Predicted NMR shifts — NorrChemica Analysis Notes, NMR Series Part 13. Navy card with three schematic spectra: measured, predicted for structure A (every signal matches) and predicted for structure B (one signal differs).

Predicted NMR shifts: how accurate ChemDraw, MestReNova and other predictors are, and why a good match does not prove the structure

Analysis Notes · NMR Series · Part 13

Predicted NMR shifts: how accurate ChemDraw, MestReNova and other predictors are, and why a good match does not prove the structure

How the predictors work, the published errors for 1H and 13C, the signals they predict badly, and which experiments distinguish two similar candidate structures

Sources C. R. Chim., Chem. Sci., Magn. Reson. Chem., J. Am. Chem. Soc., Org. Lett., Molecules, Nat. Prod. Rep., Front. Nat. Prod. Methods additivity rules, HOSE codes, neural networks, DFT calculation, DP4-type probabilities

Chemists often compare the measured spectrum of a new compound with a spectrum predicted for the proposed structure. The result depends on the accuracy of the prediction, and it gives no information about a second structure that would match equally well.

In one published comparison, MestReNova, ACD and ChemDraw predicted 1H shifts with average errors of 0.24–0.30 ppm, and the differences between the three programs were not significant.[1] In the only comparison of the three programs for 13C that we found, their root-mean-square errors on benzene derivatives were 1.8–3.4 ppm, and ChemDraw had the largest error.[2]

Errors of this size are small enough to exclude a structure that is very different, but often too large to exclude a similar isomer. In a test of 99 pairs of a drug-like molecule and a close isomer, an automated comparison based on 1H NMR or on IR alone could not distinguish the two structures in 27–49 % of the pairs.[6]

Six terms used in this article. A predicted shift is a chemical shift calculated for a proposed structure. The mean absolute error (MAE) is the average of the differences between predicted and measured shifts, all counted as positive. The root-mean-square error (RMSE, also called RMSD) is the square root of the average squared difference; it is always at least as large as the MAE, and the difference grows when a few shifts are badly predicted. A HOSE code describes the surroundings of an atom sphere by sphere, so that atoms with the same surroundings can be found in a database.[3, 4] DFT (density functional theory) is a quantum-chemical method that calculates the shielding of each nucleus from the electronic structure of the molecule.[9] DP4 is the probability that a candidate structure is the correct one, calculated from the differences between its calculated shifts and the measured shifts.[10]

Published errors of predicted shifts

Predictor and method Nucleus Average error, ppm Test set
MestReNova · combination of predictors 1H 0.24 (RMSD) 30 compounds, 396 shifts[1]
ACD · additivity rules, HOSE codes, neural networks 1H 0.25 (RMSD) the same set[1]
ChemDraw · additivity rules 1H 0.30 (RMSD) the same set[1]
NMRShiftDB · neural networks 1H 0.54 (RMSD) the same set[1]
ChemDraw 23.1 13C 3.4 (RMSE), 2.2 (MAE) 171 benzene derivatives, about 1010 shifts[2]
MestReNova 15.0.1 13C 1.9 (RMSE), 1.4 (MAE) the same set[2]
ACD 11.01 13C 1.8 (RMSE), 1.2 (MAE) the same set[2]
NMRShiftDB deep-learning model 13C 1.1 (RMSE) 100 of the same molecules[2]
Graph network trained on measured shifts 13C 1.43 (MAE) held-out measured shifts[3]
DFT with empirical corrections 1H, 13C about 0.15 and 2.5 (RMSD) protocols summarised in a 2021 study[3]

Each 1H value is the average, over 30 compounds, of the error within each compound. The 13C comparisons used different test sets, so their values cannot be compared directly. The 13C MAE values of [2] are those of its preprint version.

How a predicted shift is calculated

Every predictor starts from the drawn structure, but the programs calculate the shift in different ways (Figure 1). The method determines which molecules a program predicts well.

