Can AI-Predicted Complexes Teach Machine Learning to Compute Drug Binding Affinity?

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Machine learning-based scoring functions (MLSFs) are crucial for predicting binding affinity.
• Synthetic data augmentation is a promising approach to improve MLSF performance.
• Co-folding models offer a potential source of synthetic structural data.

Purpose of the Study:

• To evaluate the feasibility of using co-folding models for synthetic data augmentation in MLSF training.
• To determine the impact of augmented data quality on MLSF performance.
• To develop methods for identifying high-quality co-folding predictions for data augmentation.

Main Methods:

• Utilized co-folding models to generate synthetic protein complex structures.
• Trained MLSFs using both experimental and co-folding-derived data.
• Developed and applied heuristics to assess the structural quality of co-folding predictions.
• Compared MLSF performance using different data augmentation strategies.

Main Results:

• Performance gains from co-folding data augmentation are highly dependent on the structural quality of the predictions.
• Established simple heuristics can effectively identify high-quality co-folding predictions without requiring experimental structures.
• Co-folding predictions, when filtered by quality, can successfully substitute for experimental structures in MLSF training.

Conclusions:

• Co-folding models are a feasible, albeit quality-dependent, source for synthetic data augmentation in MLSF training.
• Heuristics for quality control are essential for successful data augmentation using co-folding models.
• This work provides a framework for leveraging co-folding models to enhance binding affinity prediction through data augmentation.