When the stroma rewrites a colorectal cancer cell's drug response

What the work claims

This is a primary mechanistic study built on human colorectal cancer patient-derived organoids (PDOs, three-dimensional cultures grown from a patient's own tumour cells) and the cancer-associated fibroblasts (CAFs, the tumour's dominant stromal cell) that surround them in vivo. Its central claim is unusually sharp: among the dozens of signals a fibroblast secretes, a single one, prostaglandin E2 (PGE2, a lipid made by the enzyme COX2, encoded by the gene PTGS2), is the dominant regulator of colorectal cancer cell fate.¹

Colorectal cancer cells are phenotypically plastic. They slide along an axis between a proliferative colonic stem cell state (proCSC, cycling, chemosensitive) and a slow-cycling revival colonic stem cell state (revCSC, marked by CD55, EMP1 and ANXA1, associated in prior work with metastasis and chemoresistance). The paper argues that stromal PGE2 acutely and reversibly pushes cells from proCSC toward revCSC, that only a defined subset of cancer cells (those expressing the transcriptional co-repressor DACH1) are susceptible, and that knocking out or inhibiting COX2 traps the epithelium in the chemosensitive proCSC state. That last point is offered as a mechanistic explanation for a thirty-year-old epidemiological observation: why anti-COX drugs such as aspirin reduce colorectal cancer risk and recurrence.

How it works

The load-bearing method is an intercellular CRISPR screen, and it is genuinely clever. Rather than perturb the cancer cell, the authors put Cas9 into the fibroblasts and knocked out members of the CAF secretome one at a time, then read out what happened to the co-cultured cancer epithelium at single-cell resolution across roughly 972 knockout co-culture conditions. The epithelial response was quantified as a stem cell index (SCI), a relative score subtracting the revCSC signature from the proCSC signature, so a fall in SCI means a shift toward the resistant revival state.

The screen produced a clean negative result and a clean positive one. Knocking out the classic niche factors WNT3A, EGF, NOG and RSPO1 in the fibroblast did nothing to the cancer cell's fate, and, strikingly, neither did knocking out fibroblast TGF-beta, despite its celebrated role in mouse colorectal cancer. Only four perturbations moved the needle: loss of PTGS2, IL1R1, GREM1 or HGF each pushed cells back toward proCSC. A follow-up experiment tested all sixteen combinations of the four surviving ligands in a factorial design across several PDOs, and PGE2 alone reproduced almost the entire effect of full fibroblast co-culture. The logic was then closed at both ends: PTGS2-knockout fibroblasts lost the ability to shift cell fate, and adding back PGE2 at 1 to 50 nanomolar rescued it, while aspirin (a COX inhibitor) blocked it. A single-cell time course showed PGE2 firing fast phospho-signalling within six hours, then a delayed collapse of DACH1 and a switch to revCSC that reversed once the lipid was withdrawn.

To ask whether any of this happens in real tumours, the authors turned to Xenium spatial transcriptomics on five primary human colorectal cancers. They report that COX2-positive fibroblasts sit inside a DACH1 exclusion zone roughly 200 to 300 micrometres wide: nearby epithelium is DACH1-negative and revCSC-like, more distant epithelium is DACH1-positive and proCSC-like. A reanalysis of a public patient cohort (GEO accession GSE39582, restricted to 316 untreated patients) is used to argue that the COX2-high, DACH1-low configuration tracks with the stroma-rich, poor-prognosis consensus molecular subtype.²

Where a skeptic should push

The mechanism, at the level of a signalling ligand, is well supported. The necessity-and-sufficiency pairing (COX2 knockout removes the effect, PGE2 add-back restores it) is exactly what one wants, and the rapid reversibility favours genuine state induction over the stable selection of a pre-existing resistant clone, though without lineage tracing it does not fully exclude a state-selective survival effect. That distinction matters, because a screen that merely enriches for survivors tells you less than one that actively converts cell state.

