FishCaduceus/FishCaduceus-Functional-Site-Benchmark

FishCaduceus Functional Site Benchmark

Dataset description

This dataset contains sequence-based benchmarks for four gene-annotation tasks used to evaluate FishCaduceus:

• translation initiation site (TIS) prediction
• translation termination site (TTS) prediction
• splice donor site prediction
• splice acceptor site prediction

The benchmark was designed to evaluate both within-species performance and cross-species transfer. Models are trained and selected only with labeled zebrafish (Danio rerio) data, then evaluated on a held-out zebrafish test set and six additional fish species without target-species retraining or parameter adjustment.

Dataset summary

The repository contains four task directories:

.
├── acceptor/
├── donor/
├── tis/
└── tts/

Each task directory contains:

• zebrafish training, validation, and held-out test files
• one cross-species test file for each of six additional fish species

The original CSV structure generated by the FishCaduceus benchmark pipeline has been preserved. Files were renamed and organized by task, species, and split without altering their internal contents.

Tasks

Translation initiation sites

Positive TIS examples correspond to annotated ATG translation initiation sites.

Translation termination sites

Positive TTS examples correspond to annotated TAA, TAG, or TGA translation termination sites.

Splice donor and acceptor sites

Positive splice-site examples correspond to annotated canonical GT-AG intron boundaries. The benchmark construction retained introns meeting the annotation-support and length criteria used in the FishCaduceus study.

For each task, negative examples were constructed from motif-matched and motif-free genomic positions while excluding annotated functional sites and their surrounding regions. The final benchmark uses a 1:6 positive-to-negative design. Sequences were standardized to the functional orientation before downstream modeling.

Species and evaluation splits

Species	Repository split	Role in the benchmark	Included in FishCaduceus pretraining
Danio rerio	train / validation / test	source species for training, model selection, and held-out evaluation	Yes
Ctenopharyngodon idella	test	cross-species evaluation	Yes
Carassius gibelio	test	cross-species evaluation	Yes
Culter alburnus	test	cross-species evaluation	No
Pelteobagrus fulvidraco	test	cross-species evaluation	No
Micropterus salmoides	test	cross-species evaluation	No
Larimichthys crocea	test	cross-species evaluation	No

Pelteobagrus fulvidraco is also referred to as Tachysurus fulvidraco in parts of the literature. Repository filenames follow pelteobagrus_fulvidraco.

Repository structure

.
├── acceptor/
│   ├── carassius_gibelio_test.csv
│   ├── ctenopharyngodon_idella_test.csv
│   ├── culter_alburnus_test.csv
│   ├── danio_rerio_test.csv
│   ├── danio_rerio_train.csv
│   ├── danio_rerio_validation.csv
│   ├── larimichthys_crocea_test.csv
│   ├── micropterus_salmoides_test.csv
│   └── pelteobagrus_fulvidraco_test.csv
├── donor/
│   └── [the same species and split organization]
├── tis/
│   └── [the same species and split organization]
└── tts/
    └── [the same species and split organization]

Dataset structure

Each CSV file contains the sequence examples and labels used in the corresponding FishCaduceus analysis. Column names and column order are retained exactly as generated by the original data-preparation pipeline.

The central fields used for model training are:

• a nucleotide sequence
• a binary class label indicating a positive or negative functional site

Depending on the task file, additional genomic-position or annotation fields may also be present. Users should inspect the CSV header before adapting the dataset to a new pipeline.

Sequence representation

The benchmark was constructed from 1,024-bp genomic sequence windows centered on candidate functional sites. Sequences are represented at single-nucleotide resolution and oriented consistently with the annotated functional strand.

Users applying models with a shorter context length should use a consistent centered truncation strategy.

Loading the data

With pandas

import pandas as pd

train = pd.read_csv("tis/danio_rerio_train.csv")
validation = pd.read_csv("tis/danio_rerio_validation.csv")
test = pd.read_csv("tis/danio_rerio_test.csv")

cross_species_test = pd.read_csv(
    "tis/ctenopharyngodon_idella_test.csv"
)

print(train.columns.tolist())
print(train.shape)

With Hugging Face Datasets

from datasets import load_dataset

data_files = {
    "train": "tis/danio_rerio_train.csv",
    "validation": "tis/danio_rerio_validation.csv",
    "test": "tis/danio_rerio_test.csv",
}

dataset = load_dataset("csv", data_files=data_files)
print(dataset)

A cross-species test file can be loaded separately:

from datasets import load_dataset

dataset = load_dataset(
    "csv",
    data_files={
        "test": "tis/larimichthys_crocea_test.csv",
    },
)

Benchmark protocol

The benchmark protocol used in the FishCaduceus study is:

1. train downstream models only on the Danio rerio training split;
2. select models and hyperparameters only on the Danio rerio validation split;
3. report within-species performance on the held-out Danio rerio test split;
4. evaluate the selected model directly on the other six species;
5. do not use labels from the target species for retraining or parameter adjustment.

The primary evaluation metric is area under the precision-recall curve (AUPRC), which is appropriate for the class-imbalanced test sets.

Intended uses

This dataset is intended for research on:

• fish genome annotation
• functional-site prediction
• transfer learning with DNA language models
• frozen-embedding classification
• full-model fine-tuning
• cross-species generalization
• comparative benchmarking of genomic sequence models

Limitations

• Labels are derived from reference genome annotations and therefore depend on annotation completeness and accuracy.
• The benchmark focuses on canonical TIS, TTS, and GT-AG splice-site definitions used in the FishCaduceus study.
• Negative examples are computationally constructed and do not represent every possible genomic background.
• Cross-species performance can be affected by genome assembly quality, annotation quality, phylogenetic distance, and sequence composition.
• The benchmark does not establish biological causality and should not replace experimental validation.
• The included species do not represent the full diversity of teleost fishes.

Related models

• FishCaduceus-20L-512
• FishCaduceus-28L-512
• FishCaduceus-28L-1024

Citation

The FishCaduceus manuscript is in preparation. Citation information will be added after publication.

When using this benchmark, please cite the FishCaduceus manuscript and the original genome and annotation resources used to construct the species-specific datasets.

Acknowledgements

FishCaduceus was developed for research on fish genomes at the Institute of Hydrobiology, Chinese Academy of Sciences.

Contact

Xiao-Qin Xia
Institute of Hydrobiology, Chinese Academy of Sciences
Email: xqxia@ihb.ac.cn