Cell state programs may outclass grade alone in neuroendocrine tumors
Well-differentiated gastroenteropancreatic neuroendocrine tumors have long resisted tidy classification. Two tumors can share the same organ of origin, morphology, and Ki67 grade, yet behave very differently at the bedside. One may remain indolent for years; another may metastasize, seed fibrosis, or respond poorly to therapy. The central argument of this commentary is straightforward: tumor site and grade still matter, but they may be too blunt to capture the biology that drives progression in gastroenteropancreatic neuroendocrine tumors.
A major reason for that mismatch is that current grading compresses dynamic cellular states into a few bins. The WHO system assigns an organ of origin and a proliferative grade, but the same label can mask meaningful differences in metastatic potential and treatment response. Even Ki-67 itself can shift depending on whether the count is taken from a hotspot or across the whole slide. In one multisite analysis cited in the commentary, grade assignments also varied across primary, regional, and distant sites. The result is a classification scheme that is clinically useful, yet often imprecise when oncologists need to predict where a tumor is headed next.
A paired transcriptomic and chromatin map of 38 tumors
The study under discussion assembled the largest multimodal dataset yet reported for these tumors: joint single-nucleus RNA sequencing and ATAC sequencing from 38 archival well-differentiated tumors, including 21 pancreatic NETs and 17 small-intestinal NETs. Investigators generated 217,636 transcriptomes and 74,004 chromatin accessibility profiles, then added Visium spatial transcriptomics on 19 matched specimens and Xenium imaging on 4 specimens. That design matters because it links gene expression, regulatory architecture, and tissue position in the same disease ecosystem rather than treating each as a separate problem.
An unsupervised consensus nonnegative matrix factorization approach was used to avoid forcing malignant cells into rigid categories. Instead of discrete subtypes, the tumors resolved along a continuum. Two conserved malignant programs emerged across both organ sites: a neuronal-like program and a secretory neuroendocrine program. The neuronal-like state featured synaptic and axon guidance genes such as NRXN1, DLG2, RIMS2, and KCNJ3. The secretory program carried the canonical machinery of hormone production and release, including CHGA, CHGB, PCSK1, and SCG5, along with enrichment for oxidative phosphorylation and mTORC1 signaling.
This is not a trivial re-labeling exercise. It implies that neuroendocrine tumors do not simply differ by where they arise; they also differ by how fully they retain secretory identity versus how strongly they engage neuronal-like programs linked to a different regulatory posture.
One continuum, different epigenetic wiring by organ site
The transcriptional convergence across pancreatic and small-intestinal tumors did not erase organ specificity. Matched ATAC data showed that the same broad programs were sustained by different transcription factor motifs in each site. In small-intestinal tumors, neuronal-like cells were enriched for MAX:MYC, MITF, and CLOCK motifs, while secretory cells were associated with HOXA9, AP-1, and SOX9 motifs. In pancreatic tumors, neuronal-like cells aligned with TFAP4, NR3C2, and ISL1, while secretory cells showed a different homeodomain signature including MSX2, LBX2, and EN2.
The practical implication is subtle but important. The output programs may be shared across tumor sites, but the upstream epigenetic wiring is not identical. A therapy aimed at the final state of the cell may therefore have broader relevance than one aimed at a site-specific transcription factor. In plain terms, the same malignant behavior may be assembled through different regulatory routes, and the downstream state could be the more actionable target.
Tumor neighborhoods follow the program, not random chance
Spatial analysis added another layer. Secretory-high malignant regions tended to sit deeper within the tumor interior and were embedded in fibrovascular, stromally infiltrated niches. These areas were enriched for endothelial cells, fibroblasts, and myeloid populations, and they carried elevated COL1A1, SPARC, and MMP2. Predicted signaling links in these regions included TGFB1-ITGB1, VEGFA-FLT1, and LAMA2-ITGA1.
Neuronal-like–high regions occupied a different terrain. They appeared in densely cellular zones with less stromal infiltration and a higher local proportion of immune spots in the immediate neighborhood. Predicted interactions there included CX3CL1-CX3CR1 and IL34 or CSF1 signaling through CSF1R. Xenium imaging in four specimens confirmed the same basic architectural split, placing endothelial, fibroblast, and myeloid cells adjacent to malignant cells in secretory-high regions.
The anatomy is striking.
This spatial segregation supports the idea that malignant cell state and microenvironment are biologically coupled. The tumor is not merely expressing a program; it is building a niche around that program.
Fibrosis, metastasis, and the secretory axis
The commentary also situates these findings in a familiar clinical problem: fibrosis. Neuroendocrine tumors can produce mesenteric desmoplasia that obstructs bowel or contribute to carcinoid heart disease. Fibrosis in this disease has long been linked to vasoactive peptides and serotonin acting through TGF-β1 pathways. The secretory program is a plausible source of those signals because its small-intestinal form includes TAC1, which encodes substance P, and secretory-high regions expressed TGFB1 within the fibrovascular niche.
The study did not directly test whether tumors with stronger secretory states are the ones that generate more desmoplastic complications. It does, however, provide a biologically credible framework for asking that question in existing cohorts. If the secretory state helps organize a pro-fibrotic microenvironment, then fibrosis may be more than a downstream complication; it may be part of the malignant state itself.
Stage associations leaned in the same direction. Across both tumor types, neuronal-like scores were higher in localized primary tumors, while secretory scores were higher in metastatic lesions. In pancreatic primaries, neuronal-like scores exceeded secretory scores. An external cohort of 22 pancreatic primaries from patients with localized or metastatic disease showed lower neuronal-like scores in tumors from patients with metastases, with secretory scores trending higher. The sample sizes were small and some comparisons did not survive correction, so these data are best treated as hypothesis generating rather than definitive.
What this means for trial design and biomarker strategy
The next step is not another descriptive atlas. It is biomarker-driven testing of whether this cell-state axis predicts treatment benefit. Existing trial biospecimens could be used to ask whether secretory-high tumors derive more benefit from antiangiogenic agents such as sunitinib, surufatinib, or cabozantinib. The same samples could examine relationships with somatostatin receptor 2 expression and response to 177Lu-DOTATATE. Neuronal-like tumors may be better suited to myeloid-directed strategies if their microenvironmental signaling proves functionally relevant.
That is where this work becomes clinically consequential. The study does not yet prove that a cell-state score should guide therapy selection. It does show that the malignant compartment in gastroenteropancreatic neuroendocrine tumors is organized along a conserved neuronal-to-secretory axis, and that axis is linked to chromatin regulation, tissue architecture, fibrosis-associated signaling, and metastatic enrichment. For a disease still classified largely by anatomy and Ki67, that is a meaningful shift in how the biology can be read.
What remains unresolved is whether these programs will outperform existing grading in forecasting progression, desmoplasia, or drug response in prospective cohorts. That question will determine whether the axis becomes a research framework, a translational biomarker, or both.