Narrowing the gap between machine learning scoring functions and free energy perturbation using augmented data.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Machine learning models promise fast binding affinity predictions but lack robust evaluation for drug discovery tasks like lead optimization.
• Current models struggle with ranking congeneric ligand series, limiting their real-world application.

Purpose of the Study:

• To introduce a novel attention-based graph neural network, AEV-PLIG, for improved binding affinity prediction.
• To develop a new, realistic out-of-distribution test set (OOD Test) for evaluating machine learning models.
• To rigorously assess machine learning model performance against physics-based approaches.

Main Methods:

• Developed AEV-PLIG, an attention-based graph neural network model.
• Introduced the OOD Test set for out-of-distribution evaluation.
• Benchmarked AEV-PLIG on OOD Test, CASF-2016, and a free energy perturbation (FEP) benchmark.

Main Results:

• AEV-PLIG demonstrates competitive performance compared to physics-based methods.
• Augmented data significantly improved prediction correlation and ranking on the FEP benchmark (PCC: 0.41 to 0.59, Kendall's τ: 0.26 to 0.42).
• AEV-PLIG is approximately 400,000 times faster than FEP calculations.

Conclusions:

• AEV-PLIG offers a robust and efficient machine learning approach for binding affinity prediction.
• The developed strategies bridge the performance gap with FEP calculations, enhancing drug discovery efficiency.
• AEV-PLIG provides a realistic assessment of machine learning models for drug discovery applications.