TransPeakNet for solvent-aware 2D NMR prediction via multi-task pre-training and unsupervised learning.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Spectroscopy

Background:

• Nuclear Magnetic Resonance (NMR) spectroscopy is crucial for determining molecular structure, electronic properties, and dynamics.
• Accurate NMR shift prediction aids in validating molecular structures by comparing experimental and theoretical values.
• Machine learning (ML) has advanced 1D NMR shift prediction, but 2D NMR prediction faces challenges due to limited annotated datasets.

Purpose of the Study:

• To develop an unsupervised training framework for predicting cross-peaks in 2D Heteronuclear Single Quantum Coherence (HSQC) NMR spectra.
• To address the data scarcity issue in 2D NMR prediction by utilizing unlabeled HSQC data.
• To improve the accuracy of NMR shift prediction and aid in structural elucidation.

Main Methods:

• Pretraining an ML model on an annotated 1D NMR dataset (¹H and ¹³C shifts).
• Finishing the model in an unsupervised manner using unlabeled HSQC data to generate cross-peak annotations.
• Incorporating solvent effect adjustments into the prediction model.

Main Results:

• Achieved Mean Absolute Errors (MAEs) of 2.05 ppm for ¹³C shifts and 0.165 ppm for ¹H shifts.
• Demonstrated superior performance compared to traditional methods like ChemDraw and Mestrenova.
• Algorithmic annotations showed 95.21% concordance with expert assignments on 479 HSQC spectra.

Conclusions:

• The unsupervised framework effectively predicts 2D NMR cross-peaks, overcoming data limitations.
• The model's high concordance with expert assignments highlights its potential for accurate structural elucidation.
• This approach offers significant advantages for organic chemistry, pharmaceuticals, and natural product research.