Therapeutics intelligence
Most target decisions are made on the evidence somebody had time to read.
Curasynth reads all of it. Ask a question and get a cited answer in a minute; hand it a shortlist and it will weigh each target for hours without you watching, reach a verdict against a bar you set, and tell you what nobody has measured yet — with the strength of every claim on the record.
Campaigns — new
A question that takes hours, answered while you are somewhere else.
An answer in a minute is a good start. A real decision needs every target on a shortlist weighed against every source that speaks to it — dependency, expression, clinical precedent, the papers behind the trials — and that takes hours. A campaign does it as a detached run: you close the tab, it keeps going, and the page is the report as it fills.
What comes back is not a ranking. Each target gets a belief and a plausibility — the floor and the ceiling on what the evidence supports — so the gap between them is the ignorance, stated rather than averaged away. It is read against a bar you declared before the run: triage, validate, or commit, where commit also demands the interval be narrow, because a wide interval at a high belief is an unfinished argument rather than a decision.
And the verdict is allowed to be I do not know. A campaign can come back inconclusive, or say the sources contradict each other, or that the question itself is framed wrong — each a distinct answer with its own reason, none of them a number that hides the problem.
Three things a ranking cannot give you, and a campaign reports every time:
- Coverage — the classes of evidence nobody gathered for this target. Silence is reported as silence, not as a low score.
- What would change its mind — the next evidence to get, ranked by how much it would narrow the interval.
- Refusals — a hypothesis that states nothing refutable is turned away at the frame, not carried to the end and called inconclusive.
A stage that would spend money stops and asks first, naming the most it can cost. Nothing is charged without a human saying yes.
What is in it
Seven entities, harmonized, joined and served.
Counts are distinct records the Analyst can reach today, measured from the live indexes, joined by 128 bridge tables that carry source, license and vintage on every row.
| Entity | Records |
|---|---|
| Publications full text, PubMed Central open access | 7,253,502 |
| Clinical trials ClinicalTrials.gov, current to September 2026 | 601,387 |
| Biomarkers | 94,289 |
| Targets a 3-D structure for nearly all | 80,177 |
| Diseases on the MONDO spine, nine code systems resolved to it | 24,030 |
| Drugs keyed by ingredient set | 11,758 |
| Pathways | 4,228 |
| Joined by | Rows |
|---|---|
| Paper-to-paper citations | 375,185,880 |
| Citation intent and stance markers | 208,113,080 |
| Target–disease associations | 3,148,612 |
| Target–tissue expression | 2,550,100 |
| Trial–disease links | 819,136 |
| Adverse-event signals, scored | 864,709 |
| Drug–target binding measurements | 461,919 |
Why harmonize
One target, six names, one identity — and everything it joins to.
Public sources each key their records their own way. FLT3, the kinase mutated in about a third of acute myeloid leukemias, is one accession at Ensembl, another at UniProt, a third at ChEMBL, and a handful of names in the literature. The evidence about it is real, and scattered, and does not add up until someone decides they are the same thing.
Curasynth assigns one identifier per entity and holds it stable, then declares every relationship between entities explicitly. That is what makes what else is known about this target answerable in one pass rather than seven.
What the sources call it
- EnsemblENSG00000122025
- UniProtP36888
- HGNC3765
- ChEMBLCHEMBL1974
- TTDT74312
- NamesFLT3 · CD135 · FLK-2 · STK-1
What it joins to, today
- Disease associations, scored by evidence type794
- Trials of drugs that act on it1,547
- Measured binding partners294
- Synthetic-lethal observations143
- Tissues with expression99
- Interaction partners96
- Pathways66
- Drugs with a mechanism53
- Solved structure, with its paper6JQR
What one question can reach
Six kinds of evidence, in one pass.
Most tools are deep in one modality. The point here is how much a single question can touch — and that each answer says which of these it rests on.
LiteratureWhat has been written, and what it rests on
Every open-access paper in full text, and every paper they cite resolved to a work through OpenAlex and the NIH citation collection, held or not — a citation graph with the stance of each citing sentence, and the authors and organizations behind it as pages of their own.
