Drop in a target
Name a drug target and a rare cancer (DMG at launch). Onkydra picks the right synthetic-cohort library, cell-line panel and molecular signatures for you. No specialised syntax, no nine separate tools.
In construction · Private beta soon The agentic in-silico oncology lab.
Onkydra simulates rare cancers in-silico. Drop in a drug target and it runs a virtual patient cohort, predicts who responds, and ranks what drives resistance, before any wet-lab spend.
CDK7 inhibitor · H3 K27M DMG
Predicted responder fraction 28% 95% CI 11–49% · H3.3 / TP53-mut
Top resistance pathway CDK2 bypass LINCS reversal · concordance 0.74
Ensemble agreement 0.81 ridge vs foundation model
New paper PMID 41234567 on CDK7-driven DMG. Re-ran automatically. Predicted responder fraction up 4 points.
How it works
Three steps. The verb is simulate, not search.
It simulates, not searches
A virtual patient cohort is drawn from real molecular data. Response is predicted per molecular subgroup, the top resistance pathways are ranked, and where the models disagree is surfaced as an honesty signal.
It keeps working after you log off
When ClinVar reclassifies a variant, a new CRISPR screen lands or a Phase 1 result publishes, Onkydra re-runs your target and flags what changed. Your board memo regenerates on demand.
Why it's different
A workspace that thinks like a sceptical colleague.
Simulate, don’t search
Most tools tell you what is already published. Onkydra predicts what would happen if you perturbed a target, so you can rank candidates before spending a cent in the wet lab.
Honest about its limits
When a simpler model beats the fancy one, Onkydra says so and ships the honest predictor. Where two models disagree becomes a signal you can act on, not something to hide.
Always live
Each target sits in your workspace as a living simulation that re-runs itself as the evidence moves. Come back weekly to see how your thesis is holding up.
The first indication
Starting with DMG.
Diffuse midline glioma is a rare childhood brain cancer that is almost always fatal, with survival that has barely moved in decades. The first targeted therapy was only FDA-approved in 2025, on a pooled cohort of roughly fifty patients. When real patients are that scarce, every simulated cohort matters. Onkydra draws a thousand synthetic cohorts from that hard-won data to stretch the evidence and prioritise what to test next.
≈50Real patients behind the first approval
1,000Synthetic cohorts drawn from it
DecadesOf flat survival we want to move
Lineage
Built for the Gemini XPRIZE.
Onkydra grows directly out of my rare-cancer genomics research, and is being built on Google's Gemini models for the Build with Gemini XPRIZE. It is the most ambitious thing I'm making: turning the messy reality of rare-cancer data into something a researcher can actually simulate against.
AI for regenerative biology
Request early access.
Onkydra is in active construction, starting private and close to the science. If you're a rare-cancer researcher, clinician or builder who wants in early, leave your email.
For researchers and partners. One email when the beta opens.
Events
No events scheduled right now. XPRIZE milestones, demos and talks will appear here as Onkydra moves.