PreprintPDF Available

Learning Single-Cell Perturbation Responses using Neural Optimal Transport

Authors:
Preprints and early-stage research may not have been peer reviewed yet.

Abstract

Understanding and predicting molecular responses towards external perturbations is a core question in molecular biology. Technological advancements in the recent past have enabled the generation of high-resolution single-cell data, making it possible to profile individual cells under different experimentally controlled perturbations. However, cells are typically destroyed during measurement, resulting in unpaired distributions over either perturbed or non-perturbed cells. Leveraging the theory of optimal transport and the recent advents of convex neural architectures, we learn a coupling describing the response of cell populations upon perturbation, enabling us to predict state trajectories on a single-cell level. We apply our approach, CellOT, to predict treatment responses of 21,650 cells subject to four different drug perturbations. CellOT outperforms current state-of-the-art methods both qualitatively and quantitatively, accurately capturing cellular behavior shifts across all different drugs.
Learning Single-Cell Perturbation Responses
using Neural Optimal Transport
Charlotte Bunne,
1,2,Stefan G. Stark,
1,2,3,4,Gabriele Gut,
5,
Jacobo Sarabia del Castillo,
5Kjong-Van Lehmann,
1,2,3,4,Lucas Pelkmans,
5,
Andreas Krause,
1,2,Gunnar Rätsch1,2,3,4,6,
1Department of Computer Science, ETH Zurich, Switzerland;
2AI Center, ETH Zurich, Switzerland;
3Medical Informatics Unit, University Hospital Zurich, Switzerland;
4Swiss Institute of Bioinformatics, Switzerland;
5Department of Molecular Life Sciences, University of Zurich, Switzerland;
6Department of Biology, ETH Zurich, Switzerland.
December 15, 2021
Abstract
The ability to understand and predict molecular responses towards
external perturbations is a core question in molecular biology. Techno-
logical advancements in the recent past have enabled the generation of
high-resolution single-cell data, making it possible to profile individual cells
under different experimentally controlled perturbations. However, cells are
typically destroyed during measurement, resulting in unpaired distributions
over either perturbed or non-perturbed cells. Leveraging the theory of
optimal transport and the recent advents of convex neural architectures,
we learn a coupling describing the response of cell populations upon pertur-
bation, enabling us to predict state trajectories on a single-cell level. We
apply our approach, CellOT, to predict treatment responses of 21,650 cells
subject to four different drug perturbations. CellOT outperforms current
state-of-the-art methods both qualitatively and quantitatively, accurately
capturing cellular behavior shifts across all different drugs.
1 Introduction
Characterizing and modeling perturbation responses at the single-cell level from
non-time resolved data remains one of the grand challenges of biology. It finds
applications in predicting cellular reactions to environmental stress or a patient’s
response to drug treatments. Accurate inference of perturbation responses at the
single-cell level allows us, for instance, to understand how and why individual
These authors contributed equally.
To whom correspondence should be addressed:
kjlehmann@ukaachen.de
,
lucas.pelkmans@mls.uzh.ch,krausea@inf.ethz.ch,raetsch@inf.ethz.ch.
1
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
tumor cells evade cancer therapies (Frangieh et al., 2021). More generally, it
deepens the mechanistic understanding of the molecular machinery determining
the respective responses to perturbations.
Cell responses to perturbations such as drugs are highly heterogeneous
in nature (Liberali et al., 2014), determined by many factors, including the
preexisting variability in the abundance and localization of molecular entities,
such as RNA or proteins (Shaffer et al., 2017), cellular states (Kramer and
Pelkmans, 2019), or the cellular microenvironment (Snijder et al., 2009). To
effectively predict the drug response of a patient during treatment, it is thus
crucial to incorporate the molecular subpopulation structure of the cell populations
into the analysis.
A key difficulty in learning perturbation responses is that a cell (usually) must
be destroyed to measure its state, meaning that it is only possible to measure a
cell state either before or after a perturbation is applied. The typical experimental
setup divides a set of cells into subsets to which individual perturbations are
applied. Hereby, a subset of cells remains unperturbed, allowing us to measure
the base state of the population. So while we do not have access to a set of
paired control/perturbed single-cell observations, we do have access to samples
of distributions of control/perturbed cell states.
Previous methods to approximate single-cell perturbation responses fall
short of solving this highly complex pairing problem while, at the same time,
accounting for cellular heterogeneity and the strong subpopulation structure of
cell samples. Despite incorporating cell heterogeneity, mechanistic models do
not recover cellular response trajectories, instead of predicting factors such as
cell viability or response variables in the data in order to predict drug efficacy
(Snijder et al., 2012; Berchtold et al., 2018; Green and Pelkmans, 2016). Linear
models (Dixit et al., 2016), on the other hand, are unable to capture complex and
inhomogeneous population responses upon perturbation. Current state-of-the-art
methods (Lopez et al., 2018; Lotfollahi et al., 2019; Yang et al., 2020) predict
perturbation responses via linear shifts in a learned low-dimensional latent
space. While capturing nonlinear cell-type-specific responses, their use of linear
interpolations cause them to resolve the alignment problem with the challenging
task of learning representations that are invariant to their perturbation status.
A similar matching problem was considered in Stark et al. (2020) for matching
cell populations that are profiled with different profiling technologies.
This work proposes CellOT, a novel approach to predict single-cell pertur-
bation responses by uncovering couplings between control and perturbed cell
states while accounting for heterogeneous subpopulation structures of molecular
environments. We achieve this by utilizing the theory of optimal transport, which
provides natural geometry and mathematical tools to manipulate probability
distributions. To this end, we learn a robust optimal transport map describing
how the distribution of control cells connects to the distribution of perturbed cells.
