README.md

---
license: cc-by-3.0
language:
- en
pretty_name: Character Similarity Dataset
description: "Collection of textual trait descriptions of vertebrates (primarily fish) along with the corresponding ontology based similarity measures between trait description pairs. The distance is estimated using the Phenoscape Knowledgebase as the ontology."
task_categories:
- feature-extraction
tags:
- biology
- organism
- animals
- fish
- traits
- ontology
- phenoscape
size_categories: 10K<n<100K
configs:
  - config_name: full_data
    data_files:
      - split: train
        path: processed-data/all/*_TRAINING.tsv.gz
      - split: test
        path: processed-data/all/*_ALL_NON_TRAIN.tsv.gz
    default: true
  - config_name: characiformes
    data_files:
      - split: train
        path: processed-data/characiformes/*_TRAINING.tsv.gz
      - split: test
        path: processed-data/characiformes/*_ALL_NON_TRAIN.tsv.gz
  - config_name: cypriniformes
    data_files:
      - split: train
        path: processed-data/cypriniformes/*_TRAINING.tsv.gz
      - split: test
        path: processed-data/cypriniformes/*_ALL_NON_TRAIN.tsv.gz
  - config_name: gymnotiformes
    data_files:
      - split: train
        path: processed-data/gymnotiformes/*_TRAINING.tsv.gz
      - split: test
        path: processed-data/gymnotiformes/*_ALL_NON_TRAIN.tsv.gz
  - config_name: siluriformes
    data_files:
      - split: train
        path: processed-data/siluriformes/*_TRAINING.tsv.gz
      - split: test
        path: processed-data/siluriformes/*_ALL_NON_TRAIN.tsv.gz
---

# Dataset Card for Character Similarity Dataset

<!-- Provide a quick summary of what the dataset is or can be used for. --> 

## Dataset Details
The Character Similarity Dataset is a collection of textual trait descriptions along with the corresponding ontology based similarity measures between trait description pairs. The distance is estimated using the [Phenoscape Knowledgebase](https://kb.phenoscape.org/) as the ontology. The Knowledgebase is built upon a number of OBO ontologies, most importantly the Uberon anatomy ontology.

### Dataset Description

- **Curated by:** Jim Balhoff, Soumyashree Kar, Juan Garcia, Hilmar Lapp
- **Language(s) (NLP):** English
- **Repository:** [Imageomics/char-sim](https://github.com/Imageomics/char-sim)
- **Paper:** Coming soon!

The Character Similarity Dataset is a collection of 19K textual trait descriptions of fish and other vertebrates collected from the [Phenoscape Knowledgebase](https://kb.phenoscape.org/). The dataset also contains the corresponding pairwise similarity measures between trait descriptors (i.e., maxIC, Jaccard, SimGIC). These metrics estimate semantic similarity between the ontological representation of the traits descriptions per the Phenoscape Knowledgebase. The goal is to use this pairwise similarities to inform an embedding space that preserves the structure of the underlying ontology.

### Supported Tasks and Leaderboards
Task: Aligned feature extraction. Metric: Spearman's correlation coefficient. 

| Model              | Test set |
|:-------------------|:-----------|
| [**Trait2Vec**](https://huggingface.co/imageomics/trait2vec) | **0.7057** |

<!-- Provide benchmarking results -->


## Dataset Structure

```
raw-source/
    phenex-data-merged.ofn.gz
    phenoscape-kb-tbox-classified.ttl.gz
processed-data/
    all/
        data_{percentage}p_TRAINING.tsv.gz
        data_{percentage}p_ALL_NON_TRAIN.tsv.gz
        data_{percentage}p_NON_OVERLAP.tsv.gz
    characiformes/
        data_{percentage}p_TRAINING.tsv.gz
        data_{percentage}p_ALL_NON_TRAIN.tsv.gz
        data_{percentage}p_NON_OVERLAP.tsv.gz
    cypriniformes/
        data_{percentage}p_TRAINING.tsv.gz
        data_{percentage}p_ALL_NON_TRAIN.tsv.gz
        data_{percentage}p_NON_OVERLAP.tsv.gz
    gymnotiformes/
        data_{percentage}p_TRAINING.tsv.gz
        data_{percentage}p_ALL_NON_TRAIN.tsv.gz
        data_{percentage}p_NON_OVERLAP.tsv.gz
    siruliformes/
        data_{percentage}p_TRAINING.tsv.gz
        data_{percentage}p_ALL_NON_TRAIN.tsv.gz
        data_{percentage}p_NON_OVERLAP.tsv.gz

