chebai/ChEBI25_v252

A machine learning dataset for the ChEBI ontology (v252). Samples are molecules taken from the chebi.sdf.gz file, labels are ChEBI classes those molecules belong to (based on chebi.obo). Note: Only the ChEBI subhierarchy of molecular entity, excluding molecular entity itself, is used (unlike the v248 dataset). This means that classes such as group and chemical substance are not part of the dataset.

This dataset includes two parts:

1. ChEBI25 - classes with at least 25 molecule-annotated subclasses are used as labels. Results in 2,117 labels and 192,336 samples.
2. ChEBI25-3STAR - only samples from the 3-STAR subset (=the human-annotated subset) of ChEBI are used, labels are classes with at least 25 3-STAR molecules. Results in 1,766 labels and 52,901 samples.

The dataset has been generated with the ChEB-AI library (v1.4.0).

The included splits are 80/10/10 (train/validation/test).