Utilizing recent developments of neural optimal transport (Makkuva et al., 2020),
we learn a general optimal transport coupling for each perturbation, allowing us
to predict behavioral changes of incoming single-cell samples, e.g., of another
patient, using parameterizations learned for the previous cohort. We demonstrate
CellOT’s effectiveness by deploying it to learning cellular responses to different
2
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
<latexit sha1_base64="8eXnwrIlpVC9lDJsXJA3b54Pwd4=">AAAB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0J6jHgxWME84BkCbOTTjJmdmaZmRXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEsGN9f1vb219Y3Nru7BT3N3bPzgsHR03jUo1wwZTQul2RA0KLrFhuRXYTjTSOBLYisa3M7/1hNpwJR/sJMEwpkPJB5xR66RmV49Uj/VKZb/iz0FWSZCTMuSo90pf3b5iaYzSMkGN6QR+YsOMasuZwGmxmxpMKBvTIXYclTRGE2bza6fk3Cl9MlDalbRkrv6eyGhszCSOXGdM7cgsezPxP6+T2sFNmHGZpBYlWywapIJYRWavkz7XyKyYOEKZ5u5WwkZUU2ZdQEUXQrD88ippXlaCq0r1vlquVfM4CnAKZ3ABAVxDDe6gDg1g8AjP8ApvnvJevHfvY9G65uUzJ/AH3ucPkfqPGA==</latexit>
c
<latexit sha1_base64="goeuxSzxVssdTm512/imKl9KPTI=">AAAB7XicbVBNSwMxEJ3Ur1q/qh69BIvgqexKUY8FLx4r2A9ol5JNs23abLIkWaEs/Q9ePCji1f/jzX9j2u5BWx8MPN6bYWZemAhurOd9o8LG5tb2TnG3tLd/cHhUPj5pGZVqyppUCaU7ITFMcMmallvBOolmJA4Fa4eTu7nffmLacCUf7TRhQUyGkkecEuukVk+PVH/cL1e8qrcAXid+TiqQo9Evf/UGiqYxk5YKYkzX9xIbZERbTgWblXqpYQmhEzJkXUcliZkJssW1M3zhlAGOlHYlLV6ovycyEhszjUPXGRM7MqveXPzP66Y2ug0yLpPUMkmXi6JUYKvw/HU84JpRK6aOEKq5uxXTEdGEWhdQyYXgr768TlpXVf+6WnuoVeq1PI4inME5XIIPN1CHe2hAEyiM4Rle4Q0p9ILe0ceytYDymVP4A/T5A5yWjx8=</latexit>
j
<latexit sha1_base64="04Egm20ziq88gpeWon2fenuWZ1c=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0mkqMeiF48V7Ae0oUy2m3bpbhJ3N0IJ/RNePCji1b/jzX/jts1BWx8MPN6bYWZekAiujet+O4W19Y3NreJ2aWd3b/+gfHjU0nGqKGvSWMSqE6BmgkesabgRrJMohjIQrB2Mb2d++4kpzePowUwS5kscRjzkFI2VOr0hSol93i9X3Ko7B1klXk4qkKPRL3/1BjFNJYsMFah113MT42eoDKeCTUu9VLME6RiHrGtphJJpP5vfOyVnVhmQMFa2IkPm6u+JDKXWExnYTolmpJe9mfif101NeO1nPEpSwyK6WBSmgpiYzJ4nA64YNWJiCVLF7a2EjlAhNTaikg3BW355lbQuqt5ltXZfq9Rv8jiKcAKncA4eXEEd7qABTaAg4Ble4c15dF6cd+dj0Vpw8plj+APn8wcIEo/6</latexit>
i
<latexit sha1_base64="S3W/iUmsawT7LtgFsbdsZXTOkfo=">AAAB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0J6jHgxWME84BkCbOTTjJmdmaZmRXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEsGN9f1vb219Y3Nru7BT3N3bPzgsHR03jUo1wwZTQul2RA0KLrFhuRXYTjTSOBLYisa3M7/1hNpwJR/sJMEwpkPJB5xR66RmV49Uj/dKZb/iz0FWSZCTMuSo90pf3b5iaYzSMkGN6QR+YsOMasuZwGmxmxpMKBvTIXYclTRGE2bza6fk3Cl9MlDalbRkrv6eyGhszCSOXGdM7cgsezPxP6+T2sFNmHGZpBYlWywapIJYRWavkz7XyKyYOEKZ5u5WwkZUU2ZdQEUXQrD88ippXlaCq0r1vlquVfM4CnAKZ3ABAVxDDe6gDg1g8AjP8ApvnvJevHfvY9G65uUzJ/AH3ucPmxKPHg==</latexit>
i
<latexit sha1_base64="/OqJF5gzkm0KC22mvfjOhjlLC2Y=">AAAB73icbVBNS8NAEJ3Ur1q/qh69LBbBU0lE1GPRi8cK9gPaUCbbTbt2N4m7G6GE/gkvHhTx6t/x5r9x2+agrQ8GHu/NMDMvSATXxnW/ncLK6tr6RnGztLW9s7tX3j9o6jhVlDVoLGLVDlAzwSPWMNwI1k4UQxkI1gpGN1O/9cSU5nF0b8YJ8yUOIh5yisZK7e4ApcTeQ69ccavuDGSZeDmpQI56r/zV7cc0lSwyVKDWHc9NjJ+hMpwKNil1U80SpCMcsI6lEUqm/Wx274ScWKVPwljZigyZqb8nMpRaj2VgOyWaoV70puJ/Xic14ZWf8ShJDYvofFGYCmJiMn2e9Lli1IixJUgVt7cSOkSF1NiISjYEb/HlZdI8q3oX1fO780rtOo+jCEdwDKfgwSXU4Bbq0AAKAp7hFd6cR+fFeXc+5q0FJ585hD9wPn8ACZaP+w==</latexit>
j
: cells in different states
: new cell state after perturbation
: apoptotic cell
<latexit sha1_base64="Xx16NuI/ECoStdk45JUsAVXLMg4=">AAAB73icbVDLSgNBEOz1GeMr6tHLYBA8hV0J6jHgxWME84BkCb2T2WTIzOw6MyuEkJ/w4kERr/6ON//GSbIHTSxoKKq66e6KUsGN9f1vb219Y3Nru7BT3N3bPzgsHR03TZJpyho0EYluR2iY4Io1LLeCtVPNUEaCtaLR7cxvPTFteKIe7DhlocSB4jGnaJ3U7g5QSuyNeqWyX/HnIKskyEkZctR7pa9uP6GZZMpSgcZ0Aj+14QS15VSwabGbGZYiHeGAdRxVKJkJJ/N7p+TcKX0SJ9qVsmSu/p6YoDRmLCPXKdEOzbI3E//zOpmNb8IJV2lmmaKLRXEmiE3I7HnS55pRK8aOINXc3UroEDVS6yIquhCC5ZdXSfOyElxVqvfVcq2ax1GAUziDCwjgGmpwB3VoAAUBz/AKb96j9+K9ex+L1jUvnzmBP/A+fwAG5I/u</latexit>
k
:optimal transport plan
: of perturbation k
:
: cells in different states
: new cell state after perturbation
: apoptotic cell
:optimal transport plan
: of perturbation k
:
cells after
perturbation i
cells after
perturbation j
control
cells
cell data
space
cells after
perturbation j
control cells
cells after
perturbation i
cell data
space
cells after
perturbation i
cells after
perturbation j
control
cells
cell data
space
cells after
perturbation i
cells after
perturbation j
control
cells
cell data
space
cells after
perturbation i
cells after
perturbation j
control
cells
cell data
space
: cells in different states
cells after
perturbation i
cells after
perturbation j
control
cells
cell data
space
: new cell state after perturbation
: apoptotic cell
<latexit sha1_base64="Xx16NuI/ECoStdk45JUsAVXLMg4=">AAAB73icbVDLSgNBEOz1GeMr6tHLYBA8hV0J6jHgxWME84BkCb2T2WTIzOw6MyuEkJ/w4kERr/6ON//GSbIHTSxoKKq66e6KUsGN9f1vb219Y3Nru7BT3N3bPzgsHR03TZJpyho0EYluR2iY4Io1LLeCtVPNUEaCtaLR7cxvPTFteKIe7DhlocSB4jGnaJ3U7g5QSuyNeqWyX/HnIKskyEkZctR7pa9uP6GZZMpSgcZ0Aj+14QS15VSwabGbGZYiHeGAdRxVKJkJJ/N7p+TcKX0SJ9qVsmSu/p6YoDRmLCPXKdEOzbI3E//zOpmNb8IJV2lmmaKLRXEmiE3I7HnS55pRK8aOINXc3UroEDVS6yIquhCC5ZdXSfOyElxVqvfVcq2ax1GAUziDCwjgGmpwB3VoAAUBz/AKb96j9+K9ex+L1jUvnzmBP/A+fwAG5I/u</latexit>
k
:optimal transport plan
: of perturbation k
:
É
: different cell states
: cell states after perturbation
