open-reaction-database/ord-artifacts

Open Reaction Database derived artifacts

Each dataset in the Open Reaction Database restated in shapes a query engine reads cheaply. The artifacts are derived by ord_schema.artifacts and searched by ord_schema.search. They add nothing to the source datasets, which stay the authoritative record.

Artifacts

artifact	one file per	holds
projections	source dataset	every field of every reaction as nested Parquet, with units folded into column names (setpoint_kelvin) and each molecule's structural identifiers resolved to one smiles
structures	source dataset	one row per distinct molecule: SMILES, registration hashes, pattern and Morgan fingerprints, and a serialized RDKit mol, keyed by structure_id
pivots	source dataset and repeated level	one row per element of that level, with its reaction_id, the ordinal of every enclosing level, and the element's own non-repeated fields
occurrences	source dataset and indexed path	one row per molecule occurrence at that path: reaction_id, structure_id

A structure_id is local to its dataset: it joins a projection, pivot, or occurrences file only to the structures file derived from the same source dataset.

Layout

projections/<shard>/ord_dataset-<id>.parquet
structures/<shard>/ord_dataset-<id>.parquet
pivots/<level>/<shard>/ord_dataset-<id>.parquet
occurrences/<path>/<shard>/ord_dataset-<id>.parquet

• The shard and file name are the source dataset's own, as in ord-data's data/<shard>/ord_dataset-<id>.parquet.
• The repository holds one build at a time. Every file's Parquet footer names the artifact (ord.artifact), the artifact schema version (ord.artifact_schema_version), the RDKit release it was derived under (ord.rdkit_version), and the MD5 of the source dataset it restates (ord.source_md5).
• A new artifact schema version or RDKit release replaces the build, and earlier builds stay in the repository's history. A reader needs the RDKit the build was derived under exactly, since fingerprints and registration hashes change between releases.

The dataset viewer shows the reaction ID, reaction SMILES, and identifiers of the projections, and the structures in full.

Usage

fetch_artifacts checks that the installed ord-schema and RDKit read the build before downloading it, and a repeat call fetches only what changed:

pip install 'ord-schema[search,huggingface]'

from ord_schema import huggingface
from ord_schema.search import execute

artifacts = huggingface.fetch_artifacts("ord-artifacts")
corpus = execute.Corpus(
    artifacts.projection_pattern,
    artifacts.structures_pattern,
    pivots_dir=str(artifacts.pivots),
    occurrences_dir=str(artifacts.occurrences),
)

Pass revision= a commit to pin one, or to fetch an earlier build. Each upload's commit message names the ord-data revision it was derived from, and its description names the artifact schema version and RDKit. See the search README for queries, and the artifacts README for how each artifact is derived and stamped.

Citation

Please cite the Open Reaction Database as given on the ord-data card.