of T
cells. Our strategy in general was to perturb the \gls{dms} system from the
normal operating conditions at which it was used up until this point either
through temporal modulation of activation signal or by blocking pathways of
interest using \glspl{mab}.

\section{Methods}

\subsection{DMSs Temporal Modulation}

% METHOD The concentration for the surface marker cleavage experiment was much
% higher, if that matters
\glspl{dms} were digested in active T cell cultures via addition of sterile
\product{\gls{colb}}{\sigald}{11088807001} or
\product{\gls{cold}}{\sigald}{11088858001}. Collagenase was dissolved in
\product{\gls{hbss}}{Gibco}{14025-076} or
\product{TexMACS}{\miltenyi}{170-076-307} at approximately \SI{100}{\ug\per\ml}.
This solution was added to T cell cultures at a 1:1 ratio in place of plain
media normally used to feed the cells during the regular media addition cycle at
day 4. Cultures were then incubated as described in \cref{sec:tcellculture}, and
the \glspl{dms} were verified to have been digested after \SI{24}{\hour}.

Adding \glspl{dms} was relatively much simpler; the number of \gls{dms} used per
area on day 0 was scaled up by 3 on day 4 to match the change from a 96 well
plate to a 24 well plate, effectively producing a constant activation signal.

\subsection{Mass Wytometry and Clustering Analysis}

T cells were stained using a \product{34 \gls{cytof} marker
  panel}{Fluidigm}{201322} and \product{cisplatin}{Fluidigm}{201064} which were
used according to the manufacturer’s instructions. \numrange{2e6}{3e6} stained
cells per group were analyzed on a Fluidigm Helios.

Unbiased cell clusters were obtained using \gls{spade} analysis by pooling three
representative \gls{fcs} files and running the \gls{spade} pipeline with k-means
clustering (k = 100), arcsinh transformation with cofactor 5, density
calculation neighborhood size of 5 and local density approximation factor of
1.5, target density of 20000 cells, and outlier density cutoff of
\SI{1}{\percent}\cite{Qiu2017}. All markers in the \gls{cytof} panel were used
in the analysis

\subsection{Integrin Blocking Experiments}

To block \gls{a2b1} and \gls{a2b2}, active T cell cultures with \gls{dms} were
supplemented with \product{\anti{\gls{a2b1}}}{\sigald}{MAB1973Z} and
\product{\anti{\gls{a2b2}}}{\sigald}{MAB1950Z} (both \gls{leaf}) at indicated
concentrations and timepoints. T cells were grown as described in
\cref{sec:tcellculture}.

\gls{a2b1} and \gls{a2b2} were verified to be present on active T cell cultures
by staining with \product{\anti{\gls{a2b1}}-\gls{apc}}{\bl}{328313} and
\product{\anti{\gls{a2b2}}-\gls{fitc}}{\bl}{359305} on day 6 of culture and
analyzing via a \bd{} Accuri flow cytometer.

\subsection{IL15 Blocking Experiments}

To block the \gls{il15r}, we supplemented T cell
cultures activated with \gls{dms} with either
\product{\anti{\gls{il15r}}}{Rnd}{AF247} or \product{\gls{igg} isotype
  control}{RnD}{AB-108-C} at the indicated timepoints and concentrations. T
cells were grown as otherwise described in \cref{sec:tcellculture} with the
exception that volumes were split by $\frac{1}{3}$ to keep the culture volume
constant and minimize the amount of \gls{mab} required.

To block soluble \gls{il15}, we supplemented analogously with
\product{\anti{\gls{il15}}}{RnD}{EEP0419081} or \product{\gls{igg} isotype
  control}{\bl}{B236633}.

\section{Results}

\subsection{Adding or Removing DMSs Alters Expansion and Phenotype}

We hypothesized that adding or removing \gls{dms} in the middle of an active
culture would alter the activation signal and hence the growth trajectory and
phenotype of T cells. While adding \glspl{dms} was simple, the easiest way to
remove \glspl{dms} was to use enzymatic digestion. Collagenase is an enzyme that
specifically targets collagen proteins. Since our \glspl{dms} are composed of
porcine-derived collagen, this enzyme should target the \gls{dms} while sparing
the cells along with any markers we wish to analyze. We tested this specific
hypothesis using either \gls{colb}, \gls{cold} or \gls{hbss}, and stained the
cells using a typical marker panel to assess if any of the markers were cleaved
off by the enzyme which would bias our final readout. We observed that the
marker histograms in the \gls{cold} group were similar to that of the buffer
group, while the \gls{colb} group visibly lowered CD62L and CD4, indicating
partial enzymatic cleavage (\cref{fig:collagenase_fx}). Based on this result, we
used \gls{cold} moving forward.

% FIGURE this figure is tall and skinny like me
\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/collagenase.png}

  \endgroup
  \caption[Effects Collagenase Treatment on T cells]
  {T cells treated with either \gls{colb}, \gls{cold}, or buffer and then
    stained for various surface markers and analyzing via flow cytometry.}
  \label{fig:collagenase_fx}
\end{figure*}

When either adding more \glspl{dms}, removing \glspl{dms} using \gls{cold}, or
doing nothing, we observed that, counterintuitively, cell growth seemed to be
inhibited in the \textit{added} group while the cells seemed to grow faster in
the \textit{removed} group relative to the \textit{no change} group
(\cref{fig:add_rem_growth}). Additionally, the \textit{removed} group seemed to
have a negative growth rate in the final \SI{4}{\day} of culture, indicating
that either the lack activation signal had slowed the cell growth down or that
the cells were growing fast enough to outpace the media feeding schedule. The
viability was the same between all groups, indicating that this negative growth
rate and the lower growth rate in the \textit{added} group were likely not due
to cell death (\cref{fig:add_rem_viability}). Interestingly, the \textit{added}
group had significantly higher \pth{} cells compared to the \textit{no change}
group, and the inverse was true for the \textit{removed} group
(\cref{fig:add_rem_cd4}). These results show that the growth rate and phenotype
are fundamentally altered by changing the number of \glspl{dms} temporally.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/add_remove_endpoint.png}
  \phantomsubcaption\label{fig:add_rem_growth}
  \phantomsubcaption\label{fig:add_rem_viability}
  \phantomsubcaption\label{fig:add_rem_cd4}

  \endgroup
  \caption[Endpoint results from adding/removing \gls{dms} on day 4]
  {Changing \gls{dms} concentration on day 4 has profound effects on phenotype
    and growth.
    \subcap{fig:add_rem_growth}{Longitudinal fold change}, 
    \subcap{fig:add_rem_viability}{longitudinal viability}, and
    \subcap{fig:add_rem_cd4}{day 14 \pthp{}} of T cell cultures with \glspl{dms}
    added, removed, or kept the same on day 4.
  }
  \label{fig:add_rem}
\end{figure*}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/spade_gates.png}

  \endgroup
  \caption[SPADE Gating Strategy]
  {Gating strategy for quantifying early-differentiated T cells via
    \gls{spade}.}
  \label{fig:spade_gates}
\end{figure*}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/add_remove_spade.png}
  \phantomsubcaption\label{fig:spade_msts}
  \phantomsubcaption\label{fig:spade_quant}
  \phantomsubcaption\label{fig:spade_tsne_all}
  \phantomsubcaption\label{fig:spade_tsne_stem}

  \endgroup
  \caption[SPADE and tSNE analysis temporally-modified DMS concentration]
  {Removing \glspl{dms} leads to a higher fraction of potent stem-memory T
    cells compared to both adding and not changing the \gls{dms} concentration
    at day 4.
    \subcap{fig:spade_msts}{SPADE plots of CD4, CD45RA, CD27, and CD45RO
      expression on T cells. All cells from the added, removed, or no change
      groups were pooled and clustered at once.}
    \subcap{fig:spade_quant}{T cells from SPADE plots clustered by expression in
      (\subref{fig:spade_msts}) quantified against total cell number from each
      group.}
    \subcap{fig:spade_tsne_all}{\gls{tsne} plots of all cells pooled from all
      groups.}
    \subcap{fig:spade_tsne_stem}{\gls{tsne} plots of T cells from all groups
      manually gated on \cdp{8}\cdp{27}\cdp{45RO}.}
  }
  \label{fig:spade}
\end{figure*}

We next asked what the effect of removing the \glspl{dms} would have on other
phenotypes, specifically \gls{tcm} and \gls{tscm} cells. To this end we stained
cells using a 34-marker mass cytometry panel and analyzed them using a Fluidigm
Helios. After pooling the \gls{fcs} file events from each group and analyzing
them via \gls{spade} we see that there is a strong bifurcation of CD4 and CD8 T
cells. We also observe that among CD27, CD45RA, and CD45RO (markers commonly
used to identify \gls{tcm} and \gls{tscm} subtypes) we see clear `metaclusters'
composed of individual \gls{spade} clusters which are high for that marker
(\cref{fig:spade_msts,fig:spade_gates}). We then gated each of these
metaclusters according to their marker levels and assigned them to one of three
phenotypes for both the CD4 and CD8 compartments: \gls{tcm} (high CD45RO, low
CD45RA, high CD27), \gls{tscm} (low CD45RO, high CD45RA, high CD27), and
`transitory' \gls{tscm} cells (mid CD45RO, mid CD45RA, high CD27). Together
these represent low differentiated cells which should be highly potent as
anti-tumor therapies.

