Automated identification of metabolites from nuclear magnetic resonance (NMR) spectra of mixtures remains the recognized bottleneck of NMR-based metabolomics. A 2015 pattern-recognition method (Dubey, Rangarajan, Pal, and Atreya, Analytical Chemistry) matched geometric patterns of spin-system peaks against reference data instead of fixed chemical-shift tolerances, achieving correct assignment of all 719 tested Human Metabolome Database (HMDB) metabolites on accurate data and substantial robustness to missing peaks and large shift deviations, but its geometric-hashing design requires at least three peaks per spin system to define a matching pattern, its experimental validation covered only a 16-metabolite mixture, it did not integrate multiple NMR experiments, and, like any database-matching method, it inherited every limitation of database completeness. Later work closed individual facets — multi-spectra co-analysis (COLMARm), peak-picking-free global library fitting (minimum-cost-flow), transformer-based 1D peak assignment at biofluid scale (NMRformer), overlap-tolerant spin-system characterization via 3D TOCSY J-resolved spectroscopy — while field reviews in 2022 and 2025 still describe compound identification, and especially the identification of unknowns outside reference databases, as an open bottleneck. The surviving open problem is a jointly validated identification capability that is robust to the documented failure modes at real biofluid scale, with calibrated confidence and principled treatment of metabolites absent from the database.

Background

In NMR-based metabolomics, small molecules in biofluids or cell extracts are quantified by their nuclear magnetic resonance spectra, most commonly one-dimensional proton spectra (1^{1}H 1D) and two-dimensional correlation spectra such as TOCSY (total correlation spectroscopy, which spreads the resonances of each coupled proton network across two frequency axes) and HSQC (heteronuclear single-quantum coherence, which correlates each proton with the carbon it is attached to). A spin system is the group of NMR peaks belonging to one connected set of coupled nuclei within a molecule; identifying a metabolite from a mixture spectrum amounts to matching its observed spin-system peaks against reference data such as the Human Metabolome Database (HMDB), a curated library of measured chemical shifts for human metabolites. The classical approach searches the database within a fixed chemical-shift tolerance for each peak. This fails under real conditions: chemical shifts move with sample pH, temperature, and ionic strength, peaks go missing because of low concentration or overlap, and in a real biofluid hundreds of spin systems interleave in a narrow spectral window. Dubey, Rangarajan, Pal, and Atreya (Analytical Chemistry, 2015) proposed instead to match the relative pattern of peaks — using geometric hashing, a technique from computer vision that indexes point configurations under a canonical basis so that configurations can be retrieved regardless of absolute displacement — combined with a similarity score. On simulated data for 719 HMDB metabolites the method assigned 100% correctly when the data were accurate, and it kept a high success rate under stress conditions of up to 50% missing peaks and chemical-shift deviations as large as 0.5 ppm in proton and 3 ppm in carbon dimension, conditions under which tolerance-based database search returns essentially nothing; on experimental data it was validated on a mixture of 16 metabolites measured at eight combinations of pH and temperature. The limitations that define this problem follow from the method's design and from the field's reviews: geometric hashing needs at least three peaks to define a matching basis, so spin systems with fewer observable peaks — including experimentally weak or heavily overlapped systems — cannot be matched this way; the experimental validation is small (16 metabolites) relative to real biofluids containing hundreds of overlapping spin systems; deconvolution of crowded spectra is still an open problem and there is no standard way to estimate the confidence of a match; no platform explicitly integrates evidence from multiple complementary NMR experiments into a single identification decision; and no matching method can identify metabolites that have no reference pattern, so reference-database incompleteness itself bounds any matching approach. These limitations are not specific to one algorithm: the field's reviews (Judge and Ebbels, Metabolomics, 2022; Nagana Gowda, Zhu, and Raftery, Progress in NMR Spectroscopy, 2025) classify compound identification — separating reliable identification (hypothesis testing) from annotation (hypothesis generation) — as the critical, still-open bottleneck of NMR metabolomics, with deconvolution of crowded spectra, calibration of annotation confidence, and dependence on reference-database coverage named as the core difficulties. Solving the problem thus means building and validating identification capability that survives these documented failure modes, not merely improving one scoring function.

Problem Statement

The problem is to develop automated identification of metabolites from NMR spectra of complex biofluid mixtures that is robust to the failure modes documented for tolerance-based database matching and for the 2015 pattern-recognition approach of Dubey et al., and to validate that capability on real samples. Concretely, an identification capability is sought that: (1) handles spin systems with fewer than three observable peaks as well as heavily overlapped and weak-signal spin systems, rather than excluding them; (2) maintains identification performance under realistic stress that matches or exceeds the stress robustness reported by the 2015 Dubey et al. method under the same protocol — missing peaks (up to 50% deletion) and chemical-shift deviations of the magnitude caused by pH, temperature, and ionic-strength variation (up to 0.5 ppm in the proton and 3 ppm in the carbon dimension); (3) is validated at real biofluid scale, with hundreds of overlapping spin systems and independent ground truth, rather than on a small constructed mixture; (4) attaches calibrated confidence to each identification, where a confidence value is calibrated when its stated reliability matches the observed error frequency on held-out or independent data, uses that confidence to separate verified identification from annotation, where a match counts as verified identification only when its calibrated confidence is at the 95% reliability level or above (observed error frequency at or below 5% on held-out or independent data) and otherwise is reported as annotation, and detects and characterizes metabolites absent from the reference database; and (5) combines evidence from multiple complementary NMR experiments (for example 1D proton, TOCSY, HSQC, and HSQC-TOCSY) so that the experiments jointly raise identification confidence. Requirements (1)–(5) are jointly mandatory: the problem is resolved only by a capability that satisfies all five together and is demonstrated on real biofluid samples with independent ground truth; meeting any subset — for example only a better scoring function or performance on simulated stress data — leaves the remaining facets open. The problem includes, as an explicit component, the consequence of reference-database incompleteness: since no matching method can identify what has no reference pattern, out-of-database metabolites must be detected and characterized.

