Bridging between Structure-Based and Data-Driven Affinity Prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Protein-ligand binding affinity prediction is crucial for computer-aided drug design.
• Various tools exist, including physics-based and machine learning (ML) models, but choosing the optimal one is challenging.
• Physics-based models excel with limited data, while ML models perform better with increasing target-specific data.

Purpose of the Study:

• To develop a method for smoothly transitioning between physics-based and knowledge-based predictions.
• To combine docking scores with Gaussian Process model predictions based on model uncertainty.
• To improve prediction accuracy and generalizability in drug discovery.

Main Methods:

• Developed a framework to combine predictions from physics-based (docking) and ML (Gaussian Process) models.
• The combination weighting dynamically adjusts based on the uncertainty of each model.
• Trained the Gaussian Process model on binding affinities from two industrial datasets.

Main Results:

• Combining structure-based and ML models significantly enhances prediction accuracy, particularly with limited training data.
• The model weighting shifts from docking to ML as more data becomes available.
• Structure-based methods improve generalizability to new chemical entities and increase hit rates in active learning screens.

Conclusions:

• Integrating predictions from multiple computational tools optimizes the use of limited experimental data.
• A hybrid approach ensures more robust performance compared to relying on a single model.
• This method enhances drug discovery pipelines by improving binding affinity prediction and virtual screening efficiency.