When quantifying the number of cells from each experimental group in these
phenotypes, we clearly see that the number of lower differentiated cells is much
higher in the \textit{no change} or \textit{removed} groups compared to the
\textit{added} group (\cref{fig:spade_quant}). Furthermore, the \textit{removed}
group had a much higher fraction of \gls{tscm} cells compared to the \textit{no
  change} group, which had more `transitory \gls{tscm} cells'. The majority of
these cells were \cdp{8} cells. When analyzing the same data using \gls{tsne},
we observe a higher fraction of CD27 and lower fraction of CD45RO in the the
\textit{removed} group (\cref{fig:spade_tsne_all}). When manually gating on the
CD27+CD45RO- population, we see there is higher density in the \textit{removed}
group, indicating more of this population (\cref{fig:spade_tsne_stem}).
Together, these data indicate that removing \glspl{dms} at lower timepoints
leads to potentially higher expansion, lower \pthp{}, and higher fraction of
lower differentiated T cells such as \gls{tscm}, and adding \gls{dms} seems to
do the inverse.

\subsection{Blocking Integrin Binding Does not Alter Expansion or Phenotype}

One of the reasons the \gls{dms} platform might perform better than the beads is
the fact that they are composed of gelatin, which is a collagen derivative. The
beads are simply \gls{mab} attached to a polymer resin coated onto an iron oxide
core, and thus have no analogue for collagen. Collagen domains present on the
\gls{dms} group could be creating pro-survival and pro-expansion signals to the
T cells through \gls{a2b1} and \gls{a2b2}, causing them to grow better in the
\gls{dms} system.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/integrin_1.png}
  \phantomsubcaption\label{fig:inegrin_1_overview}
  \phantomsubcaption\label{fig:inegrin_1_fc}
  \phantomsubcaption\label{fig:inegrin_1_mem}
  \phantomsubcaption\label{fig:inegrin_1_cd49}

  \endgroup
  \caption[Integrin blocking I]
  {Blocking with integrin does not lead to differences in memory or growth.
    \subcap{fig:inegrin_1_overview}{Experimental overview}
    \subcap{fig:inegrin_1_fc}{Fold change of \gls{dms}-activated T cell over
      time with each blocking condition.}
    \subcap{fig:inegrin_1_mem}{\ptmemp{} at day 14 for each blocked condition.}
    \subcap{fig:inegrin_1_cd49}{\gls{a2b1} and \gls{a2b2} expression over time.}
    `A' and `B' refer to the inclusion of \anti{\gls{a2b1}} or \anti{\gls{a2b2}}
    respectively.
  }
  \label{fig:integrin_1}
\end{figure*}

\begin{table}[!h] \centering
  \caption{Linear regression for day 14 phenotype shown in \cref{fig:integrin_1}}
  \label{tab:integrin_1_reg}
  \input{../tables/integrin_1_reg.tex}
\end{table}

We tested this hypothesis by adding blocking \glspl{mab} against \gls{a2b1}
and/or \gls{a2b2} to running T cell cultures activated using the \glspl{dms}.
These block \glspl{mab} were added at day 6 of culture when \gls{a2b1} and
\gls{a2b2} were known to be expressed\cite{Hemler1990}. We found that the fold
expansion was identical in all the blocked groups vs the unblocked control group
(\cref{fig:inegrin_1_fc}). Furthermore, we observed that the \ptmemp{} (total
and across the CD4/CD8 compartments) was not significantly different between any
of the groups (\cref{fig:inegrin_1_mem,tab:integrin_1_reg}). We also noted that
\gls{a2b1} and \gls{a2b2} were present on the surface of a significant subset of
T cells at day 6, showing that the target we wished to block was present
(\cref{fig:inegrin_1_cd49}).

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/integrin_2.png}
  \phantomsubcaption\label{fig:inegrin_2_overview}
  \phantomsubcaption\label{fig:inegrin_2_fc}
  \phantomsubcaption\label{fig:inegrin_2_mem}

  \endgroup
  \caption[Integrin blocking II]
  {Blocking with integrin does not lead to differences in memory or growth.
    \subcap{fig:inegrin_1_fc}{Fold change of \gls{dms}-activated T cell over
      time with each blocking condition.}
    \subcap{fig:inegrin_1_mem}{\ptmemp{} at day 14 for each blocked condition.}
    `A' and `B' refer to the inclusion of \anti{\gls{a2b1}} or \anti{\gls{a2b2}}
    respectively.
  }
  \label{fig:integrin_2}
\end{figure*}

\begin{table}[!h] \centering
  \caption{Linear regression for day 14 phenotype shown in \cref{fig:integrin_2}}
  \label{tab:integrin_2_reg}
  \input{../tables/integrin_2_reg.tex}
\end{table}

Since this last experiment gave a negative result, we decided to block
\gls{a2b1} and \gls{a2b2} harder by adding \glspl{mab} at more timepoints
between day 0 and day 6, hypothesizing that the majority of the signaling would
be during the period of culture where the \gls{dms} surface concentration was at
its maximum. Once again, we observed no difference between any of the blocked
conditions and the unblocked controls in regard to expansion
(\cref{fig:inegrin_2_fc}). Furthermore, none of the \ptmemp{} readouts (total,
CD4, or CD8) were statistically different between groups
(\cref{fig:inegrin_2_mem,tab:integrin_2_reg}).

Taken together, these data suggest that the advantage of the \gls{dms} platform
is not due to signaling through \gls{a2b1} or \gls{a2b2}.

\subsection{Blocking IL15 Signaling does not Alter Expansion or Phenotype}

\gls{il15} is a cytokine responsible for memory T cell survival and maintenance.
Furthermore, we observed in other experiments that it is secreted to a much
greater extend in \gls{dms} compared to bead cultures (\cref{fig:doe_luminex}).
One of our driving hypotheses in designing the \gls{dms} system was that the
higher cell density would lead to greater local signaling. Since we observed
higher \ptmemp{} across many conditions, we hypothesized that \gls{il15} may be
responsible for this, and further that the unique \textit{cis/trans} activity of
\gls{il15} may be more active in the \gls{dms} system due to higher cell
density.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/il15_blockade_1.png}
  \phantomsubcaption\label{fig:il15_1_overview}
  \phantomsubcaption\label{fig:il15_1_fc}
  \phantomsubcaption\label{fig:il15_1_viability}
  \phantomsubcaption\label{fig:il15_1_mem}

  \endgroup
  \caption[IL15 blocking I]
  {Blocking IL15Ra does not lead to differences in memory or growth.
    \subcap{fig:il15_1_overview}{Experimental overview}
    Longitudinal measurements of
    \subcap{fig:il15_1_fc}{fold change} and
    \subcap{fig:il15_1_viability}{viability} for blocked and unblocked
    conditions expanded with either beads or \glspl{dms}.
    \subcap{fig:il15_1_mem}{Flow cytometry markers for \gls{dms}-expanded T
      cells at day 14 for blocked and unblocked groups.}.
  }
  \label{fig:il15_1}
\end{figure*}

We first tested this hypothesis by blocking \gls{il15r} with either a specific
\gls{mab} or an \gls{igg} isotype control at
\SI{5}{\ug\per\ml}\cite{MirandaCarus2005}. We observed no difference in the
expansion rate of blocked or unblocked cells (this experiment also had
bead-based groups but they did not expand well and thus were not included)
(\cref{fig:il15_1_fc}). Furthermore, there were no differences in viability
between any group (\cref{fig:il15_1_viability}). We also performed flow
cytometry to asses the \ptmemp{} and \pthp{} outputs. Without even gating the
samples, simply lining up their histograms showed no difference between any of
the markers, and by extension showing no difference in phenotype
(\cref{fig:il15_1_mem}).

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/il15_blockade_2.png}
  \phantomsubcaption\label{fig:il15_2_overview}
  \phantomsubcaption\label{fig:il15_2_fc}
  \phantomsubcaption\label{fig:il15_2_viability}
  \phantomsubcaption\label{fig:il15_2_mem}

  \endgroup
  \caption[IL15 blocking II]
  {Blocking soluble IL15 does not lead to differences in memory or growth.
    \subcap{fig:il15_2_overview}{Experimental overview}
    Longitudinal measurements of
    \subcap{fig:il15_2_fc}{fold change} and
    \subcap{fig:il15_2_viability}{viability} for blocked and unblocked
    conditions expanded with \glspl{dms}.
    \subcap{fig:il15_2_mem}{Flow cytometry markers for \gls{dms}-expanded T
      cells at day 14 for blocked and unblocked groups.}.
  }
  \label{fig:il15_2}
\end{figure*}

We next tried blocking soluble \gls{il15} itself using either a \gls{mab} or an
\gls{igg} isotype control. \anti{\gls{il15}} or \gls{igg} isotype control was
added at \SI{5}{\ug\per\ml}, which according to \cref{fig:doe_luminex} was in
excess of the \gls{il15} concentration seen in past experiments by over
\num{20000} times. Similarly, we observed no difference between fold change,
viability, or marker histograms between any of these markers, showing that
blocking \gls{il15} led to no difference in growth or phenotype.

