DiffDock-Glide: A Hybrid Physics-Based and Data-Driven Approach to Molecular Docking

Area of Science:

• Computational chemistry
• Molecular modeling
• Artificial intelligence in drug discovery

Background:

• Deep learning methods like DiffDock show promise for molecular docking but face limitations with novel targets.
• Conventional docking methods remain competitive, highlighting areas for improvement in AI-driven approaches.

Purpose of the Study:

• To develop an enhanced molecular docking model, DiffDock-Glide, by integrating deep learning with established computational chemistry tools.
• To address the shortcomings of pure deep learning docking methods, particularly in generalizing to unseen protein targets.

Main Methods:

• A hybrid model, DiffDock-Glide, was developed, combining a modified diffusion model for pose generation with Glide's post-docking minimization.
• The generative process was adapted to better sample poses within protein binding pockets.
• Glide's established post-docking minimization pipeline replaced the deep learning confidence model.

Main Results:

• DiffDock-Glide demonstrated improved sampling of near-native poses on the PoseBusters dataset.
• Performance gains were particularly notable for protein targets lacking homologous sequences in the training set.
• Virtual screening against AlphaFold2-generated structures using DUD-E dataset yielded superior enrichment values compared to traditional Glide.

Conclusions:

• DiffDock-Glide represents a significant advancement in AI-driven molecular docking, enhancing accuracy and applicability to novel targets.
• The hybrid approach effectively leverages the strengths of both deep learning generative models and physics-based refinement.
• This method shows potential for more effective virtual screening and drug discovery pipelines.