AlphaGenome informed machine learning for pathogenicity prediction of germline mutations

Gradient-boosted classifier integrating gnomAD frequency, VEP annotations and 111 AlphaGenome functional features

AGVarPred is a gradient-boosted classifier that predicts the pathogenicity of germline SNVs and indels across consequence classes, combining gnomAD allele frequency, Ensembl VEP annotations and cell-type-specific AlphaGenome functional scores. Trained and tested with a gene-level split of 134,002 high-confidence ClinVar variants across 9,222 genes, it achieved an AUC-ROC of 0.9753 and an AUC-PR of 0.9606 on 12,385 test variants, well above a frequency-plus-VEP model (AUC gain of 0.1539). On five external benchmarks, AUC-ROC ranged from 0.528 on a functional-assay-derived set to 0.977 on a clinically curated deafness database. AGVarPred scored 100% of variants in these benchmarks, whereas AlphaMissense and REVEL covered only small fractions of non-missense variants, and a web interface accepts GRCh38 variants.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

Preprint, not peer reviewed. Covering all variant classes addresses a real gap, but gnomAD frequency is the model's dominant predictor and training relies on ClinVar, whose classifications already incorporate frequency: the AUC of 0.528 on the functional benchmark is a reminder of this circularity risk. Only coverage, not performance, is compared with existing predictors; to be used as supporting information, not as a calibrated PP3 criterion.

Analysis by Dr Thibaut Benquey