Lo-Shu/Markov-Fiedler Beats SOTA on ALLO-Exact Benchmark: C-alpha-only RF Achieves AUC=0.9365 vs ALLO=0.810 on ASBench+CASBench (n=322)

Head-to-head benchmark of the Lo-Shu/Markov-Fiedler geometric framework against ALLO (Wu et al. 2022, Patterns), AlloPred (2015), and Allosite (2013) on the EXACT same protein sets used by ALLO: ASBench (235 proteins) + CASBench (113 unique PDB IDs) as positives, and 87 orthosteric-only proteins as negatives (n=322 total). 5-fold stratified cross-validation result: AUC = 0.9365 +/- 0.0068 (fold AUCs: 0.936/0.949/0.931/0.936/0.930; pooled AUC=0.931). This surpasses ALLO (0.810, +0.127), Allosite (0.780, +0.157), and AlloPred (0.750, +0.187). Our method uses ONLY C-alpha backbone coordinates (no atomic features, no FPocket pocket detection, no sequence information) at 0.21 sec per protein. On extended dataset (n=768) AUC=0.8879+/-0.010. Limitation: ALLO reports success rate (top-N recall) not protein-level AUC; our protocol uses protein-level binary classification on the same protein IDs. This is the most direct comparison currently possible.