---
datasets:
- Metanova/Proteins
task_categories:
- tabular-classification
- tabular-regression
- feature-extraction
tags:
- proteins
- genomics
- bioinformatics
- metanova
- amino-acid-sequences
- protein-foldingprotein-folding
pretty_name: Metanova Proteins Dataset
size_categories:
- 1M<n<10M
---

# Dataset Card for Metanova / Proteins

## Dataset Details

### Dataset Description

- **Name:** Metanova / Proteins  
- **Curated by:** Metanova Labs ([Hugging Face profile](https://huggingface.co/Metanova))  
- **Shared by:** Metanova Labs  
- **Language(s):** Not natural language; amino acid sequences (protein sequences) using the standard 20-letter code (A, C, D, …) and possibly special/unknown letters.  
- **License:** *needs verification*  

A large collection of protein sequences hosted by Metanova Labs. The dataset contains approximately 2 million sequences. It is suitable for machine learning tasks in protein informatics such as protein language modeling, representation learning, and sequence-based function or structure inference.  

### Dataset Sources

- **Repository:** [Metanova/Proteins on Hugging Face](https://huggingface.co/datasets/Metanova/Proteins)  

---

## Uses

### Direct Use

- Training protein language models  
- Learning protein sequence embeddings for downstream tasks  
- Protein clustering, similarity search, and phylogenetic analysis  
- Transfer learning for structure or function prediction  

### Out-of-Scope Use

- Applications requiring curated annotations (e.g., detailed functional labels, structures) unless combined with external databases  
- Clinical or diagnostic decision-making without experimental validation  
- Use in tasks where data provenance or sequence redundancy control is critical without further preprocessing  

---

## Dataset Structure

- **Format:** CSV / tabular data  
- **Rows:** Individual protein sequences (~2.07 million)  
- **Columns:** (to verify in the dataset files)  

  | Field (expected) | Type | Description |
  |------------------|------|-------------|
  | `id` or `sequence_id` | string | Unique identifier for each sequence |
  | `sequence` | string | Protein sequence in single-letter amino acid code |
  | `length` | integer | Sequence length |
  | `organism` / `taxonomy` | string (optional) | Source organism or taxonomic category |
  | `annotation` | string (optional) | Functional / descriptive annotation |

- **Splits:** No predefined train/validation/test splits  

---

## Dataset Creation

### Curation Rationale

This dataset was likely created to provide a large repository of protein sequences for use in computational biology, machine learning, and protein informatics research.  

### Source Data

#### Data Collection and Processing

- Exact collection methodology is **not specified**.  
- Likely compiled from publicly available sequence repositories (e.g., UniProt, RefSeq, or metagenomic datasets).  
- Unknown whether filtering, redundancy removal, or quality control were applied.  

#### Who are the source data producers?

- Likely generated by sequencing projects and deposited in public biological databases.  
- No explicit acknowledgment of original contributors is provided.  

### Annotations

- No additional human annotations appear to be provided.  
- No metadata regarding function, structure, or localization is included.  

#### Personal and Sensitive Information

- This dataset contains only biological sequences.  
- No personal, sensitive, or private human information is present.  

---

## Bias, Risks, and Limitations

- **Bias:** Likely overrepresentation of well-studied organisms (e.g., model organisms, pathogens).  
- **Redundancy:** Dataset may contain highly similar or duplicate sequences.  
- **Annotation gaps:** Lack of metadata limits supervised tasks.  
- **Technical risk:** Models trained directly on this dataset may overfit due to redundancy or taxonomic leakage.  

### Recommendations

- Perform sequence clustering or deduplication before training.  
- Create train/validation/test splits that respect taxonomy to avoid leakage.  
- If function/structure labels are needed, map these sequences to external annotated databases.  
- Contact Metanova Labs for clarification of license before commercial use.  

---

## Citation

If you use this dataset, please cite as:


**BibTeX (generic):**
```bibtex
@misc{metanova_proteins,
  author       = {Metanova Labs},
  title        = {Proteins Dataset},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Metanova/Proteins}}
}