DiffBP: generative diffusion of 3D molecules for target protein binding.

Area of Science:

• Computational chemistry
• Drug discovery
• Structural biology

Background:

• Generating molecules with specific protein-binding capabilities is crucial for drug discovery but remains a significant challenge.
• Existing autoregressive methods generate molecules atom-by-atom, often violating physical rules due to their sequential nature and overlooking global atomic interactions.

Purpose of the Study:

• To develop a non-autoregressive generative model, DiffBP, for creating 3D molecular structures that bind to target proteins.
• To leverage protein binding sites as constraints for generating molecules with favorable properties.

Main Methods:

• DiffBP utilizes a generative diffusion model with an equivariant network to denoise element types and 3D coordinates of molecules.
• The model operates in a non-autoregressive manner, considering the entire molecule and its interaction with the protein binding site simultaneously.

Main Results:

• DiffBP demonstrates competitive performance compared to existing methods in generating molecules with high protein affinity.
• The generated molecules exhibit appropriate sizes and favorable drug-like profiles, addressing limitations of prior approaches.

Conclusions:

• DiffBP offers a promising non-autoregressive approach for de novo molecular generation, improving upon sequential methods by incorporating global atomic interactions.
• The model's ability to generate high-affinity binders with desirable properties advances the field of structure-based drug design.