---

pretty_name: MSNet Datasets
language: en
tags:
  - proteomics
  - mass-spectrometry
  - deep-learning
  - bioinformatics
  - peptide-identification
  - license mit
  - size_categories 500M<n

---

# MSNet Datasets

## Dataset Description

MSNet Datasets is a large-scale, standardized, and AI-ready collection of mass spectrometry (MS) data designed for machine learning applications in computational proteomics.

Instead of hosting raw data directly, this Hugging Face dataset serves as an **entry point and interface**, providing standardized access to externally hosted MSNet resources.

The dataset addresses key limitations of existing public repositories, such as heterogeneous metadata, inconsistent processing pipelines, and lack of benchmarking standards, by offering a unified and reproducible data representation.

![Figure1_V8](https://cdn-uploads.huggingface.co/production/uploads/69cc8d6600711b4d415ba97a/gXBqhBR3l-LNL8l_jsUDI.png)

---

## Background

Deep learning has become integral to modern proteomics, supporting tasks such as:

* Fragment ion intensity prediction
* Retention time (RT) prediction
* Peptide–spectrum match (PSM) rescoring
* De novo peptide sequencing

Despite the abundance of publicly available MS data, most repositories primarily store raw files with inconsistent metadata and processing standards, making them difficult to use directly in machine learning workflows.

---

## Motivation

Current proteomics datasets often suffer from:

* Incomplete or inconsistent metadata
* Heterogeneous preprocessing pipelines
* Limited diversity in experimental conditions
* Lack of standardized benchmarks

MSNet Datasets provides a unified, curated, and ML-ready interface to facilitate reproducible research and fair model evaluation.

---

## Data Sources

The dataset is curated from:

* Public proteomics datasets (ProteomeXchange)
* Large-scale projects (e.g., π-HuB)

A total of **114 large-scale datasets** are included, covering diverse biological contexts, instrument platforms, and fragmentation strategies.

---

## Data Processing

All datasets are systematically reprocessed using a reproducible workflow:

* Metadata standardized using SDRF
* Uniform reanalysis of raw MS data
* Peptide-spectrum match (PSM) generation
* Peak annotation including:

  * b⁺, b²⁺, y⁺, y²⁺ ions
  * With and without neutral losses
  * Multiple mass tolerance settings applied during annotation
* Harmonization into a unified structure

---

## Data Format

Processed data are stored in **Parquet format**, enabling:

* Efficient storage and compression
* Fast I/O for large-scale data
* Compatibility with PyTorch, TensorFlow, and other ML frameworks

---

## Dataset Structure

Each entry corresponds to a peptide-spectrum match (PSM) and includes:

| Column          | Description                      |
| --------------- | -------------------------------- |
| spectrum_id     | Spectrum identifier              |
| mz_array        | m/z values                       |
| intensity_array | Intensity values                 |
| precursor_mz    | Precursor m/z                    |
| charge          | Charge state                     |
| peptide         | Peptide sequence                 |
| modifications   | Post-translational modifications |
| rt              | Retention time                   |
| instrument      | Instrument type                  |
| fragmentation   | Fragmentation method             |

---

## Access

⚠️ **Note:** Hugging Face does **not host the raw data**.

Instead, data can be accessed through the following official resources:

* **Web portal:** https://msnet.ncpsb.org.cn or https://quantms.org/datasets

---

## Official Loader

We provide an official data loader for seamless integration:

👉 https://github.com/PHOENIXcenter/pi-MSnet

Supports:

* PyTorch
* TensorFlow

---

## Usage

```python
from datasets import load_dataset

# This dataset provides metadata / interface only
dataset = load_dataset("your-username/msnet")

print(dataset)
```

For full data access, please use the official loader.

---

## Use Cases

* Training deep learning models for proteomics
* PSM rescoring and confidence estimation
* De novo peptide sequencing
* Retention time and intensity prediction
* Benchmarking computational methods

---

## Next step

* PTM coverage is continuously expanding
* Some modalities (e.g., DIA, XL-MS) are fully integrated

---

## Citation

If you use MSNet Datasets, please cite the corresponding π-MSNet publication.

---

## License

MIT License

---