AI cancer driver mutation predictions are valid in real-world data.

Area of Science:

• Genomics
• Computational Biology
• Oncology

Background:

• Identifying cancer driver mutations is complex.
• Artificial intelligence (AI) shows promise beyond protein structure prediction for cancer genomics.
• The utility of AI in identifying cancer drivers requires further investigation.

Purpose of the Study:

• To evaluate computational methods for identifying cancer driver mutations.
• To validate AI-identified variants of unknown significance (VUSs) in non-small cell lung cancer (NSCLC).
• To assess the biological validity of AI predictions using patient survival and pathway analysis.

Main Methods:

• Compared computational methods using evolutionary, protein structure, and functional genomic data.
• Validated AI-annotated pathogenic VUSs by assessing their association with overall survival in two NSCLC patient cohorts.
• Analyzed mutual exclusivity of pathogenic VUSs with known oncogenic alterations at the pathway level.

Main Results:

• Methods incorporating protein structure or functional genomic data outperformed evolutionary-only methods for identifying known cancer drivers.
• AI-identified pathogenic VUSs in KEAP1 and SMARCA4 were associated with worse patient survival.
• Pathogenic VUSs demonstrated pathway-level mutual exclusivity with known oncogenic alterations, supporting biological validity.

Conclusions:

• AI-driven computational methods can effectively identify cancer driver mutations.
• Validated AI predictions enhance the understanding of tumor genetics and clinical relevance.
• AI contributes to a more comprehensive analysis of cancer-associated genetic variants.