Normalized Protein-Ligand Distance Likelihood Score for End-to-End Blind Docking and Virtual Screening

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Molecular docking is crucial for structure-based virtual screening.
• Diffusion models excel at blind docking but lack binding strength estimation.
• Protein-ligand scoring functions are needed for accurate virtual screening.

Purpose of the Study:

• Introduce a deep learning (DL) scoring function, the normalized mixture density network (NMDN) score.
• Develop an end-to-end blind docking and virtual screening protocol (DiffDock-NMDN).
• Improve pose selection and binding affinity prediction in virtual screening.

Main Methods:

• Developed the NMDN score, a DL model learning protein residue-ligand atom distance distributions.
• Integrated an interaction module for experimental binding affinity prediction.
• Created the DiffDock-NMDN protocol combining diffusion models with the NMDN score.

Main Results:

• The NMDN score shows robust performance in pose selection and virtual screening.
• DiffDock-NMDN achieved an average enrichment factor of 4.96 on the LIT-PCBA dataset.
• The protocol demonstrates effectiveness in drug discovery with limited binder information.

Conclusions:

• The NMDN score offers superior pose selection and virtual screening capabilities.
• DiffDock-NMDN is an effective protocol for real-world drug discovery scenarios.
• This work provides benchmarks and a robust DL scoring function for future research.