Preprint calibrates CADD v1.7 and AlphaGenome scores for ACMG/AMP variant classification
A bioRxiv preprint following ClinGen methodology calibrates two computational variant predictors across missense variants, finding both provide moderate evidence for pathogenicity when appropriately thresholded.
A preprint deposited on bioRxiv on 2 November 2026 reports a systematic calibration of two widely used computational variant effect predictors — CADD version 1.7 and the AlphaGenome Variant Impact (AVI) score — for use as supporting evidence under the ACMG/AMP variant classification framework.
Clinical and research laboratories increasingly apply computational predictors as PP3 (pathogenicity-supporting) or BP4 (benign-supporting) evidence when classifying variants of uncertain significance (VUS), but the accuracy of this practice depends on the score thresholds used being properly calibrated against variants of known clinical significance. The authors note that CADD is commonly applied with a single genome-wide cutoff irrespective of variant type or model version — an approach the paper argues is imprecise — and that AVI, introduced alongside the AlphaGenome model, had not previously been calibrated for ACMG/AMP use.
Following a prespecified analytical plan and the ClinGen calibration procedure, the authors evaluated both predictors on rare ClinVar missense variants with established pathogenicity classifications. Both CADD v1.7 and AVI yielded moderate evidence of pathogenicity at calibrated thresholds, with performance varying across variant classes.
This work is directly relevant to genetic-counselling and laboratory practice in variant interpretation, and complements a cluster of related calibration preprints published in late October 2026. The manuscript has not yet undergone peer review.
Sources
Read the original reporting — these are the public sources this summary draws from.
-
Primary sourcePreprint bioRxiv (Cold Spring Harbor Laboratory) · 2026-11-02Beyond a single threshold: calibration of CADD v1.7 and AlphaGenome for ACMG/AMP computational evidence across variant classes, model versions, and reference data