  • Additivity rules. The shift is a base value plus tabulated contributions of the neighbouring groups. The predictor in ChemDraw uses only this method, and additivity rules are less reliable when the effects of the groups are not additive.[1]
  • HOSE codes. The program finds atoms in its database with the same HOSE code as the atom to be predicted, and uses their measured shifts.[3, 4] When enough matching atoms are found, the spread of their shifts gives a range for the prediction.[1] The original codes contain neither the stereochemistry nor the 3D shape of the molecule.[4]
  • Neural networks. A network is trained on assigned spectra to relate the structure to the shift. NMRShiftDB uses only neural networks.[1] Recent graph networks represent the molecule as atoms and bonds, and some also as a 3D structure.[3] Their published errors apply mainly to molecules similar to those used in training.[4]
  • Combinations. MestReNova combines two predictors with a statistical weighting, and ACD combines additivity rules, HOSE codes and neural networks.[1]
  • DFT calculation. The shielding is calculated for each low-energy conformer of a 3D model. The conformers are weighted by their energies, and the shielding is converted into a shift with a reference compound and a fitted line.[9] No database is needed, but the calculation is slow: for 24 test molecules, one DFT protocol needed 19 hours of computing time and a graph network 10 seconds.[3]
Four ways to predict a chemical shift: additivity rules, HOSE codes, neural networks and DFT calculation

Figure 1. The four approaches to shift prediction described in [1], [3], [4] and [9]. Schematic drawing by us; the molecule, the spheres and the network are generic.

A database method is accurate only for molecules similar to those in its database. DFT needs no database, but its result depends on finding the conformers that the molecule adopts.

How large the error is

Two published comparisons test the commercial programs against measured spectra (Figure 2).

For 1H, a 2022 study predicted the shifts of 30 compounds with four programs and compared them with 396 measured values from the SDBS database.[1] MestReNova, ACD and ChemDraw gave average errors of 0.24–0.30 ppm, with no significant difference between them, and NMRShiftDB gave 0.54 ppm.

For 13C, a 2024 study compared ChemDraw 23.1, MestReNova 15.0.1 and ACD 11.01 on 171 benzene derivatives in CDCl3.[2] The root-mean-square error of ChemDraw was 3.4 ppm, almost twice that of MestReNova (1.9 ppm) and ACD (1.8 ppm). The deep-learning model of NMRShiftDB, which could be run on 100 of the molecules, gave 1.1 ppm, and the authors’ own graph network gave 0.9 ppm. The authors note that these differences are not necessarily significant, because they do not know which molecules the other programs were trained on.

Published average errors of predicted shifts: proton RMSD of four programs, and carbon-13 RMSE and MAE of three programs and a graph network

Figure 2. Average errors of predicted shifts in two published comparisons. Left: 1H, average over 30 compounds of the root-mean-square deviation within each compound, from [1]. Right: 13C root-mean-square error (RMSE) and mean absolute error (MAE) for 171 benzene derivatives, from [2] (RMSE) and from its preprint version (MAE); the graph network is the authors’ own model. Plotted by us from the published values. The two panels use different test sets.

The 1H averages also differ between proton types. The Supporting Information of the 2022 study lists the measured and the predicted shift of every proton, and we sorted these protons by type and calculated the average error of each program for each type (Figure 3).[1]

  • MestReNova had the lowest average error for aromatic CH, for CH next to C=O, C=C, an aromatic ring or S, and for the other aliphatic CH (0.11–0.16 ppm). ACD had the lowest for olefinic CH and for CH next to O or N (0.15–0.16 ppm).
  • ChemDraw had an average error of 0.44 ppm for olefinic and imine CH, about twice its error for any other type. Its largest errors in this group were for a CH=CH group between two naphthalene rings, an oxime CH=N and a dienone CH. Only 12 protons are in this group.
  • NMRShiftDB had the largest average error for every type except the other aliphatic CH.
  • Some large errors come from one program only. ACD predicted 2.09 ppm for the CH2 chain signal at 1.25–1.30 ppm in five long-chain compounds, and MestReNova predicted 7.51 ppm for a benzylic OCH2 group measured at 4.54 ppm.
Mean absolute error of predicted proton shifts by proton type for MestReNova, ACD, ChemDraw and NMRShiftDB, calculated by us from the per-proton data of a published comparison

Figure 3. Mean absolute error of the predicted 1H shifts of 335 protons in 28 compounds, sorted by proton type. Calculated by us from the per-proton tables in the Supporting Information of [1]; the sorting into types is ours. One compound was left out because its table gives no predicted shifts for three programs, and one because its ACD values are identical to the measured values. These averages are taken over all protons; the values in Figure 2 are averages over compounds of the root-mean-square deviation, and are therefore larger.

