license: apache-2.0
task_categories:
- fill-mask
- text-classification
tags:
- biology
Dataset Summary This repository contains datasets derived from UniProt and SwissProt databases for analyzing biological sequences. The data includes raw sequences, processed files, and benchmark datasets split using hierarchical strategies to ensure diversity and discourage model bias. The datasets are organized into different folders for balanced and unbalanced splits, with additional variations where full-length sequences are chunked into smaller sequences.
Dataset Structure The file directory is structured as follows:
UniprotAndSwissprotDatasets:
SwissprotDatasets:
BalancedSwissprot: train.csv, validation.csv
UnbalancedSwissprot: train.csv, validation.csv
**test.csv, test1.csv, test2.csv SwissprotDatasetsChunked: Contains the same data as SwissprotDatasets, but each full-length sequence is split into smaller sequences of 20 read length. BalancedSwissprot: train.csv, validation.csv UnbalancedSwissprot: train.csv, validation.csv **test.csv, test1.csv, test2.csv
UniprotDatasets:
BalancedUniprot: train.csv, validation.csv
UnbalancedUniprot: train.csv, validation.csv
**test.csv, test1.csv, test2.csv UniprotDatasetsChunked: Contains the same data as UniprotDatasets, but each full-length sequence is split into smaller sequences of 20 read length. BalancedUniprot: train.csv, validation.csv UnbalancedUniprot: train.csv, validation.csv **test.csv, test1.csv, test2.csv
Data Description
The datasets are derived from:
UniProt: Contains two main sections: TrEMBL: Automatically annotated; contains homology-based annotations with more errors. SwissProt: Manually curated with experimental evidence of functional annotations. Preprocessing Steps: Organism Filtering: Only prokaryotic organisms (bacteria and archaea) were included. Data Mapping: UniProt IDs were mapped to UniRef clusters (UniRef50, UniRef90, UniRef100) and EMBL CDS IDs. Data Cleaning: Removed records with missing or partial EC numbers. Split records with multiple EC numbers into individual entries. Dataset Splitting A hierarchical splitting strategy ensured sequences from the same cluster do not appear in more than one set:
Train, Validation, and Test Sets are derived based on UniRef50, UniRef90, or UniRef100 clusters. Special test sets include: In-Distribution (Test Set-I): Contains sequences with EC numbers present in the training sets. Out-of-Distribution (Test Set-II): Contains sequences with EC numbers absent from the training sets. Benchmarks Four benchmark datasets were created:
Benchmark-I: SwissProt+TrEMBL (unbalanced). Benchmark-II: SwissProt+TrEMBL (balanced). Benchmark-III: SwissProt only (unbalanced). Benchmark-IV: SwissProt only (balanced). Chunked Sequences ChunkedSwissProt and ChunkedUniProt folders contain the same data as their respective non-chunked counterparts but with each full-length sequence divided into smaller 20 read-length sequences. These chunked datasets are especially useful for sequence alignment and similarity-based tasks. Dataset Generation To generate the dataset, follow the instructions in the notebooks:
DatasetGeneration.ipynb: Processes raw data to create final datasets. DatasetSplitting.ipynb: Splits the dataset into training, validation, and test sets. Citation If using this dataset, please cite the original UniProt and SwissProt resources:
UniProt: https://www.uniprot.org/ SwissProt: https://www.uniprot.org/uniprotkb?query=reviewed:true License This dataset is licensed under the Apache 2.0 License.