Fluctuation structure predicts genome-wide perturbation outcomes.

Area of Science:

• Functional Genomics
• Systems Biology
• Statistical Physics

Background:

• Interpreting pooled single-cell perturbation screens is challenging.
• Current methods are either opaque deep learning models or oversimplified frameworks.
• Gene co-fluctuations in unperturbed cells offer a promising avenue for modeling perturbation responses.

Purpose of the Study:

• To present CIPHER (Covariance Inference for Perturbation and High-dimensional Expression Response), a framework for predicting transcriptome-wide perturbation outcomes.
• To leverage linear response theory and gene co-fluctuations for robust biological conclusions.

Main Methods:

• Developed CIPHER, a conceptual framework using linear response theory.
• Applied CIPHER to synthetic networks and 11 large-scale single-cell perturbation datasets (4,234 perturbations, >1.36M cells).
• Validated model performance by comparing gene covariance-based predictions with and without baseline gene covariances.

Main Results:

• CIPHER accurately recapitulated genome-wide responses to single and double perturbations by utilizing baseline gene covariance.
• Removing gene-gene covariances reduced model performance 11-fold, highlighting the importance of fluctuation structures.
• Gene-gene correlations proved transferable across independent experiments of the same cell type.
• CIPHER outperformed differential expression metrics and provided uncertainty-aware effect size estimates via Bayesian inference.
• Genome-wide responses propagated through the covariance matrix along approximately three independent gene modules.

Conclusions:

• CIPHER demonstrates the power of theoretically-grounded models in capturing complex biological responses.
• Baseline gene fluctuation patterns encode fundamental design principles crucial for understanding cellular responses.
• Harnessing gene co-fluctuations provides a robust and interpretable approach to functional genomics.