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

Background: The identification and drug targeting of cancer causing (driver) genetic alterations has seen immense improvement in recent years, with many new targeted therapies developed. However, identifying, prioritizing, and treating genetic alterations is insufficient for most cancer patients. Current clinical practices rely mainly on DNA level mutational analyses, which in many cases fail to identify treatable driver events. Arguably, signal strength may determine cell fate more than the mutational status that initiated it. The use of transcriptomics, a complex and highly informative representation of cellular and tumor state, had been suggested to enhance diagnostics and treatment successes. A gene-expression based model trained over known genetic alterations could improve identification and quantification of cancer related biological aberrations’ signal strength. Methods: We present STAMP (Signatures in Transcriptome Associated with Mutated Protein), a Graph Convolution Networks (GCN) based framework for the identification of gene expression signatures related to cancer driver events. STAMP was trained to identify the p53 dysfunction of cancer samples from gene expression, utilizing comprehensive curated graph structures of gene interactions. Predictions were modified for generating a quantitative score to rank the severity of a driver event in each sample. STAMP was then extended to almost 300 tumor type-specific predictive models for important cancer genes/pathways, by training to identify well-established driver events’ annotations from the literature. Results: STAMP achieved very high AUC on unseen data across several tumor types and on an independent cohort. The framework was validated on p53 related genetic and clinical characteristics, including the effect of Variants of Unknown Significance, and showed strong correlation with protein function. For genes and tumor types where targeted therapy is available, STAMP showed correlation with drugs sensitivity (IC50) in an independent cell line database. It managed to stratify drug effect on samples with similar mutational profiles. STAMP was validated for drug-response prediction in clinical patients’ cohorts, improving over a state-of-the-art method and suggesting potential biomarkers for cancer treatments. Conclusions: The STAMP models provide a learning framework that successfully identifies and quantifies driver events’ signal strength, showing utility in portraying the molecular landscape of tumors based on transcriptomics. Importantly, STAMP manifested the ability to improve targeted therapy selection and hence can contribute to better treatment.