<latexit sha1_base64="S3W/iUmsawT7LtgFsbdsZXTOkfo=">AAAB7XicbVDLSgNBEOz1GeMr6tHLYBA8hV0J6jHgxWME84BkCbOTTjJmdmaZmRXCkn/w4kERr/6PN//GSbIHTSxoKKq66e6KEsGN9f1vb219Y3Nru7BT3N3bPzgsHR03jUo1wwZTQul2RA0KLrFhuRXYTjTSOBLYisa3M7/1hNpwJR/sJMEwpkPJB5xR66RmV49Uj/dKZb/iz0FWSZCTMuSo90pf3b5iaYzSMkGN6QR+YsOMasuZwGmxmxpMKBvTIXYclTRGE2bza6fk3Cl9MlDalbRkrv6eyGhszCSOXGdM7cgsezPxP6+T2sFNmHGZpBYlWywapIJYRWavkz7XyKyYOEKZ5u5WwkZUU2ZdQEUXQrD88ippXlaCq0r1vlquVfM4CnAKZ3ABAVxDDe6gDg1g8AjP8ApvnvJevHfvY9G65uUzJ/AH3ucPmxKPHg==</latexit>
i
<latexit sha1_base64="5WAh8hGQBqNWWh1baUtLFoSmCxs=">AAAB+XicbVBNS8NAEN3Ur1q/oh69BItQLyWRoh4LXjxWsB/QhDLZbtu1u0nYnRRK6D/x4kERr/4Tb/4bt20O2vpg4PHeDDPzwkRwja77bRU2Nre2d4q7pb39g8Mj+/ikpeNUUdaksYhVJwTNBI9YEzkK1kkUAxkK1g7Hd3O/PWFK8zh6xGnCAgnDiA84BTRSz7b9IUgJFR9HDKH3dNmzy27VXcBZJ15OyiRHo2d/+f2YppJFSAVo3fXcBIMMFHIq2Kzkp5olQMcwZF1DI5BMB9ni8plzYZS+M4iVqQidhfp7IgOp9VSGplMCjvSqNxf/87opDm6DjEdJiiyiy0WDVDgYO/MYnD5XjKKYGgJUcXOrQ0eggKIJq2RC8FZfXietq6p3Xa091Mr1Wh5HkZyRc1IhHrkhdXJPGqRJKJmQZ/JK3qzMerHerY9la8HKZ07JH1ifP/dkkzE=</latexit>
(j)
<latexit sha1_base64="jXCPyDunvokkoB+UGMx5puOLU/8=">AAAB+XicbVBNS8NAEN34WetX1KOXYBHqpSRS1GPBi8cK9gOaECbbbbt0Nwm7k0IJ/SdePCji1X/izX/jts1BWx8MPN6bYWZelAqu0XW/rY3Nre2d3dJeef/g8OjYPjlt6yRTlLVoIhLVjUAzwWPWQo6CdVPFQEaCdaLx/dzvTJjSPImfcJqyQMIw5gNOAY0U2rY/BCmh6uOIIYT8KrQrbs1dwFknXkEqpEAztL/8fkIzyWKkArTueW6KQQ4KORVsVvYzzVKgYxiynqExSKaDfHH5zLk0St8ZJMpUjM5C/T2Rg9R6KiPTKQFHetWbi/95vQwHd0HO4zRDFtPlokEmHEyceQxOnytGUUwNAaq4udWhI1BA0YRVNiF4qy+vk/Z1zbup1R/rlUa9iKNEzskFqRKP3JIGeSBN0iKUTMgzeSVvVm69WO/Wx7J1wypmzsgfWJ8/9d+TMA==</latexit>
(i)
<latexit sha1_base64="if51Wk1neoSWccIW7XKRJD0nyH0=">AAAB+XicbVBNS8NAEN34WetX1KOXYBHqpSRS1GPBi8cK9gOaECbbbbt0Nwm7k0IJ/SdePCji1X/izX/jts1BWx8MPN6bYWZelAqu0XW/rY3Nre2d3dJeef/g8OjYPjlt6yRTlLVoIhLVjUAzwWPWQo6CdVPFQEaCdaLx/dzvTJjSPImfcJqyQMIw5gNOAY0U2rY/BCmh6uOIIYTjq9CuuDV3AWedeAWpkALN0P7y+wnNJIuRCtC657kpBjko5FSwWdnPNEuBjmHIeobGIJkO8sXlM+fSKH1nkChTMToL9fdEDlLrqYxMpwQc6VVvLv7n9TIc3AU5j9MMWUyXiwaZcDBx5jE4fa4YRTE1BKji5laHjkABRRNW2YTgrb68TtrXNe+mVn+sVxrVIo4SOScXpEo8cksa5IE0SYtQMiHP5JW8Wbn1Yr1bH8vWDauYOSN/YH3+APVNkyY=</latexit>
(k)
Figure 1: Learning single-cell perturbation responses. We aim to recover a mapping from
control cell distributions
ρc
to some perturbed cell distribution
ρi
or
ρj
by learning the
corresponding neural optimal transport map
γ
(
θk
), parameterized by
θk
, from the observed
distribution of untreated cells and the set of cells observed after the perturbation is applied.
cancer drugs and tumor combination therapies. An overview of our approach is
illustrated in Figure 1.
Optimal transport has previously been applied in the domain of single-cell
biology to uncover trajectories of single-cell reprogramming and to link rich,
non-spatially-resolved with sparse, spatially resolved measurements (Schiebinger
et al., 2019; Cang and Nie, 2020; Demetci et al., 2020; Huizing et al., 2021;
Lavenant et al., 2021; Zhang et al., 2021). Here we apply optimal transport
to a new data modality, consisting of cell morphological measurements and
multiplexed protein state measurements obtained by 4i (Gut et al., 2018) from
large populations of cancer cells exposed in vitro to different drugs used in the
clinic.
2 Background
2.1 Optimal Transport
Optimal transport plays dual roles as it induces a mathematically well-characterized
distance measure between distributions besides providing a geometry-based
approach to realizing couplings between two probability distributions. Let
µ=Pn
i=1 aiδxiand ν=Pm
j=1 bjδyjbe two discrete probability measures in Rd.
The optimal transport (OT) problem (Kantorovich, 1942) reads
W2
2(µ, ν) = inf
γΓ(µ,ν)Zkxyk2(x, y),(1)
where the polytope Γ(
a, b
)is
{γRn×m
+, γ1m
=
a, γ>1n
=
b}
describes the
set of all couplings
γ
between
µ
and
ν
. The optimal transport plan
γ
thus
corresponds to the coupling between two probability distributions minimizing
the overall transportation cost. Computing optimal transport distances in
(1)
involves solving a linear program, and thus their computational cost is prohibitive
for large-scale machine learning problems. Regularizing objective
(1)
with an
entropy term results in significantly more efficient optimization (Cuturi, 2013),
W2
2(µ, ν) = inf
γΓ(µ,ν)Zkxyk2(x, y)εH (γ),(2)
with entropy
H
(
γ
) =
Pij γij
(
log γij
1) and parameter
ε
controlling the
strength of the regularization.
Wε
2
is further differentiable w.r.t. its inputs and
3
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
thus serves as a loss function in machine learning applications.
Problem
(1)
denotes the primal formulation for the Wasserstein-2 distance.
The corresponding dual introduced by Kantorovich in 1942 is a constrained
concave maximization problem defined as
W2
2(µ, ν) = sup
(f,g)Φc
Eµ[f(x)] + Eν[g(y)],(3)
where the set of admissible potentials is Φ
c:
=
{
(
f, g
)
L1
(
µ
)
×L1
(
ν
) :
f
(
x
) +
g
(
y
)
1
2kxyk2
2
,
(
x, y
)
a.e.}
(Villani, 2003, Theorem 1.3). Villani
(2003, Theorem 2.9) further simplifies the dual problem
(3)
over the pair of
functions (f, g)to
W2
2(µ, ν) = 1
2Ekxk2
2+kyk2
2
| {z }
Cµ,ν
inf
f˜
Φ
Eµ[f(X)] + Eν[f(Y)] ,(4)
where
˜
Φ
is the set of all convex functions in
L1
(
)
×L1
(
),
L1
(
µ
)
:
=
{fis measurable
&
Rfdµ < ∞}
, and
f
(
y
) =
supxhx, yi − f
(
x
)is
f
’s con-
vex conjugate. Villani (2003, Theorem 2.9) then proves the existence of an
optimal pair (
f, f
)of lower semi-continuous proper conjugate convex functions
on Rnminimizing (3).