```

`phenex-data-merged.ofn.gz` and `phenoscape-kb-tbox-classified.ttl.gz` are raw data files built as part of the [Phenoscape Knowledgebase](https://kb.phenoscape.org/) construction pipeline. Running the processing script creates the four subset folders (`characiformes/`, `cypriniformes/`, `gymnotiformes/`, and `siruliformes/`, each an order of fish), then combines their data into the `all/` directory to create the training and test datasets. 

Note: `percentage` is the parameter passed for the percentage of the data to use for training; in this case, `percentage = 80`.

### Data Instances
Percentage is the proportion of data that will be used for training (i.e. data_{percentage}p_TRAINING.tsv.gz). In case the percentage is smaller than 100, the remaining proportion of the dataset is stored in data{percentage}p_ALL_NON_TRAIN.tsv.gz and a subset of this in data_{percentage}p_NON_OVERLAP.tsv.gz. All files follow the same structure.

<!--
Describe data files

Ex: All images are named <img_id>.png, each within a folder named for the species. They are 1024 x 1024, and the color has been standardized using <link to color standardization package>.
-->

### Data Fields

**data_{percentage}p_TRAINING.tsv.gz**: Collection of pairs of trait descriptors (i.e. `desc_1`, `desc_2`) along with corresponding ontology-based similarity metric (i.e. `maxIC`, `jaccard`, `simGIC`). Each pair of trait descriptors has a corresponding pair of identifiers (i.e. `id_1`, `id_2`) in [Phenoscape](https://kb.phenoscape.org/). The columns `character_1` and `character_2` correspond to [evolutionary characters](https://en.wikipedia.org/wiki/Character_evolution) for which the trait descriptions are states. These are used to split the dataset into disjoint train/test sets.
  - `id_1`: Phenoscape identifier for trait 1
  - `id_2`: Phenoscape identifier for trait 2
  - `maxIC`: Ontology-based similarity measure 
  - `jaccard`: Ontology-based similarity measure 
  - `simGIC`: Ontology-based similarity measure 
  - `order`: Pairwise score indices
  - `character_1`: Varying character for trait 1
  - `desc_1`: Textual trait description 1
  - `character_2`: Varying character for trait 2
  - `desc_2`: Textual trait description 2

Notes: (`id_1`, `character_1`, `desc_1`) are not fixed; `order` column will be removed in a future version.

### Data Splits
`data_{percentage}p_TRAINING.tsv.gz` is split into a proper training set and a proper testing set in the [train_model.py](https://github.com/Imageomics/char-sim/blob/feature/no-ref/ablations/train_model.py) file. 


## Dataset Creation

### Curation Rationale
The overall objective of the Phenoscape Project is to create a scalable infrastructure that enables linking descriptive phenotype observations across different fields of biology by the semantic similarity of their free-text descriptions. This linking is manual and labor intensive; see https://doi.org/10.1093/database/bav040. The dataset created by the Phenoscape project is used here to train a model which produces text embeddings aligned with ontology-derived similarity values.

### Source Data
Data was collected from the [Phenoscape Knowledgebase](https://kb.phenoscape.org/). `phenoscape-kb-tbox-classified.ttl.gz` contains the classified terminology composed of several OBO ontologies as well as generated class expressions materializing various axes of classification. `phenex-data-merged.ofn.gz` contains links from textual character trait descriptions to ontology class expressions.

#### Data Collection and Processing
The raw data is directly generated by the [Phenoscape build pipeline](https://github.com/phenoscape/pipeline). Further processing to compute ontology-based similarity scores, and to create derived files for model training, is defined by the workflow at https://github.com/Imageomics/char-sim.

#### Who are the source data producers?
The source data was originally created by the curators of the [Phenoscape Project](https://phenoscape.org), a collaborative NSF-funded project.

