ProtEnv: Protein Environment Dataset

Raw protein sequences, labels, and predicted structures for 15 downstream tasks used in ProtCompass benchmarking.

Dataset Overview

ProtEnv provides the foundational data for evaluating protein encoders across diverse biological tasks. It includes:

• Raw sequences: FASTA format protein sequences
• Labels: Task-specific annotations (regression/classification)
• Predicted structures: ESMFold-generated 3D structures (PDB format)
• Splits: Pre-defined train/test splits for reproducibility

Dataset Structure

structure_encoder_data/              # Raw sequences and labels (117GB)
├── contact_prediction/
│   ├── train.fasta
│   ├── test.fasta
│   ├── train_labels.npy
│   └── test_labels.npy
├── secondary_structure/
├── ppi_site/
├── metal_binding/
├── mutation_effect/
├── go_bp/
├── stability/
├── solubility/
├── go_mf/
├── fluorescence/                    # Note: "fluorescence" spelling maintained for compatibility
├── ec_classification/
├── subcellular_localization/
├── membrane_soluble/
├── remote_homology/
└── ppi_affinity/

predicted_structures/                # ESMFold structures (5GB compressed)
├── fluorescence.tar.gz              # 2.0GB → 444MB uncompressed
├── solubility.tar.gz                # 2.6GB → 580MB uncompressed
├── stability.tar.gz                 # 444MB → 98MB uncompressed
└── ppi_affinity.tar.gz              # 49MB → 11MB uncompressed

Task Descriptions

Protein Function Prediction

• EC Classification: Enzyme Commission number prediction (multi-class)
• GO-BP: Gene Ontology Biological Process (multi-label)
• GO-MF: Gene Ontology Molecular Function (multi-label)
• Subcellular Localization: Cellular compartment prediction (multi-class)

Protein-Protein Interactions

• PPI Site: Binding site prediction (binary per-residue)
• PPI Affinity: Binding affinity prediction (regression)

Structure Prediction

• Contact Prediction: Residue-residue contact maps (binary per-pair)
• Secondary Structure: 3-state or 8-state structure (per-residue)

Biophysical Properties

• Stability: Thermostability prediction (regression)
• Solubility: Expression solubility (binary)
• Fluorescence: GFP fluorescence intensity (regression)
• Metal Binding: Metal ion binding sites (binary per-residue)
• Membrane/Soluble: Membrane vs soluble classification (binary)

Sequence Analysis

• Remote Homology: Fold recognition (multi-class)
• Mutation Effect: Fitness effect prediction (regression)

Download Instructions

Full Dataset

# Clone the entire repository (122GB)
git lfs install
git clone https://huggingface.co/datasets/Anonymoususer2223/ProtEnv

Specific Tasks

from huggingface_hub import hf_hub_download

# Download raw sequences for a specific task
train_fasta = hf_hub_download(
    repo_id="Anonymoususer2223/ProtEnv",
    filename="structure_encoder_data/mutation_effect/train.fasta",
    repo_type="dataset"
)

# Download predicted structures
structure_tar = hf_hub_download(
    repo_id="Anonymoususer2223/ProtEnv",
    filename="predicted_structures/fluorescence.tar.gz",
    repo_type="dataset"
)

File Formats

Sequences

• Format: FASTA
• Headers: >protein_id or >protein_id|metadata
• Sequences: Standard 20 amino acids

Labels

• Format: NumPy arrays (.npy)
• Regression tasks: Float arrays
• Classification tasks: Integer arrays (class indices)
• Multi-label tasks: Binary matrices (N × num_classes)
• Per-residue tasks: 2D arrays (N × sequence_length)

Structures

• Format: PDB files (compressed as .tar.gz)
• Source: ESMFold predictions
• Quality: pLDDT scores included in B-factor column
• Note: Structures are predictions, not experimental

Usage Example

import numpy as np
from Bio import SeqIO
from huggingface_hub import hf_hub_download

# Load sequences
fasta_path = hf_hub_download(
    repo_id="Anonymoususer2223/ProtEnv",
    filename="structure_encoder_data/stability/train.fasta",
    repo_type="dataset"
)

sequences = []
ids = []
for record in SeqIO.parse(fasta_path, "fasta"):
    sequences.append(str(record.seq))
    ids.append(record.id)

# Load labels
labels_path = hf_hub_download(
    repo_id="Anonymoususer2223/ProtEnv",
    filename="structure_encoder_data/stability/train_labels.npy",
    repo_type="dataset"
)
labels = np.load(labels_path)

print(f"Loaded {len(sequences)} proteins")
print(f"First sequence: {sequences[0][:50]}...")
print(f"First label: {labels[0]}")

Dataset Statistics

Task	Train Size	Test Size	Label Type	Avg Length
Contact Prediction	25,299	40	Binary (L×L)	256
Secondary Structure	8,678	513	Multi-class (L)	208
PPI Site	15,051	1,672	Binary (L)	312
Metal Binding	5,654	629	Binary (L)	287
Mutation Effect	3,072	342	Regression	452
GO-BP	29,898	3,322	Multi-label (1,943)	394
Stability	53,614	2,512	Regression	178
Solubility	62,478	6,942	Binary	224
GO-MF	29,898	3,322	Multi-label (489)	394
Fluorescence	21,446	5,362	Regression	238
EC Classification	15,011	1,668	Multi-class (538)	382
Subcellular Localization	8,943	2,236	Multi-class (10)	493
Membrane/Soluble	3,797	423	Binary	312
Remote Homology	12,312	736	Multi-class (1,195)	209
PPI Affinity	3,899	434	Regression	156

Total: ~500K protein sequences across 15 tasks

Data Sources

All datasets are curated from public databases:

• UniProt: Protein sequences and annotations
• PDB: Experimental structures (for validation)
• CATH/SCOP: Fold classifications
• STRING: Protein-protein interactions
• Gene Ontology: Functional annotations
• Literature: Experimental measurements (fluorescence, stability, etc.)

Predicted Structures

Structures are generated using ESMFold (Lin et al., 2023) for tasks where experimental structures are unavailable:

• Fluorescence: 27,808 structures (GFP variants)
• Solubility: 69,420 structures
• Stability: 56,126 structures
• PPI Affinity: 4,333 structures

These structures enable structure-based encoder evaluation on tasks traditionally limited to sequence-only data.

Data Splits

All train/test splits are:

• Pre-defined: Ensures reproducibility across studies
• Non-overlapping: No sequence identity between train/test
• Stratified: Balanced label distributions where applicable
• Temporally split: For some tasks (e.g., mutation effect)

Known Issues

1. Spelling: "fluorescence" directory uses British spelling for historical compatibility
2. Structure quality: ESMFold predictions vary in quality (check pLDDT scores)
3. Label noise: Some experimental labels may contain measurement errors
4. Class imbalance: Some tasks have imbalanced class distributions

Citation

If you use ProtEnv, please cite:

@article{protcompass2026,
  title={ProtCompass: Interpretable Benchmarking and Task-Aware Evaluation of Protein Encoders},
  author={Your Name et al.},
  journal={NeurIPS},
  year={2026}
}

Related Resources

• Pre-computed Embeddings: ProtCompass_Embeddings on HuggingFace
• Code Repository: GitHub
• Paper: arXiv

License

MIT License - Free for academic and commercial use

Contact

For questions, issues, or data requests, please open an issue on the GitHub repository.