Simpatico: accurate and ultra-fast virtual drug screening with atomic embeddings

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Bioinformatics

Background:

• Structure-based deep learning has advanced virtual drug screening.
• Existing methods, while accurate, can be computationally intensive.

Purpose of the Study:

• To introduce Simpatico, a new method for rapid and accurate virtual drug screening.
• To leverage Representation Learning and graph neural networks for accelerated drug discovery.

Main Methods:

• Simpatico utilizes graph neural networks to generate high-dimensional embeddings for proteins and small molecules.
• These embeddings enable rapid prediction of interaction potential between drug candidates and protein pockets.

Main Results:

• Simpatico screens 600 million compounds against a protein pocket in 2.5 hours on a single GPU.
• Achieves speedups >1000x compared to state-of-the-art docking and diffusion methods, with competitive accuracy.
• Embeddings facilitate toxicity risk assessment and identification of proteins with similar binding potentials.

Conclusions:

• Simpatico offers a significant speed improvement for virtual drug screening.
• The method demonstrates competitive accuracy and broad applicability beyond binding prediction.
• Simpatico is open-source, promoting accessibility and further research.