Single-cell disentangled representations for perturbation modeling and treatment effect estimation.

Area of Science:

• Computational Biology
• Single-cell Genomics
• Systems Biology

Background:

• Single-cell sequencing captures cell states but hinders estimating individualized treatment effects (ITEs).
• Understanding cell-state-specific gene regulation after perturbations is key to biological discovery.
• Heterogeneous cellular responses necessitate methods to infer counterfactual states.

Purpose of the Study:

• To present scDRP, a generative framework for estimating ITEs from single-cell perturbation data.
• To enable the dissection of cell-state-specific gene regulation and causal relationships.
• To reveal heterogeneous mechanistic responses in biological systems.

Main Methods:

• Leveraging disentangled representation learning via a sparsity regularized beta-variational autoencoder (β-VAE).
• Separating perturbation-dependent and independent latent variables.
• Performing conditional optimal transport in latent space to infer counterfactual states and estimate ITEs.

Main Results:

• scDRP accurately estimates treatment effects and individual counterfactual responses in simulated and real single-cell data.
• The framework reveals cell type-specific functional gene module dynamics under various exposures (e.g., rhinovirus, cigarette smoke).
• scDRP identified distinct cellular patterns and functional module activation in response to interferon stimulation and CRISPR knockouts.

Conclusions:

• scDRP provides a principled computational approach to elucidate heterogeneous causal relationships from single-cell perturbation data.
• The framework enhances understanding of cellular and molecular mechanisms by revealing cell-specific responses.
• scDRP generalizes to unseen perturbation doses and combinations, offering robust biological insights.