2.2 Convex Neural Networks
In order to parameterize convex spaces such as
˜
Φ
in
(4)
, we need neural networks
which are convex w.r.t. to their inputs. One example are input convex neural
networks (ICNN) introduced by Amos et al. (2017). ICNNs are based on fully-
connected feed-forward networks that ensure convexity by placing constraints
on their parameters. An ICNN with parameters
θ
=
{bi, W z
i, W x
i}
represents a
convex function f(x;θ)and, for a layer i= 0 . . . L 1, is defined as
hi+1 =σi(Wx
ix+Wz
ihi+bi)and f(x;θ) = hL,(5)
where activation functions
σi
are convex and non-decreasing, and elements of
all
Wz
i
are constrained to be nonnegative. Despite their constraints, ICNNs are
able to parameterize a rich class of convex functions. In particular, Chen et al.
(2019) provide a theoretical analysis that any convex function over a convex
domain can be approximated in sup norm by an ICNN. Huang et al. (2021)
further extend ICNNs from fully-connected feed-forward neural networks to
convolutional neural architectures.
2.3 Neural Optimal Transport
Despite existing numerical approximations of the optimal transport distance
and the corresponding optimal coupling (Cuturi, 2013; Aude et al., 2016)
(2)
,
recent efforts have investigated neural network-based approaches as fast and
scalable approximations to
(1)
. Taghvaei and Jalali (2019) consider solving
(4)
by parameterizing
f
with an ICNN and solve for
f
at each step, which
4
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
has a high computational cost. Makkuva et al. (2020) extend this work by
approximating
f
with another ICNN
g
, which scales well but transforms the
problem into a min-max optimization task. Huang et al. (2021) introduce a novel,
OT-inspired parameterization of normalizing flows utilizing ICNNs. Korotin
et al. (2021) provide a detailed comparison of the current state of neural optimal
transport solvers. Furthermore, convex neural architectures have been utilized
to parameterize Wasserstein gradient flows (Bunne et al., 2021; Alvarez-Melis
et al., 2021; Mokrov et al., 2021).
3 Model
Recent high-throughput methods provide great insights on how cell populations
respond to various perturbations on the level of individual cells. The provided
data, however, is non-time-resolved and unaligned. Hence, snapshots taken of
biological samples before and after perturbations do not provide information
on single-cell trajectories. Perturbations might include the application of drugs
affecting molecular functions in cells, or changes in the cellular environment
causing shifts in biological signaling, thus impacting cells and their states in
various ways. In the following, we describe our approach, which uncovers
single-cell perturbation responses by predicting couplings between control and
perturbed cell states. Hereby, let
X
denote the biological data space spanned by
cell morphology and gene expression features. We then treat a cell’s response to
perturbation
k
as an evolution in a high-dimensional space of cell states
Rd
=
X
.
3.1 Recovering Perturbation Effects via Optimal Transport
Given a dataset of
n
observations
{xc
1, . . . , xc
n}, xc
i∈ X
drawn from
ρc∈ P
(
X
),
the distribution of cells before applying a perturbation, we aim to learn the
distribution of cells
ρk∈ P
(
X
)upon some perturbation
k
, given a set of separate
samples {xk
1, . . . , xk
m}, xk
i X .
Perturbation responses of cells are dynamic: after applying perturbation
k
,
cell states evolve over time and thus can be modeled as a stochastic process on
the cell data space. Despite this time-resolved nature of single-cell responses, we
only have access to the distributions of cell states before,
ρc
, and after injecting
perturbation
k
,
ρk
. We thus aim at understanding the underlying stochastic
process without access to time-resolved perturbation responses by uncovering the
coupling
γ
between
ρc
and
ρk
. Given prior biological knowledge, we can assume
that cells do not drastically alter their phenotype w.r.t. morphology and gene
expression pattern. We thus posit that the evolution of probability distributions
of single-cells upon perturbation can be modeled via the mathematical theory
of optimal transport. The coupling
γ
then corresponds to an optimal transport
plan (1) between ρcand ρk.
Following Makkuva et al. (2020), we infer the optimal coupling
γ(1)
between
ρc
and
ρt
. Thus, instead of computing a coupling individually for each pair of cell
samples using existing solvers (Cuturi, 2013), we learn a parameterized optimal
transport map using neural networks. The parameterized OT coupling then
5
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
serves as a robust predictor for cellular distribution shifts upon perturbations on
unseen samples {xc
i}n0
i=1 ρc, i.e., of another patient.
3.2 Parametrization of the Optimal Transport Coupling
Directly learning the optimal transport map in the primal
(1)
and dual
(3)
is
notoriously difficult. Instead, Makkuva et al. build upon celebrated results by
Knott and Smith (1984) and Brenier (1991), which relate the optimal solutions
for the dual form
(3)
and the primal form
(1)
, to derive a min-max formulation
replacing the convex conjugate in (4) (Makkuva et al., 2020, Theorem 3.3)
W2
2(ρc, ρk) = sup
f˜
Φ
fL1(ρk)
inf
g˜
Φ
CρckEρc[f(x)] Eρk[hy, g(y)i − f(g(y))]
| {z }
Vρck(f,g)
.(6)
We can further relax the constraint
g˜
Φ
to
L1
(
ρk
), as a function
gL1
(
ρk
)
minimizing
(6)
is convex and equal to
f
for any convex function
f
. In order
to learn the resulting optimal transport, i.e., the solution of the minimization
problem in
(6)
, Makkuva et al. (2020) parametrize both dual variables
f
and
g
using input convex neural networks (§ 2.3) (Amos et al., 2017). The resulting
approximate Wasserstein distance is thus defined as
ˆ
W2
2(ρc, ρk) = sup
φ
inf
θ
Cρck− Vρck(fφ, gθ),(7)
where
θ
and
φ
are the parameters of each ICNN. The resulting
g
θ
produces an
approximate optimal transport plan γ(g
θ×Id)#ρc.
3.3 Predicting Perturbation Effects via CellOT
The framework described above allows us to recover couplings between con-
trol
{xc
1, . . . , xc
n}
and perturbed cells
{xk
1, . . . , xk
n}
, giving insights into cellular
response trajectories upon application of a perturbation
k
. Given a set of
perturbations
K
, and sample access to the control distribution
ρc
as well as
distributions
ρk
for each perturbation
kK
,CellOT learns the optimal pair
of dual potentials (
f
φk, gθ
k
)for each perturbation
k
. Given parametrizations of
the convex potentials for each
k
,CellOT then predicts the transformation of a
control cell
xc
i
upon perturbation
k
via
ˆxk
i
=
gθ?
k
(
xc
i
), i.e., samples following
the predicted perturbed distribution
ˆρk
= (
gθ
k
)
#ρc
.CellOT thus provides
a general approach to predict state trajectories on a single-cell level, as well
as understand, how heterogeneous subpopulation structures evolve under the
impact of external factors.
4 Evaluation
We evaluate CellOT on the task of predicting single-cell drug responses for
drugs with different molecular effects, using melanoma cell lines profiled by the
4i technology (Gut et al., 2018).
6
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
4.1 Datasets
4i is an imaging technology that detects protein abundance by attaching a
fluorescent tag designed to bind to a target protein and then measuring the
fluorescence intensity of this tag. An iterative staining and washing procedure
allows for the capture of multiple tags. Additionally, an image processing pipeline
extracts morphological features, such as cell perimeter and area and detects
the cell nucleus. We considered four common cancer therapies for this works
since they target different biological processes. Erlotinib is an inhibitor of the
epidermal growth factor receptor (EGFR) tyrosine kinase, Imatinib inhibits the
Bcr-Abl tyrosine kinase, and Trametinib is an inhibitor of mitogen-activated
extracellular signal-regulated kinase 1 (MEK1) and MEK2.
We utilized a mixture of 2 melanoma tumor cell lines (ratio 1:1) in order to
image a total of 21,650 cells, of which 11,526 are in the (untreated) control state,
2,364 are treated with Erlotinib, 2,650 with Imatinib, 2,683 with Trametinib,
and 2,417 are treated with a combination of Trametinib and Erlotinib, and
48 features are extracted for each cell. 22 features are morphological, and the
remaining 26 are mean intensities of 13 protein markers detected both inside the
cell nucleus and in the cell as a whole. Finally, we perform an 80/20 train test
split for each condition and evaluate model performance on its ability to make
predictions on the unseen set of control cells. More details regarding dataset
preparation can be found in Appendix B.1.
