VariantAgent

The problem

A clinical variant interpretation analyst processes 50-200 variants per week. Each one needs:

Allele frequency lookups across population databases (gnomAD, ExAC)
Pathogenicity evidence from ClinVar
Functional predictions (Ensembl VEP, REVEL, CADD)
Literature search for variant-disease links (PubMed)
ACMG/AMP rule application across 28 evidence criteria
A defensible written justification for the final classification

Existing automation tools cover individual steps. None of them produce a full chain-of-thought report you can defend in front of a CAP/CLIA audit.

The approach

VariantAgent is a six-agent pipeline orchestrated with LangGraph. Each agent has a narrow responsibility, deterministic tools, and a structured output contract enforced with Pydantic. The orchestrator routes dynamically — QC failures skip annotation, novel variants trigger literature search, and low-confidence classifications hit a human checkpoint.

The classification step is not an LLM call. The LLM reasons about which ACMG criteria apply; a deterministic rule engine then applies the standard combination logic. This separation is the difference between a demo and something a clinical lab can actually deploy.

The agents

Agent	Responsibility	Tools
Orchestrator	Plan and route across the other five agents	LangGraph state machine
QC	Validate input quality (coverage, depth, alignment)	MultiQC, flagstat, coverage thresholds
Annotation	Query ClinVar, gnomAD, Ensembl VEP, UniProt	MCP servers (one per database)
Literature	RAG over PubMed + ACMG guidelines	ChromaDB + sentence-transformers
Classification	Apply ACMG rules deterministically	Custom rule engine
Reviewer	Self-evaluation, contradiction detection	LLM-as-judge

Why this matters

Clinical variant interpretation is the kind of workflow where a pure LLM is dangerous and a pure rule engine is incomplete. VariantAgent is my attempt at the hybrid: LLMs read evidence, a deterministic ACMG rule engine classifies, and the whole pipeline is auditable end to end. The goal is a system an analyst could actually sign off on, not a demo.

Try it live

Try VariantAgent

Paste a variant and get an ACMG classification grounded in ClinVar and gnomAD. Simplified demo, not clinical.

live

Variant (chromosome:position:ref:alt, GRCh38)

Try:

Run it locally

Every project is open source. Clone the repo, install dependencies with uv sync, drop your API keys into .env, and run the CLI. The README walks you through each step.

git clone https://github.com/deepmind11/variantagent.git
cd variantagent
uv sync
cp .env.example .env  # add your keys
uv run python -m variantagent --help