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

Area of Science:

• Computational chemistry
• Structural biology
• Drug design

Background:

• Protein-ligand pose prediction is crucial for structure-based drug design.
• Data-driven methods, including deep learning and diffusion, now surpass traditional molecular docking techniques.
• Concerns about data leakage and generalizability persist with current data-driven models.

Purpose of the Study:

• To introduce a simple, data-driven baseline method for ligand-based protein-ligand pose prediction.
• To establish a meaningful benchmark for evaluating interpolative data-driven methods.
• To assess the performance of this baseline against existing methods and benchmarks.

Main Methods:

• Developed the TEMplate-based Protein-Ligand (TEMPL) baseline.
• Utilized maximal common substructure to reference molecules.
• Employed constrained 3D embedding for pose prediction.

Main Results:

• TEMPL outperformed classic docking algorithms in an antiviral competition for SARS-CoV-2 and MERS-CoV Main Protease ligand pose prediction.
• Demonstrated good performance on the PDBBind benchmark, highlighting potential data leakage issues in deep learning methods.
• Showcased limited performance on the challenging PoseBusters benchmark.

Conclusions:

• The TEMPL baseline provides a valuable, strictly data-driven benchmark for evaluating novel pose prediction algorithms.
• Findings underscore the importance of rigorous benchmarking and challenging data splits for data-driven methods.
• The open-source TEMPL method and web application facilitate the evaluation of future pose prediction techniques.