Drop in a target. Onkydra ranks candidate strata against a co-mutation cohort anchored to 60 real H3 K27M cases (DKFZ pediatric pan-cancer n=53, CPTAC pediatric brain n=7), flags where the ridge baseline and the mechanistic layer disagree, and is built to re-run when new evidence lands, once live monitoring is wired.
Overview
A single glanceable page. Per-stratum ordering signal, where the baseline and the mechanistic layer disagree, the top resistance candidate, and the proxy status of every layer. Every number cites its source.
Private beta · Built on Gemini 3 and a Gaussian-copula cohort engine anchored to 60 real K27M cases
For forty years, diffuse midline glioma had no targeted therapy. In August 2025 it did. Three structural shifts converged within twelve months, and together they make a workspace like Onkydra possible for the first time.
Dordaviprone (Modeyso) won FDA accelerated approval in August 2025 on a pooled n=50 across two trials. The first hit in a cancer that had nothing for forty years. A new generation of rare-cancer biotech founders is being funded around indications like this, and they need in-silico target validation that matches the size of the evidence.
scPRINT-2 trained on 350M cells under a permissive licence. AlphaGenome opened for non-coding variant calls. MedGemma landed on Vertex AI. The compute, the models and the data are now genuinely available to small teams. The bottleneck moved from infrastructure to methodology.
Anthropic launched Claude Science on 2026-06-30. Multi-agent deliberation loops with adversarial critics, audit-logged tool calls, reproducible artefacts, structured citations at source. The interaction pattern is now the industry floor. Onkydra takes that quality bar and specialises it to the vertical where curated cohorts and honest CIs decide whether the report gets peer-reviewed.
Rare-cancer cohorts (DKFZ + CPTAC)
Paediatric cohort expansion · planned
Cell-line dependency · proxy
Perturbation signatures · proxy
Variant significance · planned
Target-disease atlas
Literature retrieval
Adult tumour baseline
Cancer ontology + phenotype
Regulatory anchoring
Rare-cancer cohorts (DKFZ + CPTAC)
Paediatric cohort expansion · planned
Cell-line dependency · proxy
Perturbation signatures · proxy
Variant significance · planned
Target-disease atlas
Literature retrieval
Adult tumour baseline
Cancer ontology + phenotype
Regulatory anchoring
Rare-cancer cohorts (DKFZ + CPTAC)
Paediatric cohort expansion · planned
Cell-line dependency · proxy
Perturbation signatures · proxy
Variant significance · planned
Target-disease atlas
Literature retrieval
Adult tumour baseline
Cancer ontology + phenotype
Regulatory anchoring
Planner + Honesty Critic
Genomics workhorse
Non-coding variants (AlphaGenome) · planned
Imaging (MedGemma) · planned
Perturbation model · proxy
Ensemble second model (Geneformer) · planned
Vertex AI runtime
Planner + Honesty Critic
Genomics workhorse
Non-coding variants (AlphaGenome) · planned
Imaging (MedGemma) · planned
Perturbation model · proxy
Ensemble second model (Geneformer) · planned
Vertex AI runtime
Planner + Honesty Critic
Genomics workhorse
Non-coding variants (AlphaGenome) · planned
Imaging (MedGemma) · planned
Perturbation model · proxy
Ensemble second model (Geneformer) · planned
Vertex AI runtime
THE PROBLEM
Tumour profiling at scale exists. It is just priced, built and validated for billion-dollar pharma, not for the small teams taking on the cancers pharma won't.
BENCHMARK · ILLUSTRATIVE
The gap is the problem. Onkydra draws simulated profiles over the copula joint fitted to the real anchor. Bootstrap CIs stay on the real n.
You give the workspace a target. The Planner picks the right simulation, the Biology Engine ranks the strata, the Honesty Critic checks the result for itself, and the Workspace Writer hands you the report with citations. The Delta Watcher is the sixth surface: it will watch ClinVar, DepMap and the literature and re-run when something moves the needle. That live monitoring is being wired, not running yet.
Reads your question, picks the right cohort library, DepMap subset, and LINCS slice. Deterministic: no specialised syntax, no nine separate database interfaces.
Monte Carlo draws from a Gaussian copula joint fitted to your indication's real anchor cohort, with Ledoit-Wolf shrinkage on Σ. n=60 in DMG, pooled from DKFZ (n=53) and CPTAC (n=7), and including the H3.1 cases an earlier derivation missed. Tested against held-out cases never seen during fitting: single-gene frequencies reproduce at r=0.92 on 72 held-out DKFZ cases and r=0.59 against an independent institution. Every downstream statistic propagates that real n through bootstrap CIs you can see.
Ridge regression plus cohort statistics plus PubMed retrieval, returning a structured evidence bundle, never prose. CellOracle is real here: in-silico knockouts precomputed over a DMG regulatory network built from four real scATAC-seq samples and a 40,000-cell reference, served for the 47 transcription-factor regulators that produced signal, plus ACVR1 through an explicit ID1/ID3 bridge. For every other target it abstains rather than substituting a heuristic, the readout is an uncalibrated ordering signal rather than a probability, and it is not yet expert-reviewed. The scPRINT-2 layer runs as a labelled heuristic proxy with its serving endpoint written but not yet deployed, and the DepMap and LINCS layers run on curated placeholder rows pending live ingestion. Every report declares which is which.
