What Do Single-Cell Models Already Know About Perturbations?

Area of Science:

• Computational Biology
• Single-cell Genomics
• Machine Learning

Background:

• Single-cell generative models commonly incorporate virtual cells.
• The models' capacity to represent biological perturbations is not well understood.
• Investigating implicit knowledge of perturbations in these models is crucial.

Purpose of the Study:

• To explore the implicit knowledge of perturbations within single-cell generative models.
• To develop a method for inferring and visualizing gene expression changes due to perturbations.
• To evaluate the utility of these models for in-silico perturbation simulations and pathway analysis.

Main Methods:

• Trained variational autoencoders on diverse gene expression datasets (genetic, chemical, temporal perturbations).
• Inferred perturbations by computing vector fields of infinitesimal gene expression changes using automatic differentiation.
• Probed a large-scale scVI model, scoring genes by gradient alignment with healthy-to-disease axes and using LLMs for pathway evaluation.

Main Results:

• Successfully recovered known biological transitions in various perturbation datasets (e.g., gene knockout, chemical treatment, embryogenesis).
• Identified pathways with significant relevance to type 2 diabetes in a large mouse cell atlas dataset.
• Demonstrated that gradient-based methods outperform average expression baselines for perturbation analysis.

Conclusions:

• Trained single-cell decoders inherently store rich, perturbation-relevant information.
• Automatic differentiation provides a powerful tool to access this information for in-silico simulations.
• This approach enables principled ranking of genes and pathways along disease or treatment axes without requiring explicit perturbation labels or specialized model architectures.