Using test-time augmentation to investigate explainable AI: inconsistencies between method, model and human...

Area of Science:

• Computational toxicology
• Cheminformatics
• Artificial intelligence

Background:

• Machine learning models require explainable artificial intelligence (XAI) for human-understandable interpretations.
• Text-based molecular representations are crucial for transfer learning in computational toxicity.
• Augmenting molecular representations aids in comparing model outputs for the same data.

Purpose of the Study:

• To investigate the robustness of eight XAI methods using test-time augmentation for molecular representation models in computational toxicity prediction.
• To assess the consistency and reliability of XAI explanations for identical molecular structures under different representations.

Main Methods:

• Utilized test-time augmentation on text-based molecular representations.
• Evaluated eight different explainable artificial intelligence (XAI) methods.
• Compared explanations generated for multiple representations of the same molecular structure.
• Analyzed the variance between in-domain and out-of-domain predictions.
• Assessed the importance assigned to expert-derived structural alerts.

Main Results:

• Significant differences were observed in explanations generated for different representations of the same molecule.
• Randomized models exhibited similar variance in explanations compared to standard models.
• XAI measures showed greater variance for in-domain predictions than out-of-domain predictions.
• Expert-derived structural alerts received similar importance across various conditions, irrespective of applicability domain or model randomization.
• Inconsistencies were found within models for identical molecular representations.

Conclusions:

• Current XAI methods may not reliably reflect learned parameters in text-based molecular representations, potentially indicating a reliance on tokenization.
• Test-time augmentation is a valuable tool for assessing the consistency of XAI in computational toxicology.
• Researchers should validate XAI methods by comparing them to human intuition and expert knowledge, especially when using text-based molecular representations.