Docker Image: google/deepvariant:1.5.0
| |
|---|
| Approach | CNN image classification on pileup tensors |
| Strengths | SNPs, platform-aware accuracy, simpler tuning |
| Parameters | 15 across 6 categories |
| Compute | High — GPU optional but recommended |
How It Works
DeepVariant operates in three stages:
1. make_examples
Scans the BAM file and identifies candidate variant sites. For each candidate, it constructs a pileup image — a multi-channel tensor encoding read alignments, base qualities, mapping qualities, strand information, and other signals. This is where most quality-affecting parameters live (candidate thresholds, read quality filters, read processing options).
2. call_variants
Feeds each pileup image through the pre-trained CNN model. The model outputs probabilities for three genotype classes: 0/0 (hom-ref), 0/1 (het), and 1/1 (hom-alt). The model was trained on truth sets and is specific to the sequencing platform (WGS, WES, PacBio, etc.).
3. postprocess_variants
Converts model predictions to a VCF file. Applies quality filters (QUAL thresholds), handles multi-allelic sites, and computes genotype quality (GQ) scores.
Unlike GATK/FreeBayes where hyperparameters control the statistical model, DeepVariant parameters affect what data the CNN sees and how outputs are filtered. You cannot change the model itself — but you can significantly affect accuracy by controlling inputs and output filtering.
Hyperparameters
Model Selection (Top-level)
| Parameter | Values | Default | Description |
|---|
model_type | WGS / WES / PACBIO / HYBRID_PACBIO_ILLUMINA | WGS | Selects the pre-trained model. Each model is trained on data from a specific sequencing platform. Using the wrong model for your data type will significantly degrade accuracy. |
Candidate Variant Thresholds (make_examples)
These control which genomic positions become candidates for the CNN to evaluate. Lowering thresholds = more candidates = higher sensitivity but more computation and potentially more false positives (though the CNN is the final arbiter).
| Parameter | Range | Default | Description |
|---|
vsc_min_fraction_snps | 0.0-1.0 | 0.12 | Minimum fraction of reads supporting an alternate allele to consider a site as a SNP candidate. Lower values generate more candidates, increasing sensitivity for low-VAF variants. Higher values reduce candidates, potentially missing real het variants with skewed allele balance. |
vsc_min_fraction_indels | 0.0-1.0 | 0.12 | Same as above but for indel candidates. Indels often have lower allele fractions due to alignment artifacts. |
vsc_min_count_snps | 0-50 | 2 | Minimum absolute number of reads supporting an alternate allele for SNP candidates. Works alongside vsc_min_fraction_snps — both must be satisfied. |
vsc_min_count_indels | 0-50 | 2 | Same as above but for indels. |
Read Quality Filters (make_examples)
| Parameter | Range | Default | Description |
|---|
min_mapping_quality | 0-60 | 5 | Minimum mapping quality to include a read. Reads with mapQ below this are excluded from pileup images. Higher values ensure confidently mapped reads, reducing noise in the pileup image. Lower values include more reads. The default of 5 is very permissive — the CNN can typically handle some ambiguously mapped reads. |
min_base_quality | 0-50 | 10 | Minimum base quality for counting alternate allele support. Bases below this quality are not counted toward the vsc_min_count and vsc_min_fraction thresholds (but may still appear in the pileup image). Higher values raise the bar for candidate generation. |
Read Processing (make_examples)
| Parameter | Values | Default | Description |
|---|
realign_reads | true/false | true | Perform local realignment of reads before constructing pileup images. Realignment corrects misalignments around indels, improving the pileup image quality. Disabling saves compute time but may reduce accuracy, especially for indels. Reads longer than 500bp are never realigned regardless. |
normalize_reads | true/false | false | Left-align indels in each read within the allele counter. Can improve consistency of indel representation across reads. |
keep_duplicates | true/false | false | Include reads marked as PCR duplicates. By default, duplicates are excluded. Enabling increases read depth (more data for the CNN) but introduces PCR bias. |
max_reads_per_partition | 100-5000 | 1500 | Maximum number of reads per genomic partition. At high-coverage sites, reads are downsampled to this count. The DP and AD values in the output VCF are capped by this value. |
Haplotype-Aware Calling (make_examples)
| Parameter | Values | Default | Description |
|---|
sort_by_haplotypes | true/false | false | Sort reads by haplotype (HP tag) in the pileup image. When enabled, reads from the same haplotype are grouped together, making it easier for the CNN to see variant phasing patterns. Requires either a pre-phased BAM or phase_reads=true. |
phase_reads | true/false | false | Compute read phases on-the-fly and assign HP tags. Enables haplotype-aware calling without requiring pre-phased input. The DeepVariant 1.5.0 WGS model was trained with phasing enabled. This adds compute time. |
Post-Processing (postprocess_variants)
| Parameter | Range | Default | Description |
|---|
qual_filter | 0.0-50.0 | 1.0 | QUAL score threshold below which variants are marked as FILTERED (not PASS) in the output VCF. Higher values filter out more low-confidence calls, potentially improving precision. Lower values keep more calls, improving recall. |
multi_allelic_qual_filter | 0.0-50.0 | 1.0 | Separate QUAL filter threshold for multi-allelic sites. These sites are inherently harder to call correctly. |
cnn_homref_call_min_gq | 0.0-50.0 | 20.0 | Minimum genotype quality for homozygous-reference calls from the CNN. Calls with GQ below this are given an uncertain ./. genotype. Affects reference confidence, not variant calls directly. |
use_multiallelic_model | true/false | false | Use a specialized model for resolving multi-allelic genotypes (sites with 2+ alternate alleles). Enabling may improve accuracy at complex multi-allelic sites. |
