Rain021217/clinvar-pathogenicity-prediction-dataset

ClinVar Protein Variant Pathogenicity Prediction Dataset

This dataset contains the final processed modeling table used in the ClinVar-Pathogenicity-Prediction project. It is derived from the ClinVar variant summary and the UniProt human reference proteome, and is intended for machine-learning experiments on protein variant pathogenicity prediction.

Dataset Files

• data/train.parquet: 80% stratified training split with 5,009,581 rows.
• data/test.parquet: 20% stratified test split with 1,252,396 rows.
• data/modeling_dataset_final.parquet: full processed modeling dataset with 6,261,977 rows, 20 engineered protein-variant features, and one binary label column.
• metadata/feature_names.json: ordered feature names used by the modeling pipeline.
• metadata/summary.json: row count, label counts, and column metadata.

Label Definition

• label = 1: pathogenic or likely pathogenic variants.
• label = 0: benign or likely benign variants.

Variants of uncertain significance and conflicting interpretations were excluded.

Dataset Statistics

Class	Count	Percentage
Pathogenic / likely pathogenic	311,103	4.97%
Benign / likely benign	5,950,874	95.03%
Total	6,261,977	100.00%

Splits

The default configuration provides three splits:

Split	Rows	Pathogenic / likely pathogenic	Benign / likely benign
train	5,009,581	248,882	4,760,699
test	1,252,396	62,221	1,190,175
full	6,261,977	311,103	5,950,874

The train and test files were generated with an 80/20 stratified split using sklearn.model_selection.train_test_split(test_size=0.2, stratify=label, random_state=42).

Features

The dataset includes 20 engineered protein mutation features:

• blosum62_score
• delta_hydrophobicity
• delta_charge
• delta_volume
• delta_polarity
• delta_mw
• is_nonsense
• aromatic_change
• aliphatic_change
• relative_position
• distance_to_terminus
• is_terminal
• delta_alpha_propensity
• delta_beta_propensity
• grantham_distance
• context_hydrophobicity_mean
• context_charge_sum
• context_aromatic_count
• context_proline_count
• context_glycine_count

Usage

from datasets import load_dataset

dataset = load_dataset("Rain021217/clinvar-pathogenicity-prediction-dataset")
train_df = dataset["train"].to_pandas()
test_df = dataset["test"].to_pandas()

Use dataset["full"] if you need the complete processed modeling table in one split.

Source Data

• ClinVar variant summary, December 2025 release.
• UniProt human reference proteome UP000005640_9606.

Notes and Limitations

This dataset is intended for educational and research use. It should not be used as a standalone clinical diagnostic tool. Users should follow the usage terms of the original ClinVar and UniProt source databases.

Citation

If this dataset is useful for your work, please cite the GitHub project:

@misc{li2026clinvar,
  author = {Jinpeng Li},
  title = {Machine Learning-based Prediction of Pathogenicity for Protein Variants in ClinVar Database},
  year = {2026},
  publisher = {GitHub},
  url = {https://github.com/Rain021217/ClinVar-Pathogenicity-Prediction}
}