Conditional Molecular Generation Net Enables Automated Structure Elucidation Based on ¹³C NMR Spectra and Prior...

Area of Science:

• Computational Chemistry
• Organic Chemistry
• Machine Learning

Background:

• Structure elucidation of unknown compounds using Nuclear Magnetic Resonance (NMR) is a persistent challenge in synthetic organic and natural product chemistry.
• Existing library matching methods are limited by library coverage and often neglect prior knowledge like molecular fragments.

Purpose of the Study:

• To develop an advanced computational method that integrates multiple data sources for improved structure elucidation.
• To overcome the limitations of traditional library matching by incorporating diverse chemical information.

Main Methods:

• Introduction of a conditional molecular generation net (CMGNet) that accepts ¹³C NMR spectra, molecular formulas, and molecular fragments as input conditions.
• Large-scale pretraining for molecular understanding followed by fine-tuning on two NMR spectral datasets of varying granularity.
• A generative deep learning model designed for structure elucidation tasks.

Main Results:

• CMGNet achieved a 94.17% recovery rate in the top 10 recommendations for structure elucidation.
• The model demonstrated strong performance across diverse compound classes and in structural revision tasks.
• CMGNet exhibits a profound understanding of molecular connectivities derived from ¹³C NMR, molecular formulas, and fragments.

Conclusions:

• CMGNet represents a novel deep learning-assisted approach for solving inverse problems in chemistry.
• The model's ability to integrate multiple data sources enhances the accuracy and efficiency of structure elucidation.
• This work paves the way for a new paradigm in AI-driven chemical structure determination.