Chemical Language Model Linker: Blending Text and Molecules with Modular Adapters

Area of Science:

• Computational chemistry
• Artificial intelligence
• Drug discovery

Background:

• Large language models (LLMs) and multimodal models offer potential for generating novel molecules from text descriptions.
• Current multimodal models often require training from scratch, which is computationally expensive and limits scalability.
• Existing methods do not effectively leverage high-quality pretrained models for this task.

Purpose of the Study:

• To propose a novel, lightweight adapter-based strategy named Chemical Language Model Linker (ChemLML) for conditional molecular generation from text.
• To enable the use of diverse pretrained text models for molecule generation without extensive retraining.
• To investigate the impact of molecular representations (SMILES vs. SELFIES) on generation performance.

Main Methods:

• Developed ChemLML, an adapter-based strategy that links pretrained text and molecular domain models.
• Trained relatively few adapter parameters to tailor existing text models for molecule generation.
• Evaluated performance using a filtered PubChem dataset and compared SMILES and SELFIES representations.
• Generated candidate protein inhibitors and membrane-permeable molecules for practical demonstration.

Main Results:

• ChemLML effectively blends pretrained single-domain models for conditional molecular generation.
• The choice of molecular representation significantly impacts generation performance, with SMILES often being preferable.
• Identified issues with the standard PubChem dataset for evaluation and provided a refined test set.
• Successfully generated candidate molecules with potential therapeutic and pharmacokinetic properties.

Conclusions:

• ChemLML offers an efficient and scalable approach to text-to-molecule generation by leveraging pretrained models.
• Molecular representation choice is critical for successful conditional generation.
• The developed method and dataset facilitate practical applications in drug discovery and molecular design.