CanDrivR-CS: a cancer-specific machine learning framework for distinguishing recurrent and rare variants

Missense variants-single nucleotide substitutions that result in an amino acid change in the encoded protein-play an important role in cancer. Distinguishing between recurrent and rare missense variants may reveal insights into selective pressures and functional consequences. While recurrent variants may undergo positive selection across patients, rare variants can also drive resistance or other phenotypes. However, most existing tools predict pathogenicity across broad populations and ignore tumour-specific contexts. Here, we present CanDrivR-CS, a suite of cancer-specific gradient boosting models designed to distinguish between rare and recurrent somatic missense variants.

We curated data from the International Cancer Genome Consortium (ICGC) and trained 50 cancer-specific models. These significantly outperformed a pan-cancer baseline, achieving up to 90% F1 score in leave-one-group-out cross-validation (LOGO-CV) for skin melanoma. Notably, DNA shape features ranked among the most predictive across all cancers, with recurrent variants enriched in structurally complex DNA regions such as bends and rolls-potential mutational hotspots.

All code and data are available at CanDrivR-CS GitHub repository https://github.com/amyfrancis97/CanDrivR-CS, with further advice on the installation procedure in Section 1 of the Supplementary Materials.