Benchmark Bias and Conformational Dynamics in Allosteric Site Prediction

Allosteric site prediction plays a critical role in modern drug discovery, offering opportunities to target regulatory regions with high specificity. However, most existing computational approaches rely on static protein structures and pocket detection tools such as fpocket, thereby overlooking conformational dynamics essential for allosteric regulation. Here, we present AlloDyn, a framework that integrates static pocket descriptors with dynamic features derived from both all-atom molecular dynamics (MD) simulations of apo-state proteins and AlphaFlow-generated conformational ensembles. By capturing structural flexibility, solvent accessibility, and residue-residue communication patterns at the pocket level, our approach enables a dynamic-aware representation of candidate allosteric sites. Importantly, we identify a systematic bias in current benchmarking practices, showing that applying fpocket to holo structures without removing bound allosteric modulators introduces data leakage and leads to artificially inflated performance estimates. When evaluated on properly preprocessed datasets, dynamic feature augmentation significantly improves prediction performance over static baselines. Furthermore, we demonstrate that AlphaFlow-generated ensembles achieve performance comparable to MD-derived features at a fraction of the computational cost, providing a scalable alternative for conformational sampling. Benchmarking on the D24 dataset shows that AlloDyn achieves the best balance between precision and recall, yielding the highest F1 score and MCC among evaluated methods. We show that current benchmarks overestimate performance due to data leakage, and that incorporating dynamics is key to accurate and scalable allosteric site prediction.