Fine-Tuning DiffDock-L for Allosteric Kinase Docking.

Area of Science:

• Computational Chemistry
• Structural Biology
• Drug Discovery

Background:

• Allosteric kinase inhibitors offer selectivity but pose challenges for current AI docking models.
• Existing models often mispredict allosteric ligand binding modes within the ATP-binding site.

Purpose of the Study:

• To curate a comprehensive dataset (AlloSet) for evaluating and improving AI-driven allosteric kinase inhibitor pose prediction.
• To fine-tune the DiffDock-L model for enhanced accuracy in predicting allosteric binding modes.

Main Methods:

• Curated AlloSet, a kinome-wide dataset with binding mode annotations.
• Fine-tuned the DiffDock-L model using strategies like increased dropout and molecular dynamics supersampling.
• Evaluated performance on allosteric and orthosteric ligands, comparing with AlphaFold3 and Boltz-2.

Main Results:

• The fine-tuned DiffDock-L-Allo model significantly improved pose recovery for Type III/IV allosteric binders.
• Performance on orthosteric (ATP-site) ligands was maintained.
• Targeted retraining reshaped the model's sampling distribution for better low-data binding mode prediction.

Conclusions:

• Fine-tuning diffusion-based models with specialized datasets like AlloSet is crucial for accurate allosteric kinase inhibitor pose prediction.
• This approach provides practical guidance for advancing AI in kinase structure-based drug design, especially for challenging binding modes.