Summary
A bioRxiv preprint reports a ClinVar-derived AlphaGenome Variant Impact quantile cutoff of 0.9938 for prioritising single-nucleotide variants. Nearly half of the uncertain variants assessed exceeded that threshold, but the result is a prioritisation signal, not a clinical reclassification.
A bioRxiv preprint by Hui-Qi Qu and Hakon Hakonarson of Children's Hospital of Philadelphia reports a way to use AlphaGenome’s Variant Impact score to prioritise human single-nucleotide variants (SNVs). Comparing score ranks with ClinVar classifications, the authors derived a global cutoff of 0.9938. Of 24,989 variants of uncertain significance (VUS) they assessed, 11,357—45.45%—scored above it.
The preprint, posted on October 1, 2026, describes the cutoff as a way to flag variants for further evaluation. A VUS remains uncertain under its existing classification; crossing the score threshold does not itself establish that a variant causes disease.
How the score was compared with variant classifications
An SNV is a change to a single DNA letter. ClinVar records interpretations of genetic variants, including categories such as pathogenic, likely pathogenic, uncertain, conflicting, likely benign and benign. AlphaGenome Atlas provides functional predictions for nearly all possible human SNVs, using an AlphaGenome Variant Impact (AVI) score. The study examines whether the score’s quantile rank, called AVI_QUANTILE, can help distinguish variants with different ClinVar classifications.
The authors used receiver operating characteristic (ROC) analyses to derive thresholds. ROC analysis assesses how well a measure separates groups as a decision threshold changes; the Youden-optimal cutoff is a statistical choice based on the resulting sensitivity and specificity. The preprint reports 0.9938 as the global Youden-optimal cutoff for ClinVar-derived pathogenic and likely pathogenic variants. In other words, the threshold lies at the 99.38th percentile of the AVI_QUANTILE distribution.
The authors found pathogenic and likely pathogenic variants strongly enriched near the upper end of that distribution. Cutoffs were generally similar across several major molecular consequence classes, though some noncoding categories had lower thresholds. For most well-powered gene–disease groups, the estimates were also near the global cutoff, with modest variation between genes.
What the VUS result means
The 11,357 VUS above the global cutoff had AlphaGenome-predicted functional effects in a range the authors describe as comparable to those observed among pathogenic and likely pathogenic variants. That makes the score potentially useful for prioritising which uncertain variants deserve further investigation. It does not convert those variants into pathogenic findings: the analysis uses a computational prediction alongside existing classifications, rather than reporting new clinical evidence for each variant.
The study also compared AVI_QUANTILE with SuSiE posterior inclusion probabilities in a fine-mapping analysis of steroid-dependent asthma. Fine-mapping uses genetic association data to narrow a set of variants that may be responsible for a trait; a posterior inclusion probability represents a model-based estimate for a variant’s inclusion as causal. The authors found little correspondence between that measure and AVI_QUANTILE, suggesting the functional predictions capture information distinct from statistical fine-mapping.
Together, the analyses position AVI_QUANTILE as a possible complement to other methods for variant prioritisation, rather than a replacement for clinical interpretation. The findings are reported in a preprint and the cutoff is derived from ClinVar classifications, so its practical use depends on treating it as a signal for follow-up rather than a standalone diagnosis.