Simple Causal Relationships in Gene Expression Discovered through Deep Learned Collective Variables

Abstract

Developments in high-content phenotypic screening with single-cell read-out hold the promise of revealing interactions and functional relationships between genes at the genomic scale scale. However, the high-dimensionality and noisiness of gene expression makes this endeavor highly challenging when treated as a conventional problem in causal machine learning, both because of the statistical power required and because of the limits on computational tractability. Here we take different tack, and propose a deep-learning approach that finds low-dimensional representations of gene expression in which the response to genetic perturbation is highly predictable. We demonstrate that the interactions between genes that are cooperative in these representations are highly consistent with known ground-truth in terms of causal ordering, functional relatedness, and synergistic impact on cell growth and death. Our novel, statistical physics-inspired approach provides a tractable means through which to examine the response the living cell to perturbation, employing coarse graining that reduces data requirements and focuses on identifying simple relationships between groups of genes. Author summaryUnderstanding the causal relationships between genes and the functions of a cells molecular components has long been a challenge in biology and biomedicine. With recent advancements in technologies that manipulate and measure the activity of thousands of genes at once at the single-cell level, scientists are now afforded with the opportunity to interrogate such relationships at scale. However, extracting useful information from the vast readouts of these technologies is non-trivial, in part due to their many-dimensional and noisy nature. Here we develop a machine learning model that allows for the interpretation of complex genetic perturbations in terms of a simple set of causal relations. By analyzing cooperative groups of genes identified by our model, we demonstrate the model can group genes accurately based on their biological function, their relative ordering up- or downstream in the flow of causation, and how their activities combine to affect cell growth and death. Our approach complements existing machine learning methods in providing a simple way to interpret causal mechanism governing genetic interactions and functional states of cells.