How well does the AlphaGenome Variant Impact (AVI) score separate pathogenic from benign human genetic variants? An independent evaluation on ClinVar

Background. The AlphaGenome Atlas provides a precomputed AlphaGenome Variant Impact (AVI) score for every possible single-nucleotide variant (SNV) in the human genome. Its developers report strong performance for separating pathogenic from benign ClinVar variants. Aim. We evaluated AVI independently on a reproducible, pre-specified sample of ClinVar SNVs. Methods. From the ClinVar GRCh38 VCF (file dated 28 September 2026) we sampled 3,228 SNVs with ≥2-star review status (1,337 pathogenic or likely pathogenic, 1,891 benign or likely benign), stratified by molecular consequence and capped at 10 variants per gene per class and category, retrieved their AVI scores from the Atlas API, and computed ROC-AUC and PR-AUC with gene-clustered bootstrap 95% confidence intervals. Results. In this sample, pooled ROC-AUC was 0.905 (0.892–0.917) and PR-AUC 0.817 (0.784–0.848); a fixed consequence-type ranking alone reached 0.644. The macro-average over categories with at least 30 variants of each class was 0.895 (0.867–0.918). ROC-AUC was highest for missense (0.963), synonymous (0.975) and intronic (0.980) variants, and lower for canonical splice sites (0.844), nonsense/start/stop variants (0.836) and 5′ UTR variants (0.774). In the synonymous and intronic categories, most pathogenic variants lay close to exon boundaries, so those categories are probably easier than a random sample of such variants would be. Conclusions. AVI separated the two ClinVar classes well in this sample, but ClinVar-specific confounds and possible information overlap with ClinVar (which we cannot exclude) limit what can be concluded; the results are not evidence of clinical utility. Research use only. This work makes no clinical or diagnostic claims and is not medical advice. Not peer reviewed. Independent researcher, no institutional affiliation. AI tools (Claude, Anthropic) assisted with code and drafting; the author reviewed the work and takes responsibility for its content. AlphaGenome outputs reproduced here (aggregate statistics and figures) are subject to the AlphaGenome Output Terms of Use: https://deepmind.google.com/science/alphagenome/output-terms . Not affiliated with or endorsed by Google or Google DeepMind. Code and pre-registered design: https://github.com/Arths17/avi-clinvar-eval