Learning interpretable cellular responses to complex perturbations in high-throughput screens

Abstract

Recent advances in multiplexing single-cell transcriptomics across experiments are enabling the high throughput study of drug and genetic perturbations. However, an exhaustive exploration of the combinatorial perturbation space is experimentally unfeasible, so computational methods are needed to predict, interpret and prioritize perturbations. Here, we present the Compositional Perturbation Autoencoder (CPA), which combines the interpretability of linear models with the flexibility of deep-learning approaches for single-cell response modeling. CPA encodes and learns transcriptional drug response across different cell types, doses, and drug combinations. The model produces easy-to-interpret embeddings for drugs and cell types, allowing drug similarity analysis and predictions for unseen dosages and drug combinations. We show CPA accurately models single-cell perturbations across compounds, dosages, species, and time. We further demonstrate that CPA predicts combinatorial genetic interactions of several types, implying it captures features that distinguish different interaction programs. Finally, we demonstrate CPA allows in-silico generation of 5,329 missing combinations (97.6% of all possibilities) with diverse genetic interactions. We envision our model will facilitate efficient experimental design by enabling in silico response prediction at the single-cell level.

Mapping single-cell data to reference atlases by transfer learning

“…The suitability of model organisms for disease research can be directly translated into the human context: for example, projecting mouse single-cell tumor data on a reference human patient tumor atlas may help to identify accurate tumor models that include desired molecular and cellular properties of a patient's microenvironment. Incorporating additional covariates as conditional neurons in the reference model will allow modeling of treatment response with a certain perturbation or drug 63,64 . Secondly, we envision assembling multimodal single-cell reference atlases to include epigenomic 65 , chromosome conformation 66 , proteome 51 and spatially resolved measurements.…”

Supervised Training of Conditional Monge Maps

“…We consider various high-dimensional problems arising from this scenario to evaluate the performance of CONDOT ( § 3) versus other baselines. CONDOT and ICNN OT (Makkuva et al, 2020) based on embedding E moa b. as well as E ohe , and c. CONDOT and CPA (Lotfollahi et al, 2021) based on embedding E ohe on known and unknown perturbations or combinations. Results above the diagonal suggest higher predictive performance of CONDOT.…”

“…To illustrate this problem, we consider cell populations comprising three different cell lines (A549, MCF7, and K562). As visible in Table 1, CONDOT outperforms current baselines which equally condition on covariate information such as CPA (Lotfollahi et al, 2021), assessed through various evaluation metrics. Figure 4b displays a gene showing highly various responses towards the drug Givinostat dependent on the cell line.…”

“…For more details on evaluation metrics, see § E.2. To put CONDOT's performance into perspective, we compare it to current state-of-the-art baselines (Lotfollahi et al, 2021) as well as parameterized Monge maps without context variables (ICNN OT), see § E.1. As visible in Table 1 and Fig.…”

CellDrift: Inferring Perturbation Responses in Temporally-Sampled Single Cell Data

“…For example, Augur used cross-validation scores of random forest to prioritize perturbation effects across cell types [22]. scGen and CPA applied autoencoder models to learn perturbation responses in the latent space and predict unseen scenarios [23,24]. Although machine learning approaches are powerful in analyzing high-dimensional data, interpretability in latent spaces remains a significant challenge and more so in temporal variation settings.…”