Combining evolution and protein language models for an interpretable cancer driver mutation prediction with D2Deep

The mutations driving cancer are being increasingly exposed through tumor-specific genomic data. However, differentiating between cancer-causing driver mutations and random passenger mutations remains challenging. State-of-the-art homology-based predictors contain built-in biases and are often ill-suited to the intricacies of cancer biology. Protein language models have successfully addressed various biological problems but have not yet been tested on the challenging task of cancer driver mutation prediction at a large scale. Additionally, they often fail to offer result interpretation, hindering their effective use in clinical settings. The AI-based D2Deep method we introduce here addresses these challenges by combining two powerful elements: (i) a nonspecialized protein language model that captures the makeup of all protein sequences and (ii) protein-specific evolutionary information that encompasses functional requirements for a particular protein. D2Deep relies exclusively on sequence information, outperforms state-of-the-art predictors, and captures intricate epistatic changes throughout the protein caused by mutations. These epistatic changes correlate with known mutations in the clinical setting and can be used for the interpretation of results. The model is trained on a balanced, somatic training set and so effectively mitigates biases related to hotspot mutations compared to state-of-the-art techniques. The versatility of D2Deep is illustrated by its performance on non-cancer mutation prediction, where most variants still lack known consequences. D2Deep predictions and confidence scores are available via https://tumorscope.be/d2deep to help with clinical interpretation and mutation prioritization.