Investigators calibrated the AlphaGenome Variant Impact (AVI) score for missense variants and reported evidence strength ranging from moderate to strong in both pathogenic and benign directions. The study team observed that AVI did not outperform its component score, AlphaMissense. Short, direct statement: AVI can reach substantial evidence levels for missense changes. Longer context: the authors set calibrated thresholds that map AVI values to ACMG-style PP3/BP4 evidence levels, and those thresholds produced classifications that agreed with later ClinVar updates where available.
The abstract did not specify numeric threshold values or the exact calibration procedure in this summary; the preprint and supplementary files contain those methodological details. The reader should consult the full text for precise cut points, dataset composition, and any bootstrapping or cross-validation steps the authors used.
The authors divided non-coding and synonymous variants into six classes and calibrated both AVI and SpliceAI2 across those categories. AVI achieved strong pathogenic-level evidence in the three classes immediately adjacent to exons. It reached moderate evidence—the highest estimable level reported—in proximal introns and at other synonymous sites. At deep intronic positions AVI produced only supporting-level evidence, and that supporting threshold was described as unstable by the authors.
SpliceAI2 produced a similar pattern but with important differences. The study team reported that SpliceAI2 reached the same calibrated levels for most classes but scored one evidence level lower at near-canonical positions, and it failed to reach any calibrated level in deep introns. AVI, in contrast, did register supporting-level evidence there, though at an unstable threshold. The authors also compared discrimination performance: across four of the six classes AVI and SpliceAI2 did not differ detectably in their ability to discriminate variants. In the remaining two classes, and in an independent splicing assay, SpliceAI2 showed better discrimination and greater sensitivity at 99% specificity in four classes.
The abstract does not enumerate the six classes by exact nucleotide boundaries or provide per-class numeric performance metrics; users should refer to the full manuscript for those specifics.
The calibration used two benign reference sets in parallel: ClinVar likely benign classifications and population variants. When ClinVar classifications served as the benign comparator, the authors reported that none of the reached evidence levels subsequently failed validation on later ClinVar classifications. That validation statement strengthens confidence in the mapping between score ranges and PP3/BP4 evidence levels, within the limits of available ClinVar data.
However, the cohort showed some data gaps: 14 of 30 benign evidence levels were not tested because of insufficient follow-up classifications. The authors noted a striking observation about ClinVar likely benign calls—a roughly 60-fold change in their counts at 20 nucleotides from the exon—without a concurrent change in measured splicing. That suggests potential annotation or submission biases in ClinVar that can affect calibration if ClinVar is the sole benign reference.
Using population variants as controls shifted thresholds. With population controls, both AVI and SpliceAI2 reached supporting and moderate pathogenic evidence in all six classes, but threshold strictness varied: thresholds were stricter (i.e., required stronger scores) near exons and at other synonymous sites, and lower thresholds sufficed in proximal and deep introns. The abstract does not give the numeric cutoffs for these shifts; see the supplementary materials for precise threshold values and how population allele-frequency filters were applied.
The authors disclosed that some benign evidence levels could not be evaluated because of limited later ClinVar classifications (14 of 30 benign levels untested). The abstract does not report dataset sizes, precise performance statistics, or p-values in this summary. It also does not provide the exact numeric thresholds mapping scores to PP3/BP4 levels; those are in the full preprint and supplementary files.
Open questions that remain in the preprint include the stability of supporting-level thresholds in deep introns for AVI, the clinical consequences of calibrating to ClinVar versus population variants, and how these calibrated thresholds perform across diverse gene sets or clinical testing panels—issues the authors acknowledge require cautious interpretation. The source did not claim peer review certification; it is a preprint. For users implementing these calibrated thresholds clinically, the manuscript recommends consulting the complete methods, validation tables, and supplementary materials before applying thresholds to variant interpretation workflows.