PINS: A Physics-Informed Generative Framework for De Novo Structure Elucidation from 1D NMR Spectra.
TL;DR
PINS (Physics-Informed NMR Structure elucidation model), a generative framework that explicitly bridges the gap between spectral data and molecular topology by enforcing multiphysical priors, provides a trustworthy, automated strategy for decoding novel chemical structures in data-scarce regimes.
Abstract
1D Nuclear Magnetic Resonance (NMR) spectra offer a rapid and accessible alternative to time-consuming 2D experiments, making them ideal for high-throughput structure elucidation. However, reconstructing molecular topology solely from 1D NMR spectral data remains a formidable combinatorial challenge. This difficulty arises from the loss of explicit atomic connectivity information, and existing data-driven models generate chemically invalid or hallucinated structures. Here, we present PINS (Physics-Informed NMR Structure elucidation model), a generative framework that explicitly bridges the gap between spectral data and molecular topology by enforcing multiphysical priors. By constraining the generative search space within strict physical laws, PINS effectively mitigates structural hallucinations and helps resolve structural ambiguities compared with end-to-end deep learning models. PINS achieves 100% chemical validity and surpasses the state-of-the-art (SOTA) baseline by a substantial margin of 21.6 percentage points (a 68.9% relative improvement). Furthermore, PINS demonstrates robust extrapolation capabilities on novel molecular scaffolds within the specific domain of challenging new psychoactive substances (NPS), achieving an 89.7% identification rate in forensic scenarios. Our findings underscore the necessity of physical constraints in generative AI, providing a trustworthy, automated strategy for decoding novel chemical structures in data-scarce regimes.