Machine Learning-Guided Identification of Allosteric Drug Binding Poses from Molecular Dynamics Simulations
\(^{1}\) Department of Chemistry, Western University, ON, Canada
\(^{2}\) European Laboratory for Learning and Intelligent Systems (ELLIS) Institute Finland
\(^{3}\) Department of Technical Physics, University of Eastern Finland
Eumelanin is a heterogeneous, disordered biopolymer whose structural complexity complicates the identification of ligand binding sites, particularly allosteric interactions [1]. Diverse local environments and the absence of well-defined structure hinder atomistic characterization of binding mechanisms. We frame allosteric drug binding as a high-dimensional structural problem and introduce a computational pipeline for resolving binding modes from molecular simulation data. Molecular dynamics (MD) simulations generate trajectories of carbonic anhydrase inhibitors interacting with model eumelanin aggregates; these compounds [2] are experimentally classified as strong or weak binders, providing a ground truth for analysis. To extract structure from high-dimensional MD trajectories, we apply a feature representation based on a mixed radial–angular three-particle correlation function encoding local geometric and orientational relationships [3] between ligands and eumelanin subunits. This embedding is analyzed using unsupervised machine learning — t-SNE for dimensionality reduction and HDBSCAN for clustering — enabling identification of metastable binding states without prior labelling. The pipeline resolves distinct binding modes and identifies reproducible interaction motifs across independent simulations. The resulting embeddings capture physically interpretable features of ligand–eumelanin interactions and separate binding regimes, highlighting the value of combining representation learning with physics-based simulation to characterize allosteric binding in heterogeneous systems.
[1] S. Soltani, A. Roy, A. Urtti, M. Karttunen, A Computational Investigation of Eumelanin–Drug Binding in Aqueous Solutions. Mater. Adv. 2024, 5 (2024), 5494.
[2] . Valtari, S. Kalinin, J. Jäntti, P. Vanhanen, H. Martina, T. Arun, S. Katja, V. Tapani, K. S. Vellonen, E. Toropainen, M. Ruponen, and A. Urtti, Melanin-Binding-Based Discovery of Topically Instilled Carbonic Anhydrase Inhibitors for Targeted Delivery and Prolonged Action in the Eye, Mol. Pharm. 22 (2025), 721.
[3] Davies, M.; Reyes-Figueroa, A. D.; Gurtovenko, A. A.; Frankel, D.; Karttunen, M. , Elucidating Lipid Conformations in the Ripple Phase: Machine Learning Reveals Four Lipid Populations., Biophys. J. 122 (2023), 442.