4.2 Baselines
We compare CellOT to two other baselines, both of which attempt to add
perturbation effects through the manipulation of a learned latent representa-
tion: scGen (Lotfollahi et al., 2019) computes linear shifts using latent space
arithmetic to remove the source condition and add the target condition, and the
conditional autoencoder, cAE, which has an architecture based on batch correc-
tion technique popular in the single-cell community, first introduced by Lopez
et al. (2018). Here, one-hot encodings of batch labels (treatment conditions) are
concatenated to the encoder and decoder inputs, which attempt to remove and
then add condition-specific effects. More details can be found in Appendix A.
4.3 Evaluation Metrics
Since we lack access to the ground truth set of control and treatment observations
on the single-cell level, we first analyze the effectiveness of CellOT using
evaluations that operate on the level of the distribution of real and predicted
perturbation states. Drug signatures are computed as the difference in means
between the distribution of perturbed states and control states. We then report
the
`2
-distance between the drug signatures (DS) computed on the true and
predicted distributions (
`2
(
DS
)). We additionally consider two distributional
distances: kernel maximum mean discrepancy (MMD) (Gretton et al., 2012)
and entropy-regularized Wasserstein distance
W2
2(2)
(Cuturi, 2013). MMD is
computed using the RBF kernel and averaging over the length scales 0
.
5
,
0
.
1
,
0
.
01
,
and 0.005;W2
2is computed with ε= 0.5.
7
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
4.4 Results
For each drug perturbation, all models predict the perturbed cell states from
the set of held-out of control cells. Differences between the distribution of
perturbed cells and predicted cells are shown in Table 1. CellOT significantly
outperforms all baselines on all three metrics. Qualitative assessment of the
marginal distributions of control, treated, and predicted cell states provides
further evidence for superior performance of CellOT over other approaches
(see Figure 2 for three selected features of the Imatinib condition).
Table 1: Performance assessment of CellOT compared to different baselines w.r.t. to
Wasserstein (
W2
2
,
(2)
) and MMD distances between the observed perturbed cells and predicted
responses from control cells, as well as the predictive quality of drug signatures (see § 4.3).
Model Drugs
Erlotinib Imatinib Trametinib Trametinib and Erlotinib
`2(DS)MMD W2
2`2(DS)MMD W2
2`2(DS)MMD W2
2`2(DS)MMD W2
2
scGen 0.41 0.0241 3.542 0.52 0.0361 5.811 0.56 0.0180 3.587 0.60 0.0163 3.594
cAE 0.05 0.0074 3.330 0.16 0.0200 4.512 0.37 0.0087 3.343 0.44 0.0122 3.215
CellOT 0.22 0.0013 3.619 0.12 0.0010 3.851 0.14 0.0011 2.846 0.18 0.0014 2.796
Figure 2: Marginal distributions of observed and predicted cell states for three selected features,
i.e., a measure of the eccentricity of the nucleus, as well as Sox9 intensity level inside the
nucleus and pERK intensity within the cell but outside the nucleus. The marginals of control
and Imatinib distributions correspond to the observed set of untreated cells and cells treated
with the Imatinib drug. The remaining distributions are calculated using the predictions of
each model on the unseen set of control cells.
Next, we compared UMAP projections (McInnes et al., 2018) of the perturbed
and predicted cells (see Figure 3). Predicted cells are colored by the fraction
of other predicted cells in their
k
= 100 nearest neighbors. If the distribution
of predicted cells matches the true distribution of perturbed cells, then we
would expect the nearest neighbor of each cell to be well mixed (i.e., 0.5) across
conditions. Thus, cells with values closer 1 indicate regions where the predicted
distribution does not integrate with the true perturbed distribution. We conclude
that predictions made by CellOT integrate well with measurements of real
treated cells.
Finally, the previous results argue that the distribution of cells predicted
by CellOT closely matches true distribution; however, they could have also
been replicated by a map that assigns control cells to random treated cells.
Thus, we evaluate the quality of single-cell level pairs induced from the CellOT
mapping by computing the Spearman correlation of features between the control
state and predicted drug state. The distribution of the correlation coefficient
8
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
Figure 3: UMAP projections computed on the joint set of cells perturbed by Imatinib (grey)
and predictions of each model. Model predictions are colored by the fraction of other predicted
cells in their
k
= 100 nearest neighbors in data space. Predicted cells which do not share many
neighbors of the true set of perturbed cells, but instead control cells take values of
1
.
0(blue).
Predicted cells that integrate well with true perturbed cells take values of 0.5(white).
between the control state and the predicted state across all features of all learned
maps is shown in Figure 4. The low distributional distances between CellOT-
predicted cells to the true distribution of perturbed cells in conjunction with
a high correlation of features with the control states paired to the predicted
states demonstrate that CellOT makes sound predictions on the single-cell
level, outperforming current state-of-the-art methods both qualitatively and
quantitatively.
Figure 4: Distribution of Spearman correlation coefficients between the features of control cells
and the features of its corresponding predicted state upon treatment for each considered drug
Erlotinib, Imatinib, Trametinib and the combination therapy of Trametinib and Erlotinib. Low
correlations imply unexpected significant differences in the feature states between prediction
and control, and thus a reduced accuracy of predictive power.
5 Conclusion
In this paper, we present a new framework to learn single-cell perturbation
responses. We approach the problem by learning an optimal transport map that
is parameterized by an ICNN to push-forward the distribution of control cells onto
the distribution of perturbed cells. We validate CellOT’s effectiveness through
experiments on melanoma cell lines with four different drug perturbations. In
the absence of ground truth, we provide various evaluation metrics to compare
our method to existing approaches. While operating in the original data space,
instead of relying on meaningful low-dimensional representations, CellOT
performs consistently well across all perturbations, outperforming current state-
of-the-art methods. The use of neural optimal transport to learn single-cell drug
9
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
responses makes for an exciting avenue of future work, including its use to improve
our mechanistic understanding of cell therapies, to study drug responses from
patient samples, and to better account for cell-to-cell variability in large-scale
drug discovery efforts.
Acknowledgments
We are grateful to Hugo Yèche and Ximena Bonilla for their fruitful comments,
corrections, and discussions. C.B. and A.K. received funding from the Swiss
National Science Foundation under the National Center of Competence in
Research (NCCR) Catalysis under grant agreement 51NF40 180544. L.P. is
supported by the European Research Council (ERC-2019-AdG-885579), the
Swiss National Science Foundation (SNSF grant 310030_192622), the Chan
Zuckerberg Initiative, and the University of Zurich. G.G. received funding from
the Swiss National Science Foundation and InnoSuisse as part of the BRIDGE
program as well as from the University of Zurich through the BioEntrepreneur
Fellowship. K.L. and S.G.S. were partially funded by ETH Zürich core funding
(to G.R.) and from the Tumor Profiler Initiative (to G.R.).
Declaration of Interests
G.G. and L.P. have filed a patent on the 4i technology (patent WO2019207004A1).
References
D. Alvarez-Melis, Y. Schiff, and Y. Mroueh. Optimizing Functionals on the
Space of Probabilities with Input Convex Neural Networks. arXiv Preprint,
2021.
B. Amos, L. Xu, and J. Z. Kolter. Input Convex Neural Networks. In International
Conference on Machine Learning (ICML), volume 34, 2017.
G. Aude, M. Cuturi, G. Peyré, and F. Bach. Stochastic Optimization for Large-
Scale Optimal Transport. In Advances in Neural Information Processing
Systems (NeurIPS), 2016.
D. Berchtold, N. Battich, and L. Pelkmans. A systems-level study reveals
regulators of membrane-less organelles in human cells. Molecular cell, 72(6):
1035–1049, 2018.
