← CHANGELOG
One thousand draws is a resolution parameter. Every interval comes from the sixty.
16 August 2026·6 min read·Onkydra · Faith Ogundimu
The anchor for our one wired indication is 60 real H3 K27M cases, pooled from DKFZ paediatric pan-cancer (n=53) and CPTAC paediatric brain (n=7), fetched through the cBioPortal API. Thirteen binary driver features. No treatment-response layer and no survival layer, which is the limit that matters most and the one people ask about last.
From that we fit a Gaussian copula and draw simulated co-mutation profiles over molecular phenotype space. They are Monte Carlo draws from one fitted model. They are not patients, they are not donors, and they never appear in a denominator described as either. The distinction sounds pedantic until you notice what it forbids.
Draws are resolution, not information
A thousand draws and ten thousand draws from the same fit carry exactly the same evidence. Increasing the draw count buys smoother histograms and nothing else. So resampling the draws to get a confidence interval is fiction: it produces tight intervals that pretend the field is bigger than it is, and the tightness is a function of how long you left the sampler running.
Every interval we report is bootstrapped over the 60 real cases instead, resampled with replacement. The width is set by 60, because 60 is what is known. A marginal frequency of 35% with an interval from 18 to 52 is what 60 cases support; the same number quoted to three significant figures off a thousand draws is not.
This is the same posture as the approval that made this indication fundable: dordaviprone cleared FDA accelerated approval in August 2025 on a pooled n=50. The size of the underlying evidence is the size of the field.
The joint is regularised hard, on purpose
Fitting a joint distribution at this sample size needs regularisation, and ours is not a light touch. Three techniques, named without their settings, which stay in the preprint: Ledoit-Wolf shrinkage on the correlation matrix; NORTA calibration to map binary correlations onto the latent Gaussian scale before sampling, because without it thresholding attenuates them a second time; and Higham's nearest correlation matrix when the latent matrix comes back non-positive-definite, which it does. NORTA does not preserve positive-definiteness, and that is a known property of the method rather than a bug in ours.
The consequence surprises people. The realised pairwise odds ratios in the draws will not sit on the 45-degree line against the anchor's, and they should not. Shrinking a joint and landing on the diagonal are in direct tension: reproducing the anchor's noisy pairwise structure exactly would mean fitting the noise. Systematic attenuation toward the centre is the correct behaviour, not something to tune away.
Which layer to trust
This is answerable rather than a matter of taste. Fit on the anchor alone, then test against cases held out of the fit by construction. Single-gene frequencies reproduce at r=0.92 on 72 held-out cases from the same study and r=0.59 against an independent institution. Quote 0.59 when the question is transportability, because that is the one that describes moving to somebody else's cohort. Pairwise co-occurrence does not transport reliably at all.
So the marginals are the trustworthy layer and the fine joint is a hypothesis. Both figures moved down when we corrected the anchor upward and the H3.1 cases that had been sitting in the held-out set joined the fit, which made the remaining comparison set harder. That is the correct direction for a correction to move a held-out score.
The rule this puts on the price list
A billing unit is a denominator. If we charged per simulated profile, per report or per run, we would have a commercial incentive attached to the one number our own methodology says must never be read as a sample size, and every incentive to let a customer read a thousand draws as a thousand observations.
So we cannot price that way, and it is enforced rather than intended: the prohibited bases are declared in code, the daily run allowance is declared as a ceiling on spend rather than a price basis, and a guard asserts that no consumable, credit pack or top-up exists in checkout. Pricing per report has the same problem from the other end. It prices the artefact, when the strongest thing this product can do is talk you out of an experiment, and a decision not to run something produces no document to bill for.
What the pooled entity actually is
The pooled entity is not cleanly diffuse midline glioma. Of the 53 anchor cases from the DKFZ series, 52 carry the label High-Grade Glioma, NOS, and one is a pilocytic astrocytoma. The defensible description of the anchor is 60 H3 K27M-mutant paediatric high-grade glioma cases, and a domain reviewer will open with this.