Predictive Modeling of NMR Chemical Shifts without Using Atomic-Level Annotations

Area of Science:

• Computational Chemistry
• Machine Learning
• Spectroscopy

Background:

• Accurate prediction of nuclear magnetic resonance (NMR) chemical shifts is crucial in chemistry.
• Existing machine learning methods require extensive manual atomic-level annotations, hindering large-scale applications.
• The laborious nature of manual annotation limits the efficiency of developing predictive models.

Purpose of the Study:

• To develop a weakly supervised learning method for predicting NMR chemical shifts.
• To eliminate the requirement for explicit atomic-level annotations in training data.
• To enable efficient and large-scale predictive modeling of NMR chemical shifts.

Main Methods:

• A message passing neural network (MPNN) was employed as the prediction model.
• The model was trained using a weakly supervised approach with molecular-level chemical shift annotations.
• A permutation-invariant loss function minimized the difference between predicted and actual molecular-level chemical shifts.

Main Results:

• The proposed method achieves comparable performance to fully supervised methods for 1H and 13C NMR chemical shift prediction.
• Weakly supervised learning successfully predicted chemical shifts without atomic-level annotations.
• The model aligns predicted chemical shifts with actual values in a data-driven manner.

Conclusions:

• Weakly supervised learning offers a viable alternative for NMR chemical shift prediction.
• The developed method significantly reduces the burden of data annotation for machine learning models.
• This approach facilitates broader application of machine learning in NMR spectroscopy.