Target-Specific Prediction of Ligand Affinity with Structure-Based Interaction Fingerprints

Area of Science:

• Computational chemistry and structural biology
• Drug discovery and medicinal chemistry
• Machine learning in pharmacology

Background:

• Structure-based drug design is vital for developing small molecule therapeutics.
• Advances in structural biology and machine learning aid in understanding drug potency.
• Interpretability of machine learning models remains a challenge in drug design.

Purpose of the Study:

• To evaluate inhibitor diversity and machine learning models for predicting HIV-1 protease inhibitor ligand affinity.
• To develop interpretable machine learning models for drug potency prediction.
• To identify key structural features influencing drug potency.

Main Methods:

• Hierarchical clustering to group HIV-1 protease inhibitors and identify core structures.
• Extraction of explicit protein-ligand interaction features from cocrystal structures as 3D fingerprints.
• Application of a gradient boosting machine learning model with feature attribution for affinity prediction.
• Derivation of Shapley values to explain local feature importance.

Main Results:

• A gradient boosting model accurately predicted binding affinity using explicit features.
• Hierarchical clustering revealed distinct inhibitor core structures.
• Shapley values identified specific van der Waals interactions as critical for predicted potency.

Conclusions:

• Interpretable machine learning models can predict drug binding affinity with high accuracy.
• Specific van der Waals interactions with key residues are crucial for inhibitor potency.
• Protein-specific, interpretable models can guide the optimization of small molecule drugs.