Predicting Affinity Through Homology (PATH): Interpretable binding affinity prediction with persistent homology.

Area of Science:

• Computational chemistry
• Machine learning
• Structural biology

Background:

• Accurate binding affinity prediction (BAP) is essential for structure-based drug design.
• Current BAP algorithms often lack generalizability and interpretability.

Purpose of the Study:

• To present PATH+, a novel, generalizable, and interpretable machine learning algorithm for BAP.
• To improve the speed and accuracy of BAP for drug discovery.

Main Methods:

• Developed PATH+, a machine learning algorithm leveraging computational topology.
• Visualized features captured by PATH+ for protein-ligand complexes.
• Created PATH-, a scoring function for binder/non-binder differentiation.

Main Results:

• PATH+ demonstrates comparable or superior accuracy and generalizability across datasets compared to existing BAP algorithms.
• PATH+ is the first inherently interpretable BAP algorithm.
• PATH+ is over 10 times faster than dominant computational topology algorithms.
• PATH- shows outstanding accuracy in differentiating binders from non-binders.

Conclusions:

• PATH+ offers a unique combination of interpretability, speed, and accuracy for BAP.
• The algorithm empowers topological screening of virtual inhibitor libraries.
• Open-source code is available, facilitating trust and application in drug design.