Learning on topological surface and geometric structure for 3D molecular generation.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• De novo molecular design is crucial for drug discovery but faces challenges in generating structure-specific molecules.
• Existing methods often fail to capture detailed atomic interactions, limiting their effectiveness for diverse therapeutic targets.

Purpose of the Study:

• To develop a novel computational model for highly effective, structure-specific de novo molecular design.
• To address limitations in current generative models by incorporating detailed atomic interactions and pocket surface topology.

Main Methods:

• Introduced SurfGen, a model employing two equivariant neural networks: Geodesic-GNN and Geoatom-GNN.
• Geodesic-GNN captures topological interactions on the target protein's binding pocket surface.
• Geoatom-GNN models the spatial interactions between potential drug molecules (ligands) and the pocket surface.

Main Results:

• SurfGen demonstrates superior performance compared to existing methods across multiple benchmarks.
• The model exhibits high sensitivity to pocket structure variations, enabling precise molecule generation.
• Successfully generated molecules that closely adhere to the key-and-lock principle.

Conclusions:

• SurfGen offers a significant advancement in computer-aided drug discovery by effectively learning atomic interactions for de novo design.
• The model provides a robust solution for generating molecules tailored to specific protein targets.
• SurfGen's capabilities are particularly promising for tackling challenges such as mutation-induced drug resistance.