An average does not show the few shifts that are badly wrong, and these are the shifts that lead to a wrong assignment. A large difference between the RMSE and the MAE indicates such shifts: for ChemDraw the two values were 3.4 and 2.2 ppm.[2] The most accurate model in that study, with an MAE of 0.7 ppm, still predicted 9 carbons with errors above 3 ppm.[2]

Database methods also reproduce the errors in their database.[1] One HOSE-code method gave average errors of 1.0, 1.7 and 2.7 ppm, depending on the database it used.[4] The authors of one graph network found wrong entries in NMRShiftDB, among them two carbons with exchanged assignments and a compound entered with a cyclohexyl ring in place of a phenyl ring.[3]

When the test set is the same, the methods can be compared directly. On 205 structures with 2531 measured 13C shifts, HOSE-code prediction gave an MAE of 1.58 ppm, a neural network 1.91 ppm and DFT calculation 3.29 ppm,[14] and an increment method gave 2.15 ppm.[15] The authors, from the developer of a commercial predictor, concluded that on this set DFT was less accurate than the empirical methods.[14]

A published error also depends on the test set. A compilation of results published since 1990 lists 13C MAE values from about 0.5 to 4 ppm (Figure 4).[15] The test sets differ in size and composition: one is a single molecule, and some values come from cross-validation on the data set used to build the model. An average error is too optimistic when the test molecules resemble the training molecules.[4]

Left: carbon-13 mean absolute errors of four prediction methods on one test set of 205 structures. Right: published carbon-13 mean absolute errors from 1990 to 2026

Figure 4. 13C mean absolute errors. Left: four methods on one test set of 205 structures, from [14], with the increment value from [15]. Right: published values from 1990 to 2026, plotted by us from the compilation in [15]; values given as standard deviations or root-mean-square errors, values for specialised test subsets and two values with an unstated test set are omitted.

Since 1990, the published errors for 13C have fallen, but no clear improvement is visible for 1H.[4]

A typical average error is 0.24–0.30 ppm for 1H and 1.2–3.4 ppm for 13C, depending on the program and on the measure of error. Individual shifts can be wrong by several times these values.

Where predictions fail

The largest errors appear where the shift depends on a factor that the predictor does not take into account.

  • The 3D shape. In the 1H comparison, the largest errors came from chair-shaped cyclohexane rings and from a cyclopentane ring that the prediction treated as flat.[1]
  • Stereochemistry. A predictor that does not encode stereochemistry gives the same shifts for two diastereomers. Extended HOSE codes were developed for this reason.[4]
  • Hydrogen bonds and tautomers. In the 13C study, the authors attributed several of the largest errors to an enol tautomer and to intramolecular hydrogen bonds, in 1,3-diphenyl-1,3-propanedione and 5-methoxy-3-nitrosalicylaldehyde.[2]
  • Ring systems outside the training set. The same study reports large errors for naphthalene and pyridine, which were not represented among the compounds used to train its model.[2]
  • Exchangeable protons and the solvent. 1H shifts depend on the solvent, the temperature and other conditions of the measurement,[4] and automated comparisons treat the shifts of OH and NH protons as unpredictable.[6]

Emodin shows the size of the solvent effect (Figure 5). Its 1-OH and 8-OH protons form strong intramolecular hydrogen bonds to the neighbouring carbonyl group, so the solvent has little effect on them: they resonate at 12.04–12.26 ppm in CDCl3, acetone-d6 and DMSO-d6.[5] The 3-OH proton forms no intramolecular hydrogen bond, so its shift depends on the solvent: 6.18 ppm in CDCl3, 10.21 ppm in acetone-d6 and 11.41 ppm in DMSO-d6, a range of 5.23 ppm.[5] A prediction that ignores the solvent gives one value for this proton, and one value cannot be close to all three measured shifts.

Emodin: the proton shifts of its three OH groups in deuterochloroform, acetone-d6 and DMSO-d6. The 1-OH and 8-OH signals stay near 12 ppm; the 3-OH signal moves from 6.18 to 11.41 ppm

Figure 5. 1H shifts of the three OH protons of emodin at 298 K in CDCl3, acetone-d6 and DMSO-d6, in solutions of 2 mM or less. Plotted by us from Table 1 of [5]; structure drawn by us. The 3-OH signal was very broad, and the authors added a small amount of trifluoroacetic acid to sharpen it.

