Prediction of Compound Synthesis Accessibility Based on Reaction Knowledge Graph.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in chemistry

Background:

• Deep learning generative models are increasingly used for de novo molecule design.
• Quantitative estimation of molecular synthetic accessibility (SA) is crucial for prioritizing generated structures.
• SA estimation aids in hit/lead compound prioritization and retrosynthesis analysis.

Purpose of the Study:

• To develop and evaluate deep learning models for predicting molecular synthetic accessibility.
• To compare the performance of novel models against existing SA scoring schemes.

Main Methods:

• Construction of a chemical reaction network using USPTO and Pistachio datasets.
• Identification of shortest reaction paths (SRP) to classify compounds as easy-to-synthesize (ES) or hard-to-synthesize (HS).
• Development of two synthesis accessibility models: DNN-ECFP and graph-based CMPNN.

Main Results:

• The CMPNN model achieved a ROC AUC of 0.791, outperforming SYBA (0.76), SAScore, and SCScore.
• The DNN-ECFP model also demonstrated predictive capabilities for molecular SA.
• The models leverage historical reaction knowledge for SA prediction.

Conclusions:

• The developed prediction models, particularly CMPNN, show strong performance in estimating molecular SA.
• These models can serve as valuable tools for prioritizing molecules in drug discovery and de novo design pipelines.
• Utilizing reaction network analysis and deep learning offers a promising approach for assessing synthetic accessibility.