Predicting chemical shifts with graph neural networks

Area of Science:

• Computational Chemistry
• Structural Biology
• Machine Learning

Background:

• Accurate forward models are crucial for inferring molecular structure from Nuclear Magnetic Resonance (NMR) data.
• Existing models often lack differentiability and are limited to specific molecular types, hindering advanced computational methods.
• Current empirical NMR models require extensive domain expertise and feature engineering.

Purpose of the Study:

• To develop a novel, data-driven forward model for NMR chemical shift prediction using graph neural networks (GNNs).
• To create a differentiable and versatile model capable of analyzing arbitrary molecular structures, including proteins and organic molecules.
• To enable gradient-based molecular dynamics simulations using NMR data.

Main Methods:

• Implementation of graph neural networks (GNNs) for direct prediction of chemical shifts from 3D molecular structures.
• Training GNN models on experimental NMR data, bypassing the need for manual feature engineering.
• Evaluation of model accuracy in capturing phenomena such as hydrogen bonding and secondary structure effects.

Main Results:

• The GNN model demonstrates high accuracy in predicting NMR chemical shifts across various molecular types.
• The model effectively captures complex chemical phenomena, including hydrogen bonding and secondary structure influences.
• The GNNs are computationally efficient, capable of predicting one million chemical shifts in approximately 5 seconds.

Conclusions:

• Graph neural networks offer a powerful, data-driven alternative to traditional empirical models for NMR chemical shift prediction.
• This approach enables accurate analysis of diverse molecular structures without domain-specific feature engineering.
• The developed GNNs facilitate new computational approaches for studying interacting macromolecules and molecular dynamics.