### Personal and Sensitive Information

None 

## Considerations for Using the Data
The distribution of SimGIC scores is skewed towards smaller values. This imbalance may cause the similarity of embeddings to follow the same bias. Consider subsampling to ensure uniform representation of distances. 

### Bias, Risks, and Limitations
This dataset has the biases of the Phenoscape ontology. This means the estimated models embeddings will inherit the ontology's inductive biases, coverage gaps, and evolving definitions. Biological conclusions may differ under alternative similarity metrics or other phenotype ontologies.

## Licensing Information
This dataset is licensed using a [Creative Commons Attribution 3.0 Unported license](http://creativecommons.org/licenses/by/3.0/) for the benefit of scientific pursuits. We ask that you cite the dataset and original source data using the below citations if you make use of it in your research.

## Citation

We ask that you cite the dataset and original source data using the below citations if you make use of it in your research.

**Data**
```
@misc{char-sim-data-2025,
  author = {Jim Balhoff and Soumyashree Kar and Juan Garcia and Hilmar Lapp},
  title = {Character Similarity Dataset},
  year = {2025},
  url = {https://huggingface.co/datasets/imageomics/char-sim-data},
  doi = {<doi once generated>},
  publisher = {Hugging Face}
}
```
<!--
-for an associated paper:
**Paper**
```
@article{<ref_code>,
  title    = {<title>},
  author   = {<author1 and author2>},
  journal  = {<journal_name>},
  year     =  <year>,
  url      = {<DOI_URL>},
  doi      = {<DOI>}
}
```
-->

Please be sure to also cite the original data source:

```bibtext
@ARTICLE{Balhoff2016-aw,
  title    = "The Phenoscape Knowledgebase: tools and {APIs} for computing
              across phenotypes from evolutionary diversity and model organisms",
  author   = "Balhoff, James P and {Phenoscape project team}",
  journal  = "bioRxiv",
  pages    =  071951,
  abstract = "The Phenoscape Knowledgebase (KB) is an ontology-driven database
              that combines existing phenotype annotations from model organism
              databases with new phenotype annotations from the evolutionary
              literature. Phenoscape curators have created phenotype annotations
              for more than 5,000 species and higher taxa, by defining
              computable phenotype concepts for more than 20,000 character
              states from over 160 published phylogenetic studies. These
              phenotype concepts are in the form of Entity-Quality (EQ)
              compositions which incorporate terms from the Uberon anatomy
              ontology, the Biospatial Ontology (BSPO), and the Phenotype and
              Trait Ontology (PATO). Taxonomic concepts are drawn from the
              Vertebrate Taxonomy Ontology (VTO). This knowledge of comparative
              biodiversity is linked to potentially relevant developmental
              genetic mechanisms by importing associations of genes to
              phenotypic effects and gene expression locations from zebrafish
              (ZFIN), mouse (MGI), Xenopus (Xenbase), and human (Human Phenotype
              Ontology project). Thus far, the Phenoscape KB has been used to
              identify candidate genes for evolutionary phenotypes, to match
              profiles of ancestral evolutionary variation with gene phenotype
              profiles, and to combine data across many evolutionary studies by
              inferring indirectly asserted values within synthetic
              supermatrices. Here we describe the software architecture of the
              Phenoscape KB, including data ingestion, integration of OWL
              reasoning, web service interface, and application features.",
  month    =  jan,
  year     =  2016,
  url      = "http://biorxiv.org/cgi/content/short/071951",
  doi      = "10.1101/071951",
  language = "en"
}
```


## Acknowledgements

This work was supported by the [Imageomics Institute](https://imageomics.org), which is funded by the US National Science Foundation's Harnessing the Data Revolution (HDR) program under [Award #2118240](https://www.nsf.gov/awardsearch/showAward?AWD_ID=2118240) (Imageomics: A New Frontier of Biological Information Powered by Knowledge-Guided Machine Learning). Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

<!-- You may also want to credit the source of your data, i.e., if you went to a museum or nature preserve to collect it. -->

## Dataset Card Authors 

Juan Garcia, Jim Balhoff, and Elizabeth Campolongo

## Dataset Card Contact

Please open a [Discussion on the Community Tab](https://huggingface.co/datasets/imageomics/char-sim-data/discussions) with any questions on the dataset.