Ligand Conformational Variability Enhances Machine Learning Prediction of Protein-Ligand Binding Affinity.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Accurate prediction of protein-ligand binding affinities is essential for efficient drug discovery.
• Ligand conformation significantly impacts binding affinity, posing a challenge for predictive models.
• Predicting affinities for novel compounds is difficult due to unknown binding geometries.

Purpose of the Study:

• To evaluate machine learning methods for predicting SARS-CoV-2 Mpro binding affinities.
• To address the challenge of ligand geometry sensitivity in binding affinity prediction.
• To improve the accuracy of binding affinity predictions for previously unseen compounds.

Main Methods:

• Evaluated Kernel Ridge Regression (KRR), SchNet, and Polarizable Atom Interaction Neural Network (PaiNN).
• Employed a multi-instance learning framework to handle multiple ligand conformers.
• Trained and tested models using multiple conformers for both known and unseen compounds.

Main Results:

• Incorporating multiple conformers of unseen compounds significantly enhanced prediction accuracy.
• The multi-instance learning approach outperformed models relying solely on refined individual predictions.
• All tested ML methods showed improved performance when utilizing multiple conformers.

Conclusions:

• A multi-instance learning framework effectively handles ligand conformational flexibility in binding affinity prediction.
• Considering multiple ligand conformers for unseen compounds is a more impactful strategy than refining individual models.
• This approach offers a robust method for accelerating drug discovery by improving ML-based virtual screening.