Machine-learning scoring functions to improve structure-based binding affinity prediction and virtual screening.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Bioinformatics and computational biology

Background:

• Molecular docking is crucial for predicting small molecule-target interactions, but its accuracy relies heavily on scoring functions (SFs).
• Improving SFs for binding affinity prediction and virtual screening has been a persistent challenge in computational chemistry.
• Classical SFs often use limited features and linear regression, hindering performance with larger datasets.

Approach:

• This review explores novel SFs employing machine learning (ML) regression models for enhanced accuracy.
• ML models offer flexibility by not imposing predetermined functional forms, enabling effective use of extensive experimental data.
• The study highlights the benefits of nonlinear regression and data-driven feature selection in ML-based SFs.

Key Points:

• Machine learning SFs outperform traditional methods in both binding affinity prediction and virtual screening.
• ML approaches demonstrate superior performance compared to classical linear regression with expert-selected features.
• The performance gap between ML and classical SFs is expected to widen as more training data becomes available.

Conclusions:

• Modern machine learning regression models offer a significant advancement over classical scoring functions for molecular docking.
• The data-driven, nonlinear approach of ML effectively utilizes large datasets for improved predictive performance.
• Future research directions include predicting SF reliability, generating synthetic data, and establishing SF development guidelines.