Data sources
Twenty-nine public databases, reached live and without keys, and two research datasets that ship with the product. Essarion Science pulls records — sequences, structures, variants, trial registrations, expression series — rather than summaries of them.
01The databases
These are the connectors wired into Essarion Science today, grouped as they appear in the settings sheet.
| Database | Category | What it returns |
|---|---|---|
| NCBI / PubMed | Literature & citations | PubMed biomedical literature through NCBI E-utilities. |
| Europe PMC | Literature & citations | Life-science literature, including bioRxiv and medRxiv preprints. |
| OpenAlex | Literature & citations | An open scholarly index across all disciplines. |
| PubTator3 | Literature & citations | PubMed with genes, chemicals, diseases and variants resolved to database identifiers, and the relations extracted between them. |
| bioRxiv / medRxiv | Literature & citations | Preprints. Each record is labelled as not peer reviewed and carries its published DOI when one exists. |
| UniProt | Proteomics | Protein sequence and annotation records from UniProtKB. |
| RCSB PDB | Proteomics | Experimental 3D structures and coordinate files. |
| AlphaFold DB | Proteomics | Predicted protein structures, by UniProt accession. |
| InterPro | Proteomics | Protein domains and families, with residue ranges. |
| Ensembl | Genomics & biology | Gene annotation: location, biotype and cross-references. |
| Ensembl VEP | Genomics & biology | Variant consequence per transcript: consequence class, protein change, SIFT and PolyPhen. |
| gnomAD | Genomics & biology | Population allele frequencies by ancestry. |
| GWAS Catalog | Genomics & biology | Published genetic associations, from trait to gene and from gene to trait. |
| GEO | Genomics & biology | Public gene-expression datasets and series, bulk and single-cell. |
| Human Protein Atlas | Genomics & biology | Tissue and single-cell expression, subcellular location, protein class and secretome status. |
| Reactome | Genomics & biology | Curated pathways for a gene, molecule or process. |
| BLAST (EBI) | Genomics & biology | Sequence homology search, with E-values. |
| STRING | Networks & pathways | Protein interaction networks with per-channel evidence, functional enrichment, and a test of whether a gene set forms a real module. |
| Enrichr | Networks & pathways | Gene-set enrichment against more than 200 libraries, including LINCS L1000 drug and CRISPR perturbation signatures. |
| ChEMBL | Clinical & pharma | Bioactive molecules, their structures and measured activities. |
| ClinicalTrials.gov | Clinical & pharma | Registered trials: phase, status, sponsor, intervention, and why terminated trials stopped. |
| ClinVar | Clinical & pharma | Clinical significance of germline and somatic variants. |
| Open Targets | Clinical & pharma | Target–disease association scores. |
| DGIdb | Clinical & pharma | Drug–gene interactions from several curated sources, and druggable-genome categories. Also lists a drug's known targets. |
| openFDA | Clinical & pharma | FDA drug labels, FAERS adverse-event reports and Drugs@FDA approval records. |
| Guide to PHARMACOLOGY | Clinical & pharma | Expert-curated ligands, reference tool compounds, affinities and receptor classification. |
| LINCS L1000 (L1000CDS2) | Clinical & pharma | Compounds ranked by how strongly they reverse a disease gene signature. |
| PubChem | Chemistry | Structures, identifiers and computed properties, with similarity and substructure search. |
| Patents (PubChem) | Chemistry | Patents that cite a compound. Prior art around a structure, not a Markush search. |
Higher-level tools combine several of these in one call. target.assess, for example, draws on Open Targets, the GWAS Catalog, the Human Protein Atlas, DGIdb and STRING, and variant.interpret on VEP, gnomAD, ClinVar and InterPro.
02Key-free
Every connector talks to a public API. There is nothing for you to configure and no key to obtain. Public services have their own rate limits and outages. A lookup that fails shows up as a failed step in the work log and in the turn's document, and the answer is expected to say what failed.
03Choosing sources
The Databases & skills sheet lets you switch categories or single databases on and off. Everything is on by default. Your choices are remembered in your browser and sent with every message, so a follow-up uses the settings you have now.
The switches narrow the agent's direct database lookups. Higher-level tools such as target.assess call the sources they are built on regardless, so treat the switches as a way to focus a run, not as a guarantee that a source will not be contacted.
- Narrow when the question is narrow. A structural question gains little from trial registrations.
- Leave literature on. Most answers rest on retrieved papers, and citations can only point at sources the turn actually retrieved.
04How literature is handled
Besides the literature connectors, a general search merges scholarly indexes — PubMed, Europe PMC, OpenAlex, Semantic Scholar, Crossref and arXiv — with a web search. The size of the pool it gathers rises with depth.
At Deep dive and Exhaustive, two more tools are available: one pulls tables and figure captions out of source PDFs, so effect sizes can come from tables and not only abstracts; the other builds a replication graph from the papers that cite each source.
Reading, not just retrieving
For a question that needs a systematic look at a literature, the agent has a review layer:
| Tool | What it does |
|---|---|
| review.screen | Decides include, exclude or unclear for every retrieved record against written criteria. Each decision names the one criterion that settled it and quotes the sentence that did. It will not run without criteria. |
| review.extract | Pulls the same declared fields — sample size, comparator, endpoint, effect size — out of every included record, with the quote each value came from. A field the paper does not report comes back as not reported, not filled from memory. |
| review.prisma | Draws the PRISMA 2020 flow from what the run actually did. Stages nobody ran are named as not performed, not left out. |
Every quote is checked against the record's own retrieved text. A cell whose quote is not found is shown as unverified, and the count of unverified cells is on the face of the result.
05Bundled datasets
Two research datasets ship with Essarion Science. You can browse them on the Data page, and the agent queries them locally, at no external cost, for prior art and base rates.
| Dataset | Rows | What it is |
|---|---|---|
| Negative Results Index | 62,842 trials | Registered trials on ClinicalTrials.gov that halted early or missed a primary endpoint, with the sponsor's stop reason classified into a failure taxonomy, and p-values, effect estimates and confidence intervals where results were posted. About two-thirds have no linked publication. |
| Drug Development Atlas | 101,943 drugs | One row per compound: every trial it appeared in, the furthest phase reached, years in development, safety and futility halts, enrolment, sponsors, conditions, and FDA approval status. 3,723 were approved. |
Both are derived from U.S. public-domain records (ClinicalTrials.gov and openFDA) and were last built on 2 August 2026. On the Data page you can search, filter, group and sort, and download either the whole dataset or the current view as CSV.
06Beyond the connectors
For a public resource with no connector — a specialist database, a new API, a lab's data release — the agent can fetch the URL directly. The fetch is recorded as a source, so it can be cited and checked like any other. Only public addresses are reachable.
For population-level questions, a primary-data tool reads open statistical archives such as the World Bank and Our World in Data. The set of archives it may use widens with depth.
07Provenance and credit
Every retrieved record is kept with the source it came from, and the Sources tab lists what a turn retrieved and which of those the answer used. Citations can only point at sources the turn actually retrieved; the model can choose among them but cannot add to them.
Each database is licensed by its provider on its own terms. Several, including UniProt, ChEMBL, AlphaFold DB, Reactome, STRING and the Human Protein Atlas, ask to be credited when their data is reused. If you publish results built on them, cite the sources the run lists.