Augmenting a training dataset of the generative diffusion model for molecular docking with artificial binding pockets

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Accurate prediction of small molecule poses within protein binding pockets is crucial for drug discovery.
• Existing methods face challenges in accurately modeling non-bond interactions and steric clashes.

Purpose of the Study:

• To introduce PocketCFDM, a generative diffusion model for improved small molecule pose prediction.
• To enhance model performance through a novel data augmentation technique using artificial binding pockets.

Main Methods:

• Developed PocketCFDM, a generative diffusion model for molecular docking.
• Implemented a novel data augmentation strategy by creating artificial binding pockets that mimic real protein-ligand interaction patterns.
• Utilized an algorithmic method to replicate non-bond interaction patterns in artificial pockets.

Main Results:

• PocketCFDM significantly improved small molecule pose prediction accuracy.
• The model demonstrated superior performance over DiffDock in non-bond interaction and steric clash metrics.
• PocketCFDM achieved faster inference speeds compared to DiffDock.

Conclusions:

• The integration of artificial binding pockets is an effective strategy for enhancing generative diffusion models in molecular docking.
• PocketCFDM represents a significant advancement in computational drug discovery, offering improved accuracy and efficiency.
• The PocketCFDM code and model weights are publicly available for further research and development.