Y. Brenier. Polar Factorization and Monotone Rearrangement of Vector-Valued
Functions. Communications on pure and applied mathematics, 44(4):375–417,
1991.
C. Bunne, L. Meng-Papaxanthos, A. Krause, and M. Cuturi. JKOnet: Proximal
Optimal Transport Modeling of Population Dynamics. arXiv Preprint, 2021.
10
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
Z. Cang and Q. Nie. Inferring spatial and signaling relationships between cells
from single cell transcriptomic data. Nature Communications, 11(1), 2020.
A. E. Carpenter, T. R. Jones, M. R. Lamprecht, C. Clarke, I. H. Kang, O. Friman,
D. A. Guertin, J. H. Chang, R. A. Lindquist, J. Moffat, et al. Cellprofiler: image
analysis software for identifying and quantifying cell phenotypes. Genome
biology, 7(10):1–11, 2006.
Y. Chen, Y. Shi, and B. Zhang. Optimal Control Via Neural Networks: A
Convex Approach. In International Conference on Learning Representations
(ICLR), 2019.
M. Cuturi. Sinkhorn Distances: Lightspeed Computation of Optimal Transport.
In Advances in Neural Information Processing Systems (NeurIPS), volume 26,
2013.
P. Demetci, R. Santorella, B. Sandstede, W. S. Noble, and R. Singh. Gro-
mov–Wasserstein Optimal Transport to Align Single-Cell Multi-Omics Data.
BioRxiv, 2020.
A. Dixit, O. Parnas, B. Li, J. Chen, C. P. Fulco, L. Jerby-Arnon, N. D. Marjanovic,
D. Dionne, T. Burks, R. Raychowdhury, et al. Perturb-Seq: Dissecting
Molecular Circuits with Scalable Single-Cell RNA Profiling of Pooled Genetic
Screens. Cell, 167(7):1853–1866, 2016.
C. J. Frangieh, J. C. Melms, P. I. Thakore, K. R. Geiger-Schuller, P. Ho, A. M.
Luoma, B. Cleary, L. Jerby-Arnon, S. Malu, M. S. Cuoco, et al. Multimodal
pooled perturb-cite-seq screens in patient models define mechanisms of cancer
immune evasion. Nature genetics, 53(3):332–341, 2021.
V. A. Green and L. Pelkmans. A systems survey of progressive host-cell re-
organization during rotavirus infection. Cell host & microbe, 20(1):107–120,
2016.
A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Schölkopf, and A. Smola. A
kernel two-sample test. The Journal of Machine Learning Research, 13(1),
2012.
M. Guizar-Sicairos, S. T. Thurman, and J. R. Fienup. Efficient subpixel image
registration algorithms. Optics letters, 33(2):156–158, 2008.
G. Gut, M. D. Herrmann, and L. Pelkmans. Multiplexed protein maps link
subcellular organization to cellular states. Science, 361(6401), 2018.
C.-W. Huang, R. T. Q. Chen, C. Tsirigotis, and A. Courville. Convex Potential
Flows: Universal Probability Distributions with Optimal Transport and Con-
vex Optimization. In International Conference on Learning Representations
(ICLR), 2021.
G.-J. Huizing, G. Peyré, and L. Cantini. Optimal transport improves cell-cell
similarity inference in single-cell omics data. bioRxiv, 2021.
11
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
L. Kantorovich. On the transfer of masses (in Russian). In Doklady Akademii
Nauk, volume 37, 1942.
D. P. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. In
International Conference on Learning Representations (ICLR), 2014.
M. Knott and C. S. Smith. On the optimal mapping of distributions. Journal of
Optimization Theory and Applications, 43(1), 1984.
A. Korotin, L. Li, A. Genevay, J. Solomon, A. Filippov, and E. Burnaev. Do Neu-
ral Optimal Transport Solvers Work? A Continuous Wasserstein-2 Benchmark.
arXiv Preprint, 2021.
B. A. Kramer and L. Pelkmans. Cellular state determines the multimodal
signaling response of single cells. bioRxiv, 2019.
H. Lavenant, S. Zhang, Y.-H. Kim, and G. Schiebinger. Towards a mathematical
theory of trajectory inference. arXiv preprint arXiv:2102.09204, 2021.
P. Liberali, B. Snijder, and L. Pelkmans. A hierarchical map of regulatory genetic
interactions in membrane trafficking. Cell, 157(6):1473–1487, 2014.
R. Lopez, J. Regier, M. B. Cole, M. I. Jordan, and N. Yosef. Deep generative
modeling for single-cell transcriptomics. Nature methods, 15(12):1053–1058,
2018.
M. Lotfollahi, F. A. Wolf, and F. J. Theis. scGen predicts single-cell perturbation
responses. Nature Methods, 16(8), 2019.
A. Makkuva, A. Taghvaei, S. Oh, and J. Lee. Optimal transport mapping
via input convex neural networks. In International Conference on Machine
Learning (ICML), volume 37, 2020.
L. McInnes, J. Healy, and J. Melville. UMAP: Uniform Manifold Approximation
and Projection for Dimension Reduction. arXiv Preprint, 2018.
P. Mokrov, A. Korotin, L. Li, A. Genevay, J. Solomon, and E. Burnaev. Large-
Scale Wasserstein Gradient Flows. arXiv Preprint, 2021.
G. Schiebinger, J. Shu, M. Tabaka, B. Cleary, V. Subramanian, A. Solomon,
J. Gould, S. Liu, S. Lin, P. Berube, et al. Optimal-Transport Analysis of Single-
Cell Gene Expression Identifies Developmental Trajectories in Reprogramming.
Cell, 176(4), 2019.
S. M. Shaffer, M. C. Dunagin, S. R. Torborg, E. A. Torre, B. Emert, C. Krepler,
M. Beqiri, K. Sproesser, P. A. Brafford, M. Xiao, et al. Rare cell variability
and drug-induced reprogramming as a mode of cancer drug resistance. Nature,
546(7658):431–435, 2017.
B. Snijder, R. Sacher, P. Rämö, E.-M. Damm, P. Liberali, and L. Pelkmans.
Population context determines cell-to-cell variability in endocytosis and virus
infection. Nature, 461(7263):520–523, 2009.
12
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
B. Snijder, R. Sacher, P. Rämö, P. Liberali, K. Mench, N. Wolfrum, L. Burleigh,
C. C. Scott, M. H. Verheije, J. Mercer, et al. Single-cell analysis of population
context advances RNAi screening at multiple levels. Molecular Systems Biology,
8(1):579, 2012.
S. G. Stark, J. Ficek, F. Locatello, X. Bonilla, S. Chevrier, F. Singer, G. Rätsch,
and K.-V. Lehmann. Scim: universal single-cell matching with unpaired
feature sets. Bioinformatics, 36, 2020.
T. Stoeger, N. Battich, M. D. Herrmann, Y. Yakimovich, and L. Pelkmans.
Computer vision for image-based transcriptomics. Methods, 85:44–53, 2015.
A. Taghvaei and A. Jalali. 2-Wasserstein Approximation via Restricted Convex
Potentials with Application to Improved Training for GANs. arXiv Preprint,
2019.
S. Van der Walt, J. L. Schönberger, J. Nunez-Iglesias, F. Boulogne, J. D. Warner,
N. Yager, E. Gouillart, and T. Yu. scikit-image: image processing in python.
PeerJ, 2:e453, 2014.
C. Villani. Topics in Optimal Transportation, volume 58. American Mathematical
Soc., 2003.
K. D. Yang, K. Damodaran, S. Venkatachalapathy, A. C. Soylemezoglu, G. Shiv-
ashankar, and C. Uhler. Predicting cell lineages using autoencoders and
optimal transport. PLoS Computational Biology, 16(4), 2020.
