Using protein interaction networks to identify cancer dependencies from tumor genome data

Genes required for tumor proliferation and survival (dependencies) are challenging to predict from cancer genome data, but are of high therapeutic value. We developed an algorithm (network purifying selection [NPS]) that aggregates weak signals of purifying selection across a gene’s first order protein-protein interaction network. We applied NPS to 4,742 tumor genomes to show that a gene’s NPS score is predictive of whether it is a dependency and validated 58 NPS-predicted dependencies in six cancer cell lines. Importantly, we demonstrate that leveraging NPS predictions to execute targeted CRISPR screens is a powerful, highly cost-efficient approach for identifying and validating dependencies quickly, because it eliminates the substantial experimental overhead required for whole-genome screening.