HeartVar
Aggregated evidence and a transparent ACMG/AMP classification for cardiovascular variants — with optional AI-assisted interpretation. Enter a gene and variant to begin.
HeartVar assists variant curators in generating a preliminary variant classification for variants identified in cardiovascular disease cohorts. For a given gene and variant, it queries twenty public data sources and applies the ACMG/AMP framework to produce a structured first-pass interpretation. It is designed as a curator's aid and all outputs should be independently verified.
AI interpretation is optional and off by default. Tick Include AI interpretation under the search box to enable it for a curation. Without AI, HeartVar queries every linked database and returns a preliminary, rule-based ACMG/AMP classification computed from the criteria it can evaluate deterministically, alongside the full assembled evidence. The judgement-based criteria are marked "not assessed", so the preliminary tier can understate pathogenicity but does not over-call it from missing evidence. Adding AI unlocks the full ACMG/AMP classification, point score, and clinical-narrative summary. Adding AI interpretation requires single sign-on to safely manage AI usage. The rest of the tool remains available for use without signing in.
When you enable AI, HeartVar runs the interpretation through Anthropic's Claude. This means the variant and the clinical context you enter (and any questions you ask the chat assistant) are sent to Anthropic's Claude API, a third-party service, for processing. If you would rather no data leave HeartVar, leave the box unticked and the evidence-only result is computed entirely on the server with no third-party AI call. To keep the shared AI resource available, usage is rate-limited and capped by a daily budget; if the cap is reached, HeartVar falls back to the evidence-only result for the rest of the day.
Nothing you enter is stored. The variant, phenotype, and clinical details you enter are held only in your browser for the session and processed in memory on the server to assemble evidence. They are not written to a database, logged with their content, or retained after the request completes.
If you tick Include AI interpretation, data is sent overseas. The variant, the clinical context you entered, and any questions you ask the assistant are transmitted to Anthropic's Claude API, which processes them in the United States. If you would rather nothing leaves the Institute's servers, leave the box unticked. The evidence-only result is computed entirely on our server with no third-party AI call.
Please do not enter information that identifies a patient. HeartVar needs only a gene, a variant, and a phenotype. Do not enter names, medical record numbers, dates of birth, or free-text notes that could identify an individual.
Cookies. HeartVar sets one cookie, and only if you sign in: a session cookie that keeps you signed in. HeartVar sets no advertising or analytics cookies and contains no tracking code.
Signing in shares your email address and display name from Microsoft or Google, used only to identify your session and to apply usage limits.
For everything else, including how to request access to or correction of your information, see the Institute's privacy policy and terms & conditions.
| Database | What it contributes |
|---|---|
| AlphaFold | Predicted 3D protein structure (pLDDT-scored) for the spatial variant view |
| AlphaMissense | Pre-computed missense pathogenicity predictions |
| BioGRID | Curated protein-protein interactions |
| CHDgene | VCCRI curated high-confidence CHD gene list with cardiac subtype and inheritance data |
| ClinVar | Existing pathogenicity classifications and submitter evidence |
| Ensembl VEP | Variant consequence, transcript annotation, splice impact, in silico scores (SIFT, PolyPhen, REVEL, CADD) |
| GenCC | Gene-disease classifications with submitter and mode of inheritance |
| gnomAD v4 | Population allele frequency, homozygote counts, gene constraint metrics |
| GTEx v10 | Median TPM in heart (left ventricle, atrial appendage) and artery (aorta, coronary) tissues |
| Heart of Fetal Cells | Farah et al. 2024 ("Heart of Cells", Nature 627:854); per-cell-type fetal-heart scRNA-seq expression at 9/11/13/15 PCW for congenital heart disease interpretation |
| MedGen | Gene-associated conditions and disease names (NCBI) |
| MGI (via Alliance) | Mouse orthologue, phenotypic alleles, cardiac MP terms |
| Open Targets | Gene-disease association score and evidence-type breakdown |
| PanelApp (Australia) | Australian Genomics; all cardiovascular panels |
| PMC Open Access | Open-access full text (Results/Methods/table excerpts) deepening literature based criteria |
| ProtVar (EBI) | Functional, conservation, and co-located variant annotations for missense variants |
| PubMed | Variant-specific case reports and gene-level functional study abstracts |
| PubTator3 | Normalised variant→literature retrieval recovering variant-specific papers PubMed's strict query misses |
| SpliceAI | Δ-acceptor/Δ-donor splice impact scores |
| UniProt | Protein domains, active and binding sites, natural variant annotations |
Data is updated monthly.
HeartVar reports on the GRCh38/hg38 assembly. You can enter a variant either as coding HGVS (e.g. c.886C>T, with a gene symbol) or as genomic coordinates (chromosome-position-ref-alt). HGVS input is build-independent and Ensembl resolves it directly to GRCh38. For genomic coordinates, both GRCh38/hg38 (default) and GRCh37/hg19 are accepted; a build selector appears beneath the search box as soon as you enter coordinates.
GRCh37/hg19 coordinates are automatically lifted over to GRCh38/hg38 (via Ensembl) before any annotation, so all evidence and the final classification are computed on GRCh38. When a conversion is applied, the Summary tab shows both the coordinates you entered and the lifted GRCh38 coordinates.
HeartVar applies the ACMG/AMP framework (Richards, 2015) for criteria definitions with the point-based combining rules (Tavtigian, 2020) for the final classification tier.
If you use HeartVar in your work, please cite:
Thompson, J.M., Das, D., Dunwoodie, S.L., & Giannoulatou, E. HeartVar: A LLM-Assisted Tool for Clinical Classification of Variants in Cardiovascular Disease Cohorts. [Journal / Preprint]. 2026. [DOI / URL].
HeartVar is developed by the Computational Genomics Laboratory at the Victor Chang Cardiac Research Institute, Sydney, as part of a project supported by Anthropic's AI for Science program.
For questions, feedback, or a bug report, email heartvar@victorchang.edu.au.