Three-Dimensional Convolutional Neural Networks and a Cross-Docked Data Set for Structure-Based Drug Design

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Structural bioinformatics

Background:

• Predicting protein-ligand binding affinity is crucial for drug discovery.
• Current machine learning models for this task often overestimate generalization ability.
• A lack of standardized datasets hinders reliable model comparison.

Purpose of the Study:

• Introduce the CrossDocked2020 dataset for structure-based machine learning.
• Evaluate grid-based convolutional neural network (CNN) models on this new dataset.
• Establish a standardized benchmark for protein-ligand binding affinity prediction.

Main Methods:

• Generated 22.5 million ligand poses docked into similar binding pockets using the CrossDocked2020 dataset.
• Trained and evaluated various grid-based CNN models, including densely connected CNNs.
• Performed comprehensive analysis on data partitioning, training data quality, and pose sensitivity.

Main Results:

• An ensemble of five densely connected CNNs achieved a root mean squared error of 1.42 and Pearson R of 0.612 for affinity prediction.
• The model demonstrated high performance in binding pose classification (AUC 0.956) and pose selection (68.4% accuracy).
• Demonstrated the impact of data partitioning and the benefit of including lower-quality training data.

Conclusions:

• The CrossDocked2020 dataset provides a standardized benchmark for machine learning in drug discovery.
• The developed CNN models show significant potential for accurate protein-ligand binding affinity prediction.
• This work facilitates community adoption and advances the field of computational drug design.