XSMILES

XSMILES visualizes attribution scores from explainable artificial intelligence (XAI) methods on SMILES strings and corresponding 2D molecular structures to support interpretation of model predictions in cheminformatics.

Key Features:

• Interactive Visualization: Maps attribution scores associated with both atom and non-atom SMILES tokens onto a 2D molecular diagram for coordinated inspection.
• Complex Notation Handling: Supports visualization and interpretation of non-atom SMILES tokens, including substrings like [N+], alongside atom tokens.
• Integrated Bar Chart: Coordinates a bar chart of the SMILES string with the molecular diagram to link token-level scores to structural elements.
• Customizable Parameters: Provides adjustable visualization parameters to modify the mapping and display of attribution scores.

Scientific Applications:

• Model Interpretation: Interpret how atom and non-atom tokens contribute to property predictions from machine learning models.
• Comparative Analysis: Compare attribution patterns across multiple modeling approaches to evaluate differences in explanations.
• Pattern Identification: Identify structural or token-level patterns that correlate with predicted chemical properties.

Methodology:

Applies XAI techniques to compute attribution scores for atom and non-atom tokens in SMILES strings and visualizes these scores on a coordinated 2D molecule diagram and SMILES bar chart.