PharmRL: pharmacophore elucidation with deep geometric reinforcement learning.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Protein-ligand interactions are crucial for drug design and virtual screening.
• Pharmacophores, representing favorable interactions, are typically derived from protein-ligand co-crystal structures.
• Designing pharmacophores without a known ligand presents a significant challenge.

Purpose of the Study:

• To develop an automated deep learning method for pharmacophore identification in the absence of a ligand.
• To enhance virtual screening performance and facilitate drug discovery for novel targets.

Main Methods:

• A convolutional neural network (CNN) was trained to identify potential favorable interactions within protein binding sites.
• A deep geometric Q-learning algorithm was developed to select optimal interaction points for pharmacophore generation.
• The method, named PharmRL, was evaluated on benchmark datasets (DUD-E, LIT-PCBA) and a COVID-19 dataset.

Main Results:

• PharmRL demonstrated superior virtual screening performance (F1 scores) compared to random selection on the DUD-E dataset.
• The method efficiently identified active molecules in the LIT-PCBA dataset.
• Screening the COVID moonshot dataset showed PharmRL's potential for identifying lead molecules even without prior fragment screening data.

Conclusions:

• PharmRL provides an automated solution for pharmacophore design when cognate ligands are unavailable.
• Experimental results confirm PharmRL's ability to generate functional pharmacophores.
• A Google Colab notebook is available to support the method's application.