The weak joint is the word chemorefractory. Nowhere in this study is drug killing actually measured on PGE2-induced revCSC cells. The one viability assay dosed the baseline, untreated organoid panel, and the link between the revCSC state and chemoresistance is imported from prior literature and supported here only by a correlation between proCSC/revCSC gene signatures and the LD50 of 5-fluorouracil, oxaliplatin and SN-38 across those baseline lines. No experiment compares drug response with and without PGE2, CAFs, or COX2 loss. A correlation between a transcriptional signature and a drug-sensitivity number is not a demonstration that PGE2-driven transdetermination causes the same cells to survive chemotherapy. The human cohort used to anchor the clinical story (GSE39582) is untreated, so it speaks to prognosis, not to chemotherapy response. The clinical payoff, COX2 inhibition as a chemo-sensitiser, rests on that unmeasured step plus the aspirin epidemiology, with no interventional or in vivo drug-response experiment in this paper.

Two further cautions. Dominance is scoped to the library that was perturbed; PGE2 is the dominant member of the tested CAF secretome, not proven dominant over every possible niche input, and the surprising TGF-beta null may say as much about the assay window as about human biology. And the spatial argument is observational: a 200 to 300 micrometre DACH1 exclusion zone around COX2-positive fibroblasts is consistent with short-range lipid signalling but cannot by itself isolate PGE2 as the cause, since fibroblast density, perfusion and inflammatory geography all co-vary with those cells. Hypoxia, often invoked here, is better read as an upstream inducer of COX2 than as a competing explanation. The organoids also come from a single biobank and the responsive subset is DACH1-positive by construction, so the generalisation from ten discovery lines and four factorial lines to colorectal cancer as a class should be read as a hypothesis about a subset, not a settled property of the tumour type.

Why monoculture screens miss this resistant state

The result's most consequential implication is not about aspirin. It is about the dominant industrial format for organoid drug discovery: monoculture PDO screening. A tumour organoid grown without fibroblasts already contains an intrinsic proCSC/revCSC admixture, but it never receives the stromal PGE2 that, in this study, is the switch driving cells into the refractory revCSC fate. A compound scored as a strong killer against monoculture organoids may therefore be systematically over-nominated: it was tested against a cell-state distribution the patient's COX2-positive stroma actively shifts away from. This is a concrete, mechanism-grounded version of a worry the field usually states only in the abstract, that context-free organoid screens over-predict response.

The genuine opportunity is that the paper also hands over a cheap fix. Because a single defined ligand reconstitutes almost the entire fibroblast effect, one does not need a full autologous CAF co-culture to stress-test hits for stromal escape. A PGE2-challenged arm, roughly 1 to 50 nanomolar added to an otherwise standard monoculture screen, is a testable, standardisable control condition for asking whether a candidate still works once cells are pushed toward revCSC, though its predictive value would itself need validation across independent organoid cohorts. DACH1 and PTGS2 become plausible spatial biomarkers for which tumours are stroma-plastic in the first place, and an approved drug class (NSAIDs and coxibs) becomes a candidate combination strategy, one still requiring direct chemotherapy-combination testing rather than a settled chemoprevention story.

The threats are equally concrete and worth naming. The state is acute and reversible, so it is a moving target: a screen that samples one PGE2 dose at one timepoint can still misrank compounds whose advantage is state-dependent. The fix is selective, helping only the DACH1-positive subset, so a naive PGE2 add-back would leave DACH1-negative tumours mis-modelled in a different direction. And there is a substrate-obsolescence edge for platform vendors: if a defined lipid captures most of this axis, the commercial case for elaborate patient-matched fibroblast co-culture weakens for this specific readout, though it says nothing about the matrix, immune and metabolic functions of stroma that PGE2 does not cover.

The bottom line

At the level of signalling, the claim is strong and well controlled: stromal PGE2 is the dominant tested ligand that reversibly transdeterminates DACH1-positive colorectal cancer cells from proCSC toward revCSC, and COX2 loss or inhibition blocks it. At the level of therapy, the claim is a well-motivated hypothesis, not an established result: chemoresistance is inferred from signatures and prior work rather than measured, and the clinical benefit of COX2 inhibition is borrowed from epidemiology. What would confirm it is a direct experiment, showing that PGE2-induced revCSC cells survive 5-fluorouracil, oxaliplatin or SN-38 better than their proCSC siblings in the same organoid, and that COX2 inhibition improves chemotherapy response in vivo. What would break it is a finding that the signature shift does not translate into measured drug survival, or that the spatial DACH1 gradient is driven by hypoxia rather than a PGE2 gradient. For drug discovery, the actionable takeaway needs neither to be settled first: monoculture organoid hit lists are missing a stromal axis that a cheap, defined perturbation can restore.