Benchmarking the Lo-Shu/Markov-Fiedler Framework Against SOTA Allosteric Site Predictors: 5-Fold Cross-Validation Achieves ROC-AUC=0.846 on Protein-Level Binary Classification (n=569) Using Only C-alpha Coordinates

We benchmark the Lo-Shu/Markov-Fiedler geometric framework against established allosteric site prediction methods (AlloPred, Allosite, ALLO, PASSer) on a protein-level binary classification task. Using a curated dataset of n=569 proteins (486 allosteric, 83 orthosteric-only), our RandomForest model with size-residualized C-alpha-only geometric features achieves ROC-AUC = 0.8456 +/- 0.048 and PR-AUC = 0.9745 in 5-fold stratified cross-validation (pooled AUC = 0.8426). This exceeds AlloPred (0.750) and Allosite (0.780) on ASBench Core, and is competitive with ALLO (0.810) — while using ONLY C-alpha coordinates at 0.21s per protein. We clarify that PASSer's reported AUC=0.994 addresses a different task (pocket-level ranking) and is not directly comparable. A permutation test (p=0.005) confirms the model learns genuine structural signal beyond size confound. Limitations: our negative set may differ from ASBench; a direct head-to-head on identical PDB IDs is identified as future work.