Preprint calibrates AlphaGenome and SpliceAI2 scores as graded evidence for variant classification across coding and non-coding sites

A preprint posted to bioRxiv describes the calibration of two computational variant effect prediction scores — AlphaGenome Variant Impact (AVI) and SpliceAI2 — for use as graded evidence within the ACMG/AMP variant classification framework, specifically as PP3 (pathogenic computational evidence) and BP4 (benign computational evidence) criteria.

The authors calibrated AVI across missense variants and six classes of non-coding and synonymous variants, and calibrated SpliceAI2 across the same six classes. For missense variants, AVI achieved evidence strength falling between moderate and strong in both pathogenic and benign directions, though it did not outperform its component model, AlphaMissense. For variants adjacent to exons — including those at canonical splice sites and nearby intronic positions — AVI reached strong pathogenic evidence. At proximal intronic and synonymous sites, moderate evidence was the highest estimable level.

The calibration framework follows previously established methodologies for translating computational scores into quantitative likelihood ratios, enabling structured integration into variant classification workflows. The work has direct relevance for clinical and research variant interpretation, particularly for the large proportion of variants in classes — non-coding, synonymous — that have historically been difficult to classify using experimental evidence alone. The paper is a preprint and has not yet undergone peer review; calibration estimates should be treated as preliminary until validated independently.