Fluorophore Design via Generative Modeling in a Curated Subspace.

Area of Science:

• Computational Chemistry
• Molecular Design
• Organic Chemistry

Background:

• De novo molecular design requires efficient strategies for discovering novel compounds with desired properties.
• Generative models offer potential but often require extensive data and model-specific tuning.
• Designing fluorescent molecules (fluorophores) presents unique challenges due to photophysical property requirements.

Purpose of the Study:

• To develop a fluorescence-oriented molecular design strategy using integrated generative models.
• To overcome limitations of large datasets and model-specific tuning in molecular discovery.
• To enable the design of novel, optically active molecules through a property-driven approach.

Main Methods:

• Integration of multiple generative models (ReLeaSE, MolDQN, MolFinder) within a constrained chemical subspace.
• Utilized a curated training set of atom-in-SMILES (AIS) fragments from chromophore-like molecules.
• Employed a semiheuristic objective function targeting excitation energy, oscillator strength, and fluorophore similarity.

Main Results:

• Successfully generated structurally novel and optically active molecules.
• Validated molecular designs using quantum mechanical (QM) calculations at the sTDA/CAM-B3LYP level.
• Demonstrated the efficacy of strategic data curation and property-driven integration.

Conclusions:

• Strategic data curation and integration of general-purpose models can achieve specialized discovery tasks.
• The proposed pipeline enables efficient de novo design of fluorophores.
• This approach highlights the power of combining AI with chemical constraints for molecular innovation.