TEMPL: A Template-Based Protein-Ligand Pose Prediction Baseline

Pose prediction of ligands to proteins remains a central challenge of structure-based drug design. Although data leakage and generalizability concerns remain, data-driven methods for pose prediction (i.e., based on deep learning and diffusion) now routinely outperform traditional techniques such as molecular docking. In this work, we propose a simple data-driven ligand-based baseline for pose prediction, which is based on maximal common substructure to reference molecules, followed by constrained 3D embedding. As this TEMplate-based Protein-Ligand (TEMPL) baseline is strictly data-driven, it is a particularly meaningful baseline for interpolative tasks, where physics-based methods sometimes underperform as they exploit data less directly. However, it can also highlight the added advantage of other interpolative data-driven methods that should outperform this simple approach. We applied our baseline method in the ASAP-Polaris-OpenADMET antiviral competition, achieving a result that outperformed some classic docking algorithms for the pose prediction of a series of ligands at the Main Protease of SARS-CoV-2 and MERS-CoV. Furthermore, we show that the performance of our baseline is relatively good on a protein-ligand pose prediction benchmark used for deep learning based pose prediction, PDBBind, highlighting the risk of data leakage and the necessity of challenging splits for other data-driven methods as well. We also show our baseline method has limited performance on more challenging benchmarks, such as PoseBusters. We provide our baseline method as open source software. For convenience and for nontechnical users, we also provide a web application to run the pipeline. These findings will aid in the evaluation of future pose prediction methods, especially more complex data-driven approaches that are increasing in popularity.