README.md

metadata

license: mit

VenusMutHub Dataset

VenusMutHub is a comprehensive collection of protein mutation data designed for benchmarking and evaluating protein language models (PLMs) on various mutation effect prediction tasks. This repository contains mutation data across multiple protein properties including enzyme activity, binding affinity, stability, and selectivity.

Dataset Overview

VenusMutHub includes:

• Mutation data for hundreds of proteins across diverse functional categories
• Evaluation metrics for multiple protein language models
• Raw mutation data with sequence information and fitness scores
• DOI references for all data sources

Repository Structure

The repository is organized as follows:

/mutant

Contains raw mutation data organized by functional categories:

• /activity: Enzyme activity data including kcat, km, and kcat/km measurements
• /dti_binding: Drug-target interaction binding data
• /ppi_binding: Protein-protein interaction binding data
• /selectivity: Enzyme selectivity data
• /stability: Protein stability measurements

Each CSV file in these directories follows the format:

mutant,mutated_sequence,fitness_score

Where:

• mutant: The mutation(s) in standard notation (e.g., V109I)
• mutated_sequence: The full protein sequence with the mutation(s)
• fitness_score: The experimental measurement for the specific property

/single_mutant

Contains filtered mutation data that includes only single-point mutations, maintaining the same directory structure as /mutant. Only files with at least 5 single mutations are included.

/model_evaluation

Contains evaluation metrics for various protein language models on the mutation datasets:

• accuracy.csv: Classification accuracy for each model on each dataset
• f1.csv: F1 scores for classification tasks
• ndcg.csv: Normalized Discounted Cumulative Gain for ranking tasks
• spearman.csv: Spearman correlation coefficients for regression tasks

/doi.csv

Maps each dataset to its original publication DOI for proper citation and reference.

Usage

This dataset can be used for:

1. Benchmarking protein language models on mutation effect prediction
2. Training new models for specific mutation prediction tasks
3. Comparative analysis of model performance across different protein properties
4. Meta-analysis of mutation effects across protein families

Citation

If you use the VenusMutHub dataset in your research, please cite VenusMutHub directly.

License

This dataset is released under the MIT license.