DeepVariant

DeepVariant calls genetic variants from aligned sequencing reads (BAM/CRAM) by converting read alignments into pileup image tensors and classifying them with a convolutional neural network to improve variant detection accuracy for individual and population-scale cohorts.

Key Features:

• Input formats: Accepts aligned sequencing reads in BAM and CRAM formats.
• Pileup image tensor conversion: Converts aligned reads into pileup image tensors representing local read context.
• Convolutional neural network: Uses a CNN to classify pileup tensors and call genetic variants.
• Cohort merging integration: Integrates with GLnexus for scalable merging of per-sample calls into cohort callsets.
• Callset outputs: Produces individual-level variant calls and cohort-level callsets.
• Quality evaluation metrics: Utilizes variant recall and precision and Mendelian consistency checks to assess callset quality.
• Benchmarking results: Demonstrated improved callset quality and imputation reference panel performance relative to GATK Best Practices on deeply sequenced 1000 Genomes Project (1KGP) samples.

Scientific Applications:

• Cohort-level variant generation: Producing high-quality cohort callsets for large-scale genetic studies.
• Individual sample variant calling: Generating accurate per-sample variant calls from aligned reads.
• Imputation reference panel improvement: Enhancing imputation reference panel performance through higher-quality callsets.
• Benchmarking and validation: Evaluating variant caller performance using variant recall, precision, and Mendelian consistency in trios.
• Population-scale analyses: Applied to deeply sequenced 1000 Genomes Project (1KGP) samples for population genetics studies.

Methodology:

Processes aligned reads (BAM/CRAM), converts them into pileup image tensors, classifies tensors with a convolutional neural network to call variants, integrates per-sample calls with GLnexus for cohort merging, and evaluates callset quality using variant recall/precision and Mendelian consistency checks.