Datasets

mOWL is designed to handle input in OWL format. That is, you can input OWL ontologies. A mOWL dataset contains 3 ontologies: training, validation and testing.

Built-in datasets

There are several built-in datasets related to bioinformatics tasks such as protein-protein interactions prediction and gene-disease association prediction. Datasets can be found at Datasets API docs.

For \(\mathcal{EL}^{++}\) subsumption prediction, mOWL provides the GALEN, GO and Anatomy ontologies with the 80/10/10 benchmark split distributed with the reference implementation of [jackermeier2024], also used by the TransBox paper (WWW 2025):

from mowl.datasets.builtin import GALENJackermeier2024Dataset
ds = GALENJackermeier2024Dataset()
train_ontology = ds.ontology
valid_ontology = ds.validation
test_ontology = ds.testing

The training ontologies also contain \(\mathcal{EL}^{++}\) role inclusion and role chain axioms. Models built on top of EmbeddingELModel train on them when constructed with load_role_axioms=True and their module implements the two role axiom losses (see the EL guide).

The axioms and the split are the frozen release of the reference benchmark, which distributes them as integer-indexed splits plus the class and relation index files; mOWL serves the same axioms serialized as OWL. The terms of the underlying ontologies and of the reference benchmark apply to this redistribution.

Warning

In AnatomyJackermeier2024Dataset the class IRIs are not anatomy identifiers: the reference benchmark ships no class names for ANATOMY, so every class is an opaque generated IRI under http://bio2vec.net/data/mowl/anatomy_jackermeier2024/aux/. The axioms and the split are faithful, so ranking metrics are comparable, but predictions cannot be mapped back to anatomical terms. GALEN and GO keep their original IRIs.

To access any of these datasets you can use:

from mowl.datasets.builtin import PPIYeastSlimDataset
ds = PPIYeastSlimDataset()
train_ontology = ds.ontology
valid_ontology = ds.validation
test_ontology = ds.testing

evaluation_classes = ds.evaluation_classes

Built-in datasets already contain the attribute evaluation_classes, which is used to evaluate a model on the dataset. In the PPI example, the evaluation classes correspong to ontology classes representing proteins.

Where datasets are stored

Changed in version 2.2.0: Built-in datasets used to be downloaded into the current working directory. They are now cached in a user-level directory, shared by every working directory.

A built-in dataset is downloaded once and reused afterwards. The download location is resolved by default_data_root:

from mowl.datasets import default_data_root
cache_dir = default_data_root()

This is $XDG_CACHE_HOME/mowl/datasets, falling back to ~/.cache/mowl/datasets. Set the MOWL_DATA_ROOT environment variable to override it globally:

export MOWL_DATA_ROOT=/data/shared/mowl

Or pass data_root to a single dataset:

import tempfile
from mowl.datasets.builtin import FamilyDataset
ds = FamilyDataset(data_root=tempfile.mkdtemp())

Your own dataset

In case you have your own training, validation and testing ontologies, you can turn them easily to a mOWL dataset as follows:

from mowl.datasets.base import PathDataset
ds = PathDataset("training_ontology.owl",
                 validation_path="validation_ontology.owl",
                 testing_path="testing_ontology.owl")

training_axioms = ds.ontology.getAxioms()
validation_axiom = ds.validation.getAxioms()
testing_axioms = ds.testing.getAxioms()

Note

Validation and testing ontologies are optional when using PathDataset. By default they are set to None.

Attention

Custom datasets require the implementation of the evaluation_classes attribute. This can be done as:

class CustomDataset(PathDataset):
    def __init__(self, *args, **kwargs):
        super().__init__(train_path, valid_path, test_path)

   @property
   def evaluation_classes(self):
       #################
       # your code here
       #################