D2Deep

D2Deep combines a nonspecialized protein language model with protein-specific evolutionary information to predict cancer driver mutations and enable interpretable visualization of mutational landscapes.

Key Features:

• Sequence-only prediction: D2Deep relies exclusively on protein sequence information without requiring structural data.
• Balanced somatic training: The model is trained on a balanced, somatic training set that mitigates biases related to hotspot mutations.
• Epistatic change detection: D2Deep captures intricate epistatic changes throughout the protein caused by mutations, which correlate with known mutations in clinical settings.
• Interpretable outputs: The method provides confidence scores and visualizations of mutational landscapes to aid clinical interpretation and mutation prioritization.
• Cross-domain versatility: D2Deep demonstrates performance on non-cancer mutation prediction where most variants still lack known consequences.

Scientific Applications:

• Cancer driver mutation prediction: D2Deep differentiates cancer-causing driver mutations from random passenger mutations in tumor genomic data.
• Clinical mutation prioritization: Predictions and confidence scores assist with clinical interpretation of somatic mutations.
• Non-cancer variant effect prediction: The method extends to predicting consequences of variants in non-cancer contexts where functional impacts remain unknown.

Methodology:

D2Deep integrates a nonspecialized protein language model that captures the makeup of all protein sequences with protein-specific evolutionary information that encompasses functional requirements for a particular protein.