CHEBI Chemical Classification Program Ontology (C3PO) Benchmark

Name: C3PO-v237

A benchmark for chemical classification, derived from v237 of the CHEBI ontology.

The benchmark consists of

• 177,875 Structures, each represented by a SMILES string
• 1364 Classes, each of which classify (directly or indirectly) between 25 and 5000 Structures

The CHEBI smiles property (obo:chebi/smiles) is used to derive structures. Only SMILES strings that lacked a * wildcard are used. If multiple CHEBI entries have the same SMILES, they are collapsed.

The benchmark can be further partitioned based on different properties.

C3PO-Slim is the subset of structures that have at least one xref (mapping), representing the set of classes that are more commonly used in the biological literature and in other databases.

20% of the structures are designated as being for validation purposes. The remaining 80% can be used for training and testing.

The benchmark can be downloaded as either a single JSON file, or separate CSVs for classes and structures. Summary stats files are also provided:

• dataset.json - the complete benchmark
• classes.csv - all classes
• classes_slim.csv - slim classes (any with an xref)
• structures.csv - all structures
• classes_summary.csv - summary statistics for classes
• classes_slim_summary.csv - summary statistics for classes
• structures_summary.csv - summary statistics for structures
• metadata.csv - metadata about how this was derived.

For a description of all fields, see:

• datamodel

For the code that extracts the benchmark from CHEBI, see:

• extractor