Classical 1H NMR correlation rules as diagnostic tests: likelihood ratios from 168,211 experimental spectra
- 1Mol Insight S.L.
Abstract
How much does a signal between 3.3 and 4.5 ppm say about a proton on a carbon bearing oxygen? The correlation tables used to read 1H NMR spectra state where each proton type resonates, but not how much observing, or failing to observe, a signal in that range should change a chemist’s belief. We measured this for 42 single-signal correlation rules and coupling rules transcribed from two open teaching sources, treating each rule as a diagnostic test and computing positive and negative likelihood ratios (LR+ and LR−) on two independent collections of experimental spectra with known structures: 160,038 literature peak lists from NMRexp and 8,173 atom-assigned spectra from nmrshiftdb2. Rules and scoring were frozen before the confirmatory data were examined; the sources state positions and coupling ranges, and the integration and multiplicity clauses we scored are our formalisation of what those rules imply. The rules fall into two regimes. Windows in uncrowded regions carry real evidence (aldehyde C-H LR+ 34 and 20 on the two collections, carboxylic acid O-H 11), while no window between 1.0 and 4.5 ppm exceeds LR+ 3.5 by position alone (median 1.6 and 2.3 on the two collections), because the protons of other environments occupy it: the environment a window describes accounts for a median of 20% of the assigned signals inside it. A reported integration recovers part of the crowded region (alkyl methylene LR+ 2.5 to 38), and several of the strongest effects are negative observations that the tables do not quantify (no reported signal between 6.5 and 8.0 ppm, LR− 0.011 for an aromatic C-H). The published windows are narrower than the environments they describe, and LR+ agrees between the two collections (Spearman ρ = 0.92). A coupling constant considered without regard to which signal carries it holds little information; tied to the shift region of the protons the rule concerns, in an analysis added after the confirmatory run, it gains up to thirteen-fold. In a secondary analysis in the reciprocal direction, from structure to expected shift, the additive increment rules of the standard reference tables, audited on 73,049 assigned protons, predict benzene ring protons to 0.14 ppm and protons on sp3 carbon to 0.35 ppm, with heavy-tailed errors that call for a robust likelihood. Together the results give the qualitative tables a quantitative form, measured on the information that published peak lists actually carry, that can be encoded in automated interpretation.
Related
- Classical infrared correlation rules as diagnostic tests: likelihood ratios from 6,663 experimental spectra
- Correlation-Driven Singlet--Triplet Inversion and Optical Spectra of N-Substituted Phenalenyls: Insights from PPP-FCI and TDDFT
- A trace-hierarchy theorem for shared splittings in coupled spin-½ NMR spectra
- Deposit optics, film thickness and acquisition date confound Raman spectra of dried droplets
- Vibronic Absorption and Circular Dichroism Spectra of Chiral PDI Aggregates in Free Space and Optical Cavities
BibTeX
@misc{calvo2026classical,
title = {{Classical 1H NMR correlation rules as diagnostic tests: likelihood ratios from 168,211 experimental spectra}},
author = {Kementsu Ibarguen Calvo},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15010690/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010690/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Classical 1H NMR correlation rules as diagnostic tests: likelihood ratios from 168,211 experimental spectra
AU - Kementsu Ibarguen Calvo
PY - 2026
DA - 2026/10/20
PB - ChemRxiv
DO - 10.26434/chemrxiv.15010690/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15010690/v1
AB - How much does a signal between 3.3 and 4.5 ppm say about a proton on a carbon bearing oxygen? The correlation tables used to read 1H NMR spectra state where each proton type resonates, but not how much observing, or failing to observe, a signal in that range should change a chemist’s belief. We measured this for 42 single-signal correlation rules and coupling rules transcribed from two open teaching sources, treating each rule as a diagnostic test and computing positive and negative likelihood ratios (LR+ and LR−) on two independent collections of experimental spectra with known structures: 160,038 literature peak lists from NMRexp and 8,173 atom-assigned spectra from nmrshiftdb2. Rules and scoring were frozen before the confirmatory data were examined; the sources state positions and coupling ranges, and the integration and multiplicity clauses we scored are our formalisation of what those rules imply. The rules fall into two regimes. Windows in uncrowded regions carry real evidence (aldehyde C-H LR+ 34 and 20 on the two collections, carboxylic acid O-H 11), while no window between 1.0 and 4.5 ppm exceeds LR+ 3.5 by position alone (median 1.6 and 2.3 on the two collections), because the protons of other environments occupy it: the environment a window describes accounts for a median of 20% of the assigned signals inside it. A reported integration recovers part of the crowded region (alkyl methylene LR+ 2.5 to 38), and several of the strongest effects are negative observations that the tables do not quantify (no reported signal between 6.5 and 8.0 ppm, LR− 0.011 for an aromatic C-H). The published windows are narrower than the environments they describe, and LR+ agrees between the two collections (Spearman ρ = 0.92). A coupling constant considered without regard to which signal carries it holds little information; tied to the shift region of the protons the rule concerns, in an analysis added after the confirmatory run, it gains up to thirteen-fold. In a secondary analysis in the reciprocal direction, from structure to expected shift, the additive increment rules of the standard reference tables, audited on 73,049 assigned protons, predict benzene ring protons to 0.14 ppm and protons on sp3 carbon to 0.35 ppm, with heavy-tailed errors that call for a robust likelihood. Together the results give the qualitative tables a quantitative form, measured on the information that published peak lists actually carry, that can be encoded in automated interpretation.
ER -