I (with mentors' deep supervision & support) built VarViz to stop copying variant positions between browser tabs. Interpreting a panel of missense variants means querying gnomAD, ClinVar, AlphaFold, AlphaMissense, PhyloP, and ClinGen separately for each variant, then manually collating the evidence before scoring. VarViz does all of that in one session.

Reason to build it: We kept running into the same problem in variant curation work: no single tool shows the full protein context for a panel of variants, and none apply automated ACMG/AMP classification at the same time. So built one.

What it does: Nine interactive evidence tracks aligned to the protein sequence: AlphaFold pLDDT, gnomAD v4 allele frequency, ClinVar/gnomAD mutation density, AlphaMissense mean pathogenicity per residue, ClinVar tick marks + PTM sites + CCRS constrained regions, ConSurf conservation (upload a grades file), PhyloP 100-way and 470-way + PhastCons, and UniProt domain architecture. All tracks share a common amino acid x-axis, so spatial patterns across evidence layers are immediately visible.

ACMG/AMP engine: Implements Richards et al. (2015) combinatorial rules first, then Tavtigian et al. (2020) Bayesian point scoring when no rule matches. PP3/BP4 uses the calibrated thresholds from Pejaver et al. (2022) across seven predictors. PM2 uses a prevalence-derived maximum credible AF. Per-variant dropdowns for cosegregation (PP1), de novo status (PS2/PM6), and compound heterozygosity (PM3) update the classification and plain-language comment in real time. Three evidence-independence guards prevent double-dipping.

Validation: Applied to 24 CASR missense variants with functionally validated loss-of-function (FHH1) and gain-of-function (ADH1) mechanisms from a cohort of 51,289 individuals. Sensitivity 93% for LOF variants (14/15 P/LP; Wilson 95% CI 70-99%). The protein landscape showed spatial separation between LOF and GOF variants in the AlphaMissense track, with a mean of 0.83 vs 0.39, a pattern not recoverable from any individual predictor score.

Exports: TSV with all ACMG tags, 40+ annotation columns, and analysis parameters (any classification is exactly reproducible from the file); gnomAD raw data TSV; publication-quality PDF/PNG/JPEG via cowplot.

Try it: https://varviz.org

(no install, no login, paste variants in p-notation and click Go. Results within 60 seconds.)

Code: https://github.com/metviz/varviz

Disclaimer: For research use only.

What would help most: test it on a gene and variant panel you know well and let me know what breaks, what is missing, or where the ACMG scoring diverges from your expectation. Open a GitHub issue or reply here.

Thanks, Happy to answer questions.