CNN-based two-branch multi-scale feature extraction network for retrosynthesis prediction.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry

Background:

• Retrosynthesis prediction is crucial for designing chemical synthesis routes.
• Existing models often use single molecular descriptors and fail to capture multi-scale features.
• This limits the full utilization of molecular and descriptor information.

Purpose of the Study:

• To develop a novel model for retrosynthesis prediction that addresses limitations of existing methods.
• To effectively extract and fuse multi-scale features from various molecular descriptors.
• To perform predictions without requiring expert chemical knowledge.

Main Methods:

• A convolutional neural network (CNN) was designed for multi-scale feature extraction from molecular descriptors using filters of varying sizes.
• A two-branch feature extraction layer was employed to fuse these multi-scale features.
• The model was evaluated on the USPTO-50k chemical dataset and the MetaNetX metabolic dataset.

Main Results:

• The proposed model achieved superior performance on the USPTO-50k dataset, outperforming the state-of-the-art by up to 12.2% in top-10 accuracy.
• The model demonstrated feasibility in bioretrosynthesis prediction on the MetaNetX dataset, achieving top-10 accuracy of 82.2%.
• Significant improvements were observed across top-1, top-3, top-5, and top-10 accuracy metrics.

Conclusions:

• The developed model significantly surpasses existing state-of-the-art methods for chemical retrosynthesis prediction.
• The study validates the applicability of retrosynthesis prediction models to bioretrosynthesis tasks.
• The approach effectively leverages multi-scale features for enhanced predictive power.