S. Zhang, A. Afanassiev, L. Greenstreet, T. Matsumoto, and G. Schiebinger.
Optimal transport analysis reveals trajectories in steady-state systems. bioRxiv,
2021.
13
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
Appendix
A Related Work
Consider a single-cell dataset of a binary perturbation. Let
{x1. . . xn}
,
xi∈ X
,
drawn from
ρcρk
and let
c
(
i
)
∈ {
0
,
1
}
indicate the perturbation status of a
single cell,
c(i) = (0,if xiρc
1,if xiρk.
A.1 scGen
Given representations
{z1. . . zn}
of
{x1. . . xn}
, learned by an autoencoder, with
encoder
φ
and decoder
ψ
,scGen (Lotfollahi et al., 2019) predicts a perturbation
response using latent space arithmetic. Let
¯z(l)
be the mean of representations
in condition l
¯z(l)=1
|{i:c(i) = l}| Xziδc(i)l,
the perturbed state of x0ρcis predicted as
ψ(φ(x0)¯z(0) + ¯z(1)).
A.2 cAE
The conditional autoencoder is based on a batch correction technique popular
within the single-cell community, first introduced by (Lopez et al., 2018). It
introduces condition-specific parameters into the encoder and decoder, which
attempt to remove and replace information in the data specific to their conditions.
They operate by concatenating one-hot encodings of condition labels (here,
perturbation status) to the inputs of the encoder and decoder. These encodings,
in effect, make the bias term in the first layer of the encoder and decoder a
learnable parameter specific to each condition and are thus are also considered
to learn a linear shift in latent space. Given an encoder
φ
and decoder
ψ
, the
network is trained to reconstruct cells conditioning on its true label
zi=φ(xi|c(i)),ˆxi=ψ(zi|c(i)).
Once trained, the perturbed state of x0ρcis predicted as
zi=φ(x0|0),ˆx0=ψ(zi|1).
B Dataset
B.1 Single-Cell Multiplex Data
Biologists have various powerful technologies at their disposal, capable of cap-
turing multivariate single-cell measurements. High-content imaging, particularly
14
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
when augmented by multiplexing abilities such as by Iterative Indirect Immunoflu-
orescence Imaging (4i) (Gut et al., 2018), is ideally suited to study heterogeneous
cell responses. With 4i, fluorescently labeled antibodies are iteratively hybridized,
imaged, and removed from a sample to measure the abundance and localization
of proteins and their modifications. Thus, 4i quickly generates large, spatially
resolved phenotypic datasets rich in molecular information from thousands of
treated and untreated (control) cells. Additionally to the multiplexed information
4i generates, information about cellular and nuclear morphology is routinely
extracted from microscopy images (without the need for 4i) by image analysis
algorithms (Carpenter et al., 2006).
Through multiplexing, 4i datasets are able to capture meaningful features
related to both the treatment response heterogeneity (e.g., the phosphorylation
or dephosphorylation of a kinase in a signaling pathway) and the pre-existing
cell-to-cell variability (e.g., protein levels related to different cellular states or cell
cycle phases) which my determine treatment response. Traditional high-content
imaging datasets often need to compromise between features describing either
the former or the latter and may thus struggle to provide sufficient information
to pair treated and control cells accurately.
The cells were seeded in a 384-well plate, allowed to settle and adhere
overnight. Drugs and Dimethyl sulfoxide as the vehicle control was added to
the cells the next morning and incubated for 8 hours, after which the cells were
fixed with Paraformaldehyde. Subsequently, 6 cycles of 4i were performed, for
which the images were acquired with an automated high-content microscope.
All image analysis steps were performed by our in-house platform called Tis-
sueMAPS (https://github.com/TissueMAPS). The steps included illumination
correction (Snijder et al., 2012), alignment of images from different acquisition
cycles using Fast Fourier Transform (Guizar-Sicairos et al., 2008), segmentation
of nuclei and cell outlines (Stoeger et al., 2015), as well cellular and nuclear
measurements of intensity and morphology features using the scikit-image library
(Van der Walt et al., 2014).
C Experimental Details
To train all networks, we use the Adam optimizer (Kingma and Ba, 2014).
C.1 Baselines
To tune baseline models, we use a batch size of 128 and do a grid search over the
width [16,32] and depth [2,3] of the encoder and decoder hidden layers, latent
dimension [4,8], dropout rate [0
.
0,0
.
05,0
.
1,0
.
2] and learning rate [0
.
00001,
0.0001,0.001].
For both scGen and cAE we selected a width=32, depth=2, latent dim=8,
dropout=0.05. scGen uses a learning rate of 0.001, and cAE uses a learning
rate of 0.0001. Both models are trained for 1024 epochs.
15
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
Figure 5: Full set of Imatinib marginals.
C.2 Network Architectures
As suggested by Makkuva et al. (2020), we relax the convexity constraint on
gθ
and instead, penalize its negative weights Wh
l
R(θ) = λX
Wh
lθ
max Wh
l,0
2
F.(8)
The convexity constraint on
fφ
is enforced after each update by setting negative
weights of all Wh
lφto zero. Thus the full objective then states
max
φ:Wh
l0,l
min
θfφ(gθ(y)) − hy, gθ(y)i − fφ(x) + λR(θ).
C.3 Hyperparameters
To learn the optimal transport maps, we use a batch size of 256, an ICNN
architecture of 4 hidden layers of width 64, a learning rate of 0
.
0001 (
β1
= 0
.
5,
β2
= 0
.
9) and
λ
=1. The inner loop minimizing
g
runs for 10 updates to every
update of f.
16
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
Figure 6: Full set of Erlotinib marginals.
17
.CC-BY-NC 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 15, 2021. ; https://doi.org/10.1101/2021.12.15.472775doi: bioRxiv preprint
ResearchGate has not been able to resolve any citations for this publication.
Full-text available
Article
Motivation High-throughput single-cell molecular profiling is revolutionizing biology and medicine by unveiling the diversity of cell types and states contributing to development and disease. The identification and characterization of cellular heterogeneity is typically achieved through unsupervised clustering, which crucially relies on a similarity metric. Results We here propose the use of Optimal Transport (OT) as a cell-cell similarity metric for single-cell omics data. OT defines distances to compare high-dimensional data represented as probability distributions. To speed up computations and cope with the high-dimensionality of single-cell data, we consider the entropic regularization of the classical OT distance. We then extensively benchmark OT against state-of-the-art metrics over thirteen independent datasets, including simulated, scRNA-seq, scATAC-seq and single-cell DNA methylation data. First, we test the ability of the metrics to detect the similarity between cells belonging to the same groups (e.g. cell types, cell lines of origin). Then, we apply unsupervised clustering and test the quality of the resulting clusters. OT is found to improve cell-cell similarity inference and cell clustering in all simulated and real scRNA-seq data, as well as in scATAC-seq and single-cell DNA methylation data. Availability All our analyses are reproducible through the OT-scOmics Jupyter notebook available at https://github.com/ComputationalSystemsBiology/OT-scOmics. Supplementary information Supplementary data are available at Bioinformatics online.
Full-text available
Article
Understanding how cells change their identity and behaviour in living systems is an important question in many fields of biology. The problem of inferring cell trajectories from single-cell measurements has been a major topic in the single-cell analysis community, with different methods developed for equilibrium and non-equilibrium systems (e.g. haematopoeisis vs. embryonic development). We show that optimal transport analysis, a technique originally designed for analysing time-courses, may also be applied to infer cellular trajectories from a single snapshot of a population in equilibrium. Therefore, optimal transport provides a unified approach to inferring trajectories that is applicable to both stationary and non-stationary systems. Our method, StationaryOT, is mathematically motivated in a natural way from the hypothesis of a Waddington’s epigenetic landscape. We implement StationaryOT as a software package and demonstrate its efficacy in applications to simulated data as well as single-cell data from Arabidopsis thaliana root development.
