create dataset card

README.md

@@ -0,0 +1,278 @@
+---
+license: mit
+task_categories:
+  - text-generation
+  - feature-extraction
+language:
+  - en
+tags:
+  - chemistry
+  - spectroscopy
+  - molecular-generation
+  - nmr
+  - infrared
+  - smiles
+  - deep-learning
+  - multimodal
+size_categories:
+  - 1M<n<10M
+---
+
+# NMIRacle Dataset
+
+## Dataset Description
+
+This dataset supports the **NMIRacle** (NMr-IR orACLE) framework for *de novo* molecular structure elucidation from multi-modal spectroscopic data. It contains paired molecular structures (SMILES) with simulated spectroscopic measurements (IR, ¹H-NMR, ¹³C-NMR) and fragment annotations.
+
+### Dataset Summary
+
+The dataset is organized into two subsets supporting the two-stage training paradigm of NMIRacle:
+
+1. **Pretrain Dataset** (~3.7M molecules): For fragment-to-molecule pre-training (Stage 1)
+2. **Multispectra Dataset** (~790K molecules): For spectra-to-molecule fine-tuning (Stage 2)
+
+### Supported Tasks
+
+- **Molecular Structure Generation**: Generate SMILES from spectroscopic inputs
+- **Fragment Prediction**: Predict substructure counts from spectra
+- **Multi-modal Learning**: Learn joint representations across IR and NMR modalities
+
+### Languages
+
+Chemical notation (SMILES, SMARTS)
+
+---
+
+## Dataset Structure
+
+```
+nmiracle-dataset/
+├── pretrain/                          # Stage 1: Fragment pre-training (~3.7M molecules)
+│   ├── smiles.npy                     # Molecular SMILES strings
+│   ├── substructures.h5               # Binary fragment presence vectors
+│   ├── substructure_counts.h5         # Fragment occurrence counts
+│   ├── split_indices.p                # Train/val/test split indices
+│   ├── pretrain_train_indices.npy     # Train split indices (legacy)
+│   ├── pretrain_val_indices.npy       # Validation split indices (legacy)
+│   └── pretrain_test_indices.npy      # Test split indices (legacy)
+│
+└── multispectra/                      # Stage 2: Spectra fine-tuning (~790K molecules)
+    ├── smiles.npy                     # Molecular SMILES strings
+    ├── spectra.h5                     # Multi-modal spectra (IR + ¹H-NMR + ¹³C-NMR)
+    ├── substructures.h5               # Binary fragment presence vectors
+    ├── substructure_counts.h5         # Fragment occurrence counts
+    └── split_indices.p                # Train/val/test split indices
+```
+
+---
+
+## Data Fields
+
+### Molecular Data
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `smiles.npy` | numpy array (str) | Canonical SMILES strings representing molecular structures |
+
+### Spectroscopic Data (`spectra.h5`)
+
+| Field | Shape | Description |
+|-------|-------|-------------|
+| `spectra` | (N, 21800) | Concatenated spectral intensities: IR (1800) + ¹H-NMR (10000) + ¹³C-NMR (10000) |
+
+**Spectral Modalities:**
+
+| Modality | Features | Range | Description |
+|----------|----------|-------|-------------|
+| **IR** | 1,800 | 400-4000 cm⁻¹ | Infrared absorption spectra (vibrational modes) |
+| **¹H-NMR** | 10,000 | 0-14 ppm | Proton NMR spectra (hydrogen environments) |
+| **¹³C-NMR** | 10,000 | 0-220 ppm | Carbon-13 NMR spectra (carbon backbone) |
+
+### Fragment Data
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `substructures.h5` | (N, 991) binary | Presence/absence of 991 SMARTS-defined fragments |
+| `substructure_counts.h5` | (N, 991) int | Occurrence counts of each fragment in the molecule |
+
+### Split Indices
+
+| Field | Type | Description |
+|-------|------|-------------|
+| `split_indices.p` | pickle dict | Dictionary with 'train', 'val', 'test' keys containing index arrays |
+
+---
+
+## Dataset Statistics
+
+### Pretrain Dataset (Stage 1)
+
+| Statistic | Value |
+|-----------|-------|
+| **Total molecules** | ~3,700,000 |
+| **Train split** | ~2,960,000 (80%) |
+| **Validation split** | ~370,000 (10%) |
+| **Test split** | ~370,000 (10%) |
+| **Max heavy atoms** | 35 |
+| **Element types** | C, N, O, S, F, Cl, Br, P, I (9 elements) |
+| **Fragment vocabulary** | 991 SMARTS patterns |
+| **Max fragment count** | 232 |
+
+### Multispectra Dataset (Stage 2)
+
+| Statistic | Value |
+|-----------|-------|
+| **Total molecules** | ~790,000 |
+| **Train split** | ~632,000 (80%) |
+| **Validation split** | ~79,000 (10%) |
+| **Test split** | ~79,000 (10%) |
+| **Spectral features** | 21,800 total |
+| **IR features** | 1,800 |
+| **¹H-NMR features** | 10,000 |
+| **¹³C-NMR features** | 10,000 |
+
+---
+
+## Data Preprocessing
+
+### Spectra Normalization
+
+- **IR and ¹H-NMR**: Normalized to [0, 1] range; peak shapes and relative intensities preserved
+- **¹³C-NMR**: Peak detection with 10% threshold, discretized into 80 bins (~2.75 ppm each)
+
+### Fragment Vocabulary
+
+The dataset uses a curated vocabulary of **991 SMARTS patterns** covering common organic motifs including:
+- Functional groups (hydroxyl, carbonyl, amine, etc.)
