Co-fluctuations of genes in single cells predict transcriptome-wide outcomes to perturbations

Abstract

Pooled single-cell perturbation screens represent powerful experimental platforms for functional genomics, yet interpreting these rich datasets for meaningful biological conclusions remains challenging. Most current methods fall at one of two extremes: either opaque deep learning models that obscure biological meaning, or simplified frameworks that treat genes as isolated units. As such, these approaches overlook a crucial insight: gene co-fluctuations in unperturbed cellular states can be harnessed to model perturbation responses. Here we present CIPHER (Covariance Inference for Perturbation and High-dimensional Expression Response), a conceptual framework leveraging ideas from linear response in statistical physics to model transcriptome-wide perturbation outcomes using gene co-fluctuations in unperturbed cells. We validated our approach on synthetic regulatory networks before applying it to 29 large-scale single-cell genetic perturbation datasets covering 19,003 perturbations and over 6.26M cells. Our work robustly recapitulated genome-wide responses to single and double perturbations by exploiting baseline gene covariance structure. Importantly, eliminating gene-gene covariances, while retaining gene-intrinsic variances, i.e. mean-field conditions, dramatically reduced model performance by several folds across multiple metrics, demonstrating the rich information stored within baseline fluctuation structures. Benchmarked against recent deep learning and linear baselines, CIPHER matched or exceeded the best-performing approaches with fitting a single parameter. Moreover, gene-gene correlations transferred successfully across independent studies of the same cell type, revealing stereotypic fluctuation structures. We further extended CIPHER to the inverse problem of identifying true driver perturbations, where it achieved high performance across both genetic and chemical perturbation screens through uncertainty-aware Bayesian inference. We further used CIPHER to nominate drivers of therapy resistance in melanoma and pancreatic cancer, validating its top predictions experimentally. Finally, most genome-wide responses propagated through the covariance matrix along approximately 1-3 independent and global gene modules, consistent with a low-rank structure of the underlying gene regulatory network, which we show can enable the framework's success. We have also created a package called cipher-perturb, available on PyPI, to apply the framework to any dataset, accompanied by a detailed website (https://goyallab.github.io/CIPHERWebsite/). Our study underscores the importance of theoretically-grounded models in capturing complex biological responses, highlighting fundamental design principles encoded in cellular fluctuation patterns.