Deep Learning with Geometry-Enhanced Molecular Representation for Augmentation of Large-Scale Docking-Based Virtual...

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Structure-based virtual screening (SBVS) is vital in drug discovery but struggles with ultralarge chemical libraries due to high computational costs.
• Existing machine learning methods for SBVS often neglect crucial 3D structural information by representing compounds as strings or 2D graphs.

Purpose of the Study:

• To develop a novel deep learning protocol, GEM-Screen, that leverages 3D structural information for efficient virtual screening.
• To improve the accuracy and speed of identifying hit compounds from billion-entry chemical libraries.

Main Methods:

• Proposed GEM-Screen, a deep learning protocol utilizing geometry-enhanced molecular representations of docked compounds.
• Employed an active learning strategy to train the model on a small fraction of library docking scores to predict outcomes for unseen compounds.
• Applied GEM-Screen to virtual screening campaigns against AmpC and D4 targets.

Main Results:

• GEM-Screen successfully enriched over 90% of hit scaffolds for AmpC within the top 4% of predictions.
• Over 80% of hit scaffolds for D4 were identified within the top 4% of predictions.
• Demonstrated the protocol's ability to significantly reduce the number of compounds requiring traditional docking.

Conclusions:

• GEM-Screen offers a rapid and accurate method for discovering top-scoring compounds from ultralarge chemical libraries.
• The protocol can be integrated with traditional docking to optimize computational resources, avoiding exhaustive library screening.
• This approach significantly advances the efficiency of structure-based virtual screening in drug discovery.