Learning from the ligand: using ligand-based features to improve binding affinity prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Machine learning (ML) scoring functions outperform classical methods for predicting protein-ligand binding affinity.
• Structure-based scoring functions often overlook crucial ligand chemical and topological properties.

Purpose of the Study:

• To investigate the impact of incorporating diverse ligand-based features into ML scoring functions.
• To enhance the accuracy and generalizability of binding affinity prediction models.

Main Methods:

• Developed and evaluated Random Forest (RF) models combining existing structure-based features with RDKit molecular descriptors.
• Assessed model performance on PDBbind benchmark datasets (2007, 2013, 2016).
• Investigated the effect of training set composition on model performance, including data similarity.

Main Results:

• The inclusion of ligand-based features consistently improved ML scoring function performance.
• RF models integrating ligand descriptors achieved higher Pearson correlation coefficients compared to models using only structure-based features.
• RF models utilizing solely ligand-based features demonstrated predictive power comparable to classical scoring functions.

Conclusions:

• Diverse ligand-based features are crucial for improving ML-based binding affinity prediction.
• Ligand-centric feature engineering offers a powerful strategy for developing more accurate and robust scoring functions.
• The study provides valuable insights for the development of next-generation drug discovery tools.