Scalable graph neural network for NMR chemical shift prediction

Area of Science:

• Computational Chemistry
• Machine Learning for Chemistry

Background:

• Graph neural networks (GNNs) excel at predicting nuclear magnetic resonance (NMR) chemical shifts.
• Current GNN methods face limitations due to high space complexity, restricting their use to smaller molecules.

Purpose of the Study:

• To propose a scalable GNN for accurate NMR chemical shift prediction.
• To address the space complexity limitations of existing GNN models.

Main Methods:

• Sparsified molecular graph representation using only heavy atoms and bonds as nodes and edges.
• Enhanced GNN message passing with attention mechanisms and residual connections.
• Improved readout function incorporating node-level and graph-level embeddings.

Main Results:

• The proposed GNN demonstrates higher prediction accuracy on 13C and 1H NMR datasets.
• The method exhibits improved scalability for predicting chemical shifts in large molecules.
• Reduced space complexity compared to traditional GNN approaches.

Conclusions:

• The developed scalable GNN offers a more efficient and accurate approach for NMR chemical shift prediction.
• This method expands the applicability of GNNs to larger and more complex molecular systems.
• The enhanced message passing and readout functions are key to improved performance.