From Pathogenicity to Mechanism: A Variant Interpretation Framework for Monogenic Epilepsy

Pathogenicity predictors exceed AUROC 0.97 on expert-curated ClinVar, yet the monogenic-epilepsy variant-of-uncertain-significance backlog persists because pathogenicity alone provides no direction-of-effect, no scalable mechanism-to-treatment mapping, and no evidence that updates as ClinVar grows. SeizeVar couples a consensus pathogenicity head (random forest plus ESM-2 LoRA cross-attention) to a gain-versus-loss-of-function mechanism classifier and a deterministic sodium-channel mechanism-direction rule. The framework was trained on a 49-gene epilepsy panel (n = 4576 labelled variants) and evaluated on six pairwise-disjoint held-out cohorts (n = 11 274) plus an external functional cohort with patch-clamp/TEVC labels (T2, n = 415). SeizeVar's mechanism head reaches honest leave-one-gene-out AUROC = 0.736 at 100% panel coverage and matches the proteome-wide specialist LoGoFunc on fair full-coverage comparison (0.770 vs. 0.760), whereas general-purpose pathogenicity predictors are mechanism-blind (AUROC ≤ 0.62). Applied to 29 293 epilepsy VUS, the pipeline returned 4708 consensus Likely-Pathogenic candidates, of which 1500 sodium-channel variants received predicted mechanism-direction labels (679 LoF-leaning, 821 GoF-leaning). By integrating pathogenicity, mechanism, and dynamic evidence on a shared probability scale, SeizeVar produces a mechanism-annotated prioritisation list to support-not replace-expert variant curation; the outputs are computational predictions and a natural next step is prospective clinical validation. Reclassified-VUS is released as a community benchmark for reclassification-aware evaluation.