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

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Distinguishing cancer driver mutations from passenger mutations is crucial but challenging.
• Existing homology-based predictors have biases and limitations in cancer biology.
• Protein language models show promise but haven't been applied at scale for cancer driver mutation prediction, often lacking interpretability.

Purpose of the Study:

• Introduce D2Deep, an AI method for large-scale cancer driver mutation prediction.
• Address limitations of current predictors, including biases and lack of interpretability.
• Leverage protein language models and evolutionary information for accurate mutation identification.

Main Methods:

• Developed D2Deep, combining a general protein language model with protein-specific evolutionary information.
• Utilized exclusively sequence information for prediction.
• Trained the model on a balanced somatic dataset to mitigate hotspot mutation biases.

Main Results:

• D2Deep outperforms state-of-the-art predictors in identifying cancer driver mutations.
• The method captures complex epistatic changes, correlating with clinical mutations for interpretation.
• Demonstrated versatility with successful non-cancer mutation prediction.

Conclusions:

• D2Deep offers a powerful, interpretable solution for cancer driver mutation prediction.
• The AI model mitigates biases and enhances clinical utility.
• Predictions and confidence scores are available to aid clinical interpretation and mutation prioritization.