Preprint warns that variant effect predictors are shaping the very data used to train them
A bioRxiv preprint identifies a circularity problem in clinical variant classification: computational predictors are influencing ClinVar entries that are then used to benchmark those same tools.
Researchers have posted a preprint to bioRxiv describing a classification-level circularity in the variant interpretation pipeline. Variant effect predictors (VEPs) — computational tools that estimate the pathogenicity of DNA variants — have, following recent calibration work and the adoption of points-based ACMG/AMP classification standards, accumulated enough evidential weight to shift variants out of uncertain significance (VUS status) with limited supporting data. The authors show that this creates a feedback loop: VEP predictions help determine how variants are classified in ClinVar, and those ClinVar classifications are subsequently used to train, benchmark, and calibrate the same predictors.
The finding has direct relevance to the reliability of computational variant interpretation. If VEPs are evaluated against labels they have themselves helped to generate, measures of their performance — sensitivity, specificity, calibration — may be inflated in ways that are difficult to detect from standard benchmarking alone. The preprint does not argue that VEPs are unhelpful, but calls for greater methodological awareness of how training and evaluation data are assembled, and whether computational evidence has been stripped out before benchmarks are constructed.
This work sits alongside recently published preprints calibrating AlphaGenome and SpliceAI2 scores for variant classification (covered in Genetic Current, 25 October 2026) and represents a cautionary complement to that literature. The preprint has not yet been peer-reviewed.
Plain-language version
For patients, families, and general readers. Educational only — not medical advice.
When scientists try to work out whether a change in someone's DNA is harmful or harmless, they often use computer programs to help make that judgement. A new preliminary study — not yet checked by independent reviewers — suggests there may be a problem with how these programs are tested. The programs can influence the official databases used to label DNA changes as likely harmful or harmless. Those same databases are then used to check how good the programs are. This could mean the programs appear more accurate than they really are. The researchers are calling for better methods to keep these two things separate. This is an educational summary, not medical advice. If anything here raises questions for you, please speak with your GP or a clinical professional.
Sources
Read the original reporting — these are the public sources this summary draws from.
-
Primary sourcePreprint bioRxiv (Cold Spring Harbor Laboratory) · 2026-10-27Variant effect predictors are shaping the ClinVar classifications used to train and evaluate them