% RESULT this can probably be worded more specifically in terms of the cis/trans
% action of IL15
In summary, this data did not support the hypothesis that the \gls{dms} platform
gains its advantages via the \gls{il15} pathway.

\section{Discussion}

This work provides insight for how the \gls{dms} operates and may be optimized
further. The data showing increased \pthp{} when \glspl{dms} are added and the
reverse when removed is consistent with other data we produced via \gls{doe}
showing that higher \gls{dms} concentrations lead to higher \pthp{}
(\cref{fig:doe_responses_cd4,fig:add_rem_cd4}). The difference in this case is
that we showed that altering activation signal analogously affects the \pthp{}
in the dimension of time as well as space. A similar trend was observed with
memory T cells in this aim. Our previous \gls{doe} data showed that, to a point,
lower \gls{dms} concentration leads to higher \ptmemp{}
(\cref{fig:doe_responses_mem}). In this aim, we showed that decreasing
activation signal temporally by removing \glspl{dms} leads to the same effect in
the \gls{tcm}, \gls{tscm} and `transitory' \gls{tscm} populations, (all of which
are included in the \ptmem{} phenotype). Taken together, these imply that
temporally or spatially altering the \gls{dms} concentration, and thus the
activation signal, has similar effects.

% BACKGROUND this sounds like background?
% There are several plausible explanations for the observed phenotypic differences
% between beads and DMSs. First, the DMSs are composed of a collagen derivative
% (gelatin); collagen has been shown to costimulate activated T cells via
% \gls{a2b1} and \gls{a2b2}, leading to enhanced proliferation, increased
% \gls{ifng} production, and upregulated CD25 (IL2R$\upalpha$) surface
% expression8,10,11,41,42.

While we did not find support for our hypothesis that the \gls{dms} signal
through the \gls{a2b1} and/or \gls{a2b2} receptors, we can speculate as to why
either this experiment failed and may be done better in the future, or why these
receptors may simply be irrelevant for our system.

On the first point, we did not verify that these \glspl{mab} indeed blocked the
receptor we were targeting. There has been evidence from other groups that these
particular clones work at the concentrations we used\cite{MirandaCarus2005}.
This does not necessarily mean that the \glspl{mab} we obtained were functional
in blocking their intended targets (although they were from a reputable
manufacturer, \bl). Furthermore, we can safely rule out the possibility that the
\glspl{mab} never reached their targets, as they were added immediately after
the T cells were resuspended as required for cell counting, hence their resting
clustered state was disrupted.

On the second point, the collagen domains may not even be relevant to our system
depending on the nature of the \gls{stp} coating. We intended by design for the
system to be fully coated or nearly fully-coated with \gls{stp}
(\cref{fig:stp_coating}). Thus the domains that \gls{a2b1} and \gls{a2b2} may be
targeting could be sterically hindered by a layer of \gls{stp}, and if not that,
also a layer of CD3/CD28 \glspl{mab}. The other possibility is that these
domains are simply denatured to beyond recognition due to the fabrication
process for the microcarriers we used (which involves a proprietary
cross-linking step to make the material autoclave-safe). Either of these could
be tested and verified by staining the \glspl{dms} with a fluorescently-tagged
\gls{mab} and verifying binding via confocal microscopy or indirect protein
quantification as we do for the \gls{qc} of the \gls{dms}. If this test came
back negative, we would be fairly confident that the \gls{a2b1} and \gls{a2b1}
domains are either unreachable or unrecognizable. Even if it turned out that
collagen binding domains are irrelevant in the \gls{dms} system, previous
studies show that these domains can enhance proliferation and survival, and thus
adding them along with with the \glspl{mab} could enhance T cell
expansion\cite{Aoudjit2000, Gendron2003, Boisvert2007}.

We also failed to uphold our hypothesis that the \gls{dms} system gains its
advantage via \gls{il15} signaling. There could be multiple reasons for why
blocking either \il{15} itself or its receptor would not influence the response
at all. First, it could be that \il{15} is not important in our system, which is
not likely given the importance of \il{15} in T cells expansion and particularly
memory phenotypes\cite{Lodolce1998,Kennedy2000}. Second, in the case of the
receptor it could be that that \glspl{mab} we purchased did not actually block,
which also seems unlikely given that this clone has been observed to inhibit
proliferation in the past (although like the integrin blocking experiments we
did not verify that it blocked ourselves), albeit of resting T
cells\cite{MirandaCarus2005}. Third, it could be that turnover of the receptor
was so high that there were not enough \glspl{mab} to block (the key difference
between our experiment and that of \cite{MirandaCarus2005} was that they used
resting T cells, which are not expressing protein to nearly as high of a
degree). The way to test this would be to simply titrate increasing
concentrations of \gls{mab} (which we did not do in our case because the
\gls{mab} was already very expensive in the concentrations employed for our
experiment). Fourth, the blocking the soluble protein may not have worked
because the \il{15} may have been secreted and immediately captured via
\il{15R$\upalpha$} either by the cell that secreted it or by a neighboring cell.

Regardless of whether or not \il{15} is important for the overall mechanism that
differentiates the \glspl{dms} from the beads, adding \il{15} or its receptor
complex to the surface of the \gls{dms} might produce interesting and positive
results on expansion and memory phenotype. Essentially this would turn the
\glspl{dms} into stromal cells that present \il{15}, as seen to be important in
the early work with \il{15} in mice\cite{Lodolce1998}.

% DISCUSSION not sure if this belongs here, although it might make sense to offer
% alternative explanations of why the DMSs "work" given this negative data
% Second, there is evidence that providing a larger
% contact area for T cell activation provides greater stimulation16,43; the DMSs
% have a rougher interface than the 5 µm magnetic beads, and thus could facilitate
% these larger contact areas. Third, the DMSs may allow the T cells to cluster
% more densely compared to beads, as evidenced by the large clusters on the
% outside of the DMSs (Figure 1f) as well as the significant fraction of DMSs
% found within their interiors (Supplemental Figure 2a and b). This may alter the
% local cytokine environment and trigger different signaling pathways.
% Particularly, IL15 and IL21 are secreted by T cells and known to drive memory
% phenotype44–46. We noted that the IL15 and IL21 concentration was higher in a
% majority of samples when comparing beads and DMSs across multiple timepoints
% (Supplemental Figure 18) in addition to many other cytokines. IL15 and IL21 are
% added exogenously to T cell cultures to enhance memory frequency,45,47 and our
% data here suggest that the DMSs are better at naturally producing these
% cytokines and limiting this need. Furthermore, IL15 unique signals in a trans
% manner in which IL15 is presented on IL15R to neighboring cells48. The higher
% cell density in the DMS cultures would lead to more of these trans interactions,
% and therefore upregulate the IL15 pathway and lead to more memory T cells.

\chapter{AIM 3}\label{aim3}

\section{Introduction}

% DO NOT COMMENT OUT THIS LINE: the real purpose of this aim was to appease
% Nature Biotech because they think that animal models are necessary for good
% science. This entire aim exists because of their foolishness.

The purpose of this aim was to verify that \gls{car} T cells produced using the
\gls{dms} system will show potent anti-tumor properties in a complex \invivo{}
system compared to state-of-the-art bead technology\footnote{adapted from
  \dmspaper{}}. We hypothesized that due to the increased \ptmem{} and \pth{}
phenotypes as shown in \cref{aim1}, that \gls{dms}-expanded T cells would show
longer survival and lower tumor burden than those expanded with beads. We
explored the effect of dosing at different levels and the effect of harvesting T
cells at early timepoints in the culture, which has been shown to produce
lower-differentiated T cells with higher potency\cite{Ghassemi2018}.

\section{Methods}

\subsection{CD19-CAR T Cell Generation}

\subsection{T Cell Culture}

T cells were grown as described in \cref{sec:tcellculture}.


