Poké: Fast and Reliable Neural-Network-Based Identification of Protein Ligandable Sites

Abstract

Accurate identification of ligandable pockets is of key relevance for structure-based drug design; however, their identification remains challenging, particularly for cryptic sites that become accessible only upon conformational changes. Here, we introduce Poké, a high-performance pocket scouting tool that combines energetic mapping with a neural-network-based ligandability score, enabling rapid and reliable binding site detection in both static structures and long-timescale molecular dynamics (MD) trajectories. Poké identifies low-energy regions on protein surfaces through probe-based grid energy calculations, expands and clusters these regions into coherent pocket volumes, and computes physicochemical descriptors that are integrated into a learned ligandability estimator. The scoring model is trained on pockets identified across large data sets using docking-derived binding score distributions, allowing robust ranking of candidate sites. Benchmarking on publicly available data sets demonstrates competitive pocket detection accuracy, improved ranking performance, and substantially enhanced computational efficiency compared to widely used pocket detectors. As a representative application, we apply Poké to mixed-solvent MD simulations of TEM-1 β-lactamase, showing efficient tracking of an experimentally validated cryptic pocket and its dynamic ligandability along the trajectory. Overall, Poké enables scalable, high-throughput ligandability assessment across extensive conformational landscapes.