Let me say one thing up front. What I am covering this time is a preprint, that is, a paper that has not yet gone through peer review. Because expert review has not been completed, the conclusions may change in the future. When it has been peer reviewed and formally published, I will cover it again in detail. For now, I will explain what has been made public at this point.
The subject is the dendritic cell (DC) inside tumors. Dendritic cells are the command centers of immunity: they show the cancer’s identifying marks (antigens) to T cells and tell them, “attack this.” Yet these command centers are not scattered evenly through a tumor. They gather in particular places, sit next to particular T cells, and form small “hubs.”
The paper asked three questions. What states do dendritic cells take, and are those states shared across cancer types and across species? Where in the tumor is each state found, and who are its neighbors? And when immunotherapy is given, how do those hubs change? The heart of the answer was somewhat unexpected. After treatment, the detectable changes in dendritic cell gene expression were minimal, and yet the arrangement of the cells had clearly changed.
Journal Information
- Paper title: Spatial immune hubs defined by conserved activated dendritic cells are remodeled by immunotherapy
- Type of paper: Preprint (an original research article before peer review). The header of the manuscript itself states that it “was not certified by peer review”
- Where it was posted: bioRxiv (posted February 11, 2026, version 1). 90 pages in total, 7 main figures, 11 supplemental figures, 83 references
- DOI / link: 10.64898/2026.02.10.703136
- Authors: Tomoyuki Minowa and Sunita Keshari (co-first authors with equal contribution) and others, 16 in total
- Affiliations: The University of Texas MD Anderson Cancer Center (Department of Immunology and others), Dana-Farber Cancer Institute, Ragon Institute, Broad Institute, Massachusetts Institute of Technology, Stanford University
- Impact Factor: None. bioRxiv is a preprint server, not a peer-reviewed journal, so the Impact Factor metric does not apply to it
- License: CC BY-NC-ND 4.0 (attribution, non-commercial, no derivatives). For this reason, this article neither reproduces nor redraws the paper’s figures, and explains everything in text alone
- Data and code: The paper states that the integrated human and mouse single-cell data and the processed CODEX data (AnnData files) are deposited in Zenodo (DOI: 10.5281/zenodo.18157069), and that the analysis code is on GitHub. However, when I checked for this article, the Zenodo record itself was public but access to the files was restricted
- Conflict-of-interest disclosures: Corresponding author Gubin is a paid consultant for Merck and receives an honorarium from Springer Nature as an associate editor of Nature Precision Oncology. Co-author Garry P. Nolan is a co-founder of Akoya Biosciences, an inventor on patent US9909167, and a member of the company’s scientific advisory board. Akoya is the company behind the instrument for CODEX (PhenoCycler), the core technology of this study. Linghua Wang is a member of the scientific advisory board of SELLAS Life Sciences
Who Are the Corresponding Authors?
There are two corresponding authors, both in the Department of Immunology at MD Anderson Cancer Center.
Matthew M. Gubin is an associate professor in the department. He received his PhD in microbiology and immunology from the University of Missouri–Columbia in 2012, and from 2012 to 2017 was a postdoctoral fellow with Robert Schreiber at Washington University in St. Louis. His research centers on T cell responses to neoantigens, the marks that arise from gene mutations in cancer cells; on comparing and combining neoantigen vaccines with immune checkpoint therapy; and on high-dimensional analysis of intratumoral immune cells. The mouse melanoma model used here, Y1.7LI, is a cell line the authors generated previously. The acknowledgments name Schreiber as someone who helped design the CODEX antibody panel.
Stephanie S. Watowich is a professor and deputy chair of the department. She is known for her research on STAT, the transcription factors activated by cytokines, and in particular on the role of STAT3 in myeloid cells and dendritic cells, and she is also developing vaccines that use type 1 conventional dendritic cells (cDC1). The overall picture becomes easier to grasp if you see this as a study in which an expert in dendritic cell biology teamed up with an expert in immunotherapy models.
First author Minowa is supported by overseas study fellowships from the Japanese Dermatological Association and the Uehara Memorial Foundation.
What Was It That We Did Not Know?
Dendritic cells are not a monolith. Conventional dendritic cells (conventional DC, cDC) fall into two broad groups. The usual summary is that cDC1 are good at showing antigens to CD8 T cells (the cells that directly destroy cancer cells), while cDC2 mainly deal with CD4 T cells (the cells that coordinate immunity as a whole). This is not a result of the paper; it is the textbook premise.
