Insight into neutral and disease-associated human genetic variants through interpretable predictors

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Predictive models for human nonsynonymous single nucleotide polymorphisms (SNPs) are crucial for variant interpretation in next-generation sequencing data.
• Current sequence-based predictors achieve high performance, suggesting valuable information lies within sequence data, but the specific influential features remain unclear.

Purpose of the Study:

• To develop a method for interpreting existing SNP predictors to uncover the sequence characteristics most important for distinguishing neutral from disease-associated variants.
• To provide insights into the underlying mechanisms driving variant pathogenicity.

Main Methods:

• Utilized a linear support vector machine classifier to train predictors on extensive features derived from variants and their surrounding sequences.
• Extracted feature importance from trained models to identify critical sequence properties influencing predictions.
• Analyzed features including variant data, local sequence context, evolutionary conservation, and sequence annotations.

Main Results:

• Successfully extracted feature importance from trained predictors, revealing specific sequence characteristics that are critical for variant classification.
• Demonstrated that predictor interpretation can illuminate the sequence properties driving predictions of neutral or disease-associated SNPs.
• The approach provides a means to gain deeper understanding of variant impact beyond simple prediction scores.

Conclusions:

• Predictor interpretation offers a valuable method for understanding the sequence basis of SNP pathogenicity.
• This approach can enhance the interpretation of genetic variants identified through large-scale sequencing studies.
• Insights gained can contribute to a better understanding of disease mechanisms at the molecular level.