Abstract 473: GeneMasking: A new approach for inferring the functional role of cancer driver genes and its application in prostate cancer

• Published in: Cancer research (Chicago, Ill.) Ill.), 2022-06, Vol.82 (12_Supplement), p.473-473
• Main Authors: Dhruba, Saugato Rahman, Amitay, Niv, Wolf, Lior, Ruppin, Eytan
• Format: Article
• Language: English
• ISSN: 1538-7445; 1538-7445
• DOI: 10.1158/1538-7445.AM2022-473

Introduction: Genomic events in cancer driver genes such as mutations and copy number alterations play critical roles in cancer onset and progression. Here we apply a novel biologically inspired deep learning classification method called GeneMasking to perform prostate cancer (PC) classification high accuracy. The important key novelty of GeneMasking is that it marks the genomic events that determine the classification of a given sample in a sample-specific manner. This provides us with an exciting opportunity to investigate two fundamental cancer research questions: 1.For a specific tumor sample, which cancer driver genes are functionally important, playing a detrimental role in its classification? 2.For a specific tumor sample, does an established cancer driver gene act as oncogene or tumor suppressor gene (TSG) in that sample? Methods: We applied GeneMasking to study these questions by analyzing the somatic mutation and copy number alteration (CNA) data of 1011 PC patient samples recently published by Elmarakeby et al. [Nature 598, 348-352 (2021)] and classify as primary or metastatic tumors. GeneMasking achieves an overall AUC (area under the receiver operating characteristic curve) of 0.958 in this classification task. Additionally and importantly, for each sample, GeneMasking provides a list of genes with mutation and/or CNA events that are important in classifying it. We focus on 61 known COSMIC cancer driver genes marked as important in ≥ 20 samples. A driver gene that is marked as important is inferred as oncogenic-acting in that given sample if it is amplified or mutated by an activating mutation, and vice versa for inferring it as a TSG-acting in that sample. For each gene, we compute the oncogene/TSG ratio of its inferred function across all samples where it was marked ‘important’. We then investigate whether this ratio is higher than 1 as expected for drivers traditionally labeled oncogenes, or lower, as expected for those known as TSGs. Results: Our key findings are: 1) For 44 drivers (including 5 PC-specific genes), the overall inferred oncogene/TSG ratios reassuringly match their conventional annotations as oncogenes or TSGs in COSMIC. 2) For 8 drivers, however, their inferred oncogene/TSG ratios are opposite from those listed in COSMIC. Those include MUC4, XPO1, FANCD2 etc. 3) 5 drivers are annotated as both oncogenes and TSGs in COSMIC and here we infer their specific roles in PC. Conclusions: We provide a first-of-its-kind computational