Improving variant calling using population data and deep learning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Population variant data aids single-sample variant interpretation but isn't integrated into calling.
• Current methods often filter variants, sacrificing recall for precision.

Purpose of the Study:

• To develop population-aware DeepVariant models for improved variant calling.
• To integrate population allele frequencies directly into the variant calling process.

Main Methods:

• Developed population-aware DeepVariant models using allele frequencies from the 1000 Genomes Project.
• Incorporated a new channel encoding population allele frequencies.
• Assessed performance using population-specific and diverse reference panels.

Main Results:

• Population-aware models reduced variant calling errors, improving precision and recall in single samples.
• Observed reductions in rare homozygous and pathogenic ClinVar calls cohort-wide.
• Diverse reference panels yielded the highest accuracy, outperforming population-specific panels.
• Benefits generalized to samples with ancestry different from training data.

Conclusions:

• Population-aware DeepVariant models enhance variant calling accuracy and reliability.
• Large, diverse reference panels are optimal for population-aware variant calling.
• This approach improves variant interpretation across diverse populations and ancestries.