Accurate and Rapid Ranking of Protein-Ligand Binding Affinities Using Density Matrix Fragmentation and...

Area of Science:

• Computational Chemistry
• Drug Discovery
• Molecular Modeling

Background:

• Accurate prediction of protein-ligand binding affinities is crucial for drug discovery.
• Existing methods often face challenges with efficiency, accuracy, or transferability across diverse chemical systems.

Purpose of the Study:

• To evaluate the performance of the generalized many-body expansion for building density matrices (GMBE-DM) and a physics-informed, machine learning-corrected dispersion potential (D3-ML) for ranking protein-ligand binding affinities.
• To compare these methods against a deep learning model (Sfcnn) in terms of accuracy, efficiency, and transferability.

Main Methods:

• Application of the GMBE-DM method, truncated at the one-body level with a purification scheme.
• Development and application of the D3-ML potential, incorporating machine learning corrections for dispersion interactions.
• Testing on datasets for cyclin-dependent kinase 2 (CDK2) and Janus kinase 1 (JAK1), totaling 28 ligands.

Main Results:

• GMBE-DM achieved strong correlation with experimental binding free energies (R 2 =0.84) with efficient runtimes (<5 min per complex).
• D3-ML demonstrated superior ranking performance (R 2 =0.87) with sub-second runtimes, highlighting the importance of dispersion interactions.
• The deep learning model Sfcnn exhibited lower transferability across datasets (R 2 =0.57).

Conclusions:

• GMBE-DM and D3-ML are robust and scalable computational tools for ranking protein-ligand binding affinities.
• D3-ML offers exceptional speed and accuracy, making it highly suitable for high-throughput virtual screening in drug discovery.
• The study underscores the critical role of dispersion interactions and the limitations of broadly trained neural networks in diverse chemical systems.