CONSMI: Contrastive Learning in the Simplified Molecular Input Line Entry System Helps Generate Better Molecules

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Deep learning is increasingly applied to molecular de novo design.
• SMILES representations are used for molecular generation as a text generation problem.
• Generating novel and effective molecules remains a challenge, partly due to multiple valid SMILES for a single molecule.

Purpose of the Study:

• To develop a contrastive learning framework (CONSMI) to learn more comprehensive SMILES representations.
• To address the limitation of using single SMILES representations in molecular generation.
• To improve the novelty and validity of deep learning-based molecular design.

Main Methods:

• Proposed a contrastive learning framework named CONSMI.
• Leveraged multiple SMILES representations of the same molecule as positive examples.
• Utilized other SMILES representations as negative examples for contrastive learning.

Main Results:

• CONSMI significantly enhances the novelty of generated molecules.
• Maintained high validity of generated molecules.
• Generated molecules exhibit similar chemical properties to the original dataset.
• Achieved favorable results in classifier tasks, including compound-protein interaction.

Conclusions:

• CONSMI effectively learns comprehensive SMILES representations for molecular generation.
• The framework improves molecular novelty and validity while preserving chemical properties.
• CONSMI shows potential for both generative and classification tasks in cheminformatics.