# Curasynth

> Curasynth is a therapeutics-intelligence platform built by Fitila Labs (Chicago). It harmonizes seven biomedical entities — diseases, targets, drugs, biomarkers, clinical trials, publications and pathways — onto stable identifiers, joins them with 128 bridge tables and 237 declared relationships, and works at two speeds. An AI Analyst answers a question in a minute with every claim cited to its source, license and vintage. A campaign runs for hours without supervision: it weighs a shortlist of targets against every source that speaks to them, reaches a verdict against a bar declared before the run, and reports the strength of each claim as a belief-plausibility interval, the classes of evidence nobody gathered, and which experiment would most narrow the remaining uncertainty. It runs hosted at https://app.curasynth.ai or on a single GPU workstation inside a customer's environment, and every capability is exposed as an API and over MCP.

Key facts (measured from the served platform, September 2026):

- 7,253,502 open-access papers in full text; 48,666,326 works in the citation graph (every reference resolved through OpenAlex and the NIH Open Citation Collection); 375,185,880 paper-to-paper citations; 208,113,080 citation-intent/stance markers; 10.8M authors and 580k organizations resolved through ORCID, OpenAlex and ROR.
- 601,387 clinical trials; 94,289 biomarkers; 80,177 targets (a 3-D structure for nearly all); 24,030 diseases on the MONDO spine; 11,758 drugs keyed by ingredient set; 4,228 pathways.
- 3,148,612 target–disease associations; 2,550,100 target–tissue expression rows; 461,919 drug–target binding measurements collapsed to 145,727 consensus pairs; 864,709 scored adverse-event signals from 4.2M FDA FAERS cases; 261,772 receptor-occupancy profile rows.
- 43 registered sources, each row stamped with its license/distribution class; ten integrity checks gate every release; a recovery set of 23 published findings, 19 recovered from public data.
- Campaigns: a declared, hashed template (question, evidence bar, spend ceiling, approval points) resolved against a pinned shortlist and run detached. Evidence is combined as Dempster-Shafer mass, giving each hypothesis a belief (lower bound) and plausibility (upper bound); the gap between them is stated ignorance rather than averaged away. Verdicts include inconclusive, high-conflict and frame-incomplete — the system can say "I do not know" and say why. Dependent sources combine so duplicates do not compound; evidence classes are discounted by a reliability ladder; the evidence ledger is append-only, hash-chained and refuses to load if it does not verify; results are order-invariant and replayable. Discovery and confirmation never share a cell line: a discovery campaign splits DepMap models by lineage under a declared seed and confirms only on lineages the screen never saw. Calibration is signed per evidence class by a named person on a date, and a verdict reached on unsigned calibration says so.
- Disease-aware target ranking: an eight-part composite (association evidence, tractability, chemical matter, dependency in the disease's own cell lines, tumour expression, clinical precedent for this disease, mechanistic route, collateral lethality), where cancer-only parts weigh zero outside a neoplasm.
- The Analyst reads the full text of papers by section with character offsets, and is checked by a 30-question suite that verifies a tool was called, a citation is present, a concentration names its species, and forbidden claims are absent — currently 30/30 on its accepted configuration, self-authored and not externally reviewed.
- Computed evidence: Boltz-2 co-folding on demand (predicted potency shown beside measured Ki, never pooled), receptor-occupancy and target-engagement profiles, pharmacovigilance with EBGM shrinkage, arm-level and pooled trial effects, CRISPR-dependency/copy-number/collateral-lethality analysis, consensus potency over censored measurements, cheminformatics.
- The Analyst runs on Fitila's agent runtime, which carries eleven planning strategies including a hierarchical task network for drug discovery (five phases, sixteen tasks, four pharmacological gates) and influence-diagram reasoning over the discovery funnel. Campaign stages let the planner choose what to gather while a deterministic executor does the reading, so no number a verdict rests on passes through a language model.

Curasynth is looking for strategic partners (teams inside established pharmaceutical or large biotech companies), design partners (research teams with a real program), technical partners (model and infrastructure builders), and early investors. Contact: curasynth@fitila.ai.

## Pages

- [Home](https://curasynth.ai/): what Curasynth is, what is in it, why harmonize, what one question can reach, the Analyst, computed evidence, how it is checked, where it runs, who we are looking for.
- [Data](https://curasynth.ai/data.html): the evidence graph table by table — entities, identity, every bridge family with its row count, the six evidence layers with their sources, license classes, the release checks.
- [The Analyst](https://curasynth.ai/analyst.html): how a turn works (resolve, query, open the record, answer, with the calls shown), the tools it holds, the rules it is held to, the five-step flow, the planning runtime, and how a campaign is specified — frame, arithmetic, verdict, coverage, record and calibration.
- [Models](https://curasynth.ai/models.html): Boltz-2 on demand, then each catalog-wide method and what it computes, and the methods sought from technical partners.

## Related

- [Application](https://app.curasynth.ai): the hosted product (sign-in required).
- [Fitila Labs](https://www.fitila.ai): the company.