Quantitative interpretation explains machine learning models for chemical reaction prediction and uncovers bias

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Organic synthesis

Background:

• Organic synthesis is vital for drug discovery but presents significant challenges.
• Current machine learning models for reaction prediction are often "black boxes," lacking transparency in their decision-making processes.
• This opacity hinders trust and effective use by both developers and researchers.

Purpose of the Study:

• To quantitatively interpret the Molecular Transformer, a leading model for reaction prediction.
• To develop a framework for attributing predictions to specific reactant features and training data.
• To identify and address dataset biases leading to inaccurate model performance.

Main Methods:

• Development of an interpretation framework to attribute reaction predictions.
• Analysis of model predictions to identify reliance on specific chemical features and training examples.
• Identification of "Clever Hans" predictions caused by dataset biases.
• Creation of a new, debiased dataset for reaction prediction benchmarking.

Main Results:

• The developed framework successfully attributes reaction predictions to specific molecular substructures and training set reactions.
• The study identified instances where models achieved correct predictions for incorrect reasons due to dataset biases.
• A new debiased dataset was created, offering a more accurate evaluation of model performance.

Conclusions:

• Interpretable AI is essential for advancing machine learning in organic synthesis and drug discovery.
• Addressing dataset bias is critical for developing reliable and trustworthy reaction prediction models.
• The proposed debiased dataset serves as a new standard for benchmarking future reaction prediction models.