SMICLR: Contrastive Learning on Multiple Molecular Representations for Semisupervised and Unsupervised Representation...

Area of Science:

• Computational chemistry
• Machine learning
• cheminformatics

Background:

• Molecular representation is crucial for machine learning in chemical space exploration.
• Contrastive frameworks have shown promise in representation learning across various domains.
• Existing methods may not fully leverage multimodal molecular data.

Purpose of the Study:

• To propose a novel contrastive learning framework for molecular representation using multimodal data.
• To enhance the learning of structure-property relationships for improved chemical property prediction.
• To establish a robust unsupervised representation learning method for molecules.

Main Methods:

• Developed SMILES Contrastive Learning (SMICLR), a framework jointly training graph and simplified molecular-input line-entry system (SMILES) encoders.
• Utilized contrastive learning objective on multimodal molecular data (graph and SMILES strings).
• Applied data augmentations to molecular representations to further improve performance.

Main Results:

• Reduced prediction error by 44% for energetic and 25% for electronic properties on the QM9 dataset compared to supervised baselines.
• Demonstrated competitive performance in unsupervised representation learning.
• Showcased improved accuracy through data augmentation strategies.

Conclusions:

• SMICLR effectively learns molecular representations from multimodal data, outperforming supervised methods.
• The framework offers a powerful tool for chemical space exploration and property prediction.
• SMICLR provides a strong foundation for both supervised and unsupervised learning tasks in cheminformatics.