\subsection{\Invivo{} Therapeutic Efficacy in NSG Mice Model}

% METHOD describe how the luciferase cells were generated (eg the kwong lab)
% METHOD use actual product numbers for mice
All mice in this study were male \gls{nsg} mice from Jackson Laboratories. At
day 0 (\SI{-7}{\day} relative to T cell injection), 1e6 firefly
luciferase-expressing \product{Nalm-6 cells}{ATCC}{CRL-3273} suspended in
ice-cold \gls{pbs} were injected via tail vein into each mouse. At day 7, saline
or \gls{car} T cells at the indicated doses from either bead or
\gls{dms}-expanded T cell cultures (for \SI{14}{\day}) were injected into each
mouse via tail vein. Tumor burden was quantified longitudinally via an
\gls{ivis} Spectrum (Perkin Elmer). Briefly, \SI{200}{\ug} luciferin at
\SI{15}{\mg\per\ml} in \gls{pbs} was injected intraperitoneally under isoflurane
anesthesia into each mouse and allowed to circulate for at least
\SI{10}{\minute} before imaging. Mice were anesthetized again and imaged using
the \gls{ivis}. Mice from each treatment group/dose were anesthetized, injected,
and imaged together, and exposure time of the \gls{ivis} was limited to avoid
saturation based on the signal from the saline group. \gls{ivis} images were
processed by normalizing them to common minimum and maximum photon counts and
total flux was estimated in terms of photons/second. Endpoint for each mouse was
determined by \gls{iacuc} euthanasia criteria (hunched back, paralysis,
blindness, lethargy, and weight loss). Mice were euthanized according to these
endpoint criteria using carbon dioxide asphyxiation.

\subsection{Statistics}

For the \invivo{} model, the survival curves were created and statistically
analyzed using GraphPad Prism using the Mantel-Cox test to assess significance
between survival groups.

\section{Results}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/mouse_dosing_overview.png}

  \endgroup
  \caption[Mouse Dosing Experimental Overview]
  {Overview of \invivo{} experiment to test \gls{car} T cells expanded with
    either \glspl{dms} at different doses. }
  \label{fig:mouse_dosing_overview}
\end{figure*}

\begin{table}[!h] \centering
  \caption{Results for \gls{car} T cell \invivo{} dose study}
  \label{tab:mouse_dosing_results}
  \input{../tables/mouse_dose_car.tex}
\end{table}

\subsection{DMS-expanded T Cells Show Greater Anti-Tumor Activity \invivo{}
  Compared to Beads}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/mouse_dosing_qc.png}
  \phantomsubcaption\label{fig:mouse_dosing_qc_mem}
  \phantomsubcaption\label{fig:mouse_dosing_qc_cd4}
  \phantomsubcaption\label{fig:mouse_dosing_qc_growth}

  \endgroup
  \caption[Mouse Dosing T cell Characteristics]
  {Characteristics of T cells harvested at day 14 injected into NSG
    mice at varying doses.
    Fractions of T cell subtypes in the day 14 product including
    \subcap{fig:mouse_dosing_qc_mem}{\ptmemp{}}.
    \subcap{fig:mouse_dosing_qc_cd4}{\pthp{}}, and
    \subcap{fig:mouse_dosing_qc_growth}{Fold change of T cells.}
  }
  \label{fig:mouse_dosing_qc}
\end{figure*}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/mouse_dosing_ivis.png}
  \phantomsubcaption\label{fig:mouse_dosing_ivis_images}
  \phantomsubcaption\label{fig:mouse_dosing_ivis_plots}
  \phantomsubcaption\label{fig:mouse_dosing_ivis_survival}
  \phantomsubcaption\label{fig:mouse_dosing_ivis_survival_comp}
  \phantomsubcaption\label{fig:mouse_dosing_ivis_survival_full}

  \endgroup
  \caption[Mouse Dosing IVIS and Survival Results]
  {T cells expanded with \glspl{dms} confer greater anti-tumor potency \invivo{}
    even at lower doses.
    \subcap{fig:mouse_dosing_ivis_images}{IVIS images of Nalm-6 tumor-bearing
      \gls{nsg} mice injected with varying doses of T cells}
    \subcap{fig:mouse_dosing_ivis_plots}{Plots showing quantified photon counts
      of the results from (\subref{fig:mouse_dosing_ivis_plots}).}
    \subcap{fig:mouse_dosing_ivis_survival}{Survival plots of mice}
    \subcap{fig:mouse_dosing_ivis_survival_comp}{Survival plots of mice showing
      only those that received a comparable number of \gls{car} T cells.}
    \subcap{fig:mouse_dosing_ivis_survival_full}{The same data as
      \subref{fig:mouse_dosing_ivis_survival} except showing the full time until
    euthanasia for all mice (including those that died via \gls{gvhd}). Survival
    curves were statistically analyzed using the logrank test in GraphPad
    Prism.}
  }
  \label{fig:mouse_dosing_ivis}
\end{figure*}

We asked if the higher memory/naive phenotype and more balanced CD4/CD8 ratio of
our \gls{dms}-expanded \gls{car} T cells would lead to better anti-tumor potency
in vivo compared to bead-expanded \gls{car} T cells. We also asked if this
superior anti-tumor potency would hold true at lower doses of \gls{car}
expressing T cells in the DMS group vs the bead group. To test this, we used a
human xenograft model of B cell \gls{all} by intravenously injecting \gls{nsg}
mice with \num{1e6} Nalm-6 tumor cells expression firefly
luciferase\cite{Fraietta2018}. After \SI{7}{\day} of tumor cell growth
(\cref{fig:mouse_dosing_overview}), we intravenously injected saline or three
doses (high, medium, and low) of \gls{car} T cells from either bead or \gls{dms}
cultures expanded for \SI{14}{\day}. We quantified \ptcarp{} bead and \gls{dms}
groups using the \gls{ptnl} assay (\cref{tab:mouse_dosing_results}).

Before injecting the T cells into the mice, we quantified their phenotype and
growth. We observed that for this expansion, the bead and \gls{dms} T cells
produced similar numbers of \ptmem{} T cells, and the beads even had a higher
fraction of CD45RA, which is present on lower-differentiated \glspl{tn} and
\glspl{tscm} (\cref{fig:mouse_dosing_qc_mem}). However, the \pthp{} of
the final product was higher in \gls{dms} (\cref{fig:mouse_dosing_qc_cd4}). The
\gls{dms} T cells also expanded more robustly than the beads
(\cref{fig:mouse_dosing_qc_growth}).

In the Nalm-6/\gls{nsg} xenograft model, we observed lower tumor burden and
significantly longer survival of bead and \gls{dms}-treated mice at all doses
compared to the saline groups (\cref{fig:mouse_dosing_ivis}). Importantly, at
each dose we observed that the \gls{dms}-treated mice had much lower tumor
burden and significantly higher survival than their bead-treated counterparts
(\cref{fig:mouse_dosing_ivis_survival}). When factoring the percentage T cells
in each dose that expressed the \gls{car}, we note that survival of the low
\gls{dms} dose (which had similar total \gls{car} T cells compared to the bead
medium dose and less than the bead high dose) is significantly higher than that
of both the bead medium dose and the bead high dose
(\cref{fig:mouse_dosing_ivis_survival_comp}). Overall, the Kaplan-Meier survival
of Nalm-6 tumor bearing \gls{nsg} mice shown in the
\cref{fig:mouse_dosing_ivis_survival} was up to day 40 as reported
elsewhere\cite{Fraietta2018}. However, we also included a Kaplan-Meier figure up
to day 46 (\cref{fig:mouse_dosing_ivis_survival_full}) where most of the mice
euthanized from day 40 through day 46 from \gls{dms} groups showed no or very
small fragment of spleen which was due to \gls{gvhd} responses. Similar
\gls{gvhd} responses were reported earlier in \gls{nsg} mice where the mice
injected with human \gls{pbmc} exhibited acute \gls{gvhd} between
\SIrange{40}{50}{\day} post intravenous injection\cite{Ali2012}. Notably, both
survival analyses (up to day 40 in \cref{fig:mouse_dosing_ivis_survival} and up
to day 46 in \cref{fig:mouse_dosing_ivis_survival_full}) confirmed that
\gls{dms}-expanded groups outperformed bead-expanded groups in terms of
prolonging survival of Nalm-6 tumor challenged \gls{nsg} mice.

Together, these data suggested that \glspl{dms} produce T cells that are not
only more potent that bead-expanded T cells (even when accounting for
differences in \gls{car} expression) but also showed that \gls{dms} expanded T
cells are effective at lower doses. Given the quality control data of the T
cells prior to injecting into the mice, it seems that this advantage is either
due to the higher \pthp{} or the overall fitness of the T cells given the higher
expansion in the case of \gls{dms}
(\cref{fig:mouse_dosing_qc_cd4,fig:mouse_dosing_qc_growth}). It was likely not
due to the memory phenotype given that it was actually slightly higher in the
case of beads (\cref{fig:mouse_dosing_qc_mem}).

