Essarion Science
A research scientist for biology and medicine. It plans the investigation, runs the analysis and tests its own conclusions. You ask in plain words. It works through public databases and a live notebook, checks its own answer, and shows every step it took.
01What it is
Essarion Science is an agent that works in a chat. You send a message, and it decides one step at a time what to do: search the literature, fetch a structure, query a trial registry, run code, draw a chart. It reads each result before it picks the next step. It stops when it can answer, or when it reaches the limits set for the turn.
It chooses from about a hundred tools:
- 29 public databases, reached live and without keys: PubMed, Europe PMC, UniProt, the PDB, AlphaFold DB, ChEMBL, PubChem, ClinVar, gnomAD, Open Targets, ClinicalTrials.gov and more. See Data sources.
- Two bundled datasets: 62,842 clinical trials that stopped early or missed their endpoints, and 101,943 drugs with their trial history and FDA approval status.
- A persistent Python kernel that you share with the agent, and twelve computational skills, from sequence tools and CRISPR guide design to pharmacokinetics and trial sample size. See The workbench.
- Computational engines for quantities no database holds: where a molecule could bind, what an untested compound might do, whether a dose or a trial could work. See Engines & discovery.
- Structure tools for folding, viewing, labelling and editing proteins. See Fold & Studio.
02Where you work
After you sign in at science.essarion.com, the sidebar has five places.
| Place | What it is for |
|---|---|
| Workbench | The chat. Ask a question, watch the work, read the answer. A side panel holds the notebook, the evidence and the sources. |
| Fold | Fold an amino-acid sequence, see the structure coloured by confidence, and ask about it. |
| Studio | Protein projects: one structure each, with pinned labels, a conversation, saved figures and a stack of reversible edits. |
| Data | Search, filter and download the two bundled datasets. |
| Artifacts | Figures and models from your sessions, each linked back to the chat that made it. |
Your past sessions are listed in the sidebar below these links.
03What happens in a turn
- It reads the request. Before the first tool call it shows how it understood you: what you want, what the answer feeds, what it will hand over, and any assumptions it made, with what each would cost if wrong. It does not stop to ask. If the reading is wrong, correct it in your next message.
- It works. Each step appears in a work log as it happens. For longer work it writes a plan first, with what it expects each step to show.
- It checks. Before you see the answer, the system checks that the work it describes actually ran. At Deep dive and Exhaustive it also checks that the turn did not stop with an obvious step unplayed, and that the prose rests on what was retrieved. A separate reviewer pass then compares the answer with the evidence. See Verification.
- It answers. Sentences that rest on a retrieved source or a computed number carry a citation you can open to check.
- It hands something over. When the turn did real work, it assembles a document from it: a summary, dossier, decision memo, experiment protocol, candidate set, data package or research campaign. The document downloads as Markdown.
04Depth
Depth sets how much room a turn has and how hard it pushes before it accepts its own answer. You choose it in the settings sheet by the composer. Deep dive is the default.
| Depth | Steps per turn | Use it when |
|---|---|---|
| Scan | 14 | You want a quick answer. Some heavier analysis tools are left out. |
| Deep dive | 60 | You want a real answer you can defend. The default. |
| Exhaustive | 200 | The question is worth the wait. |
Each turn also has a time limit that rises with depth. When a turn reaches either limit, it stops starting new work and writes up what it has. The workbench covers what else depth changes.
05What to keep in mind
- It is research support, not medical advice. Check the sources and the code behind any result before you rely on it.
- Engines estimate; they do not measure. Each one states what it is worth. Docking, for example, shows where a molecule can sit, and its score is not fit for ranking compounds.
- It reads public data and what you upload. It has no access to private databases or your institution's systems.
- A negative answer can be the right answer. When it cannot establish something, it says so and lists what it tried.
06Where it fits
Essarion makes three products. ResearchAnything answers questions on any subject with sources attached. Sourcipedia is an encyclopedia it keeps current. Essarion Science is built for biology and medicine, and goes past reading into analysis: it fetches the records, runs the numbers and tests the result.
Use Science when the answer depends on data, structures or computation. Use ResearchAnything when you need a cited answer on a broader question.