DRACON: disconnected graph neural network for atom mapping in chemical reactions

Area of Science:

• Computational chemistry
• Machine learning applications in chemical synthesis

Background:

• Computer-assisted synthesis prediction (CASP) remains a significant challenge.
• Existing machine learning models have limitations in accurately predicting chemical reactions.

Purpose of the Study:

• To develop and validate a generalized graph convolution neural network for predicting chemical reactions.
• To accurately identify reaction centers and product atoms in molecular graphs.

Main Methods:

• Formulating reaction prediction as node classification on disconnected graphs.
• Generalizing graph convolution neural networks for disconnected graph structures.
• Utilizing the USPTO dataset for experimental validation.

Main Results:

• The proposed model successfully predicts reaction centers and product atoms.
• Achieved excellent performance and interpretability on the USPTO dataset.
• Demonstrated that learned latent vector representations correlate with chemical reaction classes.

Conclusions:

• The developed graph convolution neural network is effective for computer-assisted synthesis prediction.
• The model provides interpretable insights into chemical reaction patterns.
• Latent space analysis reveals meaningful clustering of similar chemical reactions.