What can attribution methods show us about chemical language models?

Area of Science:

• Computational Chemistry
• Machine Learning
• cheminformatics

Background:

• Language models excel at molecular string prediction but lack explainability for practical applications.
• Explainability is crucial for understanding the rationale behind chemical model predictions.

Purpose of the Study:

• To investigate explainability techniques for chemical language models.
• To adapt and compare transformer-specific and model-agnostic input attribution methods.
• To analyze model representations for predicting aqueous solubility.

Main Methods:

• Fine-tuning a pretrained language model for aqueous solubility prediction.
• Adapting SHAP (SHapley Additive exPlanations) and a transformer-specific technique for input attribution.
• Evaluating and visualizing attributed relevance across different model variants.

Main Results:

• SHAP attributions highlighted individual electronegative atoms but not functional groups.
• Transformer-specific attributions were sparse and mapped to the model's latent space, suggesting molecular similarity features.
• Neither method directly explained predictions based on solubility-relevant functional groups.

Conclusions:

• Chemical language models may represent molecular similarity rather than explicit functional groups in their latent space.
• Understanding these internal representations is key to developing more accurate and explainable chemical AI.
• Leveraging these insights can inform the design of better chemical spaces for advanced model training.