\subsection{Beads and DMSs Perform Similarly at Earlier Timepoints}

We then asked how T cells harvested using either beads or \gls{dms} performed
when harvested at earlier timepoints\cite{Ghassemi2018}. We performed the same
experiments as described in \cref{fig:mouse_dosing_overview} with the
modification that T cells were only grown and harvested after \SI{6}{\day},
\SI{10}{\day}, or \SI{14}{\day} of expansion
(\cref{fig:mouse_timecourse_overview}). T cells were frozen after harvest, and
all timepoints were thawed at the same time prior to injection. The dose of T
cells injected was \num{1.25e6} cells per mouse (the same as the high dose in
the first experiment). All other characteristics of the experiment were the
same.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/mouse_timecourse_overview.png}

  \endgroup
  \caption[Mouse Timecourse Experimental Overview]
  {Overview of \invivo{} experiment to test \gls{car} T cells using either
    \glspl{dms} or bead harvested at varying timepoints.
  }
  \label{fig:mouse_timecourse_overview}
\end{figure*}

As was the case with the first \invivo{} experiment, T cells activated with
\glspl{dms} expanded much more efficiently compared to those expanded with beads
(\cref{fig:mouse_timecourse_qc_growth}). When we quantified the \ptcarp{} of T
cells harvested at each timepoint, we noted that the bead group had much higher
\ptcar{} expression at earlier timpoints compared to \gls{dms}, while they
equalized at later timepoints (\cref{fig:mouse_timecourse_qc_car}). In addition,
overall \ptcar{} expression decreased at later timepoints, indicating that
\gls{car} transduced T cells either grow slower or died faster compared to
untransduced cells. The \pthp{} of the harvested T cells was higher overall in
\gls{dms} expanded T cells but decreased with increasing timepoints 
(\cref{fig:mouse_timecourse_qc_cd4}). The \ptmemp{} was similar at day 6
between bead and \gls{dms} groups but the \gls{dms} group had higher \ptmemp{}
at day 14 despite the overall \ptmemp{} decreasing with time as shown elsewhere
(\cref{fig:mouse_timecourse_qc_mem})\cite{Ghassemi2018}.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/mouse_timecourse_qc.png}
  \phantomsubcaption\label{fig:mouse_timecourse_qc_growth}
  \phantomsubcaption\label{fig:mouse_timecourse_qc_car}
  \phantomsubcaption\label{fig:mouse_timecourse_qc_cd4}
  \phantomsubcaption\label{fig:mouse_timecourse_qc_mem}

  \endgroup
  \caption[Mouse Timecourse T cell Characteristics]
  {Characteristics of T cells harvested at varying timepoints injected into NSG
    mice.
    \subcap{fig:mouse_timecourse_qc_growth}{Fold change of T cells (each
      timepoint only includes the runs that were harvested at day 14).}
    Fractions of T cell subtypes in the day 14 product including
    \subcap{fig:mouse_timecourse_qc_car}{\ptcarp{}},
    \subcap{fig:mouse_timecourse_qc_cd4}{\pthp{}}, and
    \subcap{fig:mouse_timecourse_qc_mem}{\ptmemp{}}.
  }
  \label{fig:mouse_timecourse_qc}
\end{figure*}

We analyzed the tumor burden using \gls{ivis} which showed that mice that
received T cells from any group performed better than those that received only
saline (\cref{fig:mouse_timecourse_ivis}). Note that unlike the previous
experiment, many of the mice survived until day 40 at which point \gls{gvhd}
began to take effect (after euthanizing the mice at day 42, most had small or no
spleen). When comparing bead and \gls{dms} groups, the \gls{dms} T cells still
seemed superior to the bead group, at least initially (note that in this case
they had similar numbers of \ptcar{} cells). At day 6, both \gls{dms} and bead
groups seemed to eradicate the tumor initially, after which it came back after
day 21 for the bead and day 28 for the \gls{dms} group. The day 10 groups
performed somewhere in between, where they increased linearly unlike the day 6
groups but not as quickly as the day 14 groups. In the case of the \gls{dms} day
10 group, it also appeared like a few mice actually performed better than all
other groups in regard to the final tumor burden.

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/mouse_timecourse_ivis.png}
  \phantomsubcaption\label{fig:mouse_timecourse_ivis_images}
  \phantomsubcaption\label{fig:mouse_timecourse_ivis_plots}

  \endgroup
  \caption[Mouse Timecourse IVIS Results]
  {\glspl{dms} exhibit superior anti-tumor activity \invivo{} at day 14 compared
    to beads but are similar to beads at lower timepoints.
    \subcap{fig:mouse_timecourse_ivis_images}{IVIS images for day 6 to day 42 of
    mice treated with varying doses of \gls{car} T cells grown with beads or
    \glspl{dms}.}
    \subcap{fig:mouse_timecourse_ivis_plots}{Quantified dotplots of the images
      in (\subref{fig:mouse_timecourse_ivis_images}). Numbers beneath each dot
      represent the number of mice at that timepoint.},
  }
  \label{fig:mouse_timecourse_ivis}
\end{figure*}

\section{Discussion}

\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/mouse_summary.png}
  \phantomsubcaption\label{fig:mouse_summary_1}
  \phantomsubcaption\label{fig:mouse_summary_2}

  \endgroup
  \caption[Mouse Summary]
  {Summary of cells injected into mice during for
    \subcap{fig:mouse_summary_1}{the first mouse experiment} and 
    \subcap{fig:mouse_summary_2}{the second mouse experiment}. The y axis
    maximum is set to the maximum number of cells injected between both
    experiments (\num{1.25e6}). Note that the \gls{car} was quantified using a
    separate panel than the rest of the markers.
  }
  \label{fig:mouse_summary}
\end{figure*}

The total number of T cells injected for each \invivo{} experiment are shown in
\cref{fig:mouse_summary}.

When we tested bead and \gls{dms} expanded \gls{car} T cells, we found that the
\gls{dms} expanded \gls{car} T cells outperformed bead groups in prolonging
survival of Nalm-6 tumor challenged (intravenously injected) \gls{nsg} mice.
\gls{dms} expanded CAR-T cells were very effective in clearing tumor cells as
early as \SI{7}{\day} post \gls{car} T injection even at low total T cell dose
compared to the bead groups where tumor burden was higher than \gls{dms} groups
across all the total T cell doses tested here. More interestingly, when only
\gls{car}-expressing T cell doses between bead and \gls{dms} groups were
compared, \gls{dms} group had significantly higher survival effects over similar
or higher CAR expression T cell doses from bead group. All these results suggest
that the T cells in \gls{dms} groups (compared to bead group) resulted in highly
effective \gls{car} T cells that can efficiently kill tumor cells.

When comparing the total number of T cells of different phenotypes, we observed
that when comparing low-dose \gls{dms} group to the mid- bead groups (which had
similar numbers of \gls{car} T cells), the number of \ptmem{} (both with and
without CD45RA) T cells injected was much lower in the \gls{dms} group
(\cref{fig:mouse_summary_1}). This could mean several things. First, the
\ptmem{} phenotype may have nothing to do with the results seen here, at least
in this model. While this may have been the case in our hands, this would
contradict previous evidence suggesting that \gls{tn} and \gls{tcm} cells work
better in almost the same model (the only difference being Raji cells in place
of Nalm-6 cells, both of which express CD19)\cite{Sommermeyer2015}. Second, the
distribution of \gls{car} T cells across different subtypes of T cells was
different between the \gls{dms} and bead groups (with possibly higher
correlation of \gls{car} expression and the \ptmem{} phenotype). It is hard to
assess this without strong assumptions as the \gls{car} was quantified using a
separate flow panel relative to the other markers.

We can also make a similar observation for the number of \pth{} T cells injected
(\cref{fig:mouse_summary_1}). In this case, either the \pth{} phenotype doesn't
matter in this model (or the \ptk{} population matters much more), or the
distribution of \gls{car} is different between CD4 and CD8 T cells in a manner
that favors the \gls{dms} group. While in a glioblastoma model and not a B-cell
\gls{all} model, previous groups have shown that \pthp{} T cells are important
for response\cite{Wang2018}.

When testing \gls{car} T cells at earlier timepoints relative to day 14 as used
in the first \invivo{} experiment, we noted that none of the \gls{car}
treatments seemed to work as well as they did in the first experiment. However,
the total number of \gls{car} T cells was generally much lower in this second
experiment relative to the first (\cref{fig:mouse_summary}). Only the day 6
group had \gls{car} T cell numbers comparable to the weakest dose of bead cells
given in the first experiment, and these T cells were harvested at earlier
timepoints than the first mouse experiment and thus may not be safely
comparable. Furthermore, the \ptcarp{} decreased over time, which suggested that
the transduced T cells grew slower. This has been observed elsewhere and could
be due to tonic signaling\cite{GomesSilva2017}. The lower overall \gls{car}
doses may explain why at best, the tumor seemed to be in remission only
temporarily. Even so, the \gls{dms} group seemed to perform better at day 6 as
it held off the tumor longer, and also slowed the tumor progression relative to
the bead group at day 14 (\cref{fig:mouse_timecourse_ivis_plots}).

Taken together, these data suggest that the \gls{dms} platform produces T cells
that have an advantage \invivo{} over beads. While we may not know the exact
mechanism, our data suggests that the responses are unsurprisingly influenced by
the \ptcarp{} of the final product. Followup experiments would need to be
performed to determine the precise phenotype responsible for these responses in
our hands.

