Deep generative framework for modeling single-cell drug perturbation response.

Area of Science:

• Computational Biology
• Genomics
• Pharmacology

Background:

• Accurate cell-specific drug response characterization is crucial for understanding patient-level therapeutic variability.
• Heterogeneous cellular responses complicate drug effect inference from single-cell transcriptomic data.

Purpose of the Study:

• Introduce scDPR (single-cell Drug Perturbation Responses), a framework to predict drug-perturbed single-cell transcriptomic states.
• Decompose observed responses into distinct causal effects, integrating molecular features, dosage, and cell attributes.

Main Methods:

• scDPR framework with an attribute adaptation module for single-cell transcriptional shifts.
• Causal graph learning module using optimal transport (OT) to infer direct drug effects and account for confounding factors.

Main Results:

• scDPR outperforms state-of-the-art methods like chemCPA in forecasting transcriptome responses to novel compounds and pathways.
• Identifies key subpopulations contributing to drug response variability, offering insights into cellular heterogeneity.
• Achieves superior performance with reduced training time and rounds compared to mainstream models.

Conclusions:

• scDPR provides efficient computational support for large-scale drug screening and mechanism research.
• Enables accurate prediction of cell-specific drug responses and decomposition of causal effects.
• Enhances understanding of drug perturbation impacts on cellular transcriptomes.