PhyMycoProm Hard-Negative Promoter Benchmark

Dataset summary

PhyMycoProm is a processed, cross-species benchmark for binary promoter recognition in three plant and two fungal species. It contains 45,000 transcription-oriented DNA windows of 300 nucleotides: 22,500 experimentally supported promoters and 22,500 difficult negatives comprising dinucleotide-preserving shuffles, gene-body windows and near-transcription- start-site (near-TSS) windows.

This release contains the hard-negative benchmark used for all quantitative results currently reported in the main manuscript, tentatively entitled PhyMycoProm: A Parameter-Efficient Cross-Species DNA Foundation Model for Plant–Fungal Promoter Recognition. It does not contain the historical main benchmark or the 67,500-record construction union.

The data are class-balanced by design. They should not be used to estimate the natural genomic prevalence of promoters.

File

File	Rows	Columns	Description
phymycoprom.csv	45,000	40	Complete hard-negative benchmark, including sequences, labels, provenance, quality metrics and leakage-control groups

CSV SHA-256:

61492e688662060704f768f955e6e5178873d9e6b38bdf47fc3a76ff73da579a

Missing or inapplicable values are represented by empty CSV fields. No sentinel value such as -1 or NA is used.

Dataset composition

Each species contributes 4,500 promoters and 4,500 negatives. The negative quota per species is fixed at 2,250 shuffles, 1,125 gene-body windows and 1,125 near-TSS windows.

Code	Species	Kingdom	Assembly	Promoters	Shuffles	Gene body	Near TSS	Total
Ath	Arabidopsis thaliana	Plant	TAIR10	4,500	2,250	1,125	1,125	9,000
Hvu	Hordeum vulgare	Plant	MorexV3	4,500	2,250	1,125	1,125	9,000
Zma	Zea mays	Plant	B73 RefGen_v3	4,500	2,250	1,125	1,125	9,000
Sce	Saccharomyces cerevisiae	Fungus	R64-1-1	4,500	2,250	1,125	1,125	9,000
Spo	Schizosaccharomyces pombe	Fungus	ASM294v2	4,500	2,250	1,125	1,125	9,000
Total				22,500	11,250	5,625	5,625	45,000

Additional integrity statistics:

• 45,000 unique sequence_id values;
• 44,973 unique DNA strings;
• 43,893 MMseqs2 sequence clusters represented in this hard subset;
• 27,219 dependency-aware cv_group values;
• sequence length is exactly 300 nt for every record;
• the released sequences contain only A, C, G and T.

Sequence and class definitions

Positive promoters

Positive records were obtained from EPDnew. Each record covers positions [-249, +50] relative to its annotated TSS and is oriented 5′ to 3′ in the direction of transcription. Coordinates in start and end use zero-based, half-open genomic intervals; tss_coordinate is the one-based genomic TSS coordinate, and tss_index=249 is its zero-based position within the transcription-oriented sequence.

All retained promoters passed an exact comparison against the corresponding 60-nt EPDnew reference sequence spanning [-49, +10]. For Zma, a recurrent 64-bp coordinate offset was corrected only when a unique exact EPDnew reference match was found within ±10 kb; the per-record correction is retained in epd_coordinate_shift.

Difficult negatives

• shuffle: a non-identical shuffle of a promoter that preserves its exact directed dinucleotide counts. parent_sequence_id identifies the source promoter.
• gene_body: a 300-nt window sampled from a coding sequence or intron. gene_body_subtype records CDS or intron.
• near_tss: a 300-nt genomic window centred 500–2,000 bp upstream or downstream of an annotated TSS. anchor_epd_id identifies that TSS.

Every negative is paired one-to-one with a positive from the same species through matched_positive_id. Genomic negatives were matched on GC fraction with an absolute difference no greater than 0.02. For a negative sequence, index 249 may be used as the model input anchor, but it is not an annotated TSS.

