Permutation Invariant Graph-to-Sequence Model for Template-Free Retrosynthesis and Reaction Prediction

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Data-driven approaches are crucial for synthesis planning and reaction outcome prediction in computer-aided organic chemistry.
• Current natural language processing (NLP) models, while effective, struggle with the inherent structural information of molecules using SMILES representations.
• SMILES augmentation is often required to improve empirical performance, indicating limitations in standard molecular representations.

Purpose of the Study:

• To develop a novel end-to-end model, Graph2SMILES, that leverages molecular graph structures for improved chemical predictions.
• To mitigate the need for data augmentation by integrating permutation-invariant graph encoders with Transformer architectures.
• To enhance the performance of retrosynthesis and reaction outcome prediction tasks.

Main Methods:

• Developed Graph2SMILES, an end-to-end architecture combining Transformer models with molecular graph encoders.
• Utilized a directed message passing neural network (D-MPNN) within the encoder to capture local chemical environments.
• Incorporated a global attention encoder with graph-aware positional embedding to model long-range and intermolecular interactions.

Main Results:

• Graph2SMILES demonstrated improved performance on existing benchmarks for both retrosynthesis and reaction outcome prediction.
• The model acts as a drop-in replacement for standard Transformer models in molecule-to-molecule transformation tasks.
• The graph-based approach inherently handles molecular structure information, reducing reliance on data augmentation.

Conclusions:

• Graph2SMILES offers a powerful and versatile approach for advancing computer-aided organic chemistry.
• The integration of graph neural networks and Transformer models represents a significant step forward in predicting chemical reactions and planning syntheses.
• This novel architecture enhances the efficiency and accuracy of data-driven methods in organic chemistry.