Enhancing Retrosynthetic Reaction Prediction with Deep Learning Using Multiscale Reaction Classification.

Area of Science:

• Computational Chemistry
• Artificial Intelligence in Chemistry
• Drug Discovery

Background:

• Chemical synthesis planning is crucial for drug discovery.
• Machine learning and AI show promise for improving computational retrosynthetic analysis.
• Existing methods require significant expert input and can be time-consuming.

Purpose of the Study:

• To develop and validate a multiscale, data-driven approach for retrosynthetic analysis using deep highway networks (DHN).
• To enhance the prediction accuracy of chemical synthesis planning compared to conventional methods.
• To automate the extraction of reaction rules from chemical reaction data.

Main Methods:

• Developed a multiscale DHN model for retrosynthetic analysis.
• Extracted reaction rules automatically from a dataset of chemical reactions from U.S. patents.
• Employed a two-step DHN prediction process: identifying reaction groups, then predicting specific transformation rules.
• Validated the model on 40 approved drugs, comparing results against a control model trained on all extracted rules.

Main Results:

• The multiscale DHN approach achieved an 82.9% success rate in predicting valid reactants for the first retrosynthetic step.
• This represents a significant improvement over the control model, which had a 58.5% success rate.
• The model demonstrated a marked enhancement in retrosynthetic analysis accuracy.

Conclusions:

• The multiscale, data-driven approach offers a significant advancement in computational retrosynthetic analysis.
• While not outperforming expert-curated systems, it provides a robust, data-driven alternative.
• The methodology is adaptable for various artificial intelligence strategies in chemistry.