A change between two non-polar solvents also moves the carbon signals of polar groups. For 35 small, rigid molecules measured in CDCl3 and in CCl4, the root-mean-square change was 3.82 ppm for carbonyl carbons, 2.13 ppm for nitrile carbons, 0.54 ppm for sp³ carbons and 0.08 ppm for protons (Figure 6).[16] For carbonyl carbons, the change was larger than the error of the DFT calculations in the same study (1.27–1.78 ppm, PBEP86 functional). The authors found that DFT with an implicit solvent model did not reproduce the solvent effect equally well for all functional groups.[16]

Root-mean-square change of shifts between deuterochloroform and carbon tetrachloride for 35 rigid molecules, compared with the root-mean-square error of DFT calculation with the PBEP86 functional

Figure 6. Root-mean-square change of the shifts of 35 rigid molecules between CDCl3 and CCl4 (141 1H and 170 13C shifts), compared with the root-mean-square error of DFT calculation (PBEP86 functional) against the measured shifts. Plotted by us from the values in the text of [16].

Predictions are least accurate where the shift depends on the 3D shape, a hydrogen bond or the solvent, and for structures unlike those in the database. OH and NH signals should be left out of a comparison, and the solvent alone can move a carbonyl carbon by several ppm.

A good match does not prove the structure

A comparison of one measured spectrum with one predicted spectrum tests only whether the spectrum is consistent with that structure. It does not test whether a likely alternative, such as a regioisomer or a diastereomer, would be equally consistent, and in synthesis this is usually what the chemist needs to know.

A match can distinguish two candidates only at atoms where their predicted shifts differ by more than the prediction error, however accurate the predictor is on average.

Tests of automated structure verification, in which software compares the measured spectrum with the spectrum predicted for the proposed structure, show how reliable such a comparison is in practice.

  • In a 2006 test of a method based on 1H prediction, half of all spectra could be verified automatically without a false result, even when difficult structures with many exchangeable protons were included. In a blind test on a typical data set, fewer than 31 % of the structures needed evaluation by a chemist.[7]
  • In a 2007 test, a 1H–13C HSQC spectrum was added to the comparison. At least 90 % of the correct structures were identified, and the lowest reported rate at which difficult incorrect structures, mostly regioisomers, were accepted was 6 %.[8]
  • A 2025 study by a pharmaceutical company and a university built 99 pairs of a correct structure and a similar isomer from 42 drug-like molecules (Figure 7).[6] Most pairs differed in the position of a group on an aromatic or an aliphatic part of the molecule. With the method set to accept 90 % of the correct structures, 27–49 % of the pairs remained undecided with 1H NMR or IR alone, and 0–15 % with the two techniques together.
Ninety-nine pairs of a correct structure and a similar isomer: the kinds of isomer, and the share of pairs left unsolved by one technique and by proton NMR and IR together

Figure 7. The test of 99 isomer pairs in [6]. Top: the kinds of isomer, as reported; the published breakdown accounts for about 80 % of the pairs. Bottom: the share of pairs that could not be decided, at settings that accept 90 % or 95 % of the correct structures, for one technique (1H NMR or IR) and for the two together. Plotted by us from the values in the text of [6].

In some pairs, the commercial 1H software gave the correct and the incorrect structure identical scores. The authors conclude that the false-positive and false-negative rates of automated methods remain too high to replace interpretation by a chemist.[6]

A worked example: a 1,4- or a 1,5-disubstituted triazole

Triazoles made by azide–alkyne cycloaddition are an example. The thermal reaction gives both regioisomers, and the copper-catalysed reaction gives the 1,4-isomer selectively.[20] In one published experiment, 1-benzyl-4-phenyl- and 1-benzyl-5-phenyl-1,2,3-triazole were measured together in one CDCl3 sample, in a 2:1 ratio (Figure 8).[19]

  • Protons. The triazole proton resonates at 7.68 ppm in the 1,4-isomer and at 7.73 ppm in the 1,5-isomer, and the benzylic CH2 protons at 5.55 and 5.53 ppm.[19] The differences, 0.05 and 0.02 ppm, are at most a fifth of the average 1H prediction error.
  • The CH2 carbon. It resonates at 54.3 and 51.9 ppm, 2.4 ppm apart,[19] which lies within the range of average 13C prediction errors in Figure 2.
  • The triazole CH carbon. It resonates at 119.5 ppm in the 1,4-isomer and at 133.2 ppm in the 1,5-isomer, 13.7 ppm apart,[19] which is four times the largest average 13C error in Figure 2.

