meaw0415/Meaw0415

NMRGym

Benchmark on NMR Spectrum. !!! Important: !!! If you need access to the dataset, please email zhengf723@connect.hkust-gz.edu.cn

Data Sources (Before Cleaning)

Source	Records	Unique SMILES	Total Spectrums	¹H NMR	¹³C NMR
CH-NP	12,165	12,165	24,326	12,161	12,165
HMDB	1,791	896	3,278	1,566	1,712
NMRBank	148,914	142,964	297,043	148,437	148,606
NMRShiftDB 2024	41,019	39,631	50,416	18,570	31,846
NP-MRD	489,569	243,598	950,242	462,233	488,009
PubChem-NMR	1,647	1,535	2,605	1,556	1,049
SDBS	12,926	12,626	22,711	11,522	11,189
Total	708,031	430,690	1,350,621	656,045	694,576

---

Balanced Dataset Statistics (After Cleaning & Processing)

After data cleaning, filtering, quality control, and balanced scaffold splitting, the final NMRGym dataset consists of three splits:

Split	Records	Unique SMILES	¹H NMR	¹³C NMR
Train	402,676	250,822	402,676	402,676
Val	53,672	50,716	53,672	53,672
Test	80,069	74,568	80,069	80,069
Total	536,417	376,106	536,417	536,417

Dataset Visualizations

---

Quick Checklist

• [✅] Data cleaning: Heavy atom filtering and illegal smiles exclusion.
• [✅] Data Summary and Visualization.
• [✅] Data Split.
• [] 3D coord. generation. Rdkit.
• [✅] Toxic Property label generation.
• [✅] Function Group label generation.

---

NMRGym Split Dataset Format (nmrgym_spec_filtered_both_{train,test}.pkl)

This dataset contains post-QC, scaffold-split molecular entries with paired NMR spectra, structure fingerprints, functional group annotations, and toxicity labels.

Each .pkl file is a Python list of dictionaries. Each dictionary corresponds to a single unique molecule.

Example record structure

{
    "smiles": "CC(=O)Oc1ccccc1C(=O)O",
    "h_shift": [7.25, 7.32, 2.14, 1.27],
    "c_shift": [128.5, 130.1, 172.9, 20.7],
    "fingerprint": np.ndarray(shape=(2048,), dtype=np.uint8),
    "fg_onehot": np.ndarray(shape=(22,), dtype=np.uint8),
    "toxicity": np.ndarray(shape=(7,), dtype=np.int8),
}

---

Field definitions

smiles (str) Canonical SMILES string for this molecule.

h_shift (list[float]) Experimental or curated ¹H NMR peak positions in ppm.

c_shift (list[float]) Experimental or curated ¹³C NMR peak positions in ppm.

fingerprint (np.ndarray(2048,), uint8) RDKit Morgan fingerprint (radius = 2, 2048 bits). Encodes local circular substructures.

fg_onehot (np.ndarray(22,), uint8) Binary functional-group vector. 1 means the SMARTS pattern is present in the molecule. Index mapping (0→21):

1. Alcohol
2. Carboxylic Acid
3. Ester
4. Ether
5. Aldehyde
6. Ketone
7. Alkene
8. Alkyne
9. Benzene
10. Primary Amine
11. Secondary Amine
12. Tertiary Amine
13. Amide
14. Cyano
15. Fluorine
16. Chlorine
17. Bromine
18. Iodine
19. Sulfonamide
20. Sulfone
21. Sulfide
22. Phosphoric Acid

toxicity (np.ndarray(7,), int8) Seven binary toxicity endpoints (1 = toxic / positive, 0 = negative): 0. AMES (mutagenicity)

1. DILI (Drug-Induced Liver Injury)
2. Carcinogens_Lagunin (carcinogenicity)
3. hERG (cardiotoxic channel block)
4. ClinTox (clinical toxicity)
5. NR-ER (endocrine / estrogen receptor)
6. SR-ARE (oxidative stress response)

---

Train / Test meaning

• nmrgym_spec_filtered_both_train.pkl: scaffold-based training set.
• nmrgym_spec_filtered_both_test.pkl: scaffold-disjoint test set.

Scaffold split is computed using Bemis–Murcko scaffolds. This means test scaffolds do not appear in train, simulating generalization to new chemotypes instead of just new stereoisomers.

---

Minimal usage example

import pickle
from rdkit import Chem
from rdkit.Chem import AllChem

# load the dataset
test_path = "/gemini/user/private/NMRGym/utils/split_output/nmrgym_spec_filtered_both_test.pkl"
with open(test_path, "rb") as f:
    dataset = pickle.load(f)

print("num molecules:", len(dataset))
print("keys:", dataset[0].keys())
print("example toxicity vector:", dataset[0]["toxicity"])

# build 3D conformer for the first molecule
smi = dataset[0]["smiles"]
mol = Chem.MolFromSmiles(smi)
mol = Chem.AddHs(mol)  # add hydrogens for 3D
AllChem.EmbedMolecule(mol, AllChem.ETKDG())
AllChem.UFFOptimizeMolecule(mol)

# extract 3D coordinates (in Å)
conf = mol.GetConformer()
coords = []
for atom_idx in range(mol.GetNumAtoms()):
    pos = conf.GetAtomPosition(atom_idx)
    coords.append([pos.x, pos.y, pos.z])

print("3D coords for first molecule (Angstrom):")
for i, (x,y,z) in enumerate(coords):
    sym = mol.GetAtomWithIdx(i).GetSymbol()
    print(f"{i:2d} {sym:2s}  {x:8.3f} {y:8.3f} {z:8.3f}")

Notes on 3D coordinates

• We generate a single low-energy conformer using RDKit ETKDG embedding + UFF optimization.
• Coordinates are in Ångström.
• These coordinates are not guaranteed to match the experimental NMR conformer in solvent; they are intended for featurization (message passing models, geometry-aware models, etc.), not quantum-accurate geometry.

---

Downstream Benchmark Tasks

This benchmark suite evaluates AI models on multiple spectroscopy-related prediction tasks. Each task reflects a distinct molecular reasoning aspect — from direct spectrum regression to property and toxicity prediction.

1. Spectral Prediction

Method / Paper Title	Model Type / Approach	Metric(s)	Model Size	Notes
x	x	x	x	x

2. Structure Prediction (Inverse NMR)

Method / Paper Title	Model Type / Approach	Metric(s)	Model Size	Notes
x	x	x	x	x

3. Molecular Fingerprint Prediction (Spec2FP)

Method / Paper Title	Model Type / Approach	Metric(s)	Model Size	Notes
x	x	x	x	x

4. Functional Group Classification (Spec2Func)

Method / Paper Title	Model Type / Approach	Metric(s)	Model Size	Notes
x	x	x	x	x

5. Molecular Toxicity Prediction (Spec2Tox / ADMET)

Method / Paper Title	Model Type / Approach	Metric(s)	Model Size	Notes
x	x	x	x	x

---