Data sources

Species	Promoter source	Reference genome and annotation
Ath	EPDnew release 004	TAIR10; Ensembl Plants release 62
Hvu	EPDnew release 000	MorexV3; Ensembl Plants release 62
Zma	EPDnew release 001	B73 RefGen_v3; NCBI RefSeq GCF_000005005.1, Annotation Release 100
Sce	EPDnew release 002	R64-1-1; Ensembl Fungi release 62
Spo	EPDnew release 002	ASM294v2; Ensembl Fungi release 62

The benchmark was built with random seed 20260720. Main- and hard-benchmark construction records were jointly clustered with MMseqs2 18.8cc5c using at least 80% sequence identity and 80% bidirectional coverage. Low-complexity masking was disabled during clustering.

Leakage-safe evaluation

All records have split=all; this release intentionally does not define train, validation or test subsets. Do not split the data by row.

Use cv_group as the indivisible partitioning unit. It combines sequence similarity clusters with known biological and synthetic dependencies, including:

• promoters and linked records from the same gene;
• near-TSS windows and their TSS anchor;
• gene-body windows and their host gene;
• shuffled negatives and their parent promoter.

The manuscript used a fixed five-fold mapping generated on the joint construction union and projected onto this hard subset. That exact fold number is not encoded in this two-file dataset release. It should be archived with the accompanying analysis code for exact numerical reproduction. For a new evaluation, a leakage-safe grouped split can be generated as follows:

import pandas as pd
from sklearn.model_selection import StratifiedGroupKFold

df = pd.read_csv(
    "phymycoprom.csv",
    dtype={"chromosome": "string"},
)
strata = df["species"].astype(str) + "::" + df["label"].astype(str)

cv = StratifiedGroupKFold(
    n_splits=5,
    shuffle=True,
    random_state=20260720,
)

for fold, (train_index, test_index) in enumerate(
    cv.split(df, y=strata, groups=df["cv_group"])
):
    train = df.iloc[train_index]
    test = df.iloc[test_index]

All preprocessing learned from the data, including feature scaling and threshold selection, must be fitted on training records only. For held-out-species or held-out-kingdom evaluation, remove every source record whose cv_group is also present in the target partition.

Loading the data

After cloning or downloading this repository:

from datasets import load_dataset

dataset = load_dataset(
    "csv",
    data_files={"benchmark": "phymycoprom.csv"},
)["benchmark"]

print(dataset)

With pandas:

import pandas as pd

df = pd.read_csv(
    "phymycoprom.csv",
    dtype={"chromosome": "string"},
)
print(df.shape)  # (45000, 40)

Data dictionary

All fractions are unitless and lie in [0, 1]. Coordinates or identifiers that do not apply to a record type are left empty.

