End-to-end molecular structure elucidation from multimodal NMR spectra images using vision transformers

Area of Science:

• Chemistry
• Spectroscopy
• Artificial Intelligence

Background:

• Nuclear magnetic resonance (NMR) spectroscopy is crucial for molecular structure elucidation.
• Interpreting complex NMR spectra is challenging, often requiring expert knowledge and heuristic methods.
• Existing deep learning methods for spectrum-to-structure prediction often rely on peak annotations, losing valuable intensity information and limiting applicability to higher-dimensional NMR.

Purpose of the Study:

• To develop a deep learning framework, NMRViT, that directly processes raw NMR spectral signals for molecular structure elucidation.
• To enable end-to-end structure prediction from both 1D (¹H, ¹³C) and 2D (HSQC) NMR spectra.
• To establish a benchmark for HSQC-based structure prediction and evaluate transferability from simulated to experimental data.

Main Methods:

• Developed NMRViT, a spectral Vision Transformer framework operating directly on raw NMR spectral data.
• Trained the model on a large-scale simulated NMR dataset.
• Evaluated zero-shot transfer and fine-tuning performance on experimental 1D and 2D NMR datasets.
• Introduced a chemical-shift-based post-processing strategy for re-ranking candidate structures.

Main Results:

• NMRViT demonstrated strong performance on single-modality and multimodal NMR spectral inputs.
• Provided an end-to-end benchmark for HSQC-based structure prediction.
• Highlighted the simulation-experiment domain gap, which can be substantially reduced by fine-tuning with minimal experimental data.
• The post-processing strategy consistently improved candidate ranking in both zero-shot and fine-tuning scenarios.

Conclusions:

• Raw-spectrum vision transformers offer a practical framework for automated molecular structure elucidation.
• Lightweight experimental adaptation and chemically informed re-ranking enhance the utility of NMRViT.
• The developed approach advances automated structure determination from multimodal NMR data.