niccogreek/nmr-canonical-cleaned

Canonical NMR Dataset Collection — Data Card

Dataset release: v2
Canonical schema: v2
Spectral modalities: 1H and 13C resonance-level peak lists

Collection overview

This release brings several of the largest openly available processed NMR corpora used by current deep-learning methods into one model-independent schema. It combines simulated and literature-derived spectra while preserving the provenance and annotation coverage of every source.

The collection has five functional components:

1. train_val, the common representation-learning pool built from the MST-NMR, NMRexp, and NMRTrans/NMRSpec train and validation partitions;
2. test_benchmark, the union of their published test partitions after a molecule-connectivity disjoin against all three rich train/validation pools and the cleaned NMRGym pool;
3. ADMET subsets, exact molecule matches to three TDC endpoints, with the original TDC train/validation and test assignments;
4. SimNMR-PubChem / NMR-Solver, a much larger simulated shift-only component kept separate because its proton peaks contain shifts and equivalence-derived integration, while multiplicity, J coupling, and reported peak ranges are unavailable;
5. NMRGym, a smaller experimental shift-only component with paired 1H and 13C resonance lists and no supplied integration, multiplicity, J coupling, or reported peak ranges.

One Parquet row represents one spectrum record. A molecule may have multiple records when spectra originate from different sources, simulations, reports, or experimental conditions. Such records remain distinct unless their canonical structure and both exact shift lists are identical under the deduplication rule described below.

Source provenance

Component	Origin used by this collection	Reference
MST-NMR	Simulated paired spectra from the Multimodal Spectroscopic Dataset, using the processed NMRPeak LMDB release.	MSD, NMRPeak
NMRexp	Experimental records mined from chemistry Supporting Information published between 2010 and 2024, using the quality-controlled NMRPeak subset.	NMRexp, NMRPeak
NMRTrans / NMRSpec	Experimental peak tables mined from chemistry Supporting Information published between 2013 and 2025; this collection uses the released 212,440-record model dataset.	NMRTrans
SimNMR-PubChem / NMR-Solver	PubChem-scale simulated atom-level 1H and 13C shifts grouped through supplied equivalence classes.	NMR-Solver
NMRGym	Experimental paired 1H and 13C shift lists released as the scaffold-split NMRGym benchmark used by UltraNMR.	UltraNMR, source repository
ADMET labels	Ames, LD50 Zhu, and AqSolDB solubility endpoints with the official Therapeutics Data Commons splits.	Property benchmark notes

The source papers describe collection and upstream curation. This repository starts from the released processed representations and records the additional canonicalisation, split protection, and cleaning applied here.

Final files

Component	NMR records	Labels or derived features
Representation-learning pool	train_val.parquet	train_val_mol_properties.csv
Connectivity-disjoint benchmark	test_benchmark.parquet	test_benchmark_mol_properties.csv
SimNMR-PubChem shift-only pool	simnmr.parquet	simnmr_mol_properties.csv
NMRGym experimental shift-only pool	nmrgym.parquet	nmrgym_mol_properties.csv
Ames	train_val.parquet, test.parquet	Matched train/validation and test labels
LD50 Zhu	train_val.parquet, test.parquet	Matched train/validation and test labels
AqSolDB solubility	train_val.parquet, test.parquet	Matched train/validation and test labels

Each molecular-property CSV has one row per final NMR record_id. It contains RDKit descriptors, functional-group indicators, ECFP4 and MACCS fingerprints, and an explicit RDKit processing status. Each ADMET CSV has the same unique record_id set as its paired Parquet and contains the TDC molecule, label Y, and source identifier.

Canonical schema

The physical representation is Arrow/Parquet with nested lists and structs:

CanonicalRecord
├── record and molecular provenance
├── atoms: list<string>
├── h_nmr_peaks: list<ProtonPeak>
│   ├── chemical shift and integration
│   ├── raw and harmonised multiplicity
│   ├── J-coupling list
│   ├── reported or reconstructed shift interval
│   └── optional equivalence-group provenance
└── c_nmr_peaks: list<CarbonPeak>
    └── shift plus optional source-specific simulated peak properties

The corresponding model-independent Python dataclasses are CanonicalRecord, ProtonPeak, and CarbonPeak. CanonicalParquetDataset streams one or more Parquet files as validated CanonicalRecord objects, while CanonicalNMRDataset provides the in-memory equivalent for small collections.

