6: Germline Short Variant Calling with GATK4 & DeepVariant

Chapter 6: Germline Short Variant Calling with GATK4 & DeepVariant

Identifying single nucleotide polymorphisms (SNPs) and small insertions/deletions (indels) from aligned sequencing data is the cornerstone of clinical genetics and population genomics. The Broad Institute's GATK4 (Genome Analysis Toolkit) Best Practices pipeline represents the internationally standardized protocol for germline variant discovery.

1. GATK4 Best Practices Workflow Execution

# Step 1: Base Quality Score Recalibration (BQSR)
gatk BaseRecalibrator \
    -R GRCh38.fasta \
    -I sample1.marked_dup.bam \
    --known-sites dbsnp_146.hg38.vcf.gz \
    --known-sites Mills_and_1000G_gold_standard.indels.hg38.vcf.gz \
    -O sample1.recal_data.table

gatk ApplyBQSR \
    -R GRCh38.fasta \
    -I sample1.marked_dup.bam \
    --bqsr-recal-file sample1.recal_data.table \
    -O sample1.recalibrated.bam

# Step 2: HaplotypeCaller in GVCF Mode (-ERC GVCF)
gatk HaplotypeCaller \
    -R GRCh38.fasta \
    -I sample1.recalibrated.bam \
    -O sample1.g.vcf.gz \
    -ERC GVCF \
    --native-pair-hmm-threads 8

# Step 3: Import Cohort GVCFs into GenomicsDB (High-performance tile storage)
gatk GenomicsDBImport \
    --genomicsdb-workspace-path cohort_genomics_db \
    -L chr22 \
    --sample-name-map sample_map.tsv

# Step 4: Joint Genotype Across the Cohort
gatk GenotypeGVCFs \
    -R GRCh38.fasta \
    -V gendb://cohort_genomics_db \
    -O cohort_raw_variants.vcf.gz

2. Standard GATK Hard Filtering for Small Cohorts

# Filter SNPs using recommended thresholds
gatk VariantFiltration \
    -R GRCh38.fasta \
    -V cohort_raw_snps.vcf.gz \
    -O cohort_filtered_snps.vcf.gz \
    --filter-expression "QD < 2.0 || FS > 60.0 || MQ < 40.0 || MQRankSum < -12.5 || ReadPosRankSum < -8.0 || SOR > 3.0" \
    --filter-name "SNP_HARD_FILTER"

3. Deep Learning Variant Discovery with DeepVariant

Google's DeepVariant reimagines variant calling as a computer vision problem. It converts local alignments around candidate loci into multi-channel tensor images (encoding read base, quality, strand, and mismatch) and classifies genotypes using a convolutional neural network (CNN), outperforming heuristic callers in challenging regions.