Spectra of the pure isomers from two other studies give the same carbon shifts within 0.2 ppm.[18, 21] For the triazole proton of the 1,4-isomer, however, one of these studies reports 7.59 ppm,[18] which differs from the value in the mixture (7.68 ppm) by more than the two isomers differ in the same sample.

Measured shifts of the 1,4- and the 1,5-disubstituted benzyl phenyl triazole in one sample: the protons differ by 0.02 to 0.05 ppm, the CH2 carbons by 2.4 ppm and the triazole CH carbons by 13.7 ppm

Figure 8. 1-Benzyl-4-phenyl- and 1-benzyl-5-phenyl-1H-1,2,3-triazole, measured in one CDCl3 sample (2:1) at 400 MHz (1H) and 100 MHz (13C); values from the Supporting Information of [19]. The assignment of the carbon signals to the two isomers is ours, by comparison with the spectra of the pure isomers in [18] and [21]. Structures drawn by us.

A predicted 1H spectrum therefore cannot distinguish these isomers by the triazole and CH2 signals. The triazole CH carbon can, and a published method for assigning such triazoles is based on it: C-5 of the 1,4-isomer appears near 120 ppm and C-4 of the 1,5-isomer near 133 ppm, and calculations reproduce this difference.[17] An HSQC spectrum identifies this carbon by its correlation with the triazole proton.

A prediction confirms a structure only when the predicted spectra of the plausible alternatives are clearly different. In the triazole example, only the triazole CH carbon meets this condition.

Calculated shifts with a probability

When the candidates are close, for example two diastereomers, the comparison needs an accurate prediction for each candidate and a measure of how much better one candidate fits than the others.

DFT provides the first. Best-practice protocols with empirical corrections reach root-mean-square deviations of about 0.15 ppm for 1H and 2.5 ppm for 13C.[3] Much of the remaining error is systematic, so the calculated values are scaled against measured ones.[9] These errors are not smaller than those of the best database methods, but they apply in the same way to a new structure, a diastereomer or a conformer, which a database may not cover.

DP4 provides the second. It converts the errors of the calculated shifts into a probability for each candidate, using the known error distribution of the calculation.[10, 11] In the original 2010 study, this probability assigned stereochemistry more successfully than the correlation coefficient or the mean absolute error, in tests with up to 64 diastereomers of 117 molecules.[10] By the end of 2020, its successor DP4+ had been used to assign more than 200 natural and synthetic products, and a review of these studies found that it had often been used improperly.[11]

Graph networks now reach a similar accuracy in seconds. In five historical cases of a misassigned natural product, a graph network gave the lower error for the correct structure each time, and for a flexible 1,3-diol it identified the correct diastereomer in three of four cases.[3]

A worked example: hexacyclinol

In 2006, the proposed structure of the fungal metabolite hexacyclinol was disputed.[9] The 13C shifts of the proposed structure and of an alternative structure were calculated by DFT and compared with the measured shifts (Figure 9).[12]

The method was first tested on three natural products of known structure. Their calculated shifts deviated by 1.2–1.9 ppm on average, and by at most 3.8 ppm.[12] For the proposed structure of hexacyclinol, the average deviation was 6.8 ppm, five carbons were wrong by more than 10 ppm, and the largest error was 22.0 ppm. These errors are far larger than those of the method on known structures, so the proposed structure was excluded.

The alternative, which has two epoxide rings in place of an endoperoxide, gave an average deviation of 2.9 ppm and a largest error of 9.5 ppm for one conformer, and 1.8 and 5.2 ppm for a second conformer.[12] The revised structure was confirmed in the same year by synthesis and X-ray crystallography.[9, 13]

Structures of hexacyclinol: the originally proposed endoperoxide structure on the left and the revised structure with two epoxide rings on the right. Absolute differences between calculated and measured carbon-13 shifts for the 23 carbons of hexacyclinol, sorted from largest to smallest, for the originally proposed and the revised structure

Figure 9. Hexacyclinol. Top: the originally proposed structure (left) and the revised structure (right), reproduced from [9], Figure 1, under the CC BY 4.0 licence. Bottom: absolute differences between the calculated and the measured 13C shifts of the 23 carbons, sorted from the largest to the smallest, plotted by us from the tables in the Supporting Information of [12]. For the revised structure, the values are those of the first conformer after the assignments of C9 and C12 were exchanged.