Record fields

Field	Arrow type	Meaning
record_id	string, required	Stable source-qualified spectrum identifier; unique within each file.
source	string	Canonical source family.
smiles	string	Structure string selected from the source release.
smiles_canonical	string	Isomeric canonical SMILES recalculated with RDKit.
molecular_formula	string	Formula recalculated from the canonical RDKit molecule.
nmr_frequency	string | null	Reported acquisition frequency when supplied.
nmr_solvent	string | null	Reported solvent when supplied.
atoms	list<string>	Atom symbols in RDKit canonical-SMILES order, normally heavy atoms.
h_nmr_peaks	list<struct<ProtonPeak>>	Canonical proton resonances; always a list.
c_nmr_peaks	list<struct<CarbonPeak>>	Canonical carbon resonances; always a list.

RDKit regenerates smiles_canonical, formula, and atom order from the selected source SMILES. This gives every source the same structural representation and records the RDKit version in the Parquet footer.

Proton peak fields

Field	Arrow type	Unit and convention
shift	float64, required	Chemical shift in ppm.
integration	int64 | null	Number of represented protons when supplied or derivable.
multiplicity_raw	string | null	Original source label.
multiplicity	string | null	Harmonised label used across rich sources.
j_values	list<float64> | null	J couplings in Hz.
range_min	float64 | null	Lower endpoint of the reported shift interval, in ppm.
range_max	float64 | null	Upper endpoint of the reported shift interval, in ppm.
range_half_span	float64 | null	(range_max - range_min) / 2, in ppm.
equivalence_class	int64 | null	Source equivalence-group identifier for atom-level simulated data.
member_shifts	list<float64> | null	Original atom-level shifts represented by one grouped resonance.

The canonical multiplicity vocabulary is: m, d, s, dd, t, ddd, q, dt, td, br, ddt, dq, tt, quint, dddd, qd, sept, ddp, ddq, bd, dqd. Lossless aliases are normalised as p → quint, hept → sept, and brd → bd. Other supplied labels become <unk>, while their exact source text remains available in multiplicity_raw.

Carbon peak fields

Field	Arrow type	Unit and convention
shift	float64, required	Chemical shift in ppm.
integral	float64 | null	Source-supplied simulated carbon integral.
intensity	float64 | null	Source-supplied simulated intensity.
width	float64 | null	Source-supplied simulated carbon width in ppm.

The three optional carbon quantities are populated by MST-NMR. Carbon peaks from NMRexp, NMRTrans, NMR-Solver, and NMRGym carry the common shift field.

Missing values

• The two modality columns are always lists. [] means that the record has no usable peaks for that nucleus.
• In the rich MST-NMR, NMRexp, and NMRTrans files, j_values is always a list. [] means that no numerical coupling is listed for that peak.
• Shift-only sources use null for annotations outside their source representation, including multiplicity, J values, and reported ranges.
• A numerical J=0 remains a supplied value and can be masked during model preparation.
• After common cleaning, all 16,608,682 proton peaks in train_val and test_benchmark have populated positive integration, canonical multiplicity, and all three range fields.

Example streaming access:

from data.dataset import CanonicalParquetDataset

records = CanonicalParquetDataset("datasets/cleaned/train_val.parquet")

for record in records:
    h_shifts = [peak.shift for peak in record.h_nmr_peaks]
    c_shifts = [peak.shift for peak in record.c_nmr_peaks]

Run repository code with PYTHONPATH=scripts so the data package is available.

Source-specific representation

MST-NMR through NMRPeak

MST-NMR is the simulated rich source. Its NMRPeak release provides paired proton and carbon resonances. Proton centroid, nH, category, parsed J values, rangeMin, and rangeMax map directly to the shared fields. If centroid is absent, conversion uses delta, followed by the midpoint of the two range endpoints. A source peak containing only one shift receives a point interval:

range_min = range_max = shift
range_half_span = 0

MST-NMR also supplies the optional carbon integral, intensity, and width (ppm). This carbon width is a source-simulated peak property and is separate from the proton range_half_span.

Converter: convert_mst_nmr.py.

NMRexp through NMRPeak

NMRexp contains experimental spectra mined from literature and processed by NMRPeak. Its proton mapping is the same as MST-NMR, including point intervals for single reported shifts. Carbon records supply shifts only. Frequency and solvent are retained where they occur in the processed release.

