Retrosynthesis prediction with an interpretable deep-learning framework based on molecular assembly tasks.

Area of Science:

• Organic Chemistry
• Artificial Intelligence
• Computational Chemistry

Background:

• Automating retrosynthesis accelerates organic chemistry research in digital labs.
• Existing deep learning models for retrosynthesis often function as "black boxes" with limited interpretability.

Purpose of the Study:

• To develop an interpretable artificial intelligence model for automating retrosynthesis.
• To enhance the explainability and transparency of deep learning approaches in organic synthesis.

Main Methods:

• Formulating retrosynthesis as a molecular assembly process guided by deep learning.
• Utilizing a multi-sense and multi-scale Graph Transformer.
• Implementing structure-aware contrastive learning and dynamic adaptive multi-task learning.

Main Results:

• RetroExplainer outperforms state-of-the-art single-step retrosynthesis methods on 12 benchmark datasets.
• The molecular assembly process provides interpretability and quantitative attribution.
• Identified 101 multi-step retrosynthesis pathways, with 86.9% of reactions reported in literature.

Conclusions:

• RetroExplainer offers a reliable, interpretable, and high-throughput solution for organic synthesis.
• The model provides valuable insights for drug development and chemical research.
• Enhances transparency in AI-driven retrosynthesis for reproducible scientific discovery.