FlowDock: Geometric flow matching for generative protein-ligand docking and affinity prediction.

Area of Science:

• Computational Biology
• Artificial Intelligence
• Drug Discovery

Background:

• Existing generative AI models for protein-ligand structures often lack support for flexible docking, affinity estimation, and concurrent multi-ligand modeling.
• Rigorous benchmarking on pharmacologically relevant targets is limited, hindering adoption in drug discovery.

Purpose of the Study:

• To introduce FlowDock, a novel deep geometric generative model for protein-ligand structure prediction.
• To enable flexible protein-ligand docking and affinity estimation for multiple ligands simultaneously.
• To provide a robustly benchmarked tool for accelerating drug discovery through virtual screening.

Main Methods:

• Utilizes conditional flow matching (CFM) to learn direct mapping from unbound (apo) to bound (holo) protein structures.
• Generates protein-ligand complex structures with predicted confidence scores and binding affinity values.
• Employs deep geometric generative modeling for structure prediction.

Main Results:

• FlowDock achieves a 51% blind docking success rate on the PoseBusters Benchmark, outperforming single-sequence AlphaFold 3 (AF3) using only unbound protein input.
• Demonstrates strong binding pocket generalization on the DockGen-E dataset, matching single-sequence Chai-1.
• Ranked among the top-5 methods for binding affinity estimation in the 16th Critical Assessment of Techniques for Structure Prediction (CASP16).

Conclusions:

• FlowDock represents a significant advancement in generative AI for drug discovery, offering accurate and efficient protein-ligand complex modeling.
• Its ability to handle multiple ligands and provide affinity predictions makes it a valuable tool for virtual screening and target identification.
• The model's performance across multiple benchmarks validates its potential for accelerating the discovery of new therapeutics.