GatorAffinity: Boosting Protein-Ligand Binding Affinity Prediction with Large-Scale Synthetic Structural Data

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Protein-ligand binding affinity prediction is crucial for drug discovery but limited by scarce experimental data.
• Existing datasets like PDBbind are insufficient for training robust data-driven models.
• Vast amounts of affinity data are underutilized due to missing structural information.

Purpose of the Study:

• To overcome data scarcity in protein-ligand binding affinity prediction.
• To develop a highly accurate and generalizable deep learning model for affinity prediction.
• To leverage large-scale synthetic protein-ligand complex data.

Main Methods:

• Curated over 450,000 synthetic protein-ligand complexes with Kd and Ki values using the Boltz-1 model.
• Augmented data with over 1 million synthetic complexes from the SAIR database (IC50 values).
• Developed GatorAffinity, a geometric deep learning scoring function, pretrained on synthetic data and fine-tuned on PDBbind experimental data.

Main Results:

• GatorAffinity significantly outperformed state-of-the-art affinity prediction methods on a leak-proof benchmark.
• Demonstrated superior accuracy and generalizability compared to existing approaches.
• Validated that synthetic data augmentation effectively addresses data scarcity while maintaining predictive reliability.

Conclusions:

• Augmenting experimental data with large-scale synthetic complexes is a viable strategy to enhance affinity prediction.
• GatorAffinity provides a scalable and reproducible foundation for virtual screening and structure-based drug design.
• The pretrained GatorAffinity model and GatorAffinity-DB dataset are released to facilitate further research.