NMRexp is the source with partial modality coverage: an experimental report may contain only 1H or only 13C. Those records remain useful and use an empty list for the unavailable modality.

Converter: convert_nmrexp.py.

NMRTrans / NMRSpec

NMRTrans uses a processed NMRSpec collection mined from chemistry-paper Supporting Information published between 2013 and 2025. The released proton token has the form:

[shift, range_half_span, multiplicity, integration, J_values]

The field called peak_width by the upstream loader is the half-span of the reported interval:

range_min = shift - range_half_span
range_max = shift + range_half_span
range_half_span = (range_max - range_min) / 2

Carbon values are shift lists. The release contains paired modalities and does not supply solvent or frequency fields.

Converter: convert_nmrtrans.py.

SimNMR-PubChem / NMR-Solver

SimNMR-PubChem is a simulated PubChem-scale shift database used by NMR-Solver. The source stores atom-level predictions in nmr_predict, element identities in atom_index, and supplied equivalence groups in equi_class. Conversion creates one resonance per nucleus and equivalence class:

shift = mean(member_shifts)
1H integration = number of hydrogen members

For proton resonances, equivalence_class and member_shifts preserve the atom-level source information. Multiplicity, J coupling, and reported ranges are represented as unavailable annotations. Carbon equivalence groups are also reduced to their mean shift. This shift-only representation is distributed as a separate component so models can use it for large-scale shift pretraining or a staged simulated-to-rich curriculum.

Converter: convert_nmrsolver.py.

NMRGym

NMRGym is an experimental shift-only benchmark released as train, validation, and test pickle files. Conversion preserves its paired 1H and 13C shift lists and concatenates the published partitions in their original order. Unlike SimNMR-PubChem, it does not supply proton equivalence groups or integration. Integration, multiplicity, J coupling, reported ranges, solvent, and frequency therefore remain null rather than being inferred. An absent modality would remain []; after common filtering every released NMRGym row has both modalities.

This component is distributed separately from the rich training pool so it can support experimental shift-only pretraining, domain adaptation, or explicit source-aware sampling.

Converter: convert_nmrgym.py.

Processing and split rationale

The reproducible data flow is:

source releases
    → source-specific canonicalisation
    → rich train/validation and source-test merges
    → benchmark molecule disjoins / ADMET exact matching
    → common quality cleaning and exact-shift deduplication
    → molecular descriptors and collection analytics

1. Canonicalisation and merge

Each converter translates the corresponding processed release into schema v2, normalises multiplicity, recalculates molecular metadata, validates every record, and writes provenance in the Parquet footer. The Arrow writer and shared source mappings are implemented in common.py.

merge_datasets.py concatenates compatible schema-v2 files while checking schema and RDKit metadata:

• source train and validation partitions form the common train_val pool;
• source test partitions form the initial rich benchmark pool.

2. Connectivity-disjoint benchmark

The benchmark starts from the published test partitions of MST-NMR, NMRexp, and NMRTrans/NMRSpec. A molecule appearing in any of the corresponding rich train/validation pools would give a model source-dependent prior exposure, even if its spectrum or record ID differed. Therefore disjoin_benchmark_from_train.py removes benchmark records whose RDKit connectivity InChIKey occurs in the merged train/validation pool.

The already-cleaned benchmark is then passed through the same helper a second time against cleaned NMRGym. This removes experimental shift-only molecules that would otherwise appear both in NMRGym-assisted training and rich-spectrum evaluation. The second stage removes 872 benchmark records representing 841 overlapping connectivity keys: 373 MST-NMR, 428 NMRexp, and 71 NMRTrans records.

The comparison uses the connectivity block rather than the full InChIKey. It therefore applies a conservative molecule-level boundary that also groups stereoisomers sharing the same connectivity. The rich-train stage removed 26,930 records before common cleaning. The final test_benchmark is a shared test set whose molecular connectivities are absent from every MST-NMR, NMRexp, and NMRTrans train/validation partition represented in this collection and from the cleaned NMRGym component. A verification rerun against NMRGym found zero residual connectivity keys.

