FishCaduceus/FishCaduceus-Evolutionary-Constraint-Benchmark

FishCaduceus Evolutionary Constraint Benchmark

Dataset description

This dataset contains sequence-based benchmarks for evaluating whether FishCaduceus representations capture evolutionary constraint in fish genomes.

Constraint labels were derived from a 26-fish whole-genome alignment generated with Progressive Cactus. Grass carp (Ctenopharyngodon idella) was used as the primary reference genome for defining aligned and conserved positions. The resulting labeled sites were converted into fixed-length 512-bp sequence windows for zero-shot scoring, frozen-embedding classification, and cross-species evaluation.

Dataset summary

The repository contains two complementary benchmark components:

.
├── region_specific/
└── whole_genome/

• whole_genome/ contains the main training, validation, and cross-species test datasets.
• region_specific/ contains balanced test sets stratified by CDS, intron, and intergenic regions.

The original CSV structure generated by the FishCaduceus constraint pipeline has been preserved. Files were renamed and organized without altering their internal contents.

Constraint-label construction

For each reference position, bases projected from the other genomes in the 26-fish alignment were summarized using:

• n_cov: number of species with an informative aligned base
• n_match: number of informative species matching the reference base

The benchmark uses the following definitions:

Highly conserved:
n_cov >= 24 and n_match >= 24

Low conservation:
n_cov >= 24 and n_match <= 10

Positions that did not meet either definition were not used as positive or negative benchmark examples.

When a species contributed multiple aligned copies, concordant copies were merged into one species-level vote. Conflicting copies were treated as ambiguous and were excluded from n_cov and n_match.

Sequence representation

Each benchmark example is represented by a 512-bp genomic sequence window centered on the labeled position.

The positive class represents highly conserved positions, whereas the negative class represents low-conservation positions. The distributed CSV files retain the label encoding used by the original FishCaduceus analysis pipeline.

Whole-genome benchmark

The whole_genome/ directory contains:

Species	Available splits	Role
Ctenopharyngodon idella	train / validation / test	reference species for downstream training and held-out evaluation
Barbodes kweichowensis	test	cross-species evaluation
Danio rerio	test	cross-species evaluation
Megalobrama amblycephala	test	cross-species evaluation

The benchmark construction used:

Species and split	Number of examples
Ctenopharyngodon idella train	70,000
Ctenopharyngodon idella validation	10,000
Ctenopharyngodon idella test	20,000
Each cross-species test set	20,000

These files support training on grass carp and direct evaluation on other fish species without target-species retraining.

Region-specific benchmark

The region_specific/ directory contains CDS, intron, and intergenic test sets for:

• Ctenopharyngodon idella
• Danio rerio
• Megalobrama amblycephala

Each regional test file was constructed as a balanced dataset containing 5,000 highly conserved and 5,000 low-conservation examples.

The regional benchmarks are intended to test whether constraint prediction performance differs among coding and noncoding genomic contexts.

Repository structure

.
├── region_specific/
│   ├── ctenopharyngodon_idella/
│   │   ├── cds_test.csv
│   │   ├── intergenic_test.csv
│   │   └── intron_test.csv
│   ├── danio_rerio/
│   │   ├── cds_test.csv
│   │   ├── intergenic_test.csv
│   │   └── intron_test.csv
│   └── megalobrama_amblycephala/
│       ├── cds_test.csv
│       ├── intergenic_test.csv
│       └── intron_test.csv
└── whole_genome/
    ├── barbodes_kweichowensis_test.csv
    ├── ctenopharyngodon_idella_test.csv
    ├── ctenopharyngodon_idella_train.csv
    ├── ctenopharyngodon_idella_validation.csv
    ├── danio_rerio_test.csv
    └── megalobrama_amblycephala_test.csv

Dataset structure

Each CSV contains the sequence examples and binary labels used in the FishCaduceus evolutionary-constraint analyses. Column names and column order are retained exactly as generated by the original pipeline.

The core fields used for modeling are:

• a 512-bp nucleotide sequence
• a binary constraint label

Depending on the source file, additional fields related to genomic coordinates or alignment-derived statistics may also be present. Users should inspect the CSV header before adapting the dataset to a new workflow.

Loading the data

With pandas

import pandas as pd

train = pd.read_csv(
    "whole_genome/ctenopharyngodon_idella_train.csv"
)
validation = pd.read_csv(
    "whole_genome/ctenopharyngodon_idella_validation.csv"
)
test = pd.read_csv(
    "whole_genome/ctenopharyngodon_idella_test.csv"
)

cross_species_test = pd.read_csv(
    "whole_genome/danio_rerio_test.csv"
)

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

With Hugging Face Datasets

from datasets import load_dataset

data_files = {
    "train": "whole_genome/ctenopharyngodon_idella_train.csv",
    "validation": "whole_genome/ctenopharyngodon_idella_validation.csv",
    "test": "whole_genome/ctenopharyngodon_idella_test.csv",
}

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

A region-specific dataset can be loaded separately:

from datasets import load_dataset

dataset = load_dataset(
    "csv",
    data_files={
        "test": (
            "region_specific/danio_rerio/"
            "cds_test.csv"
        ),
    },
)

Benchmark protocol

The supervised transfer protocol used in the FishCaduceus study is:

1. train a classifier using only the Ctenopharyngodon idella training split;
2. perform model selection using only the Ctenopharyngodon idella validation split;
3. report held-out performance on the Ctenopharyngodon idella test split;
4. evaluate the selected classifier directly on the other species;
5. do not use target-species labels for retraining or parameter adjustment.

The dataset can also be used for zero-shot scoring by masking or otherwise scoring the center nucleotide with a pretrained DNA language model.

Intended uses

This dataset is intended for research on:

• evolutionary constraint prediction
• conservation-aware genomic representation learning
• frozen-embedding classification
• zero-shot nucleotide scoring
• cross-species transfer
• comparative genomics
• evaluation across CDS, intronic, and intergenic regions

Limitations

• Constraint classes are computational labels derived from a specific 26-fish whole-genome alignment and fixed thresholds.
• The labels are not equivalent to experimentally validated functional annotations.
• Results depend on genome assembly quality, alignment quality, reference choice, species sampling, and multi-copy handling.
• Repetitive, duplicated, poorly assembled, or rapidly evolving regions may be underrepresented or more difficult to align.
• The highly conserved and low-conservation classes represent the two ends of the alignment-based conservation spectrum and do not cover all genomic positions.
• The included test species and genomic regions do not represent the full diversity of fish genomes.

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 resources used in the 26-fish whole-genome alignment.

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