Calculated shifts are most reliable when they exclude a structure. To choose between close candidates, a probability measure is needed and, where possible, an experiment that confirms the result.

What distinguishes the candidates

A prediction ranks the candidate structures. Some 2D and heteronuclear experiments can exclude a candidate directly, because they measure connections between atoms that a regioisomer or a diastereomer does not share with the correct structure.

  • HMBC shows which protons are two or three bonds from which carbons, and therefore where a substituent is attached (Part 3).
  • 1H–15N HMBC shows which nitrogen is substituted. In ten published indazoles, it distinguished the N1 and N2 isomers without a reference compound (Part 5).
  • NOE or ROESY shows which protons are close in space, which gives the relative configuration and, in some cases, the position of a substituent (Part 5 and Part 8).
  • HSQC shows which proton is attached to which carbon, and so identifies a diagnostic carbon such as the triazole CH.

When no experiment separates the candidates, DFT shifts with a DP4-type probability give a quantitative ranking.[10, 11]

We recommend using predictions in the following order (Figure 10):

  1. Predict the spectra of every plausible structure, including the likely regioisomers and diastereomers, not only that of the expected product.
  2. Leave out OH, NH and other exchangeable protons, and note the solvent of the measurement.
  3. Compare the differences between the candidates with the typical error, about 0.3 ppm for 1H and 2–3 ppm for 13C. If every difference is smaller than this, the prediction cannot distinguish the candidates.
  4. If the candidates are not separated, measure the connections by HMBC, 1H–15N HMBC or NOE.
A sequence for using predicted shifts: predict every plausible isomer, compare the differences between them with the prediction error, and measure correlations when the differences are too small

Figure 10. The sequence that we recommend for using predicted shifts. The error values are the typical averages from Figure 2 and from the table at the start of this article.

A prediction is best used to choose which experiment to run, not to replace it.

Summary

  • Predictors use additivity rules, HOSE codes, neural networks, combinations of these, or DFT calculation, and the method determines which molecules are predicted well.
  • MestReNova, ACD and ChemDraw predicted 1H shifts with average errors of 0.24–0.30 ppm, and the error of each program differs between proton types. For 13C on benzene derivatives, their root-mean-square errors were 1.8–3.4 ppm, and ChemDraw was the least accurate. On one test set of 205 structures, DFT calculation was less accurate than the empirical methods.
  • Individual shifts can be wrong by several ppm where the shift depends on the 3D shape, hydrogen bonds, tautomers or the solvent, or where the structure is unlike those in the database. The 3-OH proton of emodin moves by 5.23 ppm between CDCl3 and DMSO-d6, and carbonyl carbons move by several ppm even between CDCl3 and CCl4.
  • A good match does not exclude a similar isomer. In 99 isomer pairs, 27–49 % remained undecided with 1H NMR or IR alone. The triazole and CH2 protons of the 1,4- and 1,5-isomers of a benzyl phenyl triazole agree within 0.05 ppm, but the triazole CH carbons differ by 13.7 ppm.
  • DFT shifts with a DP4-type probability rank close candidates. They are most reliable when they exclude a structure, as for hexacyclinol, where five carbons of the proposed structure were wrong by more than 10 ppm.
  • HMBC, 1H–15N HMBC and NOE measure connections that a regioisomer or a diastereomer does not share, so they can distinguish candidates that a prediction cannot.

Structure confirmation from NorrChemica

When a predicted spectrum leaves more than one structure possible, our NMR service records the experiments that distinguish between them. Package VI (Characterisation) comprises 1H, 13C, DEPT and gCOSY spectra, and Package VII (Full elucidation) adds gHSQC and gHMBC with a full assignment; both are quoted per sample. Package HRMS-02 combines LC-HRMS with 1H and 13C NMR, and the structure is assigned by a chemist. NOESY, ROESY and 15N experiments are quoted per sample. If you do not know which experiment answers your question, describe the question on the request form. The service is available worldwide for chemical and pharmaceutical R&D. Every report is signed and supplied with the native data. Research use only.

