SAEG: A Novel Deep Learning Architecture for Somatic Alterations

Somatic alterations, like mutations and copy number changes, driver oncogenesis and cancer progression. Their inhibition has been exploited in the clinic, with several targeted therapies approved for patients with specific mutations or amplifications. However, the response rate of these treatments remains low. The causes are several, ranging from clonal heterogeneity to off target binding. For this reason, CRISPR assays have been developed to study the exact effect of a gene’s deletion. Still, the results from them are puzzling with the same alterations responding different to knockout even in the same cellular context. For this reason, we have developed SAEG, a novel deep learning architecture for somatic alterations in cancer. Our architecture is able to model mutations and copy number alterations and protein-protein interactions to predict if a cell will be susceptible to a gene knockout. SAEG outperforms other models and we show that it learns patterns that can be traced back to the biochemical and biological properties of genes and amino acids.