Calibration of AlphaGenome Variant Impact and SpliceAI2 scores as PP3/BP4 evidence for missense, non-coding, and synonymous variants

Abstract

Computational predictions can serve as graded evidence in variant classification after calibration. We calibrated the AlphaGenome Variant Impact score (AVI) for missense variants and for six non-coding and synonymous classes, and SpliceAI2 for the six classes. For missense variants, AVI reached evidence between moderate and strong in both directions but did not outperform its component AlphaMissense. AVI reached strong pathogenic evidence in the three classes adjacent to exons. It reached moderate evidence, the highest estimable level, in proximal introns and at other synonymous sites, and supporting evidence, at an unstable threshold, in deep introns. SpliceAI2 reached the same levels, except one level lower at near-canonical positions and none in deep introns. No reached level failed validation on later ClinVar classifications, although 14 of 30 benign levels were not tested. With population variants as controls, both tools reached supporting and moderate pathogenic evidence in all six classes, at stricter thresholds near exons and at other synonymous sites, and at lower thresholds in proximal and deep introns. ClinVar likely benign classifications changed about 60-fold in number at 20 nucleotides from the exon without a change in measured splicing. AVI, a genome-wide score, did not differ detectably in discrimination from the splicing-specific SpliceAI2 in four classes; SpliceAI2 discriminated better in two classes and in a splicing assay, and was more sensitive at 99% specificity in four classes. Both scores provide calibrated evidence for non-coding and synonymous variants, but thresholds depend on whether ClinVar benign classifications or population variants are used for comparison.