\chapter{CONCLUSIONS AND FUTURE WORK}\label{conclusions}

\section{Conclusions}

This dissertation describes the development of a novel T cell expansion
platform, including the fabrication, quality control, and biological validation
of its performance both \invitro{} and \invivo{}. Development of such a system
would be meaningful even if it only performed as well as current methods, as
adding another method to the arsenal of the growing T cell manufacturing
industry would reduce the reliance on a small number of companies that currently
license magnetic bead-based T cell expansion technology. However, we
additionally show that the \gls{dms} platform expands more T cells on average,
including highly potent \ptmem{} and \pth{} T cells, and produces higher
percentages of both. If commercialized, this would be a compelling asset the T
cell manufacturing industry.

In \cref{aim1}, we develop the \gls{dms} platform and verified its efficacy
\invitro{}. Importantly, this included \gls{qc} steps at every critical step of
the fabrication process to ensure that the \gls{dms} can be made within a
targeted specification. These \gls{qc} steps all rely on common, relatively
cost-effective assays such as the \gls{haba} assay, \gls{bca} assay, and
\glspl{elisa}, thus other labs and commercial entities should be able to perform
them. The microcarriers themselves are an off-the-shelf product available from
reputable vendors, and they have a regulatory history in human cell therapies
that will aid in clinical translation\cite{purcellmain}. Both these will help
in translatability. On average, we demonstrated that the \gls{dms} outperforms
state-of-the-art bead-based T cell expansion technology in terms of total fold
expansion, \ptmemp{}, and \pthp{} by \SI{131}{\percent}, \SI{3.5}{\percent}, and
\SI{7.4}{\percent} controlling for donor, operator, and a variety of process
conditions.

In addition to larger numbers of potent T cells, other advantages of our
\gls{dms} approach are that the \glspl{dms} are large enough to be filtered
(approximately \SI{300}{\um}) using standard \SI{40}{\um} cell filters or
similar. If the remaining cells inside that \glspl{dms} are also desired,
digestion with dispase or collagenase may be used. Collagenase D may be
selective enough to dissolve the \gls{dms} yet preserve surface markers which
may be important to measure as critical quality attributes \glspl{cqa}
(\cref{fig:collagenase_fx}). Furthermore, our system should be compatible with
large-scale static culture systems such as the G-Rex bioreactor or perfusion
culture systems, which have been previously shown to work well for T cell
expansion\cite{Forget2014, Gerdemann2011, Jin2012}. 

In \cref{aim2a}, we developed a modeling pipeline that can be used by commercial
entities as the scale up this process to identify \glspl{cqa} and \gls{cpp}.
These are highly important for a variety of reasons. First, understanding
pertinent \glspl{cpp} allow manufacturers to operate their process at optimal
conditions. This is important for anti-tumor cell therapies, where the prospects
of a patient can urgently depend on receiving therapy in a timely manner.
Optimal process conditions allow T cells to be expanded as quickly as possible
for the patient, while also minimizing cost for the manufacturer. Second,
\glspl{cqa} can be used to define process control schemes as well as release
criteria. Process control, and with it the ability to predict future outcomes
based on data obtained at the present, is highly important for cell therapies
given that batch failures are extremely expensive\cite{Harrison2019}, and
predicting a batch failure would allow manufacturers to restart the batch in a
timely manner without wasting resources. Furthermore, \glspl{cqa} can be used to
define what a `good' vs `bad' product is, which will important help anticipate
dosing and followup procedures in the clinic if the T cells are administered. In
the aim, we cannot claim to have found the ultimate set of \glspl{cqa} and
\glspl{cpp}, as we used tissue culture plates instead of a bioreactor and we
only used one donor. However, we have indeed outlined a process that others may
use to find these for their process. In particular, the 2-phase modeling process
we used (starting with a \gls{doe} and collecting data longitudinally) is a
strategy that manufacturers can easily implement. Also, collecting secretome and
metabolome is easily generalized to any setting and to most bioreactors and
expansion systems, as they can be obtained with relatively inexpensive equipment
(Luminex assay, benchtop \gls{nmr}, etc) without disturbing the cell culture.

In \cref{aim2b}, we further explored additional tuning knobs that could be used
to control and optimize the \gls{dms} system. We determined that altering the
\gls{dms} concentration temporally has profound effects on the phenotype and
expansion rate. This agrees with other data we obtained in \cref{aim2a} and with
what others have generally reported about signal strength and T cell
differentiation\cite{Gattinoni2012, Lozza2008, Lanzavecchia2005, Corse2011}. We
did not find any mechanistic relationship between either integrin signaling or
\gls{il15} signaling. In the case of the former, it may be more likely that the
\glspl{dms} surfaces are saturated to the point of sterically hindering any
integrin interactions with the collagen surface. In the case of \gls{il15} more
experiments likely need to be done in order to plausibly rule out this mechanism
and/or determine if it is involved at all.

In \cref{aim3} we determined that the \glspl{dms} expand T cells that also
performed better than beads \invivo{}. In the first experiment we performed, the
results were very clearly in favor of the \glspl{dms}. In the second experiment,
even the \gls{dms} group failed to fully control the tumor burden, but this is
not surprising given the low \ptcarp{} across all groups. Also, despite this,
the \gls{dms} group appeared to control the tumor better on average for early,
mid, and late T cell harvesting timepoints. It was not clear if this effect was
due to increased \pthp{}, \ptmemp{}, or fitness of the \gls{dms}-expanded T
cells given their higher expansion rate. More data is needed to establish which
phenotype is responsible for the results we observed, as we did not include the
\gls{car} in the same panel as the other phenotype surface markers, making it
difficult to reliably say the identity of the \ptcar{} cells.

Finally, while we have demonstrated the \gls{dms} system in the context of
\gls{car} T cells, this method can theoretically be applied to any T cell
immunotherapy which responds to \acd{3}/\acd{28} \gls{mab} and cytokine
stimulation. These include \glspl{til}, virus-specific T cells, T cells
engineered to express $\upgamma\updelta$ \glspl{tcr}, $\upgamma\updelta$ T
cells, T cells with transduced-\gls{tcr}, and \gls{car}-\gls{tcr} T
cells\cite{Cho2015, Straetemans2018, Robbins2011, Brimnes2012, Baldan2015,
  Walseng2017}. Similar to \glspl{car} against CD19 used in liquid tumors, these
T cell immunotherapies would similarly benefit from the increased proliferative
capacity, metabolic fitness, migration, and engraftment potential characteristic
of naïve and memory phenotypes\cite{Blanc2018, Lalor2016, Rosato2019}. Indeed,
since these T cell immunotherapies are activated and expanded with either
soluble \glspl{mab} or bead-immobilized \glspl{mab}, our system will likely
serve as a drop-in substitution to provide these benefits.

\section{Future Directions}

There are several important next steps to perform with this work, many of which
will be relevent to using this technology in a clinical trial:

\subsection{Using GMP Materials}

While this work was done with translatability and \gls{qc} in mind, an important
feature that is missing from the process currently is the use of \gls{gmp}
materials. The microcarriers themselves are made from porcine-derived collagen,
which itself is not \gls{gmp}-compliant due to its non-human animal origins.
However, using any other source of collagen should work so long as the structure
of the microcarriers remains relatively similar and it has lysine groups that
can react with the \gls{snb} to attach \gls{stp} and \glspl{mab}. Obviously
these would need to be tested and verified, but these should not be
insurmountable. Furthermore, the \gls{mab} binding step requires \gls{bsa} to
prevent adsorption to the non-polar polymer walls of the reaction tubes. A human
carrier protein such as \gls{hsa} could be used in its place to eliminate the
non-human animal origin material, but this could be much more expensive.
Alternatively, the use of protein could be replaced altogether by a non-ionic
detergent such as Tween-20 or Tween-80, which are already used for commercial
\gls{mab} formulations for precisely this purpose\cite{Kerwin2008}. Validating
the process with Tween would be the best next step to eliminate \gls{bsa} from
the process. The \gls{stp} and \glspl{mab} in this work were not
\gls{gmp}-grade; however, they are commonly used in clinical technology such as
dynabeads and thus the research-grade proteins used here could be easily
replaced. The \gls{snb} is a synthetic small molecule and thus does not have any
animal-origin concerns.

\subsection{Mechanistic Investigation}

Despite the improved outcomes in terms of expansion and phenotype relative to
beads, we don't have a good understanding of why they \gls{dms} platform works
as well as it does. The following are several plausible hypotheses and a
proposed experiment for testing them:

\subsubsection{Cytokine Cross-talk}

As hypothesized in the beginning of this work, the \gls{dms} may derive their
advantage through increased cytokine cross-talk. While this work found that
blocking \il{15} did not lead to differences in outcome, other cytokines could
be explored in a similar vein.