7,253,502 papers in full text · 48,666,326 works in the citation graph · 375,185,880 citationsBinding and structureHow strongly, and where
Measured affinities collapsed to one consensus potency per drug–target pair, co-crystal evidence, and a 3-D structure for nearly every target, rendered in place.
461,919 measurements · 145,727 pairs · 17,353 co-crystal rowsGenetics and dependencyWhether a cell needs it
Association scores by evidence type, CRISPR dependency measured in organoids as well as on plastic, copy-number context, and published synthetic-lethal pairs.
3,148,612 target–disease scores · 250,480 synthetic-lethal rowsSafetyWhat it does to people, and what it hits
Twelve quarters of FDA adverse-event reports scored per drug and term, curated warnings, and each drug's receptor-occupancy profile screened against the safety-pharmacology panel.
4,210,000 cases · 864,709 scored signals · 43-target liability panelClinicalWhat has been tried
Every registered study linked to its diseases and drugs, with arm-level effects recovered from posted results and pooled across trials of the same comparison.
601,387 trials · 819,136 trial–disease linksExpression and pathwaysWhere it acts
Baseline expression across tissues, Gene Ontology annotation, and pathway membership with each pathway linked to the diseases it is implicated in.
2,550,100 expression rows · 1,067,629 pathway membershipsThe Analyst
Ask a question. Get back the evidence, where it came from, and what it rests on.
The Analyst reads across all seven entities at once and streams a cited answer while it works. It opens the papers themselves — not just their abstracts — and quotes the methods, the dose or the results table with the section and character offsets it read them at. Before it presents a finding as settled, it traces what the paper rests on and who has cited it since, with each citing sentence's stance: a claim that rests on one Ki reads differently from one that rests on ten.
Rank the targets behind a disease and re-weight live; assess druggability, validity, selectivity and combinations; pin what matters into a Program and export it as a cited dossier, or expose it over MCP so the AI tools your team already uses can call it.
It is held to rules the data taught us, and they are tested. One species per value, because a rat exposure once shipped as a human one. Weigh evidence only when asked. Read our own screen before repeating a published claim. A thirty-question suite runs against the live Analyst on every model change and checks something narrower than eloquence: that it looked before it answered, cited what it read, and did not make the claims we have forbidden.
The failure this catches is the dangerous one — a turn that answers fluently from memory, calls no tool, is often right, and errors nowhere.
The Analyst, its rules and its planners →Computed evidence
When the measurement does not exist, compute it — and say so.
Some of what a decision needs has never been measured. Curasynth runs Boltz-2 beside the data: co-fold a drug with a target and read the pose back — predicted potency, binder probability, interface confidence, contacts, whether the ligand sits in the pocket — or screen a drug's structural neighbors at its primary target. From a monograph, Predict binding gives a cost estimate first, runs on confirmation under a dollar ceiling you set, and lands the result as a new row.
The same discipline covers a family of models that run over the whole catalog: coupled multi-receptor occupancy on an exposure ladder for every drug, adverse-event disproportionality with shrinkage and confounder adjustment, arm-level and pooled trial effects, CRISPR-dependency and copy-number analysis, consensus potency over censored measurements, and the cheminformatics behind identity and similarity.
Every model, and what it computes →Checked, not assumed
Measured against what was published.
A recovery set holds published findings — binding affinities, deletion frequencies, dependency patterns, a collateral-lethality pair — beside what the platform measures for the same quantity from public data. Nineteen of twenty-three are recovered; the ones that are not are kept in the table and marked as such, not dropped. The table exists to be honest, not flattering.
The Analyst is held to the same standard by a thirty-question suite that runs against the live system on every model change. It checks something narrower than eloquence: that a tool was actually called, that a citation is present, that a concentration names the species it was measured in, and that the claims we have forbidden are absent — where a refusal that names a forbidden claim counts as a refusal, not as the claim. It currently passes all thirty on its accepted configuration. We wrote the questions ourselves; nobody outside has reviewed them, and the report says so.
