Learning Single-Cell Perturbation Responses using Neural Optimal Transport

Abstract

Understanding and predicting molecular responses in single cells upon chemical, genetic, or mechanical perturbations is a core question in biology. Obtaining single-cell measurements typically requires the cells to be destroyed. This makes learning heterogeneous perturbation responses challenging as we only observe unpaired distributions of perturbed or nonperturbed cells. Here we leverage the theory of optimal transport and the recent advent of convex neural architectures to present CellOT, a framework for learning the response of individual cells to a given perturbation by coupling these unpaired distributions. We achieve this alignment with a learned transport map that allows us to infer the treatment responses of unseen untreated cells. CellOT outperforms current methods at predicting single-cell drug responses, as profiled by scRNA-seq and a multiplexed protein imaging technology. Further, we illustrate that CellOT generalizes well in unseen settings by (a) predicting the scRNA-seq responses of heldout lupus patients exposed to IFN-β and (b) modeling the hematopoietic developmental trajectories of different subpopulations. We expect CellOT to lay the grounds for delineating the causes of heterogeneous single-cell responses to perturbations and predicting patient-specific drug response landscapes instead of population averages.

CODEX: COunterfactual Deep learning for thein-silicoEXploration of cancer cell line perturbations

“…Observing the latter might substantially deepen our understanding of molecular downstream effects. In this direction, e.g., Bunne et al [2023] and Dong et al [2023], attempt to identify counterfactual cell pairs using optimal transport. However, a combination of those methods with CODEX requires future research, addressing, e.g., potential biases from cell-pair selection.…”

Decoding Heterogenous Single-cell Perturbation Responses

“…Consequently, due to this "survival bias", PS probably only reflects the perturbation responses in a fraction of cells, rather than the full spectrum of perturbations. To overcome this limitation, PS can combine with recently developed prediction methods that predict the responses of perturbations, even if cells between perturbed/non-perturbed states are unevenly distributed 49 .…”

“…1. First, we perform the same quality control as in the dataset's original publication 49 . Specifically, cells are retained only if their numbers of detected genes are between 1,000 and 5,000, and their UMI counts have less than 12% mitochondrial counts.…”

Quantitative Stability of Regularized Optimal Transport and Convergence of Sinkhorn's Algorithm

“…While optimal transport theory has long been an important topic in probability theory and analysis (see, e.g., [30,45,57,58,59,68]), it recently received increasing interest in the area of machine learning and related fields (see, e.g., [3,10,15,54,31,62,66]). One of the main drivers of this increased interest is the introduction of entropic regularization for the optimal transport problem in [22], which improves computational tractability by allowing for the use of Sinkhorn's algorithm [64,65].…”