The questions everyone asks first.

No. They are Monte Carlo drawsfrom a Gaussian copula joint fitted to the indication's real anchor cohort. For DMG, that anchor is 60 H3 K27M-mutant patients pooled from Gröbner (Nature 2018) and Petralia (Cell 2020), covering both the H3.3 and H3.1 variants. The simulated profiles are statistically representative draws, not records of real people. Read the methods.

n=60 for DMG. These are the numbers above which every confidence interval in the system is computed. We bootstrap over the real anchor n, not the simulated draw count. A 1,000-draw simulated cohort drawn from n=60 still reports intervals reflecting n=60. The draw count is a Monte Carlo resolution parameter, not an information parameter. We make this visible in every report header.

No. The methodology preprint is in draft and has not been posted yet; the methodology page is the current write-up. The preprint will be released under CC-BY 4.0 so anyone can read and check the methods. The implementation itself stays proprietary and the repository is private.

Three things make the reports decision-grade rather than vibes: (1) citations at source, every claim carries an inline pointer to a PMID, a paediatric HGG cell line from a published supplement, a named escape mechanism with its citation, or a cohort statistic; (2) the Honesty Critic agent runs adversarially on every sentence and refuses ones without retrieved evidence; (3) every model call and every tool call writes an agent_logsrow you can audit afterwards. You don't have to take the report on faith; you can pull the audit log and check it.

The system surfaces that case directly. The co-occurrence audit runs on every cohort, and the methods appendix in every report states how many gene pairs were tested, the Bonferroni threshold, and how many reached it. The 45 degree agreement plot, which shows which gene pairs the copula preserved and which it didn't, is published as Figure 1 on the methodology page over the anchor cohorts; it is not generated per run. The leave-one-out cross-validation harness reports per-gene Brier scores, including the cases where the joint copula loses to a marginal baseline at small n. We'd rather ship a finding that disagrees with a published headline (e.g. STAG2+TP53 in Ewing did not reach Bonferroni in our n=222 extract) than bury it.

Onkydra is a hypothesis generator and prioritisation engine. It is not: a diagnostic, a treatment recommender, a medical device, a substitute for wet-lab validation, a substitute for clinical judgement. Every generated report carries this line in the header and the footer. You can use Onkydra to decide which target to spend your next quarter's lab budget on; you cannot use it to decide what to put in a patient.

No. We are explicit about this in the product, the documentation, the Terms of Service. Clinical decisions go through clinicians, registered diagnostics, IRB-approved trials. The system is for pre-clinical and pre-trial prioritisation, not bedside care.

One, and seven more are scoped. The one you can run is H3 K27M diffuse midline glioma, within a stated scope. The full roadmap is eight rare and paediatric indications: H3 K27M diffuse midline glioma (DMG; the flagship), Ewing sarcoma, MPNST, ATRT, high-risk MYCN-amplified neuroblastoma, anaplastic thyroid cancer, rhabdomyosarcoma, KMT2A-rearranged paediatric AML. DMG is the only indication you can run today, and its library runs within a stated scope. Ewing is built and held behind the coverage gate until DMG is proven out. The other six ship as their molecular profile libraries are curated. Roadmap.

Within a much narrower set than you might expect, and the coverage gate refuses the rest rather than answering them. What is in scope on the self-serve subscription tiers is the 45 transcription factors carrying a precomputed CellOracle knockout in the DMG network, ACVR1 through an explicitly labelled bridge to its downstream TFs, and a short list of target classes (CDK7, CDK4/6, MEK1, PI3K, ONC201, MENIN) recognised by labelled heuristic proxies. Ask for anything else and Onkydra abstains and says why, with no charge for the refused run. Targets outside that set are handled as a programme, via a scoping call.

Yes, as a programme. A programme is one named indication or target, agreed in a project plan, time-boxed, and expandable only by amendment: we curate, licence and validate the molecular profile library for that specific cancer and ship it inside the workspace for your team. It is quoted rather than bought from a page, and the floor is on the pricing page. Email hello@onkydra.com to scope.

