Comparison of Pathogenicity Prediction Tools on Somatic Variants

Area of Science:

• Genomics
• Cancer Biology
• Bioinformatics

Background:

• Genomic sequencing is vital for cancer patient management.
• Interpreting somatic variants and their pathogenicity presents challenges.
• Existing pathogenicity prediction tools were primarily developed for germline variants, not somatic ones.

Purpose of the Study:

• To benchmark the performance of common pathogenicity prediction tools for cancer somatic variants.
• To identify the most effective tools for interpreting somatic mutations in oncology.

Main Methods:

• Assembled a gold standard list of 4319 somatic single-nucleotide variants (2996 oncogenic, 1323 neutral).
• Annotated variants using common prediction tools, including dbNSFP and UMD-Predictor.
• Calculated performance metrics (e.g., Matthews correlation coefficient, AUC) to rank tools.

Main Results:

• Identified top-performing tools for somatic variants: CADD, Eigen/Eigen-PC, PolyPhen-2, PROVEAN, UMD-Predictor, and REVEL.
• Sorting Intolerant From Tolerant (SIFT) ranked in the second performance tier.
• Combining prediction tools offered only marginal performance improvements due to discordant predictions.

Conclusions:

• Several pathogenicity prediction tools demonstrate utility for somatic variants in cancer.
• The performance of tools like SIFT may be suboptimal for somatic variant interpretation.
• Further refinement or development of tools specifically for somatic variants is warranted.