Highly structured scientific data with strong ontologies, covering chemistry, materials, and the life sciences. Well suited to domain-specific RAG in research and industry.
10sources
Chemistry and materials data is some of the most structured you will ever feed a RAG system, the setup that answers questions by retrieving relevant records and then generating a reply from them. Instead of loose prose, you get molecules identified by standard codes, reactions with defined inputs and outputs, and material properties tied to measured values. Much of it is held together by ontologies, shared vocabularies that agree on what each term means and how concepts relate, so a “boiling point” or a “binding affinity” means the same thing across records. That consistency makes retrieval far more reliable than free text.
Choosing within this category depends on what your questions are about. Compound-centred databases like PubChem and ChEBI are the place to start for identifying and describing molecules. ChEMBL leans towards bioactivity and drug discovery, the Materials Project towards computed physical properties, and the Open Reaction Database towards how reactions actually run. Match the source to the question rather than reaching for the largest table.
The catch with structured scientific data is that it is rarely plain text, so you cannot just chunk it into passages. You will usually need to convert records into readable descriptions before embedding them, the step of turning text into numerical vectors a system can search by meaning. Watch the identifiers too: a molecule can be written several ways, and mixing formats quietly breaks retrieval. Licences here are generally permissive, though a few sources restrict commercial use, so check before you build a product on one. The sources below span compounds, reactions, and material properties, each strong in its own corner.
An open-source project that pulls together a diverse set of life-sciences datasets from many providers into a single linked-data graph, with a SPARQL endpoint for querying across them. The full collection is about 11 billion triples across 35 datasets, including DrugBank, PubMed, and MeSH.
A dictionary of small chemical compound molecular entities, with structure files and ontology files available for download. Particularly useful as a controlled vocabulary for chemical entity linking, where you match a mention in text to a standard identifier.
A manually curated database of bioactive molecules with drug-like properties, bringing together chemical, bioactivity, and genomic data to support drug discovery. Holds close to 2.5M compound records on nearly 2M unique chemical structures, extracted mainly from the medicinal chemistry literature.
A collection of over 350,000 crystal structure files covering organic, inorganic, and metal-organic compounds, released into the public domain under CC0.
A harmonised, standardised merge of three major computational materials databases (Materials Project, Alexandria and OQMD) into a single format of roughly 6.7 million entries with consistent property definitions.
A database of computed properties for known and predicted materials, produced through high-throughput density functional theory calculations (a physics method for modelling how electrons behave in a material). Free API access with registration.
Meta FAIR's open dataset of more than 110 million DFT calculations of inorganic materials, sampled for structural and compositional diversity. Models trained on it top the MatBench-Discovery leaderboard for stability and formation-energy prediction.
A large-scale dataset of more than 100 million high-accuracy DFT calculations spanning 83 elements, built by Meta FAIR to train machine-learning interatomic potentials for molecular systems.
An open-access chemical reaction database built to support machine learning for reaction prediction, synthesis planning, and experiment design. A collaborative effort spanning pharmaceutical companies, academia, and technology firms.
An open chemistry database run by the US National Institutes of Health. Holds over 300M substances, 100M compounds, and almost 300M recorded bioactivities, covering chemical structures, identifiers, physical properties, patents, and health, safety, and toxicity data. Available in SDF, CSV, and other formats.