NMRSCOUT: Automated Annotation of Complex Mixtures from Multidimensional NMR Spectra Using Predictive Modelling and Statistical Filtering

Fast multi-dimensional NMR methods are increasingly used in untargeted metabolomics studies. By mitigating overlap, they offer improved statistical discrimination and/or facilitated biomarker annotation compared to 1D NMR. However, annotating the resulting spectra remains difficult due to limited databases and annotation tools supporting such methods. We introduce NMRSCOUT, an automated workflow for annotating 1D and 2D NMR spectra, including conventional and fast 2D experiments. Using a user-provided library of candidate structures, NMRSCOUT predicts 1H and 13C chemical shifts via a hybrid model combining Morgan fingerprints and a graph neural network, reconstructs corresponding 1D/2D spectra, and compares them to experimental data. Statistical filtering (STOCSY, RANSY) leverages signal relationships across sample series to reduce false matches, while complementary NMR experiments are integrated into a unified annotation score. NMRSCOUT was evaluated using 20 Aspergillus chevalieri extracts analyzed via 1D 1H, 13C-INEPT, ultrafast COSY (UF-COSY), and HSQC-SYMAPS, using a library of 3,250 candidate structures. Combining all four experiments narrowed high-scoring candidates (scores in the [90-100] range) to 1.8% (57 structures), significantly reducing ambiguity. Top-ranked candidates aligned with compounds previously reported for A. chevalieri, and most of them were also observed in the LC-HRMS profiles of the same extracts. Thus, NMRSCOUT provides a versatile tool for annotation of NMR spectra from highly complex mixtures, taking advantage of the convergence of spectra from multiple NMR acquisition sequences. The NMRSCOUT software is available at https://github.com/serignebamba398pixel/NMRSCOUT.