Request NMR analysis → About the NMR service

NMR tubes made in Germany by Deutero and shipped from Finland. See the NMR tube guide →

References

  1. Deciding which is the best 1H NMR predictor for organic compounds using statistical tools. C. R. Chim. 2022, 25, 83–95. DOI
  2. Can graph machines accurately estimate 13C NMR chemical shifts of benzenic compounds? Molecules 2024, 29, 3137. DOI. The MAE values are taken from its preprint version, Preprints.org 2024, 202406.0511. DOI
  3. Real-time prediction of 1H and 13C chemical shifts with DFT accuracy using a 3D graph neural network. Chem. Sci. 2021, 12, 12012–12026. DOI
  4. Prediction of chemical shift in NMR: a review. Magn. Reson. Chem. 2022, 60, 1021–1031. DOI
  5. Solvent-dependent structures of natural products based on the combined use of DFT calculations and 1H-NMR chemical shifts. Molecules 2019, 24, 2290. DOI
  6. Towards automatically verifying chemical structures: the powerful combination of 1H NMR and IR spectroscopy. Chem. Sci. 2025, 16, 21590–21599. DOI
  7. Automated structure verification based on 1H NMR prediction. Magn. Reson. Chem. 2006, 44, 524–538 (abstract). DOI
  8. Automated structure verification based on a combination of 1D 1H NMR and 2D 1H–13C HSQC spectra. Magn. Reson. Chem. 2007, 45, 803–813 (abstract). DOI
  9. Are we still chasing molecules that were never there? The role of quantum chemical simulations of NMR parameters in structural reassignment of natural products. Front. Nat. Prod. 2024, 2, 1321043; Figure 1 reproduced under the CC BY 4.0 licence. DOI
  10. Assigning stereochemistry to single diastereoisomers by GIAO NMR calculation: the DP4 probability. J. Am. Chem. Soc. 2010, 132, 12946–12959 (abstract). DOI
  11. A critical review on the use of DP4+ in the structural elucidation of natural products: the good, the bad and the ugly. A practical guide. Nat. Prod. Rep. 2022, 39, 58–76. DOI
  12. Predicting NMR spectra by computational methods: structure revision of hexacyclinol. Org. Lett. 2006, 8, 2895–2898, with its Supporting Information. DOI
  13. Total synthesis and structure assignment of (+)-hexacyclinol. Angew. Chem. Int. Ed. 2006, 45, 5790–5792 (bibliographic record). DOI
  14. Empirical and DFT GIAO quantum-mechanical methods of 13C chemical shifts prediction: competitors or collaborators? Magn. Reson. Chem. 2010, 48, 219–229 (abstract). DOI
  15. Historical NMR shift predictions and bibliography: table of published prediction errors, nmrshiftdb2 project, version listing publications known by 28 June 2026. Link
  16. Impact of solvent interactions on 1H and 13C chemical shifts investigated using DFT and a reference dataset recorded in CDCl3 and CCl4. Phys. Chem. Chem. Phys. 2022, 24, 23551–23560. DOI
  17. Method for assigning structure of 1,2,3-triazoles. J. Org. Chem. 2012, 77, 8756–8761 (abstract). DOI
  18. A ferrocene functionalized Schiff base containing Cu(II) complex: synthesis, characterization and parts-per-million level catalysis for azide alkyne cycloaddition. Dalton Trans. 2020, 49, 6578–6586, Supporting Information. DOI
  19. Copper-catalyzed carbon–carbon bond cleavage of primary propargyl alcohols: β-carbon elimination of hemiaminal intermediates. Catal. Sci. Technol. 2015, 5, 3931–3934, Supporting Information. DOI
  20. Synthesis of deuterated 1,2,3-triazoles. J. Org. Chem. 2012, 77, 8896–8904. DOI
  21. C–H arylations of 1,2,3-triazoles by reusable heterogeneous palladium catalysts in biomass-derived γ-valerolactone. Chem. Commun. 2016, 52, 9777–9780, Supporting Information. DOI