Identity that survives a re-issue
When drug identifiers were re-minted this September, every bridge keyed on the old issue was re-identified from its own chemistry, keeping the old key beside the new one.
Old identifiers still resolveDeprecated terms follow
A disease term retired by MONDO redirects to its replacement, and a question about a disease reaches its subtypes through the hierarchy.
2,188 redirects · 348,198 descendantsEvery release, every link
Cross-references, data health, license terms and identity are verified before a vintage is promoted. A link that quietly stops working is a failure, not an empty page.
Ten integrity checks gate each publishProvenance on every row
Source, license, distribution terms and vintage travel with each record, so a result built from mixed sources reports the terms it actually inherits.
43 registered sources, each with its license termsThe calibration is signed, or it says it is not
What each class of evidence is worth is a named person's decision, recorded with a date and the reasoning. A verdict reached on a calibration nobody has signed says so on its face, rather than borrowing authority it was never given.
Signed per evidence class, dated, on the recordThe same rows give the same verdict
Every run records an append-only, hash-chained ledger of the evidence it weighed and a digest of what those rows claim. A re-run is provably weighing the same evidence, and the order it arrived in cannot change the answer.
Replayable · order-invariant · tamper-evidentWhere it runs
Hosted, or on one machine inside your environment.
Hosted
The application at app.curasynth.ai, with the Analyst, Programs, sharing and the compute service. Sign in with your work account.
Metered by useIn your environment
The whole platform — the indexes, the bridge tables, the Analyst and the structure service — runs on a single GPU workstation, over the public graph and your own corpus beside it. Nothing about your questions leaves the building.
Quantized indexes fit a 256 GB boxAs a tool
Every capability is an API, and the same handlers are served over MCP. A dossier can be consumed by an external agent as a callable tool rather than pasted into a prompt.
REST and MCP over one set of handlersWho we are looking for
The relationships a platform at this stage cannot buy.
The platform is live, the data is real, and the next capabilities are decided by who we build them with. We are looking for a small number of people in each of these roles.
A team inside an established company that has built this before
A therapeutic-area, computational or discovery-informatics group at a pharmaceutical or large biotech company. You have assembled an evidence graph in-house at least once, know what it cost, and know which parts never stay current. We would rather build the next one with you than beside you.
- What you bring
- A program to test against, the judgment to say what is missing, and whatever else a relationship warrants — a sponsored pilot, a method, a corpus behind your firewall, people who have done this at scale.
- What you get
- The platform inside your environment over your own corpus, the roadmap shaped around your workflow first, and terms that reflect being early.
A research team with a real program
A biotech, an academic lab or a translational group evaluating targets, repurposing candidates or combinations in a disease it knows deeply.
- What you bring
- The questions, the domain judgment, and honest feedback on what the Analyst gets wrong.
- What you get
- The platform over your program, the standing research programs built to your workflow first, and a say in what gets computed next.
A group that builds models or runs infrastructure
A structure-prediction, docking, ADMET or generative-chemistry group; a compute or data provider; a team that wants its method called from a cited workflow.
- What you bring
- A method with a contract, or the capacity to run one, and an interest in seeing it used against real programs.
- What you get
- A registry that resolves a method to a backend, an evidence graph to ground it, and a dossier that carries your output with its provenance to the people deciding.
People who want in early
Curasynth is a working platform built by a small team, with a wedge into the thousands of pre-Series-B biotechs that can least afford bespoke informatics and a path into the companies that already pay for it.
- What you bring
- Experience with research-tool and computational-biology companies, and a network in the teams above.
- What you get
- A product to evaluate rather than a deck, the measurements behind every claim on this page, and early terms.
Contact
Curasynth is built by Fitila Labs in Chicago.
If you are evaluating targets, repurposing candidates or trial precedent and you spend more time reconciling sources than reading them, we would like to hear what you are working on.
- Contact
- curasynth@fitila.ai
- Built by
- Fitila Labs
- Where
- Fulton Market District
Chicago, Illinois - Application
- app.curasynth.ai