MERMAID: an open source automated hit-to-lead method based on deep reinforcement learning

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• The hit-to-lead process enhances drug-like properties of initial screening hits.
• Deep learning models show promise for molecular generation in drug discovery.
• Simplified molecular input line entry system (SMILES) is a common molecular representation, but generative models struggle with validity and modification.

Purpose of the Study:

• To develop a SMILES-based generative model capable of modifying existing molecules for the hit-to-lead process.
• To address limitations of SMILES in generating valid molecules and enabling targeted modifications.

Main Methods:

• Developed a novel generative model using Recurrent Neural Network (RNN) and Monte Carlo Tree Search (MCTS).
• The model generates and inserts partial SMILES strings into existing SMILES representations.
• Utilized a dataset from the ZINC database for model training and validation.

Main Results:

• Successfully generated novel molecules with optimized drug-likeness (QED) and penalized octanol-water partition coefficient (PLogP).
• Demonstrated the model's capability to modify existing molecules, facilitating the hit-to-lead optimization.
• Validated the model's performance on a large chemical dataset.

Conclusions:

• The developed SMILES-based generative model effectively supports the hit-to-lead process.
• This approach offers a new computational tool for optimizing molecular properties in drug discovery.
• The model's ability to generate valid and optimized molecules from existing structures shows significant potential.