+- Ring systems (aromatic, aliphatic, hetero)
+- Chain patterns and substituents
+
+---
+
+## Usage
+
+### Loading with Python
+
+```python
+import numpy as np
+import h5py
+import pickle
+
+# Load SMILES
+smiles = np.load('multispectra/smiles.npy', allow_pickle=True)
+
+# Load spectra
+with h5py.File('multispectra/spectra.h5', 'r') as f:
+    spectra = f['spectra'][:]  # Shape: (N, 21800)
+
+# Split spectra into modalities
+ir_spectra = spectra[:, :1800]           # IR: 1800 features
+hnmr_spectra = spectra[:, 1800:11800]    # ¹H-NMR: 10000 features
+cnmr_spectra = spectra[:, 11800:]        # ¹³C-NMR: 10000 features
+
+# Load fragment counts
+with h5py.File('multispectra/substructure_counts.h5', 'r') as f:
+    fragment_counts = f['substructure_counts'][:]  # Shape: (N, 991)
+
+# Load splits
+with open('multispectra/split_indices.p', 'rb') as f:
+    splits = pickle.load(f)
+
+train_idx = splits['train']
+val_idx = splits['val']
+test_idx = splits['test']
+```
+
+### Using with NMIRacle Framework
+
+```python
+from nmiracle.data.datamodule import SpectralDataModule
+from nmiracle.data.tokenizer import BasicSmilesTokenizer
+
+tokenizer = BasicSmilesTokenizer()
+tokenizer.setup_alphabet(alphabet)
+
+datamodule = SpectralDataModule(
+    config=config.data,
+    tokenizer=tokenizer
+)
+datamodule.setup()
+```
+
+---
+
+## Dataset Creation
+
+### Source Data
+
+- **Pretrain molecules**: Combined from GDB-17 database (~3M) and SpectraBase (~140K), augmented with molecules from Alberts et al. dataset (~670K)
+- **Multispectra**: Derived from Alberts et al. (2024) multimodal spectroscopic dataset
+
+### Spectra Simulation
+
+All spectra are computationally simulated:
+- **IR**: Simulated infrared absorption using quantum chemical methods
+- **NMR**: Simulated using established chemical shift prediction algorithms
+
+### Fragment Extraction
+
+Fragment presence and counts are computed using RDKit's `GetSubstructMatches` function with the 991 SMARTS patterns in the vocabulary.
+
+---
+
+## Considerations for Using the Data
+
+### Social Impact
+
+This dataset supports automated molecular structure elucidation, which can:
+- **Accelerate drug discovery** by reducing manual spectral interpretation
+- **Democratize chemistry** by making structure elucidation accessible to non-experts
+- **Enable metabolomics research** through faster identification of unknown compounds
+
+### Biases
+
+- **Simulated spectra**: All spectra are computationally simulated and may not capture all experimental artifacts (noise, baseline drift, solvent effects)
+- **Chemical space coverage**: Dataset is biased toward drug-like organic molecules; may not generalize to organometallic compounds, polymers, or inorganic species
+- **Element diversity**: Limited to 9 element types (C, N, O, S, F, Cl, Br, P, I)
+
+### Limitations
+
+- **Stereochemistry**: IR and NMR spectra cannot distinguish absolute stereochemistry (enantiomers)
+- **Experimental gap**: Models trained on simulated spectra may require domain adaptation for experimental data
+- **Molecule size**: Limited to molecules with ≤35 heavy atoms
+
+---
+
+## Citation
+
+If you use this dataset in your research, please cite:
+
+```bibtex
+@article{nmiracle2025,
+  title={NMIRacle: Scalable Structure Elucidation from IR and NMR Spectra},
+  author={[Author Names]},
+  journal={arXiv preprint arXiv:XXXX.XXXXX},
+  year={2025}
+}
+```
+
+### Related Datasets
+
+- Alberts et al. (2024): Original multimodal spectroscopic dataset
+- GDB-17: Chemical universe database
+
+---
+
+## License
+
+This dataset is released under the **MIT License**.
+
+---
+
+## Dataset Curators
+
+- [Your Name] - Imperial College London
+
+## Contact
+
+For questions or issues regarding the dataset, please open an issue on the [NMIRacle GitHub repository](https://github.com/yourusername/nmiracle) or contact [your.email@example.com].