Graph convolution networks model identifies and quantifies gene and cancer specific transcriptome signatures of cancer driver events

Gil Ben Cohen, Adar Yaacov, Yishai Ben Zvi, Ranel Loutati, Natan Lishinsky, Jakob Landau, Tom Hope, Aron Popovzter, Shai Rosenberg

Abstract: Highlights•STAMP is a framework trained by classical learning and Graph Convolution Networks models to identify TP53 mutational status in multiple tumor types in TCGA data.•The models reached AUC scores of 90 % or higher on unseen test set and on the independent dataset - METABRIC.•Similar models were trained to predict over 250 important cancer driver genes' and pathways' dysfunction, using TCGA samples.•Prediction was modified to form a continuous score to quantify the drivers' dysfunction, predicting signal strength of driver events. This score shows strong correlation with survival, and with relevant mutational single base signatures.•STAMP's quantitative score stratifies IC50 values in multiple cell line experiments in GDSC.•STAMP's quantitative score stratified response to treatment in multiple clinical cohorts, improving over state-of-the-art models in several setups.