Dual-View Learning Based on Images and Sequences for Molecular Property Prediction

Area of Science:

• Computational chemistry
• cheminformatics
• drug design

Background:

• Predicting molecular properties is crucial for drug design but challenging.
• Molecular images lack explicit semantic information, while SMILES strings lack structural details.
• Integrating these representations can improve molecular property prediction.

Purpose of the Study:

• To propose a novel multimodal deep learning architecture, ISMol, for molecular property prediction.
• To leverage both molecular images and SMILES strings for enhanced predictive power.
• To evaluate ISMol's performance against single-modal approaches.

Main Methods:

• Developed ISMol, a multimodal architecture utilizing a cross-attention mechanism.
• Extracted molecular representations from both image and SMILES data.
• Trained and evaluated ISMol on 14 small molecule ADMET datasets.

Main Results:

• ISMol significantly outperformed traditional machine learning and deep learning models using single-modal representations.
• Experimental analysis confirmed ISMol's superiority, interpretability, and generalizability.
• The model demonstrated robust performance across various molecular property predictions.

Conclusions:

• ISMol provides a powerful deep learning framework for drug discovery.
• Combining image and SMILES data offers a synergistic approach to molecular property prediction.
• The developed architecture advances the field of computational drug design.