SeizeVar: training data and trained model weights for mechanism-aware variant interpretation in monogenic epilepsy

Companion data and model-weights deposit for SeizeVar, a two-layer pipeline for mechanism-aware variant interpretation in monogenic epilepsy. This record contains two compressed bundles: 1. seizevar_data.tar.gz (423 MB) — training and benchmarking inputs: - ClinVar-derived training/validation splits for the 49-gene epilepsy panel (KCNQ2 held out) - UniProt sequence cache and AlphaFold v4 structures for all panel genes - External resources: gnomAD constraint, curated mechanism/inheritance table, selectivity-filter annotations, Miyazawa-Jernigan pair-energy table 2. seizevar_models_trained.tar.gz (339 MB) — trained model weights: - Random-Forest pathogenicity head (sklearn pickle) - ESM-2 LoRA cross-attention pathogenicity head (PyTorch state dict) - Random-Forest gain-vs-loss-of-function mechanism head - ESM feature standardisation arrays (mean and std) Pipeline source code is archived separately at Zenodo (DOI: 10.5281/zenodo.19912082) and developed at https://github.com/pipi0616/seizevar. Accompanies the manuscript: Ye S, Chen P. From pathogenicity to prescription: mechanism-aware variant interpretation in monogenic epilepsy. Submitted, 2026. License: code under MIT; data and weights released under CC BY 4.0.