---
license: mit
pretty_name: DarkSpec
size_categories:
  - 1M<n<10M
tags:
  - mass-spectrometry
  - proteomics
  - de-novo-peptide-sequencing
  - semi-supervised-learning
---

# DarkSpec

DarkSpec is a curated collection of **4.5 million unlabeled tandem mass
spectra** selected from PRIDE for semi-supervised de novo peptide sequencing.
It provides quality-controlled spectra that can be used without peptide
identification labels.

DarkSpec accompanies
[SemiNovo](https://github.com/grandOrgan/Seminovo), a framework for learning de
novo sequencing models from labeled and unlabeled spectra.

## Dataset summary

| Property | Value |
|---|---:|
| Number of spectra | 4,500,000 |
| Peaks per spectrum | 150 |
| Spectrum dtype | `float32` |
| Precursor charge range | 1-10 |
| Peptide labels | None |
| Peak-array size | Approximately 5.4 GB |

The dataset contains no peptide sequences, modified sequences, protein
accessions, database-search scores, or other identification labels.

## Files and schema

```text
DarkSpec/
├── manifest.json
├── precursor.npy
└── spectra.npy
```

| File | Shape | Dtype | Description |
|---|---:|---|---|
| `spectra.npy` | `(4,500,000, 150, 2)` | `float32` | Fragment m/z and normalized intensity |
| `precursor.npy` | `(4,500,000, 2)` | `float32` | Precursor m/z and precursor charge |
| `manifest.json` | - | JSON | Shape and preprocessing metadata |

For `spectra.npy`, `spectra[i, :, 0]` stores m/z values and
`spectra[i, :, 1]` stores intensities. Spectra with fewer than 150 retained
peaks are zero padded.

## Download

```bash
pip install -U huggingface_hub
hf download PanLiu/DarkSpec \
  --repo-type dataset \
  --local-dir DarkSpec
```

## Loading

Memory mapping is recommended:

```python
import json
import numpy as np

root = "DarkSpec"
spectra = np.load(f"{root}/spectra.npy", mmap_mode="r")
precursors = np.load(f"{root}/precursor.npy", mmap_mode="r")

with open(f"{root}/manifest.json") as handle:
    manifest = json.load(handle)

mz = spectra[0, :, 0]
intensity = spectra[0, :, 1]
precursor_mz, precursor_charge = precursors[0]

print(spectra.shape)
print(precursors.shape)
print(manifest)
```

## Preprocessing

The released array store applies the following fixed preprocessing:

- precursor charge between 1 and 10;
- fragment m/z between 50.52564895 and 2500 Da;
- peaks within 2 Da of precursor m/z removed;
- relative intensity threshold of 0.01;
- at least 20 valid peaks required;
- at most 150 peaks retained;
- retained peaks sorted by m/z;
- duplicate-like spectra reduced using 32-bit spectral SimHash grouping.

The exact frozen preprocessing metadata is also stored in `manifest.json`.

## Intended use

DarkSpec is intended for:

- semi-supervised de novo peptide sequencing;
- self-supervised spectrum representation learning;
- robustness and domain-shift studies for tandem mass spectra;
- reproducible benchmarking of unlabeled-spectrum learning methods.

It is not intended to provide peptide identifications or to replace
database-search validation.

## Limitations

- DarkSpec is unlabeled, so individual peptide identities are unknown.
- PRIDE acquisition protocols, instruments, collision settings, and biological
  sources are heterogeneous.
- Quality filtering reduces obvious invalid spectra but does not guarantee that
  every retained spectrum is identifiable.
- Models trained on DarkSpec should still be evaluated on held-out labeled
  benchmarks.

## Code

Training and evaluation code is available at:

<https://github.com/grandOrgan/Seminovo>

## License

The released dataset package is provided under the MIT License. Users remain
responsible for following applicable terms associated with the original PRIDE
source projects.