| | --- |
| | license: mit |
| | task_categories: |
| | - text-classification |
| | language: |
| | - en |
| | tags: |
| | - biology |
| | - chemistry |
| | - medical |
| | pretty_name: ProteinFamilyClassification |
| | size_categories: |
| | - 1K<n<10K |
| | --- |
| | |
| | ## Dataset Description |
| |
|
| | This dataset contains a curated subset (14% of the original) of protein sequences from the **Astral SCOPe 2.08 genetic domain sequence subsets**. |
| | It is designed for **protein family classification** tasks, where the goal is to assign each amino acid sequence to its corresponding **SCOPe family**. |
| |
|
| | ### Key Features |
| | - **Source:** Derived from the SCOPe database, which provides a hierarchical classification of protein structural domains based on experimental structural data from the Protein Data Bank (PDB). |
| | - **Classes:** Seven SCOPe protein classes (a–g), covering alpha proteins, beta proteins, mixed alpha/beta proteins, multi-domain proteins, membrane proteins, and small proteins. |
| | - **Embeddings:** Precomputed embeddings generated using the **ESM-2 transformer model** from Meta AI, allowing researchers to skip the computationally expensive embedding step. |
| | - **Purpose:** Ideal for training and testing classification models on protein sequence data with minimal preprocessing. |
| |
|
| | ### Why This Dataset? |
| | Generating embeddings for the full dataset takes ~3-7 hours on typical hardware. This preprocessed version provides a ready-to-use format for quick experimentation, making it accessible to teams with limited compute resources. |
| |
|
| | ### Potential Uses |
| | - Benchmarking machine learning models for protein classification. |
| | - Experimenting with XGBoost, MLP, CNN, or other classifiers. |
| | - Teaching and demonstration purposes in bioinformatics and computational biology. |
| |
|
| | --- |
| |
|
