MERMAID

MERMAID generates and optimizes SMILES-based molecular structures to support hit-to-lead optimization in drug discovery.

Key Features:

• SMILES-Based Generative Model: A generative model operates on SMILES strings to produce valid molecular structures from existing compounds.
• Monte Carlo Tree Search and RNNs: Monte Carlo Tree Search combined with Recurrent Neural Networks (RNNs) generates partial SMILES fragments that are integrated into parent molecules.
• Optimization for Drug-Likeness: Optimizes physicochemical properties including Quantitative Estimate of Drug-likeness (QED) and Penalized LogP (PLogP).
• Validation with ZINC database: Method performance was validated using molecular datasets from the ZINC database.

Scientific Applications:

• Hit-to-Lead Optimization: Modifies screening hits into more drug-like lead compounds by generating and optimizing SMILES-based analogs.
• Lead Compound Generation: Produces chemically valid candidate analogs via partial SMILES replacement to expand and explore chemical space.
• Property-Focused Library Design: Enables generation of libraries biased toward improved QED and PLogP values for downstream evaluation.

Methodology:

Starting from an existing molecule, MERMAID uses Monte Carlo Tree Search combined with Recurrent Neural Networks (RNNs) to generate partial SMILES strings which are integrated into the original molecular framework and evaluated/optimized for QED and Penalized LogP, with validation performed on the ZINC database.