Rapid prediction of full spin systems using uncertainty-aware machine learning.

Area of Science:

• Computational Chemistry
• Machine Learning
• Spectroscopy

Background:

• Accurate simulation of solution Nuclear Magnetic Resonance (NMR) spectra is crucial for molecular structure determination.
• Traditional methods rely on heuristic techniques or ab initio computational chemistry, which can be computationally intensive or lack comprehensive prediction capabilities.

Purpose of the Study:

• To develop a novel machine learning technique for accurate prediction of chemical shift and scalar coupling parameters in NMR spectra.
• To introduce an uncertainty-aware deep learning approach for robust spectral simulations.
• To improve upon existing state-of-the-art methods in predicting NMR spectral parameters.

Main Methods:

• A novel machine learning technique combining uncertainty-aware deep learning with rapid conformational geometry estimation.
• Development of Full Spin System Predictions with UnCertainty (FullSSPrUCe) model.
• Utilizing disagreement regularization to augment experimental data with ab initio data.

Main Results:

• Achieved high accuracy in predicting chemical shift values: protons within 0.209 ppm and carbons within 1.213 ppm.
• Successfully predicted all scalar coupling values, including ³J_HH with accuracies between 0.838 Hz and 1.392 Hz.
• Demonstrated a strong correlation between uncertainty quantification and prediction accuracy, with top predictions showing significantly reduced error.

Conclusions:

• The FullSSPrUCe method offers a significant advancement in the accuracy and reliability of NMR spectral simulations.
• The uncertainty quantification provides a valuable measure of prediction confidence, essential for experimental data interpretation.
• The approach effectively handles stereoisomerism and integrates diverse data sources for improved model performance.