Full-text available
Article
Resistance to immune checkpoint inhibitors (ICIs) is a key challenge in cancer therapy. To elucidate underlying mechanisms, we developed Perturb-CITE-sequencing (Perturb-CITE-seq), enabling pooled clustered regularly interspaced short palindromic repeat (CRISPR)–Cas9 perturbations with single-cell transcriptome and protein readouts. In patient-derived melanoma cells and autologous tumor-infiltrating lymphocyte (TIL) co-cultures, we profiled transcriptomes and 20 proteins in ~218,000 cells under ~750 perturbations associated with cancer cell-intrinsic ICI resistance (ICR). We recover known mechanisms of resistance, including defects in the interferon-γ (IFN-γ)–JAK/STAT and antigen-presentation pathways in RNA, protein and perturbation space, and new ones, including loss/downregulation of CD58. Loss of CD58 conferred immune evasion in multiple co-culture models and was downregulated in tumors of melanoma patients with ICR. CD58 protein expression was not induced by IFN-γ signaling, and CD58 loss conferred immune evasion without compromising major histocompatibility complex (MHC) expression, suggesting that it acts orthogonally to known mechanisms of ICR. This work provides a framework for the deciphering of complex mechanisms by large-scale perturbation screens with multimodal, single-cell readouts, and discovers potentially clinically relevant mechanisms of immune evasion.
Full-text available
Article
Single-cell RNA sequencing (scRNA-seq) provides details for individual cells; however, crucial spatial information is often lost. We present SpaOTsc, a method relying on structured optimal transport to recover spatial properties of scRNA-seq data by utilizing spatial measurements of a relatively small number of genes. A spatial metric for individual cells in scRNA-seq data is first established based on a map connecting it with the spatial measurements. The cell–cell communications are then obtained by “optimally transporting” signal senders to target signal receivers in space. Using partial information decomposition, we next compute the intercellular gene–gene information flow to estimate the spatial regulations between genes across cells. Four datasets are employed for cross-validation of spatial gene expression prediction and comparison to known cell–cell communications. SpaOTsc has broader applications, both in integrating non-spatial single-cell measurements with spatial data, and directly in spatial single-cell transcriptomics data to reconstruct spatial cellular dynamics in tissues.
Full-text available
Article
Lineage tracing involves the identification of all ancestors and descendants of a given cell, and is an important tool for studying biological processes such as development and disease progression. However, in many settings, controlled time-course experiments are not feasible, for example when working with tissue samples from patients. Here we present ImageAEOT, a computational pipeline based on autoencoders and optimal transport for predicting the lineages of cells using time-labeled datasets from different stages of a cellular process. Given a single-cell image from one of the stages, ImageAEOT generates an artificial lineage of this cell based on the population characteristics of the other stages. These lineages can be used to connect subpopulations of cells through the different stages and identify image-based features and biomarkers underlying the biological process. To validate our method, we apply ImageAEOT to a benchmark task based on nuclear and chromatin images during the activation of fibroblasts by tumor cells in engineered 3D tissues. We further validate ImageAEOT on chromatin images of various breast cancer cell lines and human tissue samples, thereby linking alterations in chromatin condensation patterns to different stages of tumor progression. Our results demonstrate the promise of computational methods based on autoencoding and optimal transport principles for lineage tracing in settings where existing experimental strategies cannot be used.
Full-text available
Article
Accurately modeling cellular response to perturbations is a central goal of computational biology. While such modeling has been based on statistical, mechanistic and machine learning models in specific settings, no generalization of predictions to phenomena absent from training data (out-of-sample) has yet been demonstrated. Here, we present scGen (https://github.com/theislab/scgen), a model combining variational autoencoders and latent space vector arithmetics for high-dimensional single-cell gene expression data. We show that scGen accurately models perturbation and infection response of cells across cell types, studies and species. In particular, we demonstrate that scGen learns cell-type and species-specific responses implying that it captures features that distinguish responding from non-responding genes and cells. With the upcoming availability of large-scale atlases of organs in a healthy state, we envision scGen to become a tool for experimental design through in silico screening of perturbation response in the context of disease and drug treatment.
Full-text available
Article
Single-cell transcriptome measurements can reveal unexplored biological diversity, but they suffer from technical noise and bias that must be modeled to account for the resulting uncertainty in downstream analyses. Here we introduce single-cell variational inference (scVI), a ready-to-use scalable framework for the probabilistic representation and analysis of gene expression in single cells (https://github.com/YosefLab/scVI). scVI uses stochastic optimization and deep neural networks to aggregate information across similar cells and genes and to approximate the distributions that underlie observed expression values, while accounting for batch effects and limited sensitivity. We used scVI for a range of fundamental analysis tasks including batch correction, visualization, clustering, and differential expression, and achieved high accuracy for each task. © 2018, The Author(s), under exclusive licence to Springer Nature America, Inc.
Article
MOTIVATION Recent technological advances have led to an increase in the production and availability of single-cell data. The ability to integrate a set of multi-technology measurements would allow the identification of biologically or clinically meaningful observations through the unification of the perspectives afforded by each technology. In most cases, however, profiling technologies consume the used cells and thus pairwise correspondences between datasets are lost. Due to the sheer size single-cell datasets can acquire, scalable algorithms that are able to universally match single-cell measurements carried out in one cell to its corresponding sibling in another technology are needed. RESULTS We propose Single-Cell data Integration via Matching (SCIM), a scalable approach to recover such correspondences in two or more technologies. SCIM assumes that cells share a common (low-dimensional) underlying structure and that the underlying cell distribution is approximately constant across technologies. It constructs a technology-invariant latent space using an autoencoder framework with an adversarial objective. Multi-modal datasets are integrated by pairing cells across technologies using a bipartite matching scheme that operates on the low-dimensional latent representations. We evaluate SCIM on a simulated cellular branching process and show that the cell-to-cell matches derived by SCIM reflect the same pseudotime on the simulated dataset. Moreover, we apply our method to two real-world scenarios, a melanoma tumor sample and a human bone marrow sample, where we pair cells from a scRNA dataset to their sibling cells in a CyTOF dataset achieving 90% and 78% cell-matching accuracy for each one of the samples, respectively. AVAILABILITY AND IMPLEMENTATION https://github.com/ratschlab/scim. SUPPLEMENTARY INFORMATION Supplementary data are available at Bioinformatics online.
Article
Numerous fundamental biological processes require individual cells to correctly interpret and accurately respond to incoming cues. How intracellular signaling networks achieve the integration of complex information from various contexts remains unclear. Here we quantify epidermal growth factor-induced heterogeneous activation of multiple signaling proteins, as well as cellular state markers, in the same single cells across multiple spatial scales. We find that the acute response of each node in a signaling network is tightly coupled to the cellular state in a partially non-redundant manner. This generates a multimodal response that senses the diversity of cellular states better than any individual response alone and allows individual cells to accurately place growth factor concentration in the context of their cellular state. We propose that the non-redundant multimodal property of signaling networks in mammalian cells underlies specific and context-aware cellular decision making in a multicellular setting.
Article
Understanding the molecular programs that guide differentiation during development is a major challenge. Here, we introduce Waddington-OT, an approach for studying developmental time courses to infer ancestor-descendant fates and model the regulatory programs that underlie them. We apply the method to reconstruct the landscape of reprogramming from 315,000 single-cell RNA sequencing (scRNA-seq) profiles, collected at half-day intervals across 18 days. The results reveal a wider range of developmental programs than previously characterized. Cells gradually adopt either a terminal stromal state or a mesenchymal-to-epithelial transition state. The latter gives rise to populations related to pluripotent, extra-embryonic, and neural cells, with each harboring multiple finer subpopulations. The analysis predicts transcription factors and paracrine signals that affect fates and experiments validate that the TF Obox6 and the cytokine GDF9 enhance reprogramming efficiency. Our approach sheds light on the process and outcome of reprogramming and provides a framework applicable to diverse temporal processes in biology.