Classical scoring functions for docking are unable to exploit large volumes of structural and interaction data.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Scoring functions (SFs) are crucial for predicting protein-ligand binding affinity.
• Random Forest (RF)-based SFs show improved accuracy with more training data, unlike classical SFs.
• The impact of training-test similarity on SF performance is not well understood.

Purpose of the Study:

• To systematically investigate how training-test similarity affects classical and machine learning SF accuracy.
• To determine if machine learning SFs can generalize from dissimilar training data.
• To develop and evaluate a novel SF using Extreme Gradient Boosting (XGBoost).

Main Methods:

• Evaluated classical (X-Score) and machine learning (RF-Score-v3) SFs.
• Assessed performance using three similarity metrics: protein structure, protein sequence, and ligand structure.
• Developed and tested a new XGBoost-based SF (XGB-Score).

Main Results:

• Classical SFs did not improve accuracy with increased training set size or similarity.
• RF-Score-v3 outperformed X-Score, especially when trained on dissimilar complexes.
• XGB-Score demonstrated improved accuracy with training set size and outperformed other SFs.

Conclusions:

• Machine learning SFs, particularly RF and XGBoost, benefit significantly from diverse training data, including dissimilar complexes.
• The ability to learn from dissimilar data is key to the superior performance of modern SFs.
• XGBoost presents a promising new approach for developing accurate SFs in drug discovery.