Adding the vessel wall back to a tumour immunology assay

What the work claims

This is a platform paper reporting a primary result, and the two parts should be graded separately.¹ The engineering claim is that patient-derived lung adenocarcinoma organoids can be grown inside a self-assembled, perfusable human microvascular network, and that immune cells introduced through that network will extravasate and infiltrate the tumour. The biological claim, which is the one the abstract leads with, is that the resulting system "distinguishes responder and non-responder patient samples, consistent with the clinical observations", and that single-cell sequencing identifies tumour-driven hyperangiogenic signalling as the barrier to T cell entry, such that blocking PD1 and VEGF together restores infiltration.

The novel engineering move is a supporting one and easy to miss: the endothelial cells have had beta-2 microglobulin knocked out, removing surface MHC class I so that allogeneic endothelium does not itself activate the T cells being studied. That is what makes a single endothelial cell stock usable across many patients' samples.

How it works

The device has four compartments separated by micropillar arrays: one for the microvascular network, a central chamber for the organoid, side chambers for fibroblasts, and media channels. On day 0, human umbilical vein endothelial cells and normal human lung fibroblasts are seeded together in a fibrin gel. Fibroblasts are not decoration here; they secrete the factors that let endothelium organise itself. By day 3 the endothelial cells have self-assembled into a connected, perfusable network. Patient-derived lung tumour organoids go in on day 3, and the co-culture is characterised four days later.

Vessel quality was checked several ways. Network morphology by AngioTool was statistically unchanged by adding an organoid. Confocal imaging with three-dimensional rendering showed endothelial sprouts reaching toward and into the organoids, with an average vessel-to-organoid contact surface area of roughly 40 square micrometres. Perfusing 2000 kDa FITC-dextran confirmed the lumens were continuous and retained a very large tracer.

Two functional demonstrations follow. Perfused cisplatin killed organoids at 10 µM and above whether or not vasculature was present, while the peripheral microvascular cells appeared largely spared in the device even though endothelial monolayers died at 100 µM. Separately, perfusing THP-1 monocytes two hours ahead of CD8-positive T cells substantially increased how many T cells reached the tumour, reproducing the known role of myeloid cells in leukocyte recruitment.

The beta-2 microglobulin knockout was made with CRISPR-Cas9 against exon 1 and validated by flow cytometry, by TIDE analysis of the edited locus, and by Sanger sequencing. Edited endothelium formed networks indistinguishable from unedited, and provoked roughly five-fold less interferon gamma from co-cultured CD8 T cells.

The headline experiment then uses two patients. For each, organoids and tumour-infiltrating lymphocytes were derived from the same tumour. Monocytes and then the patient's own lymphocytes were perfused, with or without anti-PD1. Patient 2's lymphocytes crossed into the organoid; Patient 1's largely did not. Single-cell sequencing of 25,451 cells across nine annotated cell types found that Patient 1's organoids over-expressed VEGF and were enriched for angiogenesis and endothelial migration gene ontology terms, while Patient 1's CD8 T cells carried more PD-L1 and the T cell compartment scored higher for exhaustion. Adding an anti-VEGF antibody to anti-PD1 restored recruitment in Patient 1, and the authors report the effect as dose-dependent.

Where a skeptic should push

The load-bearing assumption is that this platform has been shown to separate patients who respond to checkpoint blockade from patients who do not. It has not, and the gap between what was measured and what is claimed is wide enough to matter.

There are two patients. More importantly, the comparator is not what the word "responder" implies. What the paper compares against is the baseline immune contexture of the original tumour: Patient 2's tumour had a median of five intraepithelial lymphocytes per ten tumour cells while Patient 1 had fewer than one, plus a matching difference in CD3-positive staining of the biopsy. I checked the full text specifically for treatment outcomes and found none. There is no named checkpoint inhibitor regimen, no RECIST assessment, no progression-free or overall survival, no prior therapy or follow-up. The supplementary table carrying the patient data is titled for demographics and clinical diagnosis, not clinical response. The paper's own most careful sentence describes the exercise as retrospective comparison with the patient's baseline tumour characteristics, which is accurate; the abstract's framing is not.

So the honest description is that the assay reproduced a phenotype that was already visible on the starting material. Tumour lymphocyte density is a predictive biomarker, not an observed response, and a model that recapitulates the biomarker it was built from has demonstrated internal consistency rather than predictive validity. This should not be read as an attack on the whole paper. Recapitulating baseline immune contexture is a legitimate first milestone for a platform, and the anti-VEGF perturbation does something no stained section can do. The specific sentence about distinguishing responders is what overreaches.

Replication is thin in the place it can least afford to be. Across the immune experiments the design runs to two or more devices per group, and in the dual blockade condition the reported sampling is at least two devices with at least one organoid per device, against at least eight organoids per device in the neighbouring anti-PD1-only arm. A dose-dependence claim resting in places on a single organoid per condition is a description of an observation, not a statistical statement. The unit of inference here is also the patient rather than the device, so running more chips from the same two donors cannot fix the underlying sample size.

There is a circularity risk in the mechanism as well. The sequencing that explains why Patient 1 differed from Patient 2 was performed on Patient 1 and Patient 2. With two samples, a post-hoc transcriptional difference can always be found and narrated; VEGF is a plausible and attractive candidate, but it was selected after the outcome was known.