An efficient test of this hypothesis would be to simply incubate T cells grown
with either bead or \glspl{dms} with a cocktail of \glspl{mab} each feeding
cycle that target the cytokines seen in \cref{fig:doe_luminex}, assuming that at
least a few of the targeted cytokines will cause a difference. The experiment
should be sized appropriately such that the second order interaction effect can
be resolved (that is, the effect of adding the cocktail conditional on the
activation method). In these terms, we hypothesize that the growth and phenotype
will be more similar between the beads and \glspl{dms} when the cocktail is
added, while the \gls{dms} will have better expansion and phenotype when the
cocktail is not added. If this experiment shows any effects, the cytokines
responsible can be resolved by testing individually (or in small pools).

One caveat with this approach is that it assumes that the \gls{mab} cocktail
will completely quench their target cytokines between each feed cycle. This assumption
can be tested by running luminex with each cocktail addition. If a given
cytokine is undetectable, this indicates that the blocking \gls{mab} completely
quenched all target cytokine at the time of addition and in the time between
feeding cycles.

\subsubsection{Interior cell phenotype}

Unlike the beads, the \glspl{dms} have interior and exterior surfaces. We
demonstrated that some T cell expand on the interior of the \glspl{dms}, and is
plausible that these cells are phenotypically different than those growing on
the exterior or completely detached from the microcarriers, and that this leads
to an asymmetric cytokine cross-talk which accounts for the population-level
differences seen in comparison to the beads.

Experimentally, the first step involves separating the \glspl{dms} from the
loosely or non-adhered T cells and digesting the \glspl{dms} with \gls{cold}
(concentrations of \SI{10}{\ug\per\ml} will completely the \glspl{dms} within
\SIrange{30}{45}{\min}) isolate the interior T cells. Unfortunately, only
\SIrange{10}{20}{\percent} of all cells will be on the interior, so the interior
group may only have cells on the order of \num{1e3} to \num{1e4} for analysis. A
good first pass experiment would be to analyze both populations with a T cell
differentiation/activation state flow panel first (since flow cytometry is
relatively cheap and doesn't require a large number of cells) to simply
establish if the two groups are different phenotypes or are in a different state
of activation. From there, more in-depth analysis using \gls{cytof} or another
high-dimensionality method may be used to evaluate differential cytokine
expression.

\subsubsection{Antibody Surface Density}

While our \gls{doe} experiments showed a relationship between activating
\gls{mab} density and number of cells, we don't know how the \gls{mab} surface
density of the \gls{dms} compares to that of the beads. In all likelihood, the
\gls{mab} density on the \gls{dms} surface is lower (given the number of total
binding sites on \gls{stp} and the number of \glspl{mab} that actually bind)
which may lead to differences in performance\cite{Lozza2008}.

Before attempting this experiment, it will be vital to improve the \gls{dms}
manufacturing process such that \gls{mab} binding is predictable and
reproducible (see below). Once this is established, we can then determine the
amount of \glspl{mab} that bind to the beads, which could be performed much like
the \gls{mab} binding step is quantified in the \gls{dms} process (eg with
ELISA, \cref{fig:dms_flowchart}). Knowing this, we can vary the
\gls{mab} surface density for both the bead and the \glspl{dms} using a dummy
\gls{mab} as done previously with the \gls{doe} experiments in \cref{aim2a}.
Using varying surface densities that are matched per-area between the beads and
\glspl{dms} we can then activate T cells and assess their growth/phenotype as a
function of surface density and the presentation method.

\subsection{Reducing Ligand Variance}

While we have robust quality control steps to quantify each step of the
\gls{dms} coating process, we still see high variance across time and personnel
(\cref{fig:dms_coating}). This is less than ideal for translation.

When investigating the \gls{mab} and \gls{stp} binding, it appears that there is
a significant variance both between and within different experiments (even
within the same operator).  The following are a list of variance sources and
potential mitigation strategies:

\begin{description}
\item[Mass loss during autoclaving --] In order to ensure a consistent reaction
  volume, we mass the tube after adding carriers and \gls{pbs} prior to
  autoclaving. Autoclaving and washing will cause variations in the liquid
  level, and these are corrected using the pre-recorded tube mass. However, this
  assumes that the mass of the tube never changes, which may or may not be true
  in an autoclave where the temperature easily causes deformation of the plastic
  tube material. This can easily be tested by autoclaving empty tubes and
  observing a mass change. If there is a mass change, it may be mitigated by
  pre-autoclaving tubes (assuming that autoclaving is idempotent with respect to
  mass loss), or alternatively we could estimate the bias by autoclaving a
  set of tubes, recording the mean mass loss, and using this to correct the tube
  mass for downstream calculations.
\item[Errors in initial microcarrier massing --] The massing of microcarriers at
  the very beginning of the process requires care due to the low target mass and
  the propensity for both the plastic tubes and microcarriers to accumulate
  static. Oddly, the biotin attachment readout does not seem to be much affected
  by the mass of carriers (\cref{fig:dms_qc_doe}); however, this merely means
  that errors in carrier mass lead to different biotin surface densities, which
  downstream causes different ratios of \gls{stp} and \gls{mab} attachment since
  these relationships are non-linear with respect to biotin surface density
  (\cref{fig:stp_coating,fig:mab_coating}) (this is in addition to the fact that
  having more or less carriers will bias the total amount of \gls{stp} and
  \gls{mab} able to bind). A quick survey of operators revealed that acceptable
  margins for error in mass range from \SIrange{2.5}{5.0}{\percent} (eg, a
  target value $X$ \si{\mg} will be accepted as $X$ at plus or minus these
  margins). These could easily be reduced and standardized via protocol.
  Additionally, we do not currently record the exact mass of microcarriers
  weighed for each batch. Knowing this would allow us to pinpoint how much of
  this variance is due to our acceptable measurement margins and what errors may
  arise from static and other instrument noise.
\item[Centrifugation after washing --] After coating the \gls{dms} with \gls{snb},
  \gls{stp}, or \glspl{mab}, they must be washed. After washing, they must be
  massed in order to ensure the reaction volume is consistent. Ideally, the
  tubes are centrifuged after washing to ensure that all liquid is at the bottom
  prior to beginning the next coating step. Upon survey, not all operators
  follow this protocol, and the protocols are not written such to make this
  obvious. Therefore, protocols will be revised followed by additional training.
\item[Accidental microcarrier removal --] When washing the microcarriers after a
  coating step, liquid is aspirated using a stripette. The carriers should be at
  the bottom of the tube during this aspiration step. Depending on the skill and
  care of the operator, carriers may be aspirated with the liquid during this
  step. If this happens, downstream \gls{qc} assays will not reflect the true
  binding magnitude, as these assays assume the number of carriers is constant.
\item[\gls{bsa} binding kinetics --] Prior to \gls{mab} addition, \gls{bsa} is
  added to the \gls{mab} to block binding to the tubes. \glspl{mab} are added
  immediately after adding the \gls{bsa}, which means the \gls{bsa} has almost
  no time to mix completely and thus the \gls{mab} could come into contact with
  the sides of the tube unshielded. In theory this could cause the \gls{mab}
  reading to be lower on the \gls{elisa} during \gls{qc}. This problem may be
  minor since significant binding would only occur if the \gls{mab}/plastic
  adhesion was quite fast and happened in the seconds prior to beginning
  agitation. However, this problem is easily mitigated by agitating the tubes
  with \gls{bsa} for several minutes prior to adding \gls{mab} to ensure even
  mixing.
\item[Improving protein detection --] While the \gls{bca} assay and \gls{elisa}
  are quite precise, they both have problems that could lead to systemic bias as
  well as increases in random noise. The \gls{bca} assay is non-specific. All
  our data shows consistent small (\SI{0.5}{\ug}) but negative readings when
  adding zero \gls{snb}, which indicates that some background protein (or
  something that behaves like a protein) may be present that the \gls{bca} assay
  is detecting. The \gls{elisa} is specific to \gls{mab}; however, in our case
  we need to run a blank (just \gls{pbs}, \gls{bsa}, and \glspl{mab} without
  carriers) and subtract this from the reading, effectively doubling the assay
  variance. Using \gls{hplc} would mitigate both of these issues. \gls{hplc} can
  specifically detect species based on differences in charge and size, so it
  will likely be able to resolve \gls{stp} without the extraneous bias
  introduced via the \gls{bca} assay. In the case of \gls{elisa} it will not
  have remove the need to run a blank, but it likely will have lower variance
  due to its automated nature.
\end{description}


\subsubsection{Surface Stiffness}

The beads and \gls{dms} are composed of different materials: iron/polymer in the
former case and cross-linked gelatin in the latter. These materials likely have
different stiffnesses, and stiffness could play a role in T cell
activation\cite{Lambert2017}.

This hypotheses will be difficult to test directly, so it is advised to
eliminate other hypothesis before proceeding here. Direct testing could be
performed using a force probe to determine the Young's modulus of each
material\cite{Ju2017}. Since the microcarriers are porous and the cells will be
interacting with the bulk material itself, the void fraction and pore size will
need to be taken into account to find the bulk material properties of the
cross-linked gelatin\cite{Wang1984}.

