Data Set Augmentation Allows Deep Learning-Based Virtual Screening to Better Generalize to Unseen Target Classes and...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Structure-based virtual screening (SBVS) is crucial for identifying drug candidates.
• Current deep learning models in SBVS often underutilize protein structural data.
• This limits their generalizability and ability to capture true binding physics.

Purpose of the Study:

• To investigate the impact of dataset augmentation on deep learning models for SBVS.
• To enhance the utilization of protein structural information in ligand binding prediction.
• To improve model generalizability and focus on physical interactions over dataset biases.

Main Methods:

• Implemented a simple dataset augmentation technique.
• Trained deep learning models using augmented data, forcing protein structure incorporation.
• Evaluated model performance on diverse protein-ligand complex datasets.

Main Results:

• Augmented models demonstrated improved generalizability on unseen data distributions.
• Models showed enhanced ability to identify key protein and ligand atoms involved in binding.
• The approach successfully shifted model focus from dataset artifacts to physical binding principles.

Conclusions:

• Dataset augmentation is an effective strategy to improve deep learning-based SBVS.
• This method enhances model interpretability by highlighting important binding site residues and ligand atoms.
• The findings pave the way for more accurate and reliable virtual screening tools.