InChINet: a self-supervised molecular representation learning framework leveraging SMILES and InChI.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for molecular modeling

Background:

• Molecular representation is crucial for AI-driven drug discovery, impacting tasks like property prediction and molecule generation.
• Existing multi-modal representation models have not utilized the International Chemical Identifier (InChI) as an input.
• The Simplified Molecular Input Line Entry System (SMILES) is a common but syntactically variable molecular representation.

Purpose of the Study:

• To develop a self-supervised molecular representation learning framework incorporating InChI.
• To enhance the robustness and performance of molecular representation learning models.
• To improve the stability of models against variations in SMILES representations.

Main Methods:

• Proposed InChINet, a self-supervised framework pre-trained on 10 million unlabeled molecules.
• Leveraged mutual information between SMILES and InChI for representation learning.
• Introduced token reordering, token masking for SMILES, and SMILES enumeration for augmentation.

Main Results:

• InChINet achieved strong performance across diverse downstream tasks.
• Demonstrated improved molecular property prediction and drug-drug interaction prediction.
• Showcased effectiveness in clustering analysis and zero-shot cross-lingual retrieval.

Conclusions:

• Integrating InChI into molecular representation learning frameworks is beneficial.
• The proposed augmentation strategies enhance model stability and performance.
• InChINet offers a promising approach for advancing AI-driven drug discovery.