[Paper Review] Predicting Cellular Responses to Novel Drug Perturbations at a Single-Cell Resolution

Leon Hetzel, Simon Böhm+4 more|arXiv (Cornell University)|Apr 28, 2022

Single-cell and spatial transcriptomics42 citations

TL;DR

ChemCPA is an encoder-decoder model that predicts single-cell gene expression responses to unseen drugs by incorporating drug molecular structure and transfer learning from bulk RNA HTS data, enabling counterfactual perturbation predictions at single-cell resolution.

ABSTRACT

Single-cell transcriptomics enabled the study of cellular heterogeneity in response to perturbations at the resolution of individual cells. However, scaling high-throughput screens (HTSs) to measure cellular responses for many drugs remains a challenge due to technical limitations and, more importantly, the cost of such multiplexed experiments. Thus, transferring information from routinely performed bulk RNA HTS is required to enrich single-cell data meaningfully. We introduce chemCPA, a new encoder-decoder architecture to study the perturbational effects of unseen drugs. We combine the model with an architecture surgery for transfer learning and demonstrate how training on existing bulk RNA HTS datasets can improve generalisation performance. Better generalisation reduces the need for extensive and costly screens at single-cell resolution. We envision that our proposed method will facilitate more efficient experiment designs through its ability to generate in-silico hypotheses, ultimately accelerating drug discovery.

Motivation & Objective

• Motivate the need to predict perturbation responses for unseen drugs at single-cell resolution due to limited scRNA-seq screening capacity.
• Develop a model that uses drug structure to generalize perturbations to unseen compounds.
• Leverage bulk RNA HTS data through transfer learning to improve single-cell generalisation.
• Provide an interpretable latent space with disentangled basal, drug, and covariate components.

Proposed method

• Introduce chemCPA, an encoder-decoder architecture with adversarial disentanglement.
• Use an additive latent space where z_i is augmented by z_c (cell line) and hat{s} z_d (drug and dose).
• Employ a perturbation network G, M, S to map molecular representations to latent perturbations.
• Freeze a pretrained molecule encoder G (RDKit features) and train M and S to produce drug perturbation z_d and scaled dose hat{s}.
• Train with adversarial classifiers to disentangle basal state z_i from drug and covariate information using gradient-penalized losses.

Experimental results

Research questions

• RQ1 Can chemCPA predict perturbation responses for drugs unseen during training by leveraging drug structure?
• RQ2 Does transfer learning from bulk RNA HTS (L1000) improve single-cell generalisation, even with different gene sets?
• RQ3 Do latent-space disentanglements enable interpretable attribution of drug versus cell-line effects?
• RQ4 How well does chemCPA perform on counterfactual predictions across dosage, cell lines, and unseen compounds?

Key findings

• ChemCPA outperforms CPA and scGen on generalisation to unseen drug-covariate combinations in single-cell data.
• Pretraining chemCPA on bulk L1000 data yields the best performance, with substantial gains in both all-gene and DEG-focused r2 scores.
• With extended gene sets, pretrained chemCPA remains robust and often surpasses baselines, whereas non-pretrained chemCPA may underperform.
• The model provides an uncertainty measure for unseen drugs that correlates with generalisation ability.
• Latent perturbation embeddings cluster by MoA, indicating meaningful biological structuring in drug representations.
• The approach enables predictions for unseen compounds through the perturbation network, unlike prior CPA which required seen drugs.