\subsection{Additional Ligands and Signals on the DMSs}

In this work we only explored the use of \acd{3} and \acd{28} \glspl{mab} coated
on the surface of the \gls{dms}. The chemistry used for the \gls{dms} is very
general, and any molecule or protein that could be engineered with a biotin
ligand could be attached without any further modification. There are many other
ligands (in addition to integrin-binding domains and \il{15} complexes as
described at the end of \cref{aim2b}) that could have profound effects on the
expansion and quality of T cells which may be utilized. The simplest next step
is to simply vary the ratio of \acd{3} and \acd{28} signal. Another obvious
example is to attach \il{15}/\il{15R$\upalpha$} complexes to the surface to
mimic \textit{trans} presentation from other cell types\cite{Stonier2010}. Other
adhesion ligands or peptides such as GFOGER could be used to stimulate T cells
and provide more motility on the \glspl{dms}\cite{Stephan2014}. Finally, viral
delivery systems could theoretically be attached to the \gls{dms}, greatly
simplifying the transduction step.

\subsection{Assessing Performance Using Unhealthy Donors}

All the work presented in this dissertation was performed using healthy donors.
This was mostly due to the fact that it was much easier to obtain healthy donor
cells and was much easier to control. However, it is indisputable that the most
relevant test cases of the \glspl{dms} will be for unhealthy patient T cells, at
least in the case of autologous therapies. In particular, it will be interesting
to see how the \gls{dms} performs when assessed head-to-head with bead-based
expansion technology given that even in healthy donors, we observed the
\gls{dms} platform to work where the beads failed
(\cref{fig:dms_exp_fold_change}).

\subsection{Translation to Bioreactors}

In this work we performed some preliminary experiments demonstrating that the
\gls{dms} platform can work in a Grex bioreactor. While an important first step,
more work needs to be done to optimize how this system will or can work in a
scalable environment using bioreactors. There are several paths to explore.
Firstly, the Grex itself has additional automation accessories which could be
tested, which would allow continuous media exchange and cytokine
administration. While this is an improvement from the work done here, it is
still a Grex and has all the disadvantages of an open system. Secondly, other
static bioreactors such as the Quantum hollow fiber bioreactor (Terumo) could be
explored. Essentially the \gls{dms} would be an additional matrix that could be
supplied to this system which would enhance its compatibility with T cells.
Finally, suspension bioreactors such as the classic \gls{cstr} or WAVE
bioreactors could be tried. The caveat with these is that the T cells only seem
to be loosely attached to the \gls{dms} throughout culture, so an initial
activation/transduction step in static culture might be necessary before moving
to a suspension system (alternatively the \gls{dms} could be coated with
additional adhesion ligands to make the T cells attach more strongly).

\onecolumn
\clearpage

\appendix
\chapter{META ANALYSIS DATABASE CODE}\label{sec:appendix_meta}

The code used to aggregate all experimental data was written primarily in
Python, with a subprocess running R in a Docker container to handle the flow
cytometry data (\cref{fig:meta_overview}). The Postgres database itself was
hosted using \gls{aws} using their proprietary Aurora implementation.

% FIGURE explain what the colors mean
\begin{figure*}[ht!]
  \begingroup

  \includegraphics{../figures/metaanalysis.png}

  \endgroup
  \caption[Meta-analysis overview]
  {Overview of strategy used for meta-analysis. Colors: notebook (pink), input
    files (green), analysis framework (blue), data store (cyan), analysis
    pipeline (orange).}
  \label{fig:meta_overview}
\end{figure*}

The code is available here: \url{https://github.gatech.edu/ndwarshuis3/mdma}.

\chapter{META ANALYSIS DONORS}\label{sec:appendix_donors}

\begin{table}[!h]
  \caption{Donors used in meta-analysis}
  \begin{subtable}[t]{\textwidth} \centering
    \caption{characteristics}
    \input{../tables/donor_chars.tex}
    \phantomsubcaption\label{tab:meta_donors_chars}
  \end{subtable}

  \bigskip

  \begin{subtable}[t]{\textwidth} \centering
    \caption{phenotype (all in percents)}
    \input{../tables/donor_phenotypes.tex}
    \phantomsubcaption\label{tab:meta_donors_phenotypes}
  \end{subtable}
  \label{tab:meta_donors}
\end{table}

\chapter{BINDING KINETICS}\label{sec:appendix_binding}

\newcommand{\lig}{\textit{ligand}}
\newcommand{\rcp}{\textit{receptor}}
\newcommand{\ligs}{\textit{ligands}}
\newcommand{\rcps}{\textit{receptors}}

% TODO make notation consistent

To model binding kinetics of either \gls{stp} or \glspl{mab} (here called
\ligs{}), each microcarrier was assumed to be a porous sphere with a given
number of binding sites for the \ligs{} (here called \rcps{}). The \rcp{}/\lig{}
reaction was assumed to be instantaneous (which is reasonable given that these
are reactions between biotin and \gls{stp} which are extremely strong). From
this, we further assumed a spherical interface within each microcarrier and
aligned at the center wherein all \rcps{} in the interior were unbound and all
on the exterior were bound. At $t=0$ this interface was assumed to start with a
radius equal to that of the microcarrier, and shrunk down to radius of zero as
\ligs{} flowed into the porous microcarriers and bound. We assumed the
concentration of \lig{} to be zero at the interface and equal to the bulk
concentration at the exterior surface of the microcarrier. Furthermore, we
assumed that the interface move slowly relative to the diffusion of \lig{} into
the microcarriers, and thus we used a quasi-steady-state model to avoid solving
a boundary value problem with two movable boundaries (the interface radius and
the concentration in bulk).

The concentration profile of \lig{} in the microcarriers is given by Fick's
Second Law in spherical coordinates assuming only radial flux and steady state.
This with the boundary conditions as stated is:

\begin{equation}
  \label{eqn:binding_ficks}
  0 = \frac{1}{r^2}\frac{d}{dr}\left(r^2\frac{dC_L}{dr}\right)
\end{equation}

\begin{equation}
  \label{eqn:binding_bc_left}
  C_L\rvert_{r_i} = 0
\end{equation}

\begin{equation}
  \label{eqn:binding_bc_right}
  C_L\rvert_R = C_{L,b}
\end{equation}

Solving \cref{eqn:binding_ficks} we find the a relation for the concentration
profile in terms of the interfacial radius:

\begin{equation}
  \label{eqn:binding_conc}
  C_L = \frac{C_{L,b}}{r(1/r_i - 1/R)}\left(\frac{1}{r_i} - \frac{1}{r}\right)
\end{equation}

Solving \cref{eqn:binding_conc} for flux, the molar flow rate into the
microcarriers is given by:

\begin{equation}
  \label{eqn:binding_molar_flow}
  Q = 4\pi R^2N\rvert_R = \frac{-4\pi DC}{1/r_i - 1/R}
\end{equation}

Using the quasi-steady-state assumption, we can now find time-dependent
equations for the interfatial radius and the bulk concentration. The interfacial
volume in terms of molar flow rate is given by:

\begin{equation}
  \label{eqn:binding_volume_change}
  C_{R,0}\frac{dV_i}{dt} = -Q
\end{equation}

Substituting volume of a sphere and applying the chain rule:

\begin{equation}
  \label{eqn:radial_radial_change}
  \frac{dr_i}{dt} = \frac{-Q}{4\pi r_i^2C_{R,0}}
\end{equation}

The change in bulk concentration is simply given by:

\begin{equation}
  \label{eqn:radial_conc_change}
  \frac{dC_{L,b}}{dt} = \frac{-nQ}{V}
\end{equation}

Substituting \cref{eqn:binding_molar_flow} into \cref{eqn:radial_radial_change}
and \cref{eqn:radial_conc_change} yields \cref{eqn:stp_diffusion_1} and
\cref{eqn:stp_diffusion_2}.

The \gls{stp} binding kinetic profile was fit and calculated using the following
MATLAB code. Note that the \inlinecode{geometry} parameter was varied so as to
minimize the \inlinecode{SSE} output.

\lstinputlisting{../code/diffusion_stp.m}

The geometric diffusivity from above (the \inlinecode{geometry} variable) was
used in the below code to obtain the reaction profile for the \gls{mab} binding
step. The model is the same except for the parameters which were changes to
reflect the \gls{mab} coating process.

\lstinputlisting{../code/diffusion_mab.m}

\chapter{WASHING KINETICS CODE}\label{sec:appendix_washing}

The wash steps for the \gls{dms} were modeled using the following code:

\lstinputlisting{../code/microcarrier_diffusion_washing.m}

Complete output from this code is shown below:

\input{../code/washing_out.tex}

\chapter{REFERENCES}
\renewcommand{\chapter}[2]{} % noop the original bib section header

\bibliography{references}

\bibliographystyle{naturemag}

\end{document}