Assessing Boltz-2 Performance for the Binding Classification of Docking Hits.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• The Boltz-2 cofolding model offers new capabilities for predicting protein-ligand interactions.
• Accurate prediction of binding affinity is crucial for efficient drug discovery.
• Evaluating model performance on challenging datasets is essential for reliable application.

Purpose of the Study:

• To assess the performance of the Boltz-2 cofolding model on a difficult dataset of ultralarge-virtual-screening hits.
• To investigate the reliability of Boltz-2's affinity predictions by challenging them with biological perturbations.
• To determine if Boltz-2's predictions are based on physical principles or potentially hidden features.

Main Methods:

• Application of the Boltz-2 model to predict protein-ligand structures and binding affinities.
• Testing Boltz-2's ability to discriminate true from false positives in virtual screening.
• Evaluating affinity prediction robustness against target mutations and target shuffling experiments.

Main Results:

• Boltz-2 significantly outperformed existing scoring functions in identifying true positives from docking poses.
• Affinity predictions were largely independent of pose quality and chemical similarity.
• Model performance remained insensitive to key binding site mutations and target exchanges, suggesting potential issues.

Conclusions:

• Boltz-2 demonstrates strong performance in pose discrimination but raises concerns regarding the physical basis of its affinity predictions.
• The model's insensitivity to biological perturbations suggests it may rely on non-physical features.
• Further investigation is needed to understand the underlying mechanisms driving Boltz-2's affinity predictions for reliable drug discovery.