Prediction and Interpretable Visualization of Retrosynthetic Reactions Using Graph Convolutional Networks

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Organic Synthesis

Background:

• Data-driven methods for retrosynthetic analysis are advancing with machine learning.
• Challenges remain in prediction accuracy and the interpretability of neural network models.
• Practical adoption by chemists requires addressing these limitations.

Purpose of the Study:

• To enhance retrosynthetic reaction prediction accuracy.
• To improve the interpretability of data-driven prediction models.
• To facilitate the practical use of AI in chemical synthesis.

Main Methods:

• Developed an interpretable prediction framework using graph convolutional networks (GCN).
• Integrated gradients (IG) were employed for visualizing prediction contributions.
• Compared performance against extended-connectivity fingerprint methods.

Main Results:

• The proposed GCN model demonstrated superior performance in balanced accuracies compared to existing fingerprint-based approaches.
• Integrated gradients successfully visualized and highlighted atoms crucial for the predicted reaction.
• The framework offers enhanced capability for retrosynthetic reaction prediction.

Conclusions:

• The interpretable GCN framework effectively addresses key challenges in data-driven retrosynthesis.
• Visualization of atom contributions enhances model transparency for chemists.
• This approach promotes wider adoption of machine learning in synthetic chemistry.