Predicting Affinity Through Homology (PATH): Interpretable Binding Affinity Prediction with Persistent Homology.

Area of Science:

• Computational chemistry and structural biology
• Machine learning in drug discovery
• Topological data analysis for biomolecular interactions

Background:

• Accurate binding affinity prediction is essential for structure-based drug design.
• Computational topology, specifically algebraic topology, has shown promise in representing protein-ligand interactions.
• Previous topological methods lacked interpretability and had high computational complexity, limiting their application.

Purpose of the Study:

• To develop a faster and more interpretable algorithm for predicting protein-ligand binding affinity using persistent homology.
• To introduce novel topological features, internuclear persistent contours (IPCs) and persistence fingerprints, for binding affinity prediction.
• To present the PATH (Predicting Affinity Through Homology) algorithm, comprising PATH+ and PATH-, for enhanced drug design.

Main Methods:

• Developed the fastest known algorithm for computing persistent homology features for protein-ligand complexes, independent of protein size.
• Introduced internuclear persistent contours (IPCs) and persistence fingerprints for interpretable feature representation.
• Implemented the PATH algorithm, utilizing shallow regression trees on persistence fingerprints (PATH+) and IPCs (PATH-).

Main Results:

• Achieved a significant improvement in time complexity for computing persistence fingerprints, independent of protein size.
• Demonstrated that PATH+ achieves comparable performance to state-of-the-art methods using significantly fewer features and offers interpretability.
• Benchmarked PATH against established methods, showing comparable or superior performance with reduced overfitting.

Conclusions:

• The developed topological approach and PATH algorithm provide an efficient, interpretable, and accurate method for binding affinity prediction.
• Persistence fingerprints effectively capture binding-relevant structural information, generalizing across datasets.
• PATH represents a significant advancement in computational drug design, offering an open-source solution.