Family-aware germline variant caller that genotypes a child and both parents from one stacked read pileup, extending DeepVariant to trios and duos.
No providers recorded yet. Browse all providers
A child's genome is almost entirely a recombination of its parents' genomes, so a trio carries redundant evidence: a weakly supported heterozygous call in the child becomes far more credible when one parent clearly carries the same allele. Classical family-aware callers exploit this with explicit Mendelian transmission priors layered on per-sample calls, which must also allow for sequencing error, mapping error and the rare de novo mutations that violate inheritance.
DeepTrio, developed at Google, drops the hand-coded genetics. It extends DeepVariant so that the reads from a child and one or two parents are stacked into a single pileup tensor, and a convolutional network genotypes the site from all of them at once. Mendelian inheritance is never given to the model; it learns from Genome in a Bottle truth labels how much parental evidence should shift a call, and how to recognise a de novo variant that has no support in either parent. DeepTrio shipped with pretrained models in DeepVariant v1.1 in December 2020 and was described in an April 2021 bioRxiv preprint, revised in 2026 with an improved de novo method and additional sequencing technologies; it has not been peer reviewed.
Two models are trained per sequencing platform: a child model, which reads the middle rows of the pileup as primary evidence, and a parent model, which calls whichever parent is placed in the top rows. Users run fixed checkpoints on new families without retraining.
The v1.10 Illumina WGS child model takes a 140 × 221 pileup with seven channels (read base, base quality, mapping quality, strand, variant support, mismatch to reference and fragment length); the PacBio models use nine channels, and the ONT models a taller 300-row pileup. The v1.10 WGS child checkpoint holds about 22 million parameters. Models are trained on Genome in a Bottle samples HG001–HG007, NA12891 and NA12892, sequenced on NovaSeq, HiSeq X, HiSeq 4000 and PacBio Sequel II with PCR-free and PCR-positive preparations. Training uses chromosomes 1–19, selects checkpoints on chromosomes 21–22 and holds out chromosome 20 for evaluation. The 50× BAMs are downsampled to 35×, 25× and 15×, and extra examples downsample only the parents, so the models generalise across coverage and uneven family depths. The current WGS parent model trains on about 457 million examples.
On the HG002–HG003–HG004 trio, DeepTrio outperforms DeepVariant, GATK4 with and without CalculateGenotypePosteriors, dv-trio and Octopus on both Illumina and PacBio HiFi, with the largest margin at low coverage: at 20× it roughly matches DeepVariant at 28–30×, and at 15× it beats both GATK4 pipelines at 35×. The v1.10 release reports chromosome 20 F1 on HG002 of 0.9980 (SNP) and 0.9970 (indel) for Illumina WGS, and 0.9997 and 0.9963 for PacBio HiFi. For de novo variants, DeepTrio has higher recall than the other trio-aware callers, though slightly lower precision than Octopus at 35×.
DeepTrio is built for rare-disease genomics and other family studies in which the proband and parents are sequenced together, including dominant disorders caused by de novo variants and recessive disorders requiring a variant from each parent. Its robustness at reduced depth lets studies sequence parents more shallowly than the child. Per-sample gVCFs can be merged across many families with GLnexus to build joint cohort callsets. It runs from the deeptrio Docker image on CPUs or GPUs.
DeepTrio showed that relationships between samples can be learned from data rather than encoded as statistical priors, and that exposing biologically related evidence to a deep variant caller measurably improves it. Its limits are explicit. On Oxford Nanopore data DeepVariant still calls indels more accurately, because ONT's systematic indel errors recur across family members and reinforce false evidence. Trio-aware calling also trades some de novo recall for precision, so the authors suggest pairing it with single-sample DeepVariant when maximal de novo sensitivity matters. Like DeepVariant, it calls only diploid germline SNPs and small indels.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.