PharmacoNet: deep learning-guided pharmacophore modeling for ultra-large-scale virtual screening

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery

Background:

• Ultra-large-scale virtual screening is crucial for early-stage drug discovery.
• Deep learning models for binding affinity estimation face generalization challenges due to limited training data.
• Molecular docking is a traditional but computationally intensive method.

Purpose of the Study:

• To introduce PharmacoNet, a deep learning framework for automated, ultra-fast pharmacophore modeling.
• To enhance generalization capability in virtual screening across diverse chemical spaces and targets.
• To provide a fast and accurate alternative to traditional docking and existing deep learning scoring models.

Main Methods:

• Developed PharmacoNet, a deep learning framework for protein-based pharmacophore modeling.
• Implemented a parameterized analytical scoring function to evaluate ligand potency.
• Validated generalization ability across unseen targets and ligands.

Main Results:

• PharmacoNet demonstrated extreme speed and reasonable accuracy compared to docking and other deep learning models.
• Successfully identified selective inhibitors against cannabinoid receptors from 187 million compounds in 21 hours on a single CPU.
• Achieved high generalization ability, crucial for vast chemical space exploration.

Conclusions:

• PharmacoNet represents a significant advancement in deep learning for pharmacophore modeling.
• The framework offers a powerful and efficient solution for ultra-fast virtual screening in drug discovery.
• Highlights the untapped potential of deep learning in accelerating the identification of novel drug candidates.