A multimodal framework for comprehensive driver variant prediction in cancer.

Area of Science:

• Genomics
• Proteomics
• Computational Biology

Background:

• Cancer genomes harbor numerous mutations, but distinguishing driver mutations from passenger mutations is difficult.
• Accurate identification of driver variants is crucial for understanding tumor development and guiding therapeutic strategies.
• Current methods face challenges in integrating diverse data types for precise driver variant classification.

Purpose of the Study:

• To develop an accurate and interpretable multimodal model for classifying cancer driver variants.
• To integrate DNA sequence, protein 3D structure, and cancer omics data for enhanced variant analysis.
• To improve the identification of actionable driver mutations for targeted cancer therapies.

Main Methods:

• Introducing ModVAR, a multimodal model combining DNA sequences, predicted protein structures, and cancer omics data.
• Utilizing pre-trained models (DNAbert2, ESMFold) and a self-supervised approach for omics profiles.
• Evaluating performance on validated driver variants using classification metrics, molecular docking, and analysis of intrinsically disordered regions.

Main Results:

• ModVAR demonstrates high accuracy in identifying validated cancer driver variants across benchmarks.
• The model prioritizes variants with therapeutic potential, supported by molecular docking analyses.
• Structural predictions facilitate modeling of variants in intrinsically disordered protein regions, with protein structure being the most influential modality.
• Generated a publicly available dataset of 3,971,946 variant annotations.

Conclusions:

• ModVAR effectively integrates sequence, structure, and omics data for driver variant discovery.
• The model offers robust performance, aids in hypothesis generation and target discovery, and advances personalized cancer therapy.
• Provides a large-scale resource to accelerate cancer research and the development of novel therapeutic strategies.