About this article. Sources 1, 2, 3, 4, 5, 6, 9, 16 and 20 were read in their open-access full text, the tables of the Supporting Information of 12 were read for every calculated and measured shift, the characterisation data of the two triazoles were read in the Supporting Information of 18, 19 and 21, and the per-proton tables and structure drawings of the Supporting Information of 1 were read in full; for 24 of the 28 compounds used, these values reproduce the per-compound errors published in the article. These were read through a retrieval tool that returns the requested passages and tables; this is not equivalent to reading each article from beginning to end. The compilation 15 was read in full in the browser. Sources 7, 8, 10, 14 and 17 were read as abstracts, the passages of 11 that we cite were read in the full text, and 13 is cited from its bibliographic record. The following are our own reasoning and arithmetic: the argument that a match must be compared with the predictions for the plausible alternatives; the statement that a predictor without stereochemistry gives the same values for two diastereomers; the statement that one predicted value cannot be close to all three measured shifts of the emodin 3-OH proton; the statement that DFT errors apply in the same way to structures that a database may not cover; the assignment of the triazole carbon signals in the mixture to the two isomers, the differences between them and their comparison with the prediction errors; the statements on HSQC; the count of five carbons with errors above 10 ppm and the sorting in Figure 9; the sorting of the protons of [1] by type, using the assignment drawings of its Supporting Information, and the averages in Figure 3; the selection of values plotted in Figure 4; the typical error values in Figure 10; and the sequence of steps that we recommend. Figures 1 and 10 are our own drawings. Figures 2 and 4 to 9 are plotted by us from published values, Figure 3 is calculated by us from published per-proton values, the emodin and triazole structures are our drawings, and the hexacyclinol structures are reproduced from source 9. Our service is described from the service page of our shop. The sources were read on 16 November 2026.

NMR Series

  1. Part 1 — Reading 1H and 13C NMR
  2. Part 2 — Reading 19F, 31P and 11B NMR
  3. Part 3 — Reading 2D NMR: COSY, HSQC and HMBC
  4. Part 4 — Quantitative NMR: how a purity figure is made
  5. Part 5 — 15N NMR: which nitrogen is alkylated, protonated or coordinated?
  6. Part 6 — NMR tubes by experiment: which tube for which measurement
  7. Part 7 — Quadrupolar nuclei in NMR: broad lines and missing carbons
  8. Part 8 — Two sets of NMR signals: rotamers, diastereomers or an impurity
  9. Part 9 — NMR yield and isomer ratios: relaxation and the delay between pulses
  10. Part 10 — Silicon NMR: the glass signal, the negative NOE and M, D, T, Q units
  11. Part 11 — Solid-state NMR: magic-angle spinning, cross-polarisation and what a CP-MAS spectrum counts
  12. Part 12 — Enantiomeric excess by NMR: when NMR is the best method, the chiral auxiliaries and the detection limit for the minor enantiomer
  13. Part 13 — Predicted NMR shifts: how accurate ChemDraw, MestReNova and other predictors are (this article)
  14. Part 14 — Phosphorus NMR in biology: phospholipids, phosphorylated proteins, oligonucleotide drugs and phosphates in food
  15. Part 15 — Multiplets in ¹H NMR: coupling trees, second-order patterns and two signals from one CH₂
  16. Part 16 — NMR for cosmetics: adulterated essential oils, free formaldehyde, aloe vera, rancid lipsticks and mineral oil in lip products
  17. Part 17 — Polymers by NMR: composition, sequence, chain ends and branches, and what a spectrum of many chains can and cannot measure
  18. Part 18 — Benchtop NMR: what a permanent-magnet spectrometer does well, and when a sample needs high field
  19. Part 19 — NOESY, ROESY and TOCSY: regioisomers, E and Z isomers, anomers, and the signals of one compound in a mixture
  20. Part 20 — Rapid-injection NMR: starting a reaction inside the spectrometer, how fast the spectra can follow it, and the intermediates it has revealed
  21. Part 21 — Fluorine NMR for PFAS: total organofluorine, what targeted LC-MS/MS does not detect, and the detection limit
  22. Part 22 — Polysiloxanes by proton NMR: silicone oil in medicines, food and cosmetics, the reactive groups before curing, and the rubber network after it

© 2026 NorrChemica. All rights reserved.
Content may not be reproduced without written permission.

Back to blog

Leave a comment

Please note, comments need to be approved before they are published.