Discordant calls across genotype discovery approaches elucidate variants with systematic errors

Area of Science:

• Genomics and Bioinformatics
• Population Genetics
• Clinical Variant Interpretation

Background:

• High-throughput sequencing data are crucial for clinical variant interpretation and population genetics.
• Reference datasets like gnomAD are often considered definitive but may contain systematic errors.
• Discordant genotype calls across different discovery methods can introduce bias and false positives.

Purpose of the Study:

• To describe and characterize the phenomenon of discordant genotype calls in large-scale genetic datasets.
• To identify specific variants in gnomAD exhibiting discordant calls that require cautious analysis.
• To develop a predictive metric and machine learning classifier for identifying discordant variants in other datasets.

Main Methods:

• Analysis of genotype calls across different discovery approaches within gnomAD.
• Characterization of error modes, focusing on heterozygous vs. homozygous calls.
• Development and training of a machine learning classifier using gnomAD data to predict discordant variants.

Main Results:

• Systematic differences in genotype calls were observed for some variants passing standard filters in gnomAD.
• The most common error mode involves heterozygous calls in one approach and homozygous reference calls in another.
• Discordant sites are generally shared across ancestries, though discovery power varies by population.
• Characteristic variant features can predict discordant behavior, enabling the development of a predictive classifier.

Conclusions:

• Discordant genotype calls in reference datasets can lead to technological artifacts mistaken for biological signals.
• A list of identified discordant sites in gnomAD is provided for cautious use in downstream analyses.
• The developed metric and classifier can help identify and mitigate the impact of discordant variants in future genetic studies.