chebai/ChEBI25-3STAR

A machine learning dataset for the ChEBI ontology (v248). This dataset includes 1,808 labels and 51,670 samples. Samples are molecules taken from the chebi.sdf.gz file. Only molecules in the 3-STAR subset of ChEBI are used. Labels are classes which have at least 25 3-STAR subclasses with molecule annotations.

The dataset has been generated with the ChEB-AI Graph library (v1.1.0).

Dataset structure

 data/chebi_v248/
  ├── raw/
  │   ├── chebi.obo              # ChEBI ontology in OBO format, taken from the [ChEBI FTP server](https://ftp.ebi.ac.uk/pub/databases/chebi/archive/rel248/ontology/)
  │   └── chebi.sdf.gz           # ChEBI molecules, taken from the [ChEBI FTP server](ftp.ebi.ac.uk/pub/databases/chebi/archive/rel248/SDF/)
  │
  ├── splits_chebi248.csv        # train/val/test assignment for every molecule (stratified 80/10/10 split)
  │                              #   columns: id, split
  │                              #   id = ChEBI molecule ID (integer)
  │
  └── ChEBI25_3_STAR/            # dataset variant
      └── processed/
          ├── classes.txt        # target ChEBI class IDs, one per line
          │                      # defines the label set used for classification
          │
          ├── data.pkl           # main molecule table as a pandas DataFrame
          │                      #   indexed by ChEBI ID; columns include
          │                      #   SMILES string and multi-label target vector
          │
          └── atomfgreader_withfgedges_withgraphnode/
              │                  # PyTorch Geometric dataset cache
              │                  # directory name encodes the reader configuration:
              │                  # atom+FG node features, FG-level edges, graph-level node
              │
              ├── data.pt        # serialized list of PyG Data objects (one per molecule)
              │
              └── properties/    # pre-computed feature tensors
                  ├── AtomType_one_hot.pt
                  ├── AtomAromaticity_bool.pt
                  ├── AtomCharge_one_hot.pt
                  ├── AtomHybridization_one_hot.pt
                  ├── AtomNumHs_one_hot.pt
                  ├── AtomNodeLevel_one_hot.pt
                  ├── AtomFunctionalGroup_one_hot.pt
                  ├── AugRDKit2DNormalized_asis.pt
                  ├── BondType_one_hot.pt
                  ├── BondAromaticity_bool.pt
                  ├── BondInRing_bool.pt
                  ├── BondLevel_one_hot.pt
                  ├── NumAtomBonds_one_hot.pt
                  ├── IsFGAlkyl_bool.pt
                  ├── IsHydrogenBondAcceptorFG_bool.pt
                  └── IsHydrogenBondDonorFG_bool.pt

Usage

This dataset can be used to train Graph Neural Networks with the ChEB-AI Graph library. For usage examples, see the README A model trained on this dataset is available here.