HeartVar is a research tool for preliminary ACMG/AMP classification of variants found in cardiovascular disease cohorts. Given a gene, variant, and clinical context, it queries around 20 public databases in parallel and assembles the evidence into a per-criterion view. Classification is hybrid. Nineteen criteria are decided programmatically in Python from structured data; seven that require interpretation are decided by a single LLM call, which is optional and off by default.
| Entry | What it is |
|---|---|
backend/ |
The FastAPI service, the ACMG engine (acmg/), the per-source clients (clients/), shipped reference JSON (data/) and the test suite (tests/). |
static/ |
Frontend CSS and JavaScript, the vendored 3Dmol viewer, and acmg_constants.json. |
index.html |
The whole single-page frontend. |
scripts/ |
Deploy-host tooling that builds the local data caches. |
Dockerfile |
Runtime image. Dockerfile.builder is the data-mirror build job. |
.env.example |
Every environment variable. |
The code is MIT, see LICENSE. The data the build scripts fetch is not covered by that licence and several sources carry their own terms. The full per-source list is in scripts/README.md.
HeartVar: An LLM-Assisted Tool for Clinical Classification of Variants in Cardiovascular Disease Cohorts. Jamie-Lee Thompson, Debjani Das, Sally L Dunwoodie, Eleni Giannoulatou. bioRxiv 2026.09.10.750569; doi: https://doi.org/10.64898/2026.09.10.750569
Supported by Anthropic's AI for Science program. Built on public resources from Ensembl/EMBL-EBI, the Broad Institute (gnomAD, SpliceAI), Google DeepMind (AlphaMissense), NCBI (ClinVar, PubMed, PubTator3, MedGen), UniProt, EBI ProtVar, GTEx, PanelApp Australia, GenCC, the Alliance of Genome Resources/MGI, BioGRID, Open Targets, ClinGen (eRepo) and AlphaFold. The CHDgene gene list is maintained by Victor Chang Cardiac Research Institute.