Colorectal cancer's epigenome loses a third of itself in organoid culture - and that may be the most important number in the paper

What the work claims

This is a primary result, still a preprint and not yet peer-reviewed, from a group that spans the Istituto Nazionale dei Tumori in Milan, the University of Milan and the Institute of Cancer Research in London (Pietrantonio, Sottoriva and colleagues). It makes three claims of different strengths. First, from simultaneous single-nucleus ATAC- and RNA-sequencing (the 10x Genomics Multiome assay, which profiles chromatin accessibility and gene expression in the same individual cells) on 33 primary colorectal cancers, 11 metastases and 7 distant normal colon samples from 37 patients, the authors catalogue somatic chromatin accessibility alterations - SCAAs - and show these epigenetic changes are substantially more recurrent across patients than somatic mutations or copy-number alterations¹. Second, they show tumour-adjacent epithelium that looks genetically normal already carries tumour-like inflammatory chromatin changes - a field effect with implications for what "normal" comparison tissue means. Third, and most functional: they use CRISPR interference (CRISPRi) screens in patient-derived organoids to prove that a small set of these recurrently altered regulatory elements are genuine cancer dependencies, required for tumour cell survival in vitro¹.

How it works

The design has two halves. In the tissue half, the authors pseudo-bulk epithelial cells from each specimen and call peaks differentially accessible against pooled normal colon epithelium, nominating SCAAs per patient. Because single-cell ATAC data is too sparse to call locus-level variants in individual cells, this pseudo-bulk approach is the field's standard, and the authors validate their copy-number calls against matched whole-genome sequencing. The striking pattern: only 16 genes are mutated in at least a quarter of the 36 patients with sufficient tumour cells for recurrence analysis (APC and TP53 above 70 percent, KRAS above 25 percent), while roughly 12 times as many genes - 399 - carry promoter SCAAs in more than 8 of 36 patients. Promoter SCAAs also show a two- to three-fold higher odds of being shared between random tumour pairs than shared mutations or copy-number changes, and metastases share even more, consistent with positive selection of particular epigenetic states during progression. Matched samples from the same patient (five pre- and post-treatment primary pairs, two synchronous multi-site pairs) shared on average 37 percent of SCAAs, versus 19 percent for random unmatched pairs, so these alterations are reasonably stable within a patient over time.

The second tissue finding concerns what the authors call "pseudonormal" epithelium: genomically unperturbed cells found inside 12 of 33 primary specimens, about 14,000 cells in total, which look normal by chromatin accessibility but sit transcriptionally close to tumour cells. These cells display an AP-1-driven stress program (FOS, JUNB) and loss of cell-cycle-exit and circadian regulators (KLF6, MXI1, BHLHE40 and BHLHE41), changes the authors attribute to extrinsic inflammatory signalling from the neighbouring tumour rather than to cell-autonomous mutation. In other words, the tumour's epigenetic influence extends into tissue that every genomic test would call clean.

In the functional half, the authors engineer four colorectal cancer organoids (three microsatellite-stable, one microsatellite-instable) to express dCas9 fused to the dual repressor KRAB-MeCP2, a construct that silences the genomic region a guide RNA points at without cutting DNA. They screen a custom library of 850 guides targeting 112 promoters and 41 enhancers that gain accessibility in at least 25 percent of the cohort, plus 52 known essential and 43 non-essential control genes and 50 non-targeting guides. Guides that drop out over 11 to 28 days mark regulatory elements whose silencing kills the cells. The screen separates its control sets cleanly after about three weeks, and a small set of element-targeting guides deplete almost as strongly as essential-gene guides. The recurrent hits: promoters of DNA2 (a known DNA-damage-response essential gene), the membrane-integrity regulators ATP11A and ANXA3, the developmental transcription factor OTX1, and distal enhancers of TGIF1 and MSX2¹.

Where a skeptic should push

The most load-bearing assumption is that an organoid's chromatin state is a faithful proxy for the patient's tumour chromatin state - because the entire functional half of the paper is executed in organoids, and this is exactly where the authors' own data push back hardest. Comparing two matched tissue-organoid pairs, only 33 percent (C017) and 43 percent (C107) of tissue SCAAs were retained in the corresponding organoids, with Pearson correlations of 0.33 and 0.46. A third to a half of the catalogue the screen is drawn from does not survive a transition the field makes thousands of times a week. The authors are commendably transparent about this; most screening papers never check.

