Bidirectional Graphormer for Reactivity Understanding: Neural Network Trained to Reaction Atom-to-Atom Mapping Task.

Area of Science:

• Computational chemistry
• Artificial intelligence in chemistry
• Machine learning for chemical reactions

Background:

• Accurate atom-to-atom mapping (AAM) is crucial for understanding and predicting chemical reactions.
• Existing AAM methods often rely on sequential representations of molecules, which may not fully capture complex structural information.
• There is a need for advanced algorithms that can directly process molecular graph structures for improved AAM accuracy.

Purpose of the Study:

• To introduce GraphormerMapper, a novel algorithm for reaction atom-to-atom mapping.
• To leverage graph transformer neural networks and BERT for direct molecular graph processing.
• To demonstrate the superiority of GraphormerMapper compared to existing state-of-the-art AAM algorithms.

Main Methods:

• Developed GraphormerMapper, integrating a graph transformer for feature extraction and a BERT network for chemical transformation learning.
• Employed a transformer neural network designed for direct processing of molecular graphs (atoms and bonds).
• Utilized Bidirectional Encoder Representations from Transformers (BERT) for learning chemical transformations.

Main Results:

• GraphormerMapper demonstrated superior performance in atom-to-atom mapping.
• Benchmarking against IBM RxnMapper, the previously best-performing AAM algorithm, showed significant improvements.
• The algorithm's effectiveness was validated on a "Golden" benchmarking dataset.

Conclusions:

• GraphormerMapper represents a significant advancement in reaction atom-to-atom mapping technology.
• Directly processing molecular graphs with graph transformers and BERT offers enhanced accuracy for AAM.
• The developed algorithm sets a new benchmark for computational approaches to chemical reaction analysis.