The line-by-line comparison is on the pricing page, and this is the shape of the ladder:

• Free · €0 free forever: Everything you need to find out where Onkydra stops, before paying anything. It grants no compute: you can read what the workspace produced for somebody else, and you can ask whether it would answer for you.
• Academic lab · €300 per lab: The whole workspace for a whole lab, priced per lab rather than per person. A research group of three to fifteen churns people by design, and a seat price charges it for doing that.
• Commercial seat · €800 per seat: The same engine and the same evidence as the academic lab, licensed for work a company owns. The step up is a right rather than a feature: this is the tier that lets the output inform a commercial programme, and it brings the paperwork procurement asks for.
• Commercial team · €1,600 per team: Three seats, a higher ceiling on what the team can run and watch, and a review cadence. Extra seats are flat, because volume discounts barely exist on this shelf.
• Programme · from €40,000 per year: One named indication or target, curated end to end, time-boxed, and expandable only by amendment. This is the tier where somebody stands behind the answer rather than handing you a tool and wishing you luck.

A lab is charged per lab and not per seat, because a research group loses and gains people by design and a seat price charges it for doing so. A programme is one named indication or target, agreed in a project plan rather than self-declared, time-boxed, and expandable only by amendment. Two things the table does not hide: shared programme memory across a team, drift alerts and weekly auto-rerun are in build and not yet live, and the workspace is single-user at present, so there is no seat model behind it yet.

It is configured, and it cannot be started yet, because checkout is not open. When it opens: 7 days on the Academic lab monthly subscription, full feature access. Stripe Checkout will collect a card up front, and nothing is charged until the 7 days are up: cancel any time inside the trial and you won't be billed. One trial per customer, so if you have used it before, a new subscription starts billing immediately. No annual plan, no commercial tier and no programme offers a free trial. A plan charged a year up front is a materially different offer, and granting a trial on one by accident is a mis-sale rather than a generosity; instead, drop us a line for a guided walkthrough.

You manage cancellations from the Stripe-hosted Customer Portal. One click. No retention dark patterns. Your data stays available for export via /workspace/api/me/export and can be deleted entirely via /workspace/api/me/delete per GDPR.

Annual is 17% off monthly on every tier with a monthly rate, on the toggle on the pricing page. Monthly is the default and the headline, because a monthly transaction is a small one and an annual one is a purchase order. Checkout is not open yet, so nothing can be bought on either cadence today. Foundation sponsorship is described at /foundations, and its three tiers currently carry NO PRICE. Their prices were withdrawn and not yet replaced. Sponsorship is scoped per programme until it is rebuilt on whatever the academic tier settles at. The foundation dashboard, sponsor attribution on every report, and self-serve seat claiming are in build and not yet live. The foundations it is designed for are ChadTough, DIPG Advocacy, and Cure JM-aligned programmes.

No. Onkydra v1 deliberately does not accept patient-identifiable data. The system operates entirely on derived statistical distributions from public datasets (cBioPortal, as the interface to the DKFZ pediatric pan-cancer and CPTAC pediatric brain studies, plus the CCMA CRISPR and drug-screen deposits, GEO and PubMed). Future versions may support bring-your-own clinical data inside a covered-entity BAA with Google Cloud; not in v1.

GET /workspace/api/me/export returns every row keyed to your account (runs, traces, cohort manifests, billing audit trail) as one JSON bundle with a lawful- retention notice covering Stripe-side invoice retention under GDPR Article 17(3)(b). POST /workspace/api/me/delete is a two-phase HMAC-confirmed flow that cancels any active Stripe subscription, deletes every uid-keyed Postgres row, and deletes the Firebase Auth user.

You do. Onkydra ships the report as your output; we don't claim derivative rights on the analysis your subscription paid for. The upstream sources keep their own terms, and those terms are not uniform: cBioPortal's data hub publishes under the Open Database Licence, which carries a share-alike condition; the GEO records behind the cell-state reference state no licence at all; and PubMed metadata is public domain. We honour the attributions in the citation layer, and data status names each source, its licence, and the questions still open on it.

