Predicted NMR shifts

Implementation notes and historical checks. Open the visual guide →.

NMR prediction estimates ¹H and ¹³C chemical shifts for one molecule from measured reference data. It runs on your computer, and each value links to its atom in the drawing.

Use it

Select an atom of the molecule. If the drawing holds only one molecule, you can leave the selection empty.
Open Properties → Predict NMR…. The results open in a floating palette at the right edge of the window.
Choose ¹H or ¹³C.
Click a row or a stick label to select its atom in the drawing, together with any hydrogens drawn on it.
Turn on Show labels to write the results on the drawing. The menu next to it offers Atom numbers (#3), ppm (1.280 ppm) or Both (#3 · 1.280 ppm).
Choose Copy or Export… to keep the shifts with the method, data version, conditions, limitations and data credit.

Move or dock the palette

Drag the title to move the palette. Drag the corner glyph (◢) to resize it.
Dock moves the report into the inspector, under Predict → NMR. Undock returns it to the palette. Clicking the Predict tab also docks the report.
In the floating palette, − collapses the results and + expands them.
× closes the report and removes its labels from the drawing.
Opening, moving or docking the report never changes the zoom or the view. You can keep drawing while it is open.

Read the results

A ¹H row such as C #3 · 2H is the group of hydrogens on atom 3. A ¹³C row such as C #3 is one carbon.
The Predicted shifts plot marks each predicted shift with a stick on a ppm axis that runs from high to low.
▸ Details adds a line to each row, for example Shell 4 · 17 references · SD 0.064 ppm. Shell is how far around the atom the match reaches, references is the number of independent reference molecules, and SD is the spread of their values. SD is not an error bar.
Details also shows the method, data version, conditions, limitations and data credit, and offers Hide plot / Show plot. In the floating palette it adds Reset position and the − / + size buttons.
The footer names the nucleus and the solvent, for example ¹H · grouped per atom · CDCl₃.

What the plot and labels show

The plot marks chemical shifts only. Stick heights do not show intensity or integration.
J couplings, multiplets and second-order effects are not predicted.
Nearby sticks share one label. Clicking the label again selects its next atom.
Labels move with their atoms. An atom inside a collapsed abbreviation is labeled at the abbreviation. Labels never appear in exports or the Undo history.
A soft teal halo marks a selected atom that has a prediction.

When results update

A chemical change clears the results and labels. Moving atoms, changing depth or drawing styles, and collapsing or expanding abbreviations keep them.
To predict another molecule, select it and choose Predict NMR… again.
Each document keeps its own report, palette position, docking and label format while it is open.

Limits

Molecules up to 128 atoms of H, C, N, O, F, P, S, Cl, Br and I. Charged molecules, radicals, metals, multi-center attachments and labels for less common isotopes, such as ²H or ¹³C, are not supported.
References are CDCl₃ measurements recorded at 273–323 K. They are pooled without temperature or concentration correction.
Only hydrogens on carbon get a value. O–H, N–H and other exchangeable hydrogens do not.
Several hydrogens on one atom are one group value, not separate peaks.
E/Z, relative configuration and stereoisomers are not distinguished. Diastereotopic hydrogens get the same value.
A row with too little reference data stays empty and says why. Even simple molecules can be partly covered: in the ¹H result for ethanol only the CH₃ group gets a value.
Results are estimates, not assignments or simulated spectra.

Implementation notes

Method

The offline method follows the nuclear-rooted spherical matching and longest-sphere median fallback described in Ask Ernö (2016). It uses an original, versioned ReShiki encoder (reshiki-hose-v1), rather than the original self-learning assignment loop. No Java, MySQL, network request, API key or runtime training is required. Its numbers come from the separately licensed measured nmrshiftdb2 subset; the raw unassigned spectra from the Ask Ernö paper are not used as assigned reference data.

The request uses the complete connected graph of the selected molecule, including atoms behind collapsed abbreviations. Expand an abbreviation to see an internal site directly; the table keeps its original atom ID. With no selection, the drawing must contain exactly one molecular component. The first version supports neutral closed-shell ordinary-bond organic molecules.

For ¹H, one row represents the attached proton group of a parent drawing atom. Implicit, explicit attached and explicitly bonded hydrogens are counted consistently. Each group is linked to its parent atom and any explicit H atom IDs. Carbon-bound proton groups receive a reference-group median when supported. Exchangeable O–H/N–H rows remain without a numerical shift. Multiple attached H are an unresolved group median, not a claim that they form one peak. The connectivity-only 2D encoder ignores E/Z and relative configurations; it does not distinguish stereoisomers or assign individual diastereotopic hydrogen shifts. Equivalent sites have the same environment and prediction, but remain separate linked rows. For ¹³C, each carbon receives one atom-linked row with the same matching and limitations.

Matching starts at sphere radius 4 around the nucleus, falling back to radius 3 then 2. At least two independent molecule/connectivity groups must support an entry. Shell reports the radius actually used; references counts independent reference molecule groups. SD is the sample standard deviation of their observed environment-group medians. It describes reference dispersion and is not a calibrated error bar or confidence interval. The export also includes the observed minimum and maximum.

Coverage and accuracy

Repeated spectra, symmetrical sites and stereoisomer records do not inflate support: they contribute one median per molecular connectivity/environment. Held-out experimental validation keeps duplicate molecules and stereoisomers on the same side of the split. The measured coverage/error and preprocessing counts are recorded in validation.json. Coverage is measured among accepted, assigned carbon/H-group targets with recorded supported conditions; excluded chemistry and unknown-condition spectra are counted separately in the audit. The held-out ¹H group-median MAE is 0.307 ppm at 92.5% coverage; ¹³C MAE is 2.05 ppm at 60.2% coverage. These are errors among supported accepted targets. Those measurements come from the same source release, not an independent acquisition cohort; the metrics evaluate parent-site group medians. The paper’s published error is not claimed as ReShiki’s accuracy.

The conditions filter can leave familiar compounds partly unsupported. Ethanol’s bundled result supports its methyl group, while its CH₂ group lacks sufficient qualified references and its exchangeable OH is unsupported. The source ethanol CDCl₃ proton spectrum has unreported temperature and was excluded. Ethyl acetate has three supported non-exchangeable H groups and four supported carbon sites in this release.

Panel and labels

The palette opens at the full window’s right edge, also when the inspector is shown. Reset position returns it there without changing the camera. Docking restores the previous inspector tab when undocked. Plot sticks have equal height. Labels do not assign individual protons of an unresolved group; the table and footer name the nucleus and the group. Running prediction again clears the previous labels. Results are per document, and a background result cannot land in another document. Copied and exported text includes a prediction label, atom IDs, the method, data version, conditions, limitations and attribution. Predictions are not stored as experimental assignments or full simulated spectra. See the floating-panel design and desktop review.