Cross-Modal Retrieval Between ¹³C NMR Spectra and Structures Based on Focused Libraries.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Spectroscopy

Background:

• Compound identification relies on matching ¹³C NMR spectra with spectral libraries.
• Previous deep contrastive learning systems (CReSS) faced limitations with large, redundant libraries and lack of unknown structures.

Purpose of the Study:

• To enhance structure-elucidation accuracy in compound identification.
• To address limitations of traditional spectral libraries in deep learning models.
• To develop a more efficient method for cross-modal retrieval between ¹³C NMR spectra and chemical structures.

Main Methods:

• Replaced traditional large libraries with focused libraries generated by CMGNet.
• Employed a deep contrastive learning system (CReSS) with focused libraries.
• Introduced SAmpRNN, a recurrent neural network, to amplify focused libraries.
• Evaluated the combined model on 6,471 ¹³C NMR spectra.

Main Results:

• The combined model achieved a Top-10 accuracy of 54.03%, significantly outperforming CReSS with a random library (9.17%).
• SAmpRNN amplification increased structure-identification accuracy in 70.0% of 30 random cases.
• Focused libraries (CFLS) provided more accurate candidate structures than traditional libraries.

Conclusions:

• Focused libraries significantly improve cross-modal retrieval accuracy for ¹³C NMR spectra and structures.
• The combined CReSS and CMGNet model, enhanced by SAmpRNN, offers a powerful tool for compound identification.
• This approach provides a more accurate and efficient alternative to traditional library matching methods.