lamm-mit/protein-secondary-structure-nppe2

Protein Secondary Structure Prediction Dataset (NPPE-2)

Dataset Description

This dataset is used for the Protein Secondary Structure Prediction competition from the Deep Learning and Generative AI course at IIT Madras.

Task

Predict the secondary structure of proteins at the residue level using:

• —Q8 (sst8): Eight-state secondary structure classification (DSSP notation)
• —Q3 (sst3): Three-state secondary structure classification

Evaluation Metric

Harmonic mean of macro F1 scores: $$\text{score} = \frac{2 \cdot F1{Q8} \cdot F1{Q3}}{F1{Q8} + F1{Q3}}$$

Dataset Statistics

Sequence Length Statistics

• —Minimum: 20 residues
• —Maximum: 1,632 residues
• —Mean: 242.9 residues

Label Distribution

Q8 (Eight-State) Labels

Q3 (Three-State) Labels

File Structure

├── train.csv      # Training data with sequences and labels
├── test.csv       # Test sequences (no labels)
└── sample_submission.csv  # Submission format

Column Descriptions

train.csv:

• —id: Unique sequence identifier
• —seq: Amino acid sequence (20 standard amino acids)
• —sst8: Q8 secondary structure labels
• —sst3: Q3 secondary structure labels

test.csv:

• —id: Unique sequence identifier
• —seq: Amino acid sequence

Usage

from datasets import load_dataset

dataset = load_dataset("neuralninja110/protein-secondary-structure-nppe2")

# Access training data
train_data = dataset['train']
print(train_data[0])

Citation

If you use this dataset, please cite:

@misc{nppe2_2025,
  author = {Rahul Ashok},
  title = {Protein Secondary Structure Prediction Dataset (NPPE-2)},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/neuralninja110/protein-secondary-structure-nppe2}}
}

Author

Rahul Ashok Indian Institute of Technology Madras 23f1001177@ds.study.iitm.ac.in