Flags where ridge and the foundation model disagree. Hard-fails on claims without retrieved source. Cross-checks the cohort's co-occurrence structure with a DISCOVER-inspired audit. The disagreement is the product.
Consumes the evidence bundle. Cannot emit a sentence without a citation id. Renders the memo into your live workspace, and hands you the whole run (memo, agent trace, cohort manifest, citation ids) as a JSON download. A typeset PDF of the memo is in build, not yet live.
Live monitoring is being wired and is not live yet. It will watch ClinVar, DepMap and PubMed for changes on your targets, re-run the pipeline when something moves the needle, and flag shifts in the ranking. The drift calculation is written and tested; the scheduler that triggers it is not running.
Onkydra launches with eight rare and paediatric cancers, starting with diffuse midline glioma, where real cohorts are smallest and synthetic ones matter most.
Two ways to buy. A researcher gets a cited ranking of candidate strata by an uncalibrated ordering signal + ranked resistance candidates + a declared proxy status per evidence layer in 180 seconds per run. A team gets drift alerts and re-runs when ClinVar or DepMap moves (in build, not yet live). An engagement gets a memo reviewed by a domain expert before delivery. What you pay for is the artefact, not the session time.
For PhD students, postdocs, junior PIs. One workspace, up to three active targets in the launch indication. Memo export, fourteen-day free trial, and weekly auto-rerun (in build, not yet live).
For a lab or a Seed-to-Series-A rare-cancer R&D team. Everything in Researcher, plus ten active targets. Multi-seat access for up to five people and shared programme memory are in build, not yet live: the workspace is single-user today, with no team or seat model behind it. Drift alerts when the evidence moves and scheduled reporting are also in build.
Every Researcher seat ships with the DMG anchor cohort. Diffuse midline glioma is the only indication live today, and its library is draft, pending DMG-expert review. The rest are on the roadmap in the order below. Curating an indication means licensing and validating a whole molecular-profile library, so it is scoped as a custom indication engagement rather than a monthly add-on. Tell us which one you need and it moves up the queue.
Diffuse midline glioma
IncludedH3 K27M anchor, n=60, DKFZ + CPTAC
Included
Ewing sarcoma
In the queueEWSR1 fusion, adolescent + young-adult anchor
Roadmap
MPNST
In the queueNF1-driven, malignant peripheral nerve sheath
Roadmap
ATRT
In the queueSMARCB1 loss, atypical teratoid rhabdoid
Roadmap
High-risk neuroblastoma
In the queueMYCN-amplified, paediatric
Roadmap
Anaplastic thyroid
In the queueBRAF V600E, TP53 co-mutation
Roadmap
Rhabdomyosarcoma
In the queuePAX-FOXO1 fusion, paediatric
Roadmap
Pediatric AML (KMT2Ar)
In the queueKMT2A-rearranged, infant + paediatric
Roadmap
A pack leaves the queue once its anchor cohort has been curated and validated end-to-end through the run pipeline. Diffuse midline glioma is the only pack live today, and its library is draft, pending DMG-expert review. Every other indication listed here is roadmap, not shipped, and ships in order.
Candidate strata ranked by an uncalibrated ordering signal, each with a descriptive 10th to 90th percentile band, the top-3 ranked resistance candidates, a declared proxy status for every evidence layer, and a citation-backed narrative. Every claim linked to a real source, every cohort frequency resampled over the real anchor n.
A live lab that will watch ClinVar, DepMap, and the literature between visits and re-run when something shifts the lead stratum's ordering signal by more than five points or drops critic agreement below 0.7. The drift calculation is written and tested. Live monitoring is being wired, so it is not running yet.
A board-ready memo reviewed by a domain expert before delivery, custom indication parameterisation curated for your specific cancer, explicit acceptance criteria agreed up front, and a follow-up analysis. The engagement, not the seat.
Patient-advocacy foundations can sponsor Researcher seats in bulk from €59 per seat / month. See the sponsorship tiers.
Onkydra is two engines working in concert. Hydra runs the five stages that assemble the report. Chimera samples the co-mutation profiles those stages run on. Both are built for cancers where the incumbent stack has too few patients to work with.
Five-stage pipeline with a self-check
Five stages run in sequence: Planner, Cohort Architect, Biology Engine, Honesty Critic, Workspace Writer. The Honesty Critic flags where the ridge baseline and the mechanistic layer disagree, and declares which layers ran as proxies. It refuses to let a claim ship without a real source. Cut one head off, two grow back.
See the pipelineSimulated co-mutation profiles
Every draw is a simulated co-mutation profile, sampled from a Gaussian copula fitted to n=60 real H3 K27M DMG cases from DKFZ and CPTAC, Ledoit-Wolf shrinkage on Σ. One thousand Monte Carlo draws. Every CI resampled over the real 60, not the simulated 1,000. Same posture as the FDA's 2025 dordaviprone approval on n=50 with 22% ORR: the size of the underlying evidence is the size of the field.
Read the methodologyFounder-product fit
Onkydra is built by Faith Ogundimu, a cancer-genomics researcher, for the teams developing the drugs pharma won't. The workspace answers the questions Faith asks in her own PhD work: which subgroup responds, what breaks first under selection, what does the literature already know that we missed.
Research use onlySee scope
We are onboarding a small group of biotech and research teams working on rare cancers. Join the waitlist and we will reach out with a first run for your programme.