I also want to correct a reading I initially found tempting. It is natural to treat the anti-VEGF rescue as proof that the block was physical, a transport failure rather than an immunological one. That inference does not hold, and I have dropped it. VEGF is directly immunosuppressive: it impairs dendritic cell maturation, expands regulatory T cells and drives CD8 exhaustion including checkpoint upregulation. Blocking it could restore recruitment by normalising vessel permeability, by restoring endothelial adhesion molecule display such as ICAM-1 and VCAM-1, or by simply relieving suppression of the T cells themselves. The paper cannot separate these, and its own data show both mechanisms present in the same sample, since Patient 1's T cells were already more exhausted. The experiment that would separate them is cheap: perfuse healthy donor T cells into Patient 1's vascularized organoid. If fresh, unexhausted cells still fail to enter, the block is architectural.

Two further limits. Human umbilical vein endothelium is macrovascular, venous, fetal and not from lung. Organ-specific endothelial identity governs the adhesion molecule repertoire and chemokine display that control where leukocytes cross, which is precisely this paper's endpoint. That is an acceptable substrate for demonstrating that a network forms and perfuses; it is the weakest available substrate for a claim about tissue-specific T cell trafficking, and primary human lung microvascular endothelial cells are commercially available. Separately, a 2000 kDa tracer shows the lumen is continuous and holds something very large. It says little about permeability to a small molecule such as cisplatin, and less about whether the barrier presents a physiologically realistic obstacle to a migrating cell. A size ladder of tracers would be more informative. On that note, the apparent sparing of endothelium from cisplatin in the device is at least as easily explained by differences in delivered dose under perfusion as by genuine vascular protection, and measuring local drug concentration would settle it.

Consequences for immuno-oncology models

The non-obvious contribution is not that this model is more realistic. It is that it restores a step that conventional assays silently delete, and in doing so makes a distinction addressable that was previously invisible.

When T cells are pipetted into a well containing a tumour organoid, a failure to kill can only be read one way, as a failure of the T cells. In a patient, that same failure has at least two very different causes: the T cell never reached the tumour, or it reached the tumour and was exhausted once there. Those imply opposite interventions. The first is a vascular and trafficking problem, arguing for agents that normalise vessels or restore adhesion and chemokine signalling. The second is a T cell problem, arguing for checkpoint blockade, cytokine support or a different cell product. A plate cannot tell them apart because it has removed the barrier whose crossing defines the difference. A perfusable endothelial network reinstates that barrier and therefore makes the two failure modes separable in principle, even though this particular paper does not yet separate them.

That reframes what such a platform is for. Its natural application is not another drug sensitivity number. It is a potency and trafficking assay for cell therapies, where the product is itself a cell that must navigate to a target. For CAR T and tumour-infiltrating lymphocyte products, the ability to reach solid tumour tissue is a known and largely unmodelled failure mode, and a system that scores extravasation separately from cytotoxicity is measuring something no current release assay captures.

The beta-2 microglobulin knockout is worth thinking about carefully, because it is a genuine advance and a genuine compromise at once. It converts a patient-specific reagent problem into a catalogue product: one endothelial stock, usable across donors, without the allogeneic MHC class I response contaminating every immune readout. That is what makes the platform scalable, and scalability is the difference between a demonstration and an assay. But the standardisation is partial in a way the paper does not weigh. Removing beta-2 microglobulin removes MHC class I only. Class II, minor histocompatibility antigens and other endothelial alloantigens remain, so this endothelium is still foreign to both the tumour and the lymphocytes, and the residual mismatch is uncontrolled and unquantified. More to the point for these specific readouts, endothelium is a real antigen-presenting surface for CD8 cells, so deleting its MHC class I also deletes a genuine axis of endothelial control over T cell activation and trafficking. The usual objection, that beta-2 microglobulin null cells become natural killer targets through missing-self recognition, does not bite here because no NK cells were perfused, but it would bite immediately in any innate immunity application.

The threat is a commercial and epistemic one rather than a technical one. A platform that reproduces baseline tumour lymphocyte density is reproducing a measurement a pathologist can already make on a slide, more cheaply and with a far larger evidence base behind it. If systems like this are marketed on responder prediction while the supporting evidence is agreement with a baseline biomarker in two patients, the field acquires a category of assay that looks validated, is not, and is expensive. The way out is unglamorous and specific: run the organoid assay before treatment, treat the patient, and compare the assay to what the patient actually did, in enough patients to support the word predictive. Until that exists, the correct claim for this class of platform is mechanistic dissection, which is a real and sufficient contribution on its own.

The bottom line

Established: patient-derived lung tumour organoids can be cultured inside a perfusable human microvascular network without disrupting it; monocytes increase CD8 T cell recruitment through that network; beta-2 microglobulin can be knocked out of the endothelium without impairing network formation while cutting interferon gamma release roughly five-fold; and in two patients, autologous lymphocyte infiltration under anti-PD1 tracked the baseline lymphocyte density of the original tumours.

Asserted rather than demonstrated: that the platform distinguishes clinical responders from non-responders, since no treatment outcome data appear anywhere in the paper; that tumour VEGF is the causal barrier, which is a post-hoc explanation fitted to two samples; and that the anti-VEGF rescue identifies a transport-limited mechanism, which the experiment cannot separate from direct immunosuppression.

What would confirm the strong version: a prospective cohort large enough to compare assay output against measured clinical response; perfusion of healthy donor T cells into a non-infiltrated patient's organoid to test whether the block survives replacement of the lymphocytes; and a repeat on primary lung microvascular endothelium to show the trafficking result is not an artefact of umbilical vein cells. What would break it: finding that infiltration in this system tracks the starting lymphocyte density and nothing further, which would make the device an expensive route to a number already available from a stained section.

Background on the vascular problem in engineered tissue is in our vascularization page, and the wider stream is indexed at research analysis.