DENVIS: Scalable and High-Throughput Virtual Screening Using Graph Neural Networks with Atomic and Surface Protein...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Virtual screening accelerates drug discovery by identifying potential drug candidates.
• Traditional docking algorithms use physics-based simulations but can be slow.
• Existing machine learning methods often require pre-computed binding poses, limiting their speed.

Purpose of the Study:

• To introduce Deep Neural Virtual Screening (DENVIS), an end-to-end virtual screening pipeline using graph neural networks (GNNs).
• To evaluate DENVIS's performance and speed compared to existing virtual screening methods.
• To demonstrate DENVIS's ability to scale for large-scale virtual screening.

Main Methods:

• Developed an end-to-end pipeline using graph neural networks (GNNs) for virtual screening.
• Modeled protein pockets using atomic and surface features.
• Employed model ensembles and data augmentation with artificial negative sampling.

Main Results:

• DENVIS demonstrated competitive performance against docking-based, machine learning-based, and hybrid algorithms on benchmark databases.
• DENVIS achieved significantly faster screening times (higher throughput) by eliminating the intermediate docking step.
• Compared to sequence-based machine learning models with similar speeds, DENVIS showed superior performance.

Conclusions:

• DENVIS offers a computationally efficient and scalable solution for virtual screening.
• The method achieves competitive to state-of-the-art performance in drug discovery.
• DENVIS has the potential to screen billions of molecules with minimal resources.