PharmRL: Pharmacophore elucidation with Deep Geometric Reinforcement Learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Molecular interactions are crucial for drug design and virtual screening.
• Pharmacophores, representing favorable interactions in protein binding sites, are typically identified from co-crystal structures.
• Identifying pharmacophores without a bound ligand presents a significant challenge.

Purpose of the Study:

• To develop a novel deep learning method for identifying pharmacophores in protein binding sites, independent of ligand presence.
• To enhance virtual screening performance and accelerate the identification of active molecules and lead compounds.
• To provide an accessible tool for researchers through a Google Colab notebook.

Main Methods:

• A Convolutional Neural Network (CNN) was trained to detect potential favorable interactions within protein binding sites.
• A deep geometric Q-learning algorithm was developed to select optimal interaction points, forming a pharmacophore.
• The method was validated on benchmark datasets (DUD-E, LIT-PCBA) and a relevant drug discovery dataset (COVID moonshot).

Main Results:

• The developed method demonstrated superior prospective virtual screening performance (F1 scores) on the DUD-E dataset compared to random selection.
• Experiments on the LIT-PCBA dataset confirmed the method's efficiency in identifying active molecules.
• The approach proved effective in identifying potential lead molecules for COVID-19 drug discovery, even without fragment screening data.

Conclusions:

• The novel deep learning approach successfully identifies pharmacophores without requiring bound ligands, offering a significant advancement in computational drug discovery.
• This method enhances virtual screening accuracy and efficiency, facilitating the discovery of novel therapeutic agents.
• The provided Google Colab notebook ensures broad accessibility and usability of this innovative technique.