End-to-end protein-ligand complex structure generation with diffusion-based generative models

Area of Science:

• Computational Biology
• Structural Biology
• Drug Discovery

Background:

• Three-dimensional structures of protein-ligand complexes are vital for understanding molecular interactions and for drug design.
• Existing computational methods for modeling these complexes are often limited by their reliance on pre-existing protein structures and struggle with high-dimensional data.

Purpose of the Study:

• To develop an efficient, end-to-end method for modeling protein-ligand complexes that overcomes the limitations of existing approaches.
• To enable accurate modeling of a wider range of protein-ligand complexes, particularly when the protein structure is unknown.

Main Methods:

• Introduced an equivariant diffusion-based generative model.
• The model learns the joint distribution of ligand and protein conformations.
• Conditioning is based on the ligand's molecular graph and a protein sequence representation from a pre-trained protein language model.

Main Results:

• The developed protein structure-free model successfully generates diverse and accurate structures of protein-ligand complexes.
• The model demonstrates capability in predicting correct binding poses.
• The end-to-end approach shows particular effectiveness when the ligand-bound protein structure is unavailable.

Conclusions:

• The study demonstrates the effectiveness and generative power of an end-to-end framework for complex structure modeling using diffusion models.
• This framework is expected to advance the modeling of protein-ligand complexes, with potential for broad applications and further improvements.