Computational representations of protein-ligand interfaces for structure-based virtual screening.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Machine learning in bioinformatics

Background:

• Structure-based virtual screening (SBVS) is crucial for identifying potential drug candidates.
• Molecular docking, a key SBVS technique, relies on predicting protein-ligand binding modes and affinities.
• Machine learning (ML) models can enhance SBVS accuracy by learning from protein-ligand interface representations.

Purpose of the Study:

• To review computational methods for representing protein-ligand interfaces in SBVS.
• To discuss the impact of these representations on ML model performance.
• To highlight current trends and future directions in computational drug discovery.

Main Methods:

• Review of traditional methods using handcrafted fingerprints and descriptors for interface representation.
• Exploration of recent deep learning approaches for automatic feature extraction from interfaces.
• Analysis of case studies applying diverse computational representations to ML models.

Main Results:

• Deep learning-based feature extraction offers an end-to-end representation for protein-ligand interfaces.
• Various computational representations significantly influence the performance of ML models in SBVS.
• Case studies demonstrate the practical application and effectiveness of different representation strategies.

Conclusions:

• Automatic feature extraction via deep learning is a growing trend in SBVS.
• Areas for improvement include model interpretability, data management, data quality/quantity, and hyperparameter optimization.
• Incorporating factors like water molecules and protein flexibility is essential for advancing SBVS accuracy.