Transfer learning leverages cancer genomics for in silico CRISPR/Cas9 genetic interaction screen surrogate model

Abstract

Genetic interactions, defined as non-additive combinatorial genetic perturbation effects, reveal functional relationships between genes. Genome-wide CRISPR/Cas9 loss-of-function screens can effectively probe for genetic interactions by identifying differential genetic dependencies between co-isogenic "query mutant" cell lines, but are expensive compared to gene expression profiling. In this work, we explore methods for predicting cell line genetic dependencies from gene expression to develop an expression-based pre-screening method for unscreened query mutants. Along with a small dataset (n = 60) of CRISPR screens in co-isogenic query mutant human HAP1 cells (with matching expression profiles), we use transfer learning methods in order to extract knowledge from external cancer genomics datasets. We used linear representation learning via Model-Agnostic Meta-Learning to leverage the Cancer Dependency Map, a dataset of CRISPR screens on cancer cell lines with diverse genetic backgrounds and lineages. This yielded a prediction performance (correlation) that more than doubled the prediction performance of baselines and approached experimental reproducibility estimates. Additionally, to leverage abundant pan-cancer transcriptomes from The Cancer Genome Atlas, we used variational autoencoders to perform representation learning on the input space to enhance genetic dependency prediction models, but failed to observe an improvement over baselines. More broadly, this work demonstrates the efficacy of transfer learning and meta-learning for biological data, and demonstrates how machine learning can not only interpret high-throughput experiments, but also predict their results and guide future ones.