Second, the screen's significance threshold is a false-discovery rate of 0.2, which is lenient by genome-scale screening convention, and the hit set rests on four organoid models from a single centre. The OTX1, TGIF1 and MSX2 dependencies are biologically coherent - developmental reprogramming factors are plausible cancer addictions - but coherence is not replication, and nothing here is tested in vivo. Third, the dependency claim is defined by viability in culture; whether silencing these elements kills cells inside a tumour, with its nutrient stress and immune pressure, is open. Fourth, as a preprint this has not passed peer review, and several numbers in the text shift slightly between sentences (the pseudonormal fraction is described as 12 of 33 specimens in one place and 13 of 33 in another), which suggests the manuscript is still being tightened. Weight the descriptive atlas heavily, the dependency list provisionally.

Note also what the medium does. The organoids are grown in Matrigel with Wnt-3a, R-spondin-1, EGF and a cocktail of niche factors - a high-Wnt, growth-factor-rich environment. Wnt and TCF/LEF activity is precisely the axis on which these tumours are already abnormal, and culture in saturating Wnt is known to flatten the very lineage and stress programs this paper measures. Some of the lost third of the epigenome is presumably selection; some is almost certainly medium.

What chromatin drift does to organoid drug screens

For the organoid field, this paper is simultaneously a capability demonstration and a calibration instrument. The capability is real and worth taking seriously: dCas9-KRAB-MeCP2 CRISPRi in patient-derived organoids turns the epigenome into a druggable screenable surface. Every oncology program now chasing epigenetic targets - writers and erasers beyond the blunt DNMT and HDAC inhibitors, but also the regulatory elements themselves - currently depends on cell lines that have spent decades adapting to plastic. The authors note their hit genes show limited dependency in large pan-cancer cell-line screens, consistent with earlier work showing organoids carry a broader dependency spectrum than cell lines²; if that holds, enhancer-level CRISPRi in organoids is a genuinely new target-discovery engine, run in a patient-matched, three-dimensional, human context¹.

The calibration is the uncomfortable part. Organoid drug response is read as a phenotype, but the phenotype sits on a chromatin substrate that culture measurably rewires. If a third of tissue epigenetic state is gone by the first passage, then epigenetic-dependence findings from organoids - including this paper's own - are conditional on culture conditions, and drugs whose mechanism touches chromatin (the entire epigenetic-drug class, but also hormone, differentiation and stress-pathway agents) may be screened against a target that is partly an artifact of Wnt-saturated Matrigel. The practical consequence is that organoid QC as currently practised - mycoplasma, morphology, a genotyping panel, maybe a transcript check - is blind to the variable this paper shows matters. A reasonable near-term fix is cheap: profile a small set of sentinel SCAAs per model at establishment and after expansion, and report them alongside drug-response data, so at least drift is visible even if it cannot be eliminated.

The pseudonormal finding cuts the other way, at organoid realism rather than organoid convenience. The tumour's inflammatory field effect lives in genetically normal epithelium interacting with the microenvironment, and no epithelial-only organoid carries it. Colorectal organoids are excellent tumour epithelium and terrible tumour neighbourhood - which is precisely the gap co-culture, assembloid and immune-reconstituted systems are trying to close, and this paper supplies single-cell reference maps for what a faithful reconstruction would need to preserve. The threat to existing programs is quiet: two labs screening the same drug on colorectal organoids may disagree not because either erred, but because their cultures drifted in different directions - and neither would know, because nobody measures it.

The bottom line

Established: across 37 patients, recurrent chromatin accessibility alterations in colorectal cancer outnumber and out-recur recurrent mutations; they are stable within patients over time; tumour-adjacent genetically normal epithelium carries a tumour-linked inflammatory chromatin program; and at least five regulatory elements (around DNA2, ATP11A, ANXA3, OTX1, TGIF1 and MSX2) score as survival dependencies in four organoid models under CRISPRi. Hypothesis, not yet established: that these dependencies hold in vivo, generalize across centres and culture protocols, and are therapeutically accessible. What would confirm the organoid screening angle: dependency replication across many more PDO lines and medium formulations, retention statistics reported as standard QC, and in vivo validation of at least the ATP11A or MSX2 axis. What would break it: systematic evidence that the retained third of the epigenome is the culture-adapted, least-tumour-like third.