AlphaGenome informed machine learning for pathogenicity prediction of germline mutations
Accurate interpretation of germline variants remains a central challenge in clinical genomics. Existing in silico pathogenicity predictors often provide incomplete coverage or restricted to missense variants from the coding region, leaving the variants from splice sites, -UTR, and intronic regions under-scored. We present AGVarPred, a gradient-boosted classifier designed to predict the pathogenicity of germline mutants, with a broad consequence-class coverage across SNVs and indels. Here, the population allele frequency from gnomAD, functional consequence annotations from Ensembl VEP, and cell-type-specific functional genomic scores from the AlphaGenome platform were integrated to develop the classifier. Using a gene-level split of 134,002 high-confidence ClinVar variants across 9,222 genes, AGVarPred achieved an AUC-ROC of 0.9753 (95% CI 0.9731-0.9775) and a AUC-PR of 0.9606 on 12,385 scored internal test variants. A paired DeLong analysis showed that the full model significantly outperformed an allele-frequency-plus-VEP model ({Delta}AUC = 0.1539, 95% CI 0.1489-0.1589; p < 1 x 10-300), supporting a substantial incremental contribution from AlphaGenome-derived features. Performance remained strong when restricted to SNVs (AUC-ROC 0.9558) and was also high for indels (AUC-ROC 0.9532). On five external benchmarks, AUC-ROCs ranged from 0.528 on a functional-assay-derived benchmark to 0.977 on a clinically curated deafness database, reflecting its dependence on benchmark composition and label source. A separate analysis of the VIP Database, whose pathogenic labels are ClinVar-derived, showed consistent performance (AUC-ROC of 0.892) in held-out genes. AGVarPred scored 100% of evaluated variants in the external benchmarks, whereas AlphaMissense and REVEL covered only small fractions of non-missense variants. Model interpretation indicated that gnomAD allele frequency was the dominant individual predictor; the 111 AlphaGenome features provided complementary functional signal, with splicing and chromatin assays contributing to the largest share of the AlphaGenome SHAP mass. A web-based interface for AGVarPred is available at https://pankaj357.github.io/AGVarPred/ to facilitate the submission of normalized GRCh38 germline variants and retrieval of pathogenicity predictions using the frozen AGVarPred model.