SIMORGH: Ensemble-Aware Geometric Deep Learning for Apo-State and Cryptic Ligand Binding Site Prediction

Abstract

Most computational methods identify protein–ligand binding sites from ligand-bound (holo) protein structures, where the binding pocket is already preorganized. Although convenient for benchmarking, this setting differs from the practical drug discovery scenario, in which binding sites must be inferred from ligand-free (apo) proteins. In fact, binding-competent conformations may represent only a subset of the accessible structural ensemble, while transient and cryptic pockets can emerge through conformational fluctuations. Consequently, prediction from a single static structure is often insufficient. Here we present SIMORGH, an ensemble-aware SE(3)-equivariant geometric learning framework that predicts ligand binding sites by integrating information across protein conformational ensembles. SIMORGH employs a two-stage architecture consisting of a structure-level equivariant encoder that learns geometric representations from individual conformations and a lightweight ensemble-level aggregation module that combines them into residue-level binding predictions while scaling near-linearly with both protein and ensemble size. We evaluate SIMORGH across multiple complementary settings, including the PLINDER apo benchmark, molecular dynamics trajectories spanning more than 200 cryptic-pocket proteins, and established community benchmarks. SIMORGH consistently outperforms existing static and dynamics-aware methods under the challenging ligand-free setting, while identifying cryptic binding pockets, maintaining strong prediction consistency between paired apo and holo structures, and scaling efficiently to large multi-chain complexes. To facilitate reproducible, transparent, and accessible use, Simorgh is available through the Galaxy Europe server (https://usegalaxy.eu).