Glossary

DeepVariant

A deep convolutional neural network developed by Google that transforms the task of variant calling into an image classification problem by encoding aligned sequencing reads into pileup images.

Get in touch Learn more

Product manager reviewing autonomous task execution dashboard on laptop, completed tasks visible, casual work session.

VARIANT CALLING ARCHITECTURE

What is DeepVariant?

A deep convolutional neural network developed by Google that transforms the task of variant calling into an image classification problem by encoding aligned sequencing reads into pileup images.

DeepVariant is a deep convolutional neural network that reframes germline variant calling as an image classification task. Instead of relying on hand-crafted statistical models, it converts aligned reads from a BAM file into a multi-channel pileup image encoding—a tensor where red, green, and blue channels represent base identities, quality scores, and strand information at a candidate locus. This image is then processed by an Inception-v2 architecture to classify genotypes.

Trained on gold-standard Genome in a Bottle (GIAB) truth sets, DeepVariant learns to distinguish true single nucleotide polymorphisms and indels from sequencing artifacts like strand bias and homopolymer indel errors directly from data. The model outputs diploid genotype likelihoods for each candidate variant, achieving state-of-the-art accuracy without requiring manual feature engineering or Base Quality Score Recalibration preprocessing steps.

ARCHITECTURAL INNOVATIONS

Key Features of DeepVariant

DeepVariant reimagines variant calling as an image classification problem, leveraging a deep convolutional neural network to identify genetic variants from pileup images with state-of-the-art accuracy.

Pileup Image Encoding

Transforms aligned sequencing reads into a multi-channel RGB image at each candidate variant locus. The encoding maps reference bases, read bases, quality scores, and strand information into distinct color channels, allowing the CNN to learn visual features associated with true variants versus sequencing artifacts. This representation captures mapping quality, base quality, and allele balance in a spatially structured format.

Input Channels

221×100

Image Dimensions (px)

Inception-Based CNN Architecture

Employs a modified Inception v2 convolutional neural network trained to classify pileup images into three genotype states: homozygous reference, heterozygous variant, and homozygous alternate. The architecture uses multi-scale convolutional filters to capture both local read pileup patterns and broader genomic context, enabling robust discrimination of true variants from systematic errors like strand bias artifacts and homopolymer indel errors.

EXPLORE

Candidate Variant Generation

Pre-filters genomic loci to identify positions with evidence of variation before neural network classification. Uses a simple heuristic caller to scan aligned reads for:

Mismatched bases relative to the reference
Insertion or deletion evidence in CIGAR strings
Read depth anomalies suggesting copy number changes This two-stage approach dramatically reduces computational cost by limiting CNN inference to only candidate sites.

Allele-Specific Read Partitioning

Separates sequencing reads into allele-supporting groups before pileup image generation. Reads are assigned to reference or alternate allele bins based on alignment evidence, enabling the model to visualize allele balance patterns. This partitioning is critical for detecting heterozygous variants where the variant allele fraction (VAF) should approximate 50%, distinguishing true heterozygous calls from low-frequency sequencing errors.

Multi-Class Genotype Likelihoods

Outputs calibrated posterior probabilities for all three possible diploid genotype states (hom-ref, het, hom-alt) rather than a simple variant/no-variant binary. These probabilities serve as diploid genotype likelihoods that can be directly integrated into downstream joint genotyping and haplotype phasing pipelines. The softmax-normalized outputs provide well-calibrated confidence scores suitable for Variant Quality Score Recalibration (VQSR).

Genotype Classes

99.9%+

GIAB Concordance

Transfer Learning Across Sequencing Platforms

Demonstrates robust generalization by training on Genome in a Bottle (GIAB) truth sets and transferring to diverse sequencing chemistries including Illumina, Pacific Biosciences, and Oxford Nanopore. The image-based representation abstracts away platform-specific error modes, allowing a single architecture to achieve state-of-the-art accuracy across short-read and long-read technologies without platform-specific retuning.

DEEPVARIANT INSIGHTS

Frequently Asked Questions

Explore the core concepts behind Google's deep learning-based variant caller, from its unique image-based approach to its practical deployment in clinical genomics pipelines.

DeepVariant is a deep convolutional neural network (CNN) developed by Google that transforms the task of variant calling into an image classification problem. Instead of relying on hand-crafted statistical models, it encodes aligned sequencing reads around a candidate variant site into a multi-channel pileup image. This RGB-like tensor represents reference bases, alternate alleles, base quality scores, mapping qualities, and strand information as distinct visual features. The trained CNN then classifies the image to determine the genotype—homozygous reference, heterozygous, or homozygous alternate. By leveraging transfer learning from the Inception architecture, DeepVariant learns complex, non-linear dependencies in the data that traditional Bayesian methods often miss, achieving state-of-the-art accuracy in benchmarks like the Genome in a Bottle (GIAB) competition.

Enabling Efficiency, Speed & Accuracy

Intelligent Analysis, Decision & Execution

We build AI systems for teams that need search across company data, workflow automation across tools, or AI features inside products and internal software.

Talk to Us

Search across company data

Give teams answers from docs, tickets, runbooks, and product data with sources and permissions.

Useful when people spend too long searching or get different answers from different systems.

Enterprise searchRAGPermissions

Automate internal workflows

Use AI to route work, draft outputs, trigger actions, and keep approvals and logs in place.

Useful when repetitive work moves across multiple tools and teams.

AI agentsWorkflow automationGovernance

Add AI to products and internal tools

Build assistants, guided actions, or decision support into the software your team or customers already use.

Useful when AI needs to be part of the product, not a separate tool.

AI integrationDecision supportModel routing

About the author

Prasad Kumkar

CEO & MD, Inference Systems

Prasad Kumkar is the CEO & MD of Inference Systems and writes about AI systems architecture, LLM infrastructure, model serving, evaluation, and production deployment. Over 5+ years, he has worked across computer vision models, L5 autonomous vehicle systems, and LLM research, with a focus on taking complex AI ideas into real-world engineering systems.

His work and writing cover AI systems, large language models, AI agents, multimodal systems, autonomous systems, inference optimization, RAG, evaluation, and production AI engineering.

Limited slotsGet a Free AI Consultation

How We Work

Custom AI workflows for your Business

One-fit-all AI don't work for modern businesses. At Inferensys, we aim to understand your business & custom requirements; which we use to define most efficient agentic workflows, the data, and the tools for your business.

Talk to Us

DeepVariant

What is DeepVariant?