README.md

# CAPSUL: A Comprehensive Human Protein Benchmark for Subcellular Localization

### 📊 Dataset Specifications

This repository contains the complete dataset of the paper: **"CAPSUL: A Comprehensive Human Protein Benchmark for Subcellular Localization"** *Accepted by **ICLR 2026***.

Here we introduce the protein dataset used in our **CAPSUL** benchmark evaluation, with comprehensive 3D information and fine-grained localization annotations. Data is derived from AlphaFold2, UniProt, and the Human Protein Atlas (HPA).

### 📂 Data Structure

```
CAPSUL/
├── AlphaFold_pdb.tar      # Original AlphaFold2 PDB files
├── coordinate.tar         # Extracted C-alpha (Cα) coordinates
└── localization/          # Annotations and sequence features
    ├── union.csv               # 20-class annotations (UniProt + HPA), used for evaluation in the paper
    ├── uniprot_detail.csv      # Original raw annotations (UniProt)
    ├── hpa_detail.csv          # Original raw annotations (HPA)
    ├── uniprot_20.csv          # 20-class annotations (UniProt)
    └── hpa_20.csv              # 20-class annotations (HPA)
```

*Note:* The implementation code is available at https://github.com/getbetter-hyccc/CAPSUL

### 🔍 Detailed Component Breakdown

- **`AlphaFold_pdb.tar`**: Contains the full original PDB files predicted by AlphaFold2. These include atomic coordinates, pLDDT scores, and complete protein metadata.
- **`coordinate.tar`**: To facilitate faster training for geometric models, we provide pre-extracted C-alpha (Cα) coordinates. This reduces I/O overhead by focusing on the protein backbone trajectory.
- **`localization/`**: Includes both the original raw labels and our integrated 20-class fine-grained labels. These annotations are specifically curated to evaluate model performance across diverse subcellular compartments, whose processing details can be found in our paper.