3. ADMET matching and disjoin

The ADMET workflow begins from the official TDC train_val and test partitions. admet_overlap_audit.ipynb derives full RDKit InChIKeys for property and NMR structures. Full keys are used here because stereochemical distinctions can affect measured properties.

extract_annotated_peaks.py materialises all exact NMR matches for Ames, LD50 Zhu, and AqSolDB solubility. The union of 6,858 matched NMR record_id values is removed from the common pretraining pool before its quality cleaning. This separation prevents a downstream labelled spectrum from also occurring in train_val.

Multiple spectra of one molecule remain separate supervised examples. Repeated property rows are resolved at molecular-identity level:

• agreeing Ames labels and the one agreeing repeated LD50 value are collapsed;
• ten LD50 identities with discordant labels are removed, affecting 8 train/validation and 5 test NMR records;
• the solubility matches contain no repeated property rows.

4. Common quality cleaning

filter_dataset.py applies the same simple physical and structural checks to every rich canonical file. A record is retained when at least one modality contains peaks and:

• all shifts are finite;
• proton shifts lie in [-5, 20] ppm;
• carbon shifts lie in [-50, 300] ppm;
• each modality contains at most 60 peaks;
• each proton peak contains at most six J values;
• J values are finite and non-negative;
• each supplied proton integration is positive;
• the canonical SMILES contains one connected fragment.

The filter keeps supplied J=0 values. Molecular weight, logP, TPSA, and drug-likeness descriptors do not participate in record acceptance.

The filter removed 31,250 records from the ADMET-disjoint train/validation pool and 1,390 from the connectivity-disjoint benchmark. Reason counts can overlap when one record violates more than one condition:

Reason	Train/validation	Benchmark test
Exact duplicate shift signature	20,081	239
Non-positive supplied 1H integration	6,358	665
More than 60 13C peaks	2,015	215
Multi-fragment canonical SMILES	1,511	163
1H shift outside range	1,048	89
13C shift outside range	335	29
More than six J values in one peak	2	0

The same physical and structural rules are applied to the combined NMRGym source before publication. They retain 265,095 of 269,999 records and remove 4,904 records. Overlapping reason counts are: 2,067 proton-shift violations, 1,475 multi-fragment structures, 1,167 records with more than 60 proton peaks, 294 with more than 60 carbon peaks, and 64 carbon-shift violations. NMRGym has no rich annotations to validate or exact duplicate shift signatures among the retained rows.

5. Exact-shift deduplication

Deduplication targets identical extracted spectra rather than all records of the same molecule. Its exact identity key is:

canonical SMILES
+ sorted exact 1H shift list
+ sorted exact 13C shift list

No rounding or tolerance is applied. An unavailable modality contributes an empty shift list. A deterministic hash narrows candidate groups, after which the actual SMILES and both shift lists are compared, so hash equality alone never removes a row.

Within one exact group, selection prefers the record with more populated proton integration, multiplicity, J-list, and range annotations, followed by fewer missing or <unk> annotations and the lexicographically smallest record_id. This preserves simulated/experimental pairs, replicate spectra, and condition-dependent measurements whenever their shifts differ.

The pre-cleaning audit of 1,868,476 ADMET-disjoint train/validation rows found 20,128 exact groups, all pairs: 16,037 NMRexp–NMRTrans pairs and 4,091 internal MST-NMR pairs. After validity filters, 20,081 duplicate rows remained eligible for removal.

6. Molecular features and analytics

calculate_mol_properties.py runs after cleaning, so descriptor rows align with final record IDs. It calculates exact molecular weight, RDKit logP, TPSA, HBA, HBD, rotatable bonds, fraction Csp3, aromatic atom fraction, nine SMARTS functional-group flags, radius-2 2,048-bit Morgan/ECFP4 fingerprints, and the 166 usable MACCS keys.

analyze_cleaned_datasets.py reproduces the molecular and peak analyses for train_val, test_benchmark, NMRGym, SimNMR-PubChem, and the three ADMET cohorts. Summary CSVs use every record. Violin plots use a deterministic sample of at most 20,000 records per group; only their visible range is limited to Q1 − 2.5×IQR through Q3 + 2.5×IQR.

Dataset-specific composition

Rich train/validation collection

train_val.parquet combines the original train and validation partitions of the three rich sources after removing exact ADMET matches and applying common cleaning. It contains 1,837,226 records and supports representation learning, source-aware sampling, and later task-specific split construction.

Connectivity-disjoint benchmark

test_benchmark.parquet contains the cleaned source-test records that pass both connectivity disjoins. It contains 179,239 records. The source balance and molecular-property distributions closely follow the train/validation collection, making it suitable for cross-model comparison without an obvious composition shift introduced by the disjoin step.