Google Cloud Platform. The database is Postgres on Cloud SQL in europe-west2(London), with multi-tenant row-level security, no client-side database access, per-secret IAM on the runtime service account, and encryption at rest and in transit. Its automated backups are in Google’s EU multi-region, the application tier is in europe-west4(Netherlands), and Gemini inference runs on Vertex AI’s global endpoint, which Google may serve from any supported region. London is in the UK rather than the EEA, so this is not an EEA-only stack: the sub-processor register names every region and every third party we send anything to.

withZeta.ai is the closest vertical match, same buyer profile, same "rare-cancer AI co-scientist" story. But the shape of the offering is different in three ways. First, indication depth: withZeta covers 438 cancer types via ontology (Orphanet, NCI Thesaurus, HPO); Onkydra ships per-indication anchor cohorts curated end-to-end (n=60 real DMG cases from DKFZ + CPTAC, live today; an n=222 Ewing anchor is built and held behind the coverage gate, and six more indications are queued in the roadmap). Their ontology tells you the disease exists; our cohort lets you rank strata against real co-mutation structure. Second, model layer: withZeta's purpose-built modules are PredictBBB (blood-brain-barrier permeability) and ETHER0 (chemistry reasoning, Apache 2.0 from FutureHouse), chemistry-forward. Ours is transcriptomics-forward: a stated rule set for the per-stratum ordering, reported alongside a mechanistic perturbation layer rather than settled by it. That layer is a real CellOracle in-silico knockout over a DMG regulatory network, precomputed for 45 transcription factors and abstaining beyond them. Third, distribution: withZeta ships from Lantern Pharma, which also runs its own oncology pipeline; Onkydra is an independent tool vendor, not a drug developer.

No, and the honest answer takes four lines. First, vertical vs horizontal: Claude Science is a general-purpose workbench for any scientific discipline; Onkydra ships a pre-loaded molecular profile library anchored to a real pooled patient cohort (DMG today, n=60 from DKFZ and CPTAC, running within a stated scope). Second, a built-in adversarial check: the Honesty Critic labels every layer, and every row of the predictions export, as a real model run, a curated registry or an assumption we stated. Where two layers disagree it reports the disagreement and does not resolve it, because an assumption cannot settle a model run and neither is authoritative. Third, the perturbation model is different: the NVIDIA BioNeMo bundle inside Claude Science includes Evo 2, Boltz-2, and OpenFold3, none of which predict a transcriptomic perturbation, which is what our mechanistic layer does. Fourth, evidence-size honesty: every Onkydra report bootstraps its CIs over the real anchor n (n=60 for DMG), not over the simulated draws. Claude Science's reproducibility promise is code-level (every artefact is reproducible from the code snapshot); ours is statistical (every CI is honest about the real n). Different depth of field, same interaction quality bar.

Because a Claude Science user searching for “H3 K27M diffuse midline glioma” would spend the first 90 minutes of their session pulling and normalising the DKFZ and CPTAC pediatric K27M cohort by hand, another hour deciding whether Ledoit-Wolf shrinkage on Σ is the right regularisation for n=60 and 13 genes, and would then have to convince a peer reviewer their ad-hoc setup was sound. Onkydra ships all of that pre-curated, with the methodology documented. The academic lab tier, at €300 per lab for the whole group, is what those first three hours cost.

Not at the buyer level. Claude Science's launch customers are BMS (30,000 employees), Genentech, Novartis, Regeneron, AbbVie, Sanofi. Onkydra is built and priced for a single rare-cancer research group at €300 per lab, a Seed-to-Series-A R&D team at €1,600 per team, and founders commissioning a named programme, and it has not sold to any of them yet: the workspace is in private beta and open to one address. The buyer, price shelf, sales motion, and procurement gate are all different. Anthropic will also not compete with us on a rare-cancer vertical, because a horizontal Anthropic-scale platform will always add value on the axis with the largest customer bases, not on an indication accounting for a few hundred new patients per year worldwide. One indication is wired here today, H3 K27-altered diffuse midline glioma, and its library runs within a stated scope: an n=60 anchor, 13 driver features, and no outcome layer to calibrate against.

