XSMILES: interactive visualization for molecules, SMILES and XAI attribution scores

Area of Science:

• Computational Chemistry
• Cheminformatics
• Artificial Intelligence

Background:

• Explainable artificial intelligence (XAI) is increasingly used in chemistry for property prediction.
• Visualizing attribution scores on molecular diagrams is challenging for non-atom tokens in SMILES strings.
• SMILES notation complexity necessitates advanced techniques for analyzing token-based explanations.

Purpose of the Study:

• To introduce XSMILES, an interactive visualization technique for exploring XAI attribution scores.
• To facilitate the interpretation of explanations associated with both atom and non-atom tokens in SMILES strings.
• To improve the understanding of model behavior in chemical applications.

Main Methods:

• Developed XSMILES, an interactive visualization tool.
• Integrated 2D molecule diagrams with bar charts representing SMILES strings.
• Enabled visualization of attribution scores for atom and non-atom SMILES tokens.

Main Results:

• Demonstrated the evaluation and interpretation of SMILES attributions through interactive visualization.
• Showcased the utility of XSMILES in two distinct use cases.
• Validated the ability to represent scores on both molecular diagrams and SMILES string representations.

Conclusions:

• XSMILES aids data scientists in understanding and comparing AI models.
• The tool offers adaptable visualization parameters and platform integration.
• XSMILES facilitates pattern identification and attribution comparison for model development and communication.