1O - An automatic learning platform to annotate variants of unknown significance

Background

Genomic profiling of tumors is increasingly used for diagnosis and treatment, thereby fostering precision oncology. However, accurately interpreting tumor genomic alterations remains challenging, particularly due to the high fraction of variants of unknown significance (VUS) observed in cancer patients.

Methods

We developed an Automatic Learning Platform for precision oncology, a computational framework that leverages mutation data from thousands of patient tumor samples to identify cancer driver genes and mutations through signals of positive selection. This platform integrates IntOGen, a computational approach that systematically produces a compendium of mutational cancer driver genes, and BoostDM, a machine learning method for identifying cancer driver mutations.

Results

We present the recent application of this Automatic Learning Platform to over 33,000 cancer genomes across 271 patient cohorts representing 87 tumor types. We identified the genes driving tumorigenesis in each cancer type, generating a global list of 633 cancer driver genes. Then, we built machine learning models that learn feature combinations defining driver mutations in a gene- and tumor type-specific manner. We produced high-quality models to identify the driver mutation in 115 cancer genes across various tumor types, resulting in 515 gene/tumor type combination models. We compared our methodology with ClinVar and OncoKB, two widely used mutation interpretation databases, using the GENIE cohort. Out of 262,010 somatic mutations in cancer genes in this cohort, 35% and 59% were annotated in these databases, respectively, while 81% were classified by our platform, demonstrating its capability to reclassify VUS. This data-driven platform is included in the Cancer Genome Interpreter (CGI), a tool for automatically interpreting cancer patient genomic data by identifying oncogenic mutations and biomarkers of drug response linked to these mutations.

Conclusions

We developed an Automatic Learning Platform to identify cancer driver genes and mutations across tumor types and incorporated it into CGI to annotate VUS. Continuous improvement of our platform through accessing newly sequenced tumors will enhance CGI’s interpretation capabilities for future cancer patients.

Editorial acknowledgement

Clinical trial identification

Legal entity responsible for the study

IRB Barcelona.

Funding

European Union's Horizon Europe Program - Reference: 101057509; Instituto de Salud Carlos III and Fondo Europeo de Desarrollo Regional (FEDER) - Reference: IMP/00019.

Disclosure

All authors have declared no conflicts of interest.