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

SeizeVar: consensus pathogenicity head (random forest + ESM-2 LoRA) coupled to a gain/loss-of-function mechanism classifier and a deterministic sodium-channel rule

Pathogenicity predictors exceed 0.97 AUROC, yet the monogenic-epilepsy VUS backlog persists because pathogenicity alone gives neither direction of effect nor a link to treatment. SeizeVar couples a consensus pathogenicity head (random forest + ESM-2 LoRA cross-attention) to a gain/loss-of-function mechanism classifier and a deterministic sodium-channel rule. Trained on a 49-gene epilepsy panel (4,576 variants), it is evaluated on six independent cohorts (11,274 variants) and an external functional cohort. Applied to 29,293 epilepsy VUS, the pipeline returns 4,708 likely-pathogenic candidates, including 1,500 sodium-channel variants with predicted mechanism direction.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

A relevant framework addressing a genuine clinical need: moving beyond a pathogenicity score to predict direction of effect (gain vs loss of function), which is decisive for treatment choice in channelopathy epilepsies. Mechanistic performance remains modest (AUROC ~0.74) and needs prospective validation, but the direction — from classification toward actionable mechanism — is right.

Analysis by Dr Thibaut Benquey