Dataset Card for neuralbioinfo/PhaStyle-SequenceDB

phastyle Sequence Database

A collection of bacteriophage nucleotide sequences and metadata for training and evaluating phage lifestyle prediction models. Available splits support both strict-holdout and standard-holdout experiments.

Dataset Features

Name	Type	Description
sequence_id	int64	Unique integer identifier for each sequence
dataset	string	Source collection name (see “Splits” below)
class_label	string	Lifestyle label: "temperate" or "virulent"
y	int64	Binary label: 0 = temperate, 1 = virulent
source_fasta_id	string	Original FASTA record ID
orientation	string	Strand orientation: "forward" or "reverse_complement"
test_fastaid	string	FASTA ID used in the test split (if applicable)
L_seq	int64	Sequence length in base pairs
source_description	string	Free‐text description of isolate or environment
Genome type	string	“dsDNA”, “ssDNA”, etc.
Family	string	Taxonomic family
subds_seqid	int64	Sub‐dataset sequence index (internal use)
sequence	string	Nucleotide sequence

Splits

• BACPHLIP_TRAINING (1,798 examples, 116 MB)
  • Used for strict-holdout training (excludes any Escherichia-infecting phages or ≥ 80 % ANI to test set).
• BACPHLIP_VALIDATION (316 examples, 25 MB)
  • Escherichia phages held out for validation in the strict-holdout setting.
• BACPHLIP_ALL (2,114 examples, 142 MB)
  • Used for standard-holdout training (no ANI or host exclusions).
• ESCHERICHIA (394 examples, 22 MB)
  • Guelin collection: experimentally validated Escherichia phages.
• BASEL (412 examples, 27 MB)
  • BASEL collection: environmental E. coli isolates.
• EXTREMOPHILE (16 examples, 0.8 MB)
  • Phages from deep‐sea, acidic, and arsenic‐rich environments.

Dataset Creation

The sequences in this dataset were gathered from various sources, including the BACPHLIP database and curated collections of phages from extreme environments. Each sequence was carefully segmented into smaller fragments (512bp or 1022bp) to simulate real-world scenarios where phage sequences are often fragmented. The training data excludes Escherichia sequences, which are used in the test set to evaluate model generalization capabilities.

Intended Uses

This dataset is intended for use in phage lifestyle prediction tasks using genomic language models such as ProkBERT. The segmented sequences allow models to generalize well even with fragmented or out-of-sample data. It is particularly useful for applications in ecological and clinical settings where understanding phage behavior is critical.

Citing this work

If you use the data in this package, please cite:

@Article{ProkBERT2024,
  author  = {Ligeti, Balázs and Szepesi-Nagy, István and Bodnár, Babett and Ligeti-Nagy, Noémi and Juhász, János},
  journal = {Frontiers in Microbiology},
  title   = {{ProkBERT} family: genomic language models for microbiome applications},
  year    = {2024},
  volume  = {14},
  URL={https://www.frontiersin.org/articles/10.3389/fmicb.2023.1331233},       
    DOI={10.3389/fmicb.2023.1331233},      
    ISSN={1664-302X}
}