CASPRE Training Data
Experimental 13C chemical shifts used to train and test CASPRE were aggregated from curated datasets assembled in the Wishart laboratory and from BMRB, HMDB, JEOL and NMRShiftDB2. Only compounds with explicit solvent information were included. Each dataset below contains the compounds (NP-MRD, HMDB or NMRShiftDB2 identifiers and, where available, compound names), their solvent, structures (SDF) and experimental 13C chemical shift assignments.
Training sets
| Dataset | Molecules | 13C shifts | |
|---|---|---|---|
| CDCl₃ – JEOL natural products (NP-MRD) | 9,114 | 214,980 | Download |
| H₂O – HMDB, NP-MRD and NMRShiftDB2 | 253 | 3,889 | Download |
| MeOH – NMRShiftDB2 (August 2026 release) | 2,098 | 22,770 | Download |
The MeOH training set contains 228 molecules with a reported CD3OD solvent and 1,870 molecules assigned to MeOH by a solvent classifier.
Holdout sets
| Dataset | Molecules | 13C shifts | |
|---|---|---|---|
| NMRShiftDB2 holdout set – CDCl₃ | 37 | 338 | Download |
| NMRShiftDB2 holdout set – MeOH | 106 | 2,314 | Download |
| Holdout Set 1 (NP-MRD and HMDB) – H₂O | 64 | 1,067 | Download |
| Holdout Set 2 (NP-MRD and JEOL) – CDCl₃ | 21 | 379 | Download |
The holdout molecules were not used to train the CASPRE models.
HOSE-code reference database
| Dataset | Molecules | 13C shifts | |
|---|---|---|---|
| Full database: JEOL (9,170 / 219,064), NMRShiftDB2 (19,197 / 215,462) and HMDB (253 / 3,889), all solvent labels | 28,620 | 438,415 | Download |
| Solvent-matched entries used by CASPRE: CDCl₃ (10,808 / 245,793), MeOH (246 / 5,430) and H₂O (296 / 4,282) | 11,350 | 255,505 | Download |
CASPRE's database search matches the local environment (HOSE code) of each carbon against this solvent-aware reference database of experimental 13C shifts, and only uses entries whose solvent matches the query. The full database also contains entries with other or unreported solvents.
File format
Each zip file contains a README and, for its dataset:
- <dataset>.sdf.gz structures with explicit hydrogens, one record per spectrum (title line = molecule ID)
- <dataset>_molecules.csv record, molecule_id, compound_name, inchikey, smiles, solvent, n_shifts
- <dataset>_shifts.csv record, molecule_id, compound_name, solvent, atom_index, shift_ppm (one row per assigned carbon)
record is the 1-based position of the structure in the SDF file and atom_index the 0-based atom index within that structure. CDCl3 and MeOH shifts are referenced to TMS; aqueous shifts to DSS or TSP. The HOSE-code database files also give the source database and spectrum of each record, and the atom environment keys (Morgan_0 to Morgan_5, radius 0–5) used for the lookup.
Citing CASPRE:
- Sajed T, Saha S, Lee BL, Berjanskii M, Sayeeda Z, Wang F, Gautam V, Wishart DS. Accurate, Robust, Solvent-aware Prediction of 13C NMR Chemical Shifts of Organic Compounds Using a Hybrid Machine Learning Approach. Submitted.