Efficient and Explainable Virtual Screening of Molecules through Fingerprint-Generating Networks Integrated with...

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Artificial intelligence for molecular modeling

Background:

• Virtual screening is crucial for identifying potential drug candidates.
• Existing machine learning models often lack efficiency and explainability.
• Developing novel architectures for enhanced drug screening is an ongoing challenge.

Purpose of the Study:

• To develop and optimize a novel machine learning architecture for drug screening.
• To create an efficient, explainable, and performant tool for ligand classification.
• To improve the prediction of ligand binding affinity compared to contemporary methods.

Main Methods:

• A novel, stitched neural network architecture combining graph convolution-based fingerprints with artificial neural networks was developed.
• The model was trained and assessed using two standardized virtual screening databases and molecules from the ZINC database.
• Experiments involved binary classification of ligands based on docking scores against six drug-design relevant proteins.

Main Results:

• The developed architecture demonstrated superior performance in learning molecular features and predicting ligand binding affinity.
• The model proved more efficient in screening molecules, retaining top-hit candidates.
• Compared to Morgan fingerprints, the neural fingerprint-based model showed better ligand retention and higher filtering rates.
• Model explainability revealed accurate emphasis on critical chemical substructures and atoms influencing predictions.

Conclusions:

• The novel machine learning architecture offers an efficient and explainable approach to drug screening.
• This method enhances the identification of high-affinity ligands, outperforming existing techniques.
• The model's explainability provides insights into molecular interactions, aiding drug design.