PRISM, a Bidirectional NMR Foundation Model: From Peak List to Structure and Back

Determining molecular structures from nuclear magnetic resonance (NMR) spectra is a central challenge in organic chemistry, natural-product discovery, and drug development. Established computational approaches are robust within their design scope, but direct spectra-to-structure models have relied largely on simulated spectra or narrow chemical subsets, limiting robustness on real data. We introduce PRISM (Peak-Resolved Integrative Spectral Model), a unified transformer encoder-decoder bridging experimental spectra and molecular structures bidirectionally: spec2mol translates ¹H and ¹³C peak lists into SMILES strings, while mol2spec predicts full spectra from SMILES. Spec2mol integrates six chemistry-aware modules that embed domain knowledge into learning. Trained on 1.19 million experimental NMR-structure pairs spanning 49 atom types, spec2mol achieves 60.1% stereo-aware (72.5% stereo-unaware) top-1 exact match accuracy, 16.7 percentage points above the strongest published deep learning (DL) models under identical conditions on the same test set. Mol2spec attains ¹H MAE of 0.036 ppm and ¹³C MAE of 0.358 ppm in under 0.3 s per molecule. Fine-tuned on photochemical reactions, spec2mol achieved 69.2% (85.2%) top-1 accuracy on real spectra. Together, PRISM is a practical decision-support system that proposes and ranks candidate structures for expert verification, and a fast forward predictor for reaction analysis and virtual screening. PRISM is freely available at https://prism.uni-muenster.de/.