Dataset	Source	Records	Unique canonical SMILES	With 1H	With 13C	With both
Train/validation	MST-NMR	699,526	699,469	699,526	699,526	699,526
Train/validation	NMRexp	965,921	965,921	844,532	824,631	703,242
Train/validation	NMRTrans / NMRSpec	171,779	171,779	171,779	171,779	171,779
Train/validation	All	1,837,226	1,772,773	1,715,837	1,695,936	1,574,547
Benchmark test	MST-NMR	72,068	72,068	72,068	72,068	72,068
Benchmark test	NMRexp	94,887	94,887	82,927	80,753	68,793
Benchmark test	NMRTrans / NMRSpec	12,284	12,284	12,284	12,284	12,284
Benchmark test	All	179,239	178,635	167,279	165,105	153,145

NMRexp accounts for all single-modality records. MST-NMR and NMRTrans remain paired throughout both rich files.

ADMET property subsets

Each endpoint keeps the official TDC split and contains spectra drawn from all three rich sources:

Endpoint	Split	MST-NMR	NMRexp	NMRTrans	Total
Ames	Train/validation	827	1,104	374	2,305
Ames	Test	143	198	68	409
LD50 Zhu	Train/validation	1,148	1,035	315	2,498
LD50 Zhu	Test	265	198	52	515
AqSolDB solubility	Train/validation	1,538	1,418	457	3,413
AqSolDB solubility	Test	333	278	71	682

The same molecule can have several NMR records, so record counts exceed unique property-SMILES counts:

Endpoint	Task	Train/validation records / molecules	Test records / molecules	Target summary, train / test
Ames	Binary classification	2,305 / 1,641	409 / 311	Positive: 36.9% / 42.5%
LD50 Zhu	Regression	2,498 / 1,853	515 / 396	2.12 ± 0.67 [−0.34, 5.51] / 2.37 ± 0.80 [0.29, 5.14]
AqSolDB solubility	Regression	3,413 / 2,499	682 / 525	−2.44 ± 1.87 [−10.10, 1.63] / −2.95 ± 2.05 [−8.70, 1.10]

Target definitions and units follow the corresponding TDC releases.

SimNMR-PubChem / NMR-Solver shift-only component

The source-level schema-v2 Parquet contains 105,764,812 simulated records. Common cleaning removes 255,196 records, leaving simnmr.parquet with 105,509,616 records. Its approximate 97,936,772 unique canonical SMILES indicate that it is a large molecule-level simulation collection with a smaller number of repeated records. Both modalities are non-empty in 105,476,128 records (99.968%).

NMRGym experimental shift-only component

The combined canonical source contains 269,999 records. Common filtering leaves nmrgym.parquet with 265,095 records and exactly 265,095 unique canonical SMILES. Every retained record has both 1H and 13C shift lists. The published-split provenance retained in record_id comprises 211,874 train, 26,715 validation, and 26,506 test records; this release exposes their cleaned ordered union as one source-aware pretraining component.

Analytics

Detailed descriptive statistics, source comparisons, figures, annotation coverage, NMRGym and SimNMR shift-only analyses, and links to every generated CSV are collected in the analytics report. The report also documents how to reproduce the analysis from the full project repository.

Use considerations

The primary learning unit is a structured resonance-level peak list linked to a molecular structure. The rich collection supports models that consume chemical shifts together with integration, multiplicity, ranges, and J couplings. SimNMR-PubChem supports simulated shift-set pretraining, while NMRGym supplies experimental shift-only examples for domain support and simulated-to-experimental curricula. The ADMET files support frozen-encoder probes and supervised property-prediction studies with exact spectrum-to-label alignment.

Literature-mined records vary in solvent, field strength, and reporting practice. Source-aware evaluation is therefore informative alongside aggregate metrics. Carbon integral, intensity, and width are MST-specific simulated quantities. RDKit canonicalisation standardises the supplied structure representation; salts, protonation states, tautomers, and missing stereochemical information retain the distinctions present in the source SMILES and identity rules described above.

Further implementation rationale and source audits are documented in Datasets.md, Canonicalization_Implementation_Notes.md, Multiplicity analysis.md, Dataset_Filtering_and_Processing.md, Dataset Analysis.md, and Properties Dataset.md.