Current cancer driver variant predictors learn to recognize driver genes instead of functional variants

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Distinguishing cancer driver variants from passenger variants is crucial for understanding tumorigenesis and advancing precision oncology.
• Existing bioinformatics methods face challenges in accurately identifying driver variants due to inherent complexities in cancer data.

Purpose of the Study:

• To investigate the assumptions underlying current bioinformatics methods for driver variant identification.
• To address data construction biases that hinder machine learning models from learning true variant-level functional effects.

Main Methods:

• Analysis of assumptions and data set biases in driver/passenger variant prediction.
• Development of a novel, bias-minimized data set containing both driver and passenger variants across genes.
• Proposal of a weighting procedure to eliminate gene-specific effects for accurate variant-level evaluation.

Main Results:

• Demonstrated that data set construction biases, where drivers map to few genes and passengers to many, allow machine learning models to achieve high performance by recognizing driver genes rather than variant effects.
• Showcased a significant drop in predictor performance on a novel, less biased data set, indicating a correction of over-optimistic assessments.
• Confirmed that evaluating the ability of predictors to model functional effects of single variants is an ongoing challenge.

Conclusions:

• Current machine learning models for driver variant identification are significantly influenced by data set biases, leading to inflated performance metrics.
• A novel data set and weighting procedure are proposed to mitigate bias and accurately assess predictor capabilities in modeling variant functional effects.
• The accurate prediction of functional effects for individual variants remains an open problem in cancer genomics.