References

  1. Abhinav Dubey, Annapoorni Rangarajan, Debnath Pal, Hanudatta S. Atreya, 'Pattern Recognition-Based Approach for Identifying Metabolites in Nuclear Magnetic Resonance-Based Metabolomics', Analytical Chemistry 87(14):7148–7155, 2015. DOI: 10.1021/acs.analchem.5b00990, https://doi.org/10.1021/acs.analchem.5b00990
  2. Kerem Bingol, Da-Wei Li, Bo Zhang, Rafael Brüschweiler, 'Comprehensive Metabolite Identification Strategy Using Multiple Two-Dimensional NMR Spectra of a Complex Mixture Implemented in the COLMARm Web Server', Analytical Chemistry 88(24):12411–12418, 2016. DOI: 10.1021/acs.analchem.6b03724, https://doi.org/10.1021/acs.analchem.6b03724
  3. Amrinder Singh, Abhinav Dubey, Satish K. Adiga, Hanudatta S. Atreya, 'Phase modulated 2D HSQC-TOCSY for unambiguous assignment of overlapping spin systems', Journal of Magnetic Resonance 286:10–16, 2018. DOI: 10.1016/j.jmr.2017.11.005, https://doi.org/10.1016/j.jmr.2017.11.005
  4. Tao Huang, Pengyu Chen, Bin Liu, Xing Li, Xinqiao Lv, Kaifeng Hu, 'NPid: an Automatic Approach to Rapid Identification of Known Natural Products in the Crude Extract of Crabapple Based on 2D 1H–13C Heteronuclear Correlation Spectra of the Extract Mixture', Analytical Chemistry 92(16):10996–11006, 2020. DOI: 10.1021/acs.analchem.9b05363, https://doi.org/10.1021/acs.analchem.9b05363
  5. Michael T. Judge, Timothy M. D. Ebbels, 'Problems, principles and progress in computational annotation of NMR metabolomics data', Metabolomics 18(12):102, 2022. DOI: 10.1007/s11306-022-01962-z, https://doi.org/10.1007/s11306-022-01962-z
  6. Leonhard Lücken, Nico Mitschke, Thorsten Dittmar, Bernd Blasius, 'Network Flow Methods for NMR-Based Compound Identification', Analytical Chemistry 97(9):4832–4840, 2025. DOI: 10.1021/acs.analchem.4c01652, https://doi.org/10.1021/acs.analchem.4c01652
  7. Zhouao Zhou, Xinli Liao, Xu Qiu, Yue Zhang, Jiyang Dong, Xiaobo Qu, Donghai Lin, 'NMRformer: A Transformer-Based Deep Learning Framework for Peak Assignment in 1D 1H NMR Spectroscopy', Analytical Chemistry 97(1):904–911, 2025. DOI: 10.1021/acs.analchem.4c05632, https://doi.org/10.1021/acs.analchem.4c05632
  8. Caihong Bai, Xiu Gu, Yu Yang, Xiaomei Yu, Kaifeng Hu, 'Characterizing the Spin System Using 3D TOCSY J-Resolved Spectroscopy to Facilitate Deep Annotation of Individual Peaks within Interleaved Multiplets in High-Resolution 1D 1H Spectra of Cellular Metabolic Mixtures', Analytical Chemistry 97(25):13040–13047, 2025. DOI: 10.1021/acs.analchem.4c06938, https://doi.org/10.1021/acs.analchem.4c06938
  9. G. A. Nagana Gowda, Wentao Zhu, Daniel Raftery, 'NMR-based metabolomics: Where are we now and where are we going?', Progress in Nuclear Magnetic Resonance Spectroscopy 150–151:101564, 2025. DOI: 10.1016/j.pnmrs.2025.101564, https://doi.org/10.1016/j.pnmrs.2025.101564
  10. Alessia Vignoli, Stefano Cacciatore, Leonardo Tenori, 'Deriving three one dimensional NMR spectra from a single experiment through machine learning', Nature Communications 16:10159, 2025. DOI: 10.1038/s41467-025-65294-x, https://doi.org/10.1038/s41467-025-65294-x
  11. Chaoxing Liu, Longxiang Chen, Ziqiao Chen, Zexuan Zhang, Zhong Chen, Yulan Lin, 'NMRVPF: A High-Precision Method for Automatic Identification and Quantitative Analysis of Mixtures Based on 1D NMR Spectra', Analytical Chemistry 98(5):3603–3613, 2026. DOI: 10.1021/acs.analchem.5c04763, https://doi.org/10.1021/acs.analchem.5c04763