What the biobank makes checkable
The Sanger Institute biobank holds 256 organoids grown from colorectal, oesophageal, pancreatic, stomach and ovarian tumours, donated with consent at UK hospital sites in Birmingham, Cambridge, Glasgow, London and Southampton. What stands out here is the methodological choice rather than the count. Where enough material existed, DNA was sequenced from the organoid, from the patient's blood and from the original tumour, so each model can be compared against the thing it came from and watched for drift as it is passaged. That is what turns a collection into a reference: not that the models are good, but that a later user can check how good, on the same evidence, without asking the authors.[1]
On top of that base the team ran CRISPR screens across 162 of the models and identified thousands of dependencies, some shared by many cancers and some tied to a tumour type. Joined to genomic and clinical annotation, those screens produced 1,733 links between a dependency and a specific feature such as a DNA change or a treatment history. The organoids matched the genetic features of the source tumours closely, and in some cases models grown from the same patient before and after treatment showed how resistance had been acquired. That last design is the strongest thing in the paper, because a matched pair from one patient controls for the variation between patients that swallows most comparisons of this kind.[1]
The composition of the collection
The companion resource is broader and reports its own audit. Ten years of work under the Human Cancer Models Initiative produced 665 models from 2,780 donors across 25 cancer types, of which 522 carry comprehensive clinical data and 153 are of rare cancers. Its composition is stated plainly, which is to the project's credit: 71 models come from participants of non-European ancestry, 43 from paediatric or adolescent donors, and 23 per cent of the successful models are of rare cancer types. More useful than any of those counts is what the team did with 421 matched tumour-model pairs: genetic concordance came out at 97.8 per cent and epigenetic concordance at 95 per cent, and the analysis defines what goes with the discordant cases instead of setting them aside.[2]
Put the two side by side and the boundary of the claim appears, in two directions. A dependency called in 162 organoids drawn from five tumour types and five UK hospital sites, screened with CRISPR, is well founded for those tumours and those patients; whether it holds for a rare subtype, or in a population represented by 71 of 665 models, is a separate question neither resource answers. The second boundary is the dish. The compendium's single-nucleus sequencing found subsets of models in which culture conditions significantly shape cell states, and the CRISPR screens in the biobank identified essential genes specific to organoids. A gene a cell needs only because of how it is being grown is a real measurement and a poor guide to a patient. The narrower reading of the same evidence is that many dependencies follow a single driver mutation and travel with the genotype, independently of donor and flask, in which case the composition gap bites hardest on the calls that lean on tissue context and prior treatment, which is where these 1,733 links mostly sit.[1], [2]
What would count as failure?
Yesterday this column looked at a catalyst whose reported relationship did not survive being run in four laboratories, and asked what a claim owes to repetition. Here the equivalent test is unusually well specified, because the models, the sequences and the screen results are being released. A dependency that reverses between growth formats, or between early and late passages of the same model, would not be an embarrassment; it would be the measurement telling everyone which calls are properties of the tumour and which are properties of the dish. The resource is worth having precisely because that test can now be run by someone who did not build it.[1], [3]