← Về thư mục

DeepVariant - Usage Guide

Overview

Deep learning variant calling using Google's DeepVariant for high-accuracy germline SNP and indel detection from Illumina, PacBio HiFi, or Oxford Nanopore data.

Prerequisites

# Docker (recommended)
docker pull google/deepvariant:1.6.1

# GPU version
docker pull google/deepvariant:1.6.1-gpu

# DeepTrio (separate image tag)
docker pull google/deepvariant:deeptrio-1.6.1

# Singularity
singularity pull docker://google/deepvariant:1.6.1

Input is a sorted, indexed, duplicate-marked BAM/CRAM. Do NOT run BQSR first -- DeepVariant learned its error model from raw base qualities and BQSR slightly lowers its accuracy.

Quick Start

Tell your AI agent what you want to do: - "Call variants from my WGS BAM using DeepVariant" - "Run DeepVariant on exome data with target regions" - "Set up DeepVariant with GPU acceleration" - "Generate GVCFs for joint calling with GLnexus" - "Call de-novo variants in my trio with DeepTrio" - "Should I use DeepVariant, GATK, or DRAGEN for this cohort?"

Example Prompts

Basic Variant Calling

"Call variants from my whole genome BAM using DeepVariant"

"Run DeepVariant on sample.bam with 16 threads"

Exome/Targeted

"Run DeepVariant on my exome data with the capture BED file"

"Call variants only in my target regions using DeepVariant WES model"

Long Reads

"Call variants from my PacBio HiFi reads with DeepVariant"

"Run DeepVariant with the ONT_R104 model for Nanopore data"

Multi-Sample

"Generate GVCFs with DeepVariant for joint calling"

"Set up DeepVariant + GLnexus for my cohort of 20 samples"

Family / Trio

"Run DeepTrio on my proband-father-mother BAMs and flag de-novo variants"

"Joint genotype my DeepTrio gVCFs and check Mendelian consistency"

GPU Acceleration

"Run DeepVariant with GPU support for faster processing"

What the Agent Will Do

1. Verify input BAM is aligned, sorted, indexed, and duplicate-marked (and that BQSR was NOT applied)
2. Select the model type matching the sequencing instrument (WGS, WES, PACBIO, ONT_R104, HYBRID_PACBIO_ILLUMINA)
3. Set up the Docker/Singularity command with volume mounts, one-shot or three-stage
4. Run DeepVariant with an appropriate --num_shards; use GPU only for call_variants
5. Generate VCF and, when a cohort is planned, gVCF for GLnexus joint calling
6. For families, run DeepTrio and merge the trio gVCFs with GLnexus
7. Produce variant statistics and, when a truth set is available, a stratified hap.py/vcfeval benchmark

Tips

• Match the model to the instrument, not the analysis goal -- the wrong --model_type silently degrades accuracy without erroring
• Do NOT apply GATK hard filters or VQSR to DeepVariant output; the CNN already filtered it (FILTER = PASS / RefCall). Threshold on QUAL/GQ only
• Do NOT run BQSR upstream; it slightly lowers DeepVariant accuracy
• Always specify --regions for exome/targeted data to save time
• GPU accelerates only the call_variants stage; make_examples and postprocess_variants are CPU-bound (scale with --num_shards)
• Generate gVCFs with --output_gvcf if planning multi-sample joint calling
• Use GLnexus (not GATK GenotypeGVCFs) for joint genotyping DeepVariant gVCFs
• Use DeepTrio for de-novo calling; naive trio subtraction produces mostly false de-novos
• Ti/Tv around 2.0-2.1 (WGS) or 3.0-3.3 (WES) indicates high-quality SNP calls
• DeepVariant is trained on GIAB then benchmarked on GIAB; weight held-out/non-GIAB-ancestry performance and validate on characterized material before clinical use

Model Types

Model	Data Type
WGS	Illumina whole genome
WES	Illumina exome/targeted
PACBIO	PacBio HiFi
ONT_R104	Oxford Nanopore R10.4+
HYBRID_PACBIO_ILLUMINA	Samples with both HiFi and Illumina

Usage Patterns

Basic WGS

docker run -v "${PWD}:/data" google/deepvariant:1.6.1 \
    /opt/deepvariant/bin/run_deepvariant \
    --model_type=WGS \
    --ref=/data/reference.fa \
    --reads=/data/sample.bam \
    --output_vcf=/data/output.vcf.gz \
    --output_gvcf=/data/output.g.vcf.gz \
    --num_shards=16

Exome/Targeted Sequencing

docker run -v "${PWD}:/data" google/deepvariant:1.6.1 \
    /opt/deepvariant/bin/run_deepvariant \
    --model_type=WES \
    --ref=/data/reference.fa \
    --reads=/data/exome.bam \
    --regions=/data/targets.bed \
    --output_vcf=/data/output.vcf.gz \
    --num_shards=8

GPU Acceleration

docker run --gpus all -v "${PWD}:/data" google/deepvariant:1.6.1-gpu \
    /opt/deepvariant/bin/run_deepvariant \
    --model_type=WGS \
    --ref=/data/reference.fa \
    --reads=/data/sample.bam \
    --output_vcf=/data/output.vcf.gz

Multi-Sample with GLnexus

1. Generate gVCFs for each sample with --output_gvcf
2. Joint genotype with GLnexus:

docker run -v "${PWD}:/data" quay.io/mlin/glnexus:v1.4.1 \
    /usr/local/bin/glnexus_cli --config DeepVariantWGS \
    /data/*.g.vcf.gz | bcftools view - -Oz -o cohort.vcf.gz

Quality Control

# Statistics
bcftools stats output.vcf.gz > stats.txt

# Filter low quality
bcftools view -i 'QUAL>20 && FMT/GQ>20' output.vcf.gz -Oz -o filtered.vcf.gz

# Ti/Tv ratio (expect ~2.0-2.1 for WGS)
bcftools stats output.vcf.gz | grep TSTV

Resource Requirements

Data Type	Memory	Time (30x)
WGS	64 GB	4-6 hours
WES	32 GB	30 min
WGS + GPU	32 GB	1-2 hours

Related Skills

• variant-calling/variant-calling - Engine-selection table (DeepVariant vs GATK vs DRAGEN) and lightweight bcftools calling
• variant-calling/gatk-variant-calling - GATK HaplotypeCaller alternative with joint calling and VQSR/VETS
• variant-calling/joint-calling - GATK reference-confidence joint genotyping, the alternative to GLnexus
• variant-calling/vcf-basics - View and query VCF output
• variant-calling/filtering-best-practices - Post-call filtering for callers that expose hard-filter annotations (not DeepVariant)
• variant-calling/vcf-statistics - QC metrics for called variants
• long-read-sequencing/clair3-variants - Long-read alternative, especially for ONT R9.4