Field	Type	Definition
sequence_id	string	Unique stable record identifier
species	category	Species code: Ath, Hvu, Zma, Sce or Spo
scientific_name	string	Full binomial species name
kingdom	category	plant or fungus
assembly	string	Reference-genome assembly
chromosome	string, nullable	Chromosome or scaffold for genomic records; empty for synthetic shuffles
start	integer, nullable	Zero-based inclusive interval start
end	integer, nullable	Zero-based exclusive interval end
coordinate_system	category	0-based_half-open for genomic records or synthetic for shuffles
strand	category	Transcriptional orientation (+ or -); . for synthetic shuffles
gene_id	string	Associated promoter, host or parent gene identifier
epd_id	string, nullable	EPDnew promoter identifier; populated for positives
tss_coordinate	integer, nullable	One-based genomic TSS coordinate for positives
tss_index	integer, nullable	Zero-based TSS index within the oriented positive sequence; always 249 when populated
sequence	string	Uppercase, transcription-oriented 300-nt DNA sequence
length	integer	Sequence length; always 300
label	integer	Binary target: 1 promoter, 0 negative
negative_type	category, nullable	shuffle, gene_body or near_tss; empty for positives
gene_body_subtype	category, nullable	CDS or intron for gene-body negatives
gc_content	float	Fraction of canonical bases that are G or C
n_fraction	float	Fraction of ambiguous/non-ACGT bases; zero in this release
repeat_fraction	float	Fraction of bases marked as repeats by source soft masking or repeat annotation
low_complexity_fraction	float	Fraction flagged by the benchmark's short-DNA low-complexity mask
repeat_or_low_complexity_fraction	float	Fraction in the union of repeat and low-complexity masks
matched_positive_id	string, nullable	One-to-one GC-matched positive identifier for each negative
gc_difference	float, nullable	Absolute GC-fraction difference from matched_positive_id
same_chromosome_match	boolean, nullable	Whether a genomic negative and its matched positive are on the same chromosome
nearest_tss_distance	integer, nullable	Distance in bp from the genomic window centre to the nearest annotated TSS; zero for positives
gene_overlap_bp	integer, nullable	Number of interval bases overlapping an annotated gene
parent_sequence_id	string, nullable	Parent positive identifier for a shuffled negative
anchor_epd_id	string, nullable	EPDnew anchor for a near-TSS window or shuffled parent
evidence_score	float, nullable	EPDnew expression/evidence score for a positive promoter
epd_reference_match	boolean, nullable	Whether the positive passed the exact 60-nt EPDnew reference check
epd_coordinate_shift	integer, nullable	Coordinate correction in bp applied before positive-window extraction
cluster_id	string	MMseqs2 connected-component sequence-cluster identifier
cv_group	string	Dependency-aware, leakage-control partitioning group
split	category	Always all; no predefined train/test split
source	string	Immediate sequence source or derivation method
source_release	string	EPDnew or reference-genome source release
dataset_membership	category	main;hard for positives shared during construction or hard for hard-only negatives

dataset_membership=main;hard does not indicate that the record was evaluated on the historical main benchmark. It records that the same positive promoter was shared by both construction sets; every row in this file belongs to the hard benchmark.

Quality control

The project validation checks cover:

• species, class and negative-type quotas;
• sequence length and DNA alphabet;
• promoter orientation, TSS indexing and exact EPDnew reference matching;
• GC, ambiguous-base, repeat and low-complexity thresholds;
• one-to-one positive–negative matching;
• exact dinucleotide-count preservation for shuffled negatives;
• MMseqs2 membership and dependency-aware grouping;
• absence of cv_group leakage in the recorded project folds;
• output hashes.

Selection thresholds were an ambiguous-base fraction below 0.05 and a repeat-or-low-complexity fraction below 0.20. The final released sequences contain no ambiguous bases.

Intended uses and limitations

Appropriate uses include promoter-versus-hard-negative classification, leakage-aware model comparison, species-held-out evaluation and plant–fungal transfer analysis.

Important limitations:

• only five species are represented;
• positives depend on EPDnew annotations and the listed reference assemblies;
• constructed negatives do not reproduce the full prevalence or diversity of genome-wide non-promoter sequence;
• class balance is artificial;
• correlated records must be handled through cv_group;
• performance on this benchmark is not evidence of a causal regulatory mechanism or validated motif function.

No human participants, clinical records or personally identifiable data are included.

Licence and redistribution status

The license: other metadata is intentional. No project-level open-data licence is asserted by this draft package. Before public release, the authors must verify the reuse and redistribution terms applicable to EPDnew, Ensembl-derived and NCBI-derived sequence material, then select a compatible repository licence. Do not replace this field with CC0, CC BY or another open licence until those rights have been confirmed.

Citation

Repository record: https://huggingface.co/datasets/YDXX/PhyMycoProm.

A DOI and final DataCite-style citation have not yet been assigned. After the manuscript metadata are finalized, add the verified creators, publication year, dataset version and persistent identifier here. Until then, cite the repository URL together with the accessed commit or revision; do not cite a local filesystem path.