The MCP server is built. It exposes eight read-only tools. Three answer questions about coverage and maturity: onkydra_list_indications, onkydra_describe_indication and onkydra_check_target_across_indications. Five more read the perturbation lab: onkydra_states, onkydra_models, onkydra_perturbations, onkydra_experiments and onkydra_outcomes. They answer questions about coverage and maturity and hand back a link; none of them starts a run, because a run takes minutes and costs real model time, so that stays inside the workspace where it can be metered and shown honestly. Distribution over ownership: a Claude user asking about a rare-cancer target gets our answer about what is and is not grounded, inside their session.

Complementary. Onkydra is built on the Gemini 3 family, so improvements in Gemini's scientific reasoning flow through to us directly. Gemini for Science bundles AlphaFold, AlphaGenome, and 30+ databases; Onkydra bundles a copula cohort engine, a linear baseline fitted on measured expression and held out by sample, PubMed retrieval, and a mechanistic perturbation layer (CellOracle in-silico knockouts are real over a DMG regulatory network, precomputed for 45 transcription-factor regulators plus ACVR1 through an explicit ID1/ID3 bridge, and abstaining for every other target; the dependency rows serve measured CCMA CRISPR betas where the screen covers the gene and an explicit null where it does not, and the escape layer is a literature-derived shortlist with citations and no score), and rare-cancer indication libraries (one selectable today, DMG, running within a stated scope, with the rest on the roadmap). Different scope; we are a legitimate downstream use of Google's platform.

Founded by Faith Ogundimu, a cancer-genomics PhD researcher at the Royal College of Surgeons in Ireland (Dublin). Onkydra is currently built and run by Faith alone. DMG is the only indication you can run today, within a stated scope; the remaining seven rare-cancer indications are roadmap, not shipped.

Gemini 3.1 Pro for all three stages that call a model: the Planner, the Honesty Critic and the Workspace Writer. The Biology Engine calls no model at all, which is why its audit row records none. CellOraclefor mechanistic perturbation, which is real: 45 transcription-factor knockouts precomputed against a DMG base gene regulatory network built from four real scATAC-seq samples unioned with CellOracle's published human promoter network, served as an artifact, and abstaining outright for anything that is not a transcription-factor regulator in that network. The cells it simulates on are a mixed cohort rather than a pure DMG one: about half come from a record that is both high-grade glioma and a stated midline site, the rest are hemispheric, site-unstated, or ependymoma comparators from the original study, and H3 K27M status is recorded on no record in that deposit. The full breakdown is on the state atlas and in the capability registry. Alongside it, two independent stated rule sets score the same strata: one from the response model and one whose terms come from published target-class pathway mechanism. They are reported side by side so their disagreement is visible rather than averaged away. Nothing from DepMapis ingested and no DepMap lookup happens. Eleven of the seventeen dependency rows serve a beta score measured in that row's own cell line by the CCMA pooled CRISPR screen, released under CC BY 4.0; the other six serve an explicit null and say which of two reasons it is. The curated values those rows used to carry were nulled on 2026-08-12, when reading all three cited papers in full found none of the seventeen traceable to the paper it cited. No LINCS L1000 or CMap signature query has ever been run, so that layer is a literature-derived escape shortlist with citations and no score. PubMed retrieval via Gemini text-embedding-005 embeddings in pgvector is live. MedGemmaruns on Cloud Run with an L4, scale to zero, and describes the morphology of an H&E tile: it is asked for a description and never for a diagnosis, a grade or a subtype, and the prompt is fixed server-side so it cannot be asked for one. Non-coding variant interpretation has no supplier here at all: AlphaGenome is non-commercial on both routes, the API and the weights, so a commercial product cannot use it. Nothing else we would trust fills that gap, so the gap stays open and named.

Not currently. The implementation, including the cohort sampler and the agent orchestration layer, is proprietary and the repository is private. What we intend to make public is the methodology: a preprint under CC-BY 4.0 describing the methods in enough detail to be checked and reproduced. That preprint is still in draft and has not been posted. This keeps the science checkable while letting the business sustain itself.

The anchor cohorts are built from public datasets and the methodology is documented, so the analysis is not a black box that disappears with us. Your export bundle (/workspace/api/me/export) gives you everything we hold for your account in one JSON file. Worst case, you lose the workspace UX; you don't lose the science or your history.