4. Small Variant Calling (Germline SNVs & Indels)

Broad Institute’s GATK HaplotypeCaller calls germline SNPs and indels via local de-novo haplotype assembly of active regions. It evaluates candidate haplotypes using Pair-HMM likelihood models, generating GVCF records for joint cohort genotyping.

# 1. Run HaplotypeCaller in per-sample GVCF mode
gatk --java-options "-Xmx16g" HaplotypeCaller \
  -R GRCh38.fa \
  -I sample.dedup.bam \
  -O sample.g.vcf.gz \
  -ERC GVCF \
  --dbsnp dbsnp_146.hg38.vcf.gz \
  --native-pair-hmm-threads 8

# 2. Single-sample Genotyping (or joint call via GenomicsDBImport)
gatk GenotypeGVCFs \
  -R GRCh38.fa \
  -V sample.g.vcf.gz \
  -O sample.genotyped.vcf.gz

Google DeepVariant transforms aligned reads into multi-channel image tensors (base, quality, strand, mismatch) and feeds them into an Inception convolutional neural network to classify genotypes, consistently winning PrecisionFDA benchmark awards.

# 1. Run DeepVariant via Docker/Singularity for WGS
singularity exec --bind /data:/data \
  docker://google/deepvariant:1.6.0 \
  /opt/deepvariant/bin/run_deepvariant \
  --model_type=WGS \
  --ref=/data/GRCh38.fa \
  --reads=/data/sample.dedup.bam \
  --output_vcf=/data/deepvariant_calls.vcf.gz \
  --output_gvcf=/data/deepvariant_calls.g.vcf.gz \
  --num_shards=16

# 2. For PacBio HiFi or ONT data, change model_type:
# --model_type=PACBIO (or --model_type=ONT_R104)

Clair3 is a hybrid deep-learning variant caller specifically designed for long reads. It couples a pileup model for rapid candidate scanning with a full-alignment bi-directional LSTM network for resolving complex multi-base indels and homopolymer tracts.

# 1. Run Clair3 on Oxford Nanopore Q20+ R10.4.1 BAM
run_clair3.sh \
  --bam_fn=ont_aligned.sorted.bam \
  --ref_fn=GRCh38.fa \
  --threads=16 \
  --platform="ont" \
  --model_path=/models/ont/r1041_e82_400bps_sup_v420 \
  --output=clair3_ont_out \
  --include_all_ctgs \
  --enable_phasing

# 2. Index output VCF
tabix -p vcf clair3_ont_out/merge_output.vcf.gz

PEPPER-Margin-DeepVariant is a joint pipeline developed by UCSC and Google. PEPPER finds candidate variants using recurrent neural networks; Margin phases haplotype blocks; DeepVariant generates final high-accuracy genotypes from Oxford Nanopore reads.

# 1. Run containerized PEPPER-Margin-DeepVariant
singularity exec --bind /data:/data \
  docker://kishwars/pepper_deepvariant:r0.8 \
  run_pepper_margin_deepvariant call_variant \
  -b /data/ont_aligned.bam \
  -f /data/GRCh38.fa \
  -o /data/pepper_out \
  -p pepper_sample \
  -t 16 \
  --ont_r10_q20 \
  --phased_output

FreeBayes is a Bayesian genetic variant detector that bypasses positional alignment assumptions by evaluating haplotype windows. It natively supports arbitrary organism ploidy, pooled sequencing, and complex micro-inversions.

# 1. Run FreeBayes with parallel regions and ploidy control
freebayes \
  -f GRCh38.fa \
  -p 2 \
  --min-alternate-fraction 0.15 \
  --min-alternate-count 3 \
  --min-coverage 10 \
  sample.dedup.bam \
  | bcftools view -Oz -o freebayes_raw.vcf.gz

# 2. Filter variants by quality and depth
bcftools filter -i 'QUAL >= 30 && DP >= 15' freebayes_raw.vcf.gz \
  -Oz -o freebayes_filtered.vcf.gz
tabix -p vcf freebayes_filtered.vcf.gz

Octopus is a haplotype-based Bayesian caller that integrates machine-learned error models with combinatorial haplotype generation, excelling in repetitive DNA, segmental duplications, and micro-homology regions.

# 1. Run Octopus with forest error model
octopus \
  -R GRCh38.fa \
  -I sample.dedup.bam \
  -o octopus_calls.vcf.gz \
  --threads 16 \
  --forest-model /opt/octopus/models/germline.model \
  --organism-ploidy 2

tabix -p vcf octopus_calls.vcf.gz