Thousands of uncertain monogenic diabetes variants get a ranked list, and an ablation shows how much of the accuracy is curator logic rebuilt
Synopsis
Many variants found in monogenic diabetes genes remain variants of uncertain significance because evidence is insufficient or conflicting. A pipeline trained on 7,706 confidently classified single nucleotide variants across 22 such genes reached ROC-AUC 0.9989 and AUPRC 0.9970 on 771 held-out variants, with recall 1.00 on the pathogenic class. Feature ablation showed that part of this accuracy reflects overlap with the ACMG criteria used to assign ClinVar labels, yet on non-missense variants, where computational evidence could not have been directly invoked, the model still beat a majority-class baseline, 99.0% versus 87.0%; it then ranked 6,306 uncertain variants with per-variant explanations.
Interpretation
Uncertain variants are no longer queued by a single predictor score; consequence severity, population frequency, conservation and several computational predictions are combined into one interpretable pathogenicity probability, with the top three driving features recorded for each variant. ClinVar aggregates interpretations from submitting laboratories, but a large fraction of catalogued variants remain unresolved; general-purpose predictors such as SIFT, PolyPhen-2, CADD and REVEL cover the whole genome, offer limited transparency about what drives an individual prediction, and lack calibration for a specific disease gene set. The training cohort held 7,706 confidently classified single nucleotide variants, 5,690 benign-like and 2,016 pathogenic-like; across five-fold stratified cross-validation the four classifiers spanned a narrow AUPRC band of 0.9868 to 0.9890, with differences smaller than the cross-fold standard deviation.
The high scores are not all new biological signal; part of the performance reproduces the judgment a curator makes under ACMG criteria, and category-by-category ablation quantifies that share. Earlier work had documented circularity in variant-effect models trained on ClinVar labels with the same evidence categories as features; this analysis separates the effect across evidence categories. Computational-prediction and conservation features alone gave test accuracy 0.9533, consequence and frequency features alone gave 0.9105, and the full feature set gave 0.9896; on the non-missense subset (n=614), where computational evidence could not have been directly invoked, the model reached 99.0% accuracy against an 87.0% majority-class baseline.
The model was applied to rank 6,306 uncertain variants, and each ranking carries an explanation intended to support expert review rather than replace classification. These variants had no confident pathogenic or benign conclusion and remained in the uncertain category awaiting manual review. Predicted pathogenicity probabilities had mean 0.694 and median 0.873, with 47.08% above 0.9 and 72.66% above 0.5; top-ranked candidates spanned GCK, HNF1A, LMNA, WFS1, PDX1, ABCC8, INS, BLK, RFX6 and HNF4A.
Perspective
The analysis covers 22 genes with established roles in monogenic diabetes and includes only GRCh38, germline, single nucleotide variants with ClinVar review status of at least one star; structural variants, indels and records below that review threshold fall outside its scope. The output is positioned as triage support for expert review, and the author states it does not replace functional validation or multidisciplinary variant-classification review. Proposed next steps include temporal validation against historical ClinVar releases, adding protein structural and gene-constraint features, extending to structural variants, benchmarking directly against meta-predictors such as REVEL and BayesDel on this gene set, and building an interactive interface.
ClinVar classifications evolve as new evidence arrives, so the training labels are a single snapshot; variant_summary.txt does not record which ACMG criteria individual submitters invoked, so circular and independent evidence cannot be fully separated, and circularity tied to consequence severity and allele frequency may persist within the non-missense subset. Annotations came from REST services rather than a version-pinned local installation; raw responses were archived, but live re-querying may differ as underlying databases update. Gene identity was deliberately excluded to avoid gene-specific label-frequency shortcuts, which may reduce sensitivity to true gene-specific constraint such as dosage sensitivity.