Genesis v8: A C-alpha-only Geometric Framework for Protein-Level Allosteric Propensity Classification
Genesis v8 presents a lightweight, interpretable framework for protein-level allosteric propensity classification using five C-alpha-only geometric descriptors: Fiedler eigenvalue (lambda2), Lo-Shu lattice transfer energy (LTE), convex-hull concavity score (CHC), backbone rigidity fraction (frac_rigid), and pocket-region flexibility (pocket_flex). Four-tier validation: Internal CV (n=768): standard AUC=0.898 [0.871,0.921]; homology-aware AUC=0.701 External positive — AlloBench blind test (n=430): Sensitivity=80.2% @0.50, 88.8% @0.30 Specificity pilot (n=220, exploratory lower bound): AUC=0.76, BalAcc=0.71 @0.80 Mechanistic spatial (Route 2, n=101): hinge-to-site 11.7 A vs 15 A random; p=0.0022 Proteome feasibility: 20,158 human AlphaFold2 structures screened in <3 h on CPU. No GPU required. v8 is the first Genesis version with fully separated four-tier validation and honest negative-set contamination analysis. Previous versions: v2 (10.5281/zenodo.20187149), v3 ESM2 Route-A (10.5281/zenodo.20202158). Route 2 companion: 10.5281/zenodo.20251581.
Authors
- Yao-Kai Kao
Institutions
- Allen Institute for Brain Science (US)
- Taipei Institute of Pathology (TW)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-05-17
- DOI
- https://doi.org/10.5281/zenodo.20252223
- Primary Topic
- Machine Learning in Bioinformatics
- Type
- preprint