Inside tumors, however, a further layer called the “activation state” is superimposed on this. The paper focuses on two of them.
The first is the CCR7⁺ DC. These are dendritic cells that have matured and are ready to migrate; they express LAMP3, FSCN1, CD40 and CD83, and at the same time express immunosuppressive molecules such as PD-L1, PD-L2 and CD200. Different research groups have called them by different names: “DC3,” “mregDC,” and “LAMP3⁺ DC.”
The second is the ISG⁺ DC. These are dendritic cells that strongly express the set of genes induced by interferon (interferon-stimulated gene, ISG). In mice, they have been reported to strongly promote T cell-mediated antitumor immunity and to be required for an optimal response to immune checkpoint blockade.
👦 Student: Isn’t it confusing to have three or four names for the same thing?
🧬 Dr. Exotaro: It is. It is like the same station being called by a different name by each railway company. Unless you merge everything into a single map, you cannot even discuss whether “Company A’s station” and “Company B’s station” are in the same place. The first half of this paper is exactly that work of integration.
Three points were unresolved. Are these activation states conserved across cancer types, and also between humans and mice? Where in the tumor are they located? And how are they involved in the response to immune checkpoint therapy (ICT)?
The second point in particular faced a technical barrier. Single-cell RNA sequencing (scRNA-seq) is a method that breaks a tissue apart and reads the cells one by one. It tells you the type and state of each cell in detail, but the moment the tissue is dissociated, the information about “who it was next to” is lost.
👦 Student: Isn’t it enough to know the cell types?
🧬 Dr. Exotaro: That would be like talking about soccer with nothing but the player roster. You know who is there, but you do not know the formation. Whether the playmaker is in a position to pass to a given teammate cannot be seen without a diagram of where everyone stands. This paper is a study that tried to make both the roster and the formation diagram.
What Did This Paper Find?
The study builds up three methods in layers. Let us go through them in order.
Method ①: Pooling Single-Cell Data to Build the “Roster”
The authors first collected 28 publicly available scRNA-seq datasets of human solid tumors, covering 12 cancer types and normal tissues. So as not to miss any dendritic cells, they picked up not only the cells that the automated annotation tool ScType called myeloid dendritic cells, but also those it called macrophages, monocytes, and ISG-expressing immune cells, and built an integrated human dataset of 276,163 cells. From this they extracted only the clusters expressing dendritic cell markers (FLT3, ZBTB46, ITGAX), removed contaminating T cells and B cells, and integrated the data once more. Plasmacytoid dendritic cells (pDC) were excluded from this study.
Pooling data generated under different experimental conditions in different studies requires correction for batch effects. Here, RPCA (reciprocal PCA) in Seurat V5 was used. It is like overlaying maps made by separate survey teams by relying on shared landmarks. Because it would be a problem if the conclusions changed with the correction method, the authors confirmed that the heterogeneous cDC2 clusters were reproduced with three other methods: Harmony, FastMNN and scVI. The number of clusters was decided using an index called SC3 stability, which measures whether the way the cells split remains stable as the resolution is changed.
The result was an atlas of 29,554 cells from 353 samples and 12 tissues, of which 17,092 cells were tumor-infiltrating dendritic cells. Unsupervised clustering divided them into 8 clusters.
- C2_cDC1 (XCR1, CADM1, CLEC9A, BATF3)
- C4_CCR7⁺ DC (LAMP3, CCR7, FSCN1, CD40, CD274)
- 6 cDC2/DC2-like clusters (CD1C, CLEC10A, SIRPA). Of these, C0, C1, C6 and C7 are “inflammatory DC2” that strongly express IL1B, C3 is a non-activated state, and C5 is the ISG⁺ DC
In C5_ISG⁺ DC, the IFN-α and IFN-γ signaling pathways were enriched and JAK-STAT pathway activity was also elevated. The authors interpret this as an activation state of cDC2 that have been exposed to interferon.
Comparing across cancer types, the gene expression patterns of cDC1, ISG⁺ DC and CCR7⁺ DC were well conserved. The cDC2/DC2-like population, by contrast, was the most numerous and also the most heterogeneous (as assessed by the ROGUE score), and included cells whose character changed depending on the cancer type.
Sex differences were also examined. Sex information was available for 63.2% of patients (38.6% male, 61.4% female), and no sex-specific cluster was found. When cancers with a strongly skewed sex ratio, such as breast, uterine, ovarian and kidney cancer, were excluded, only a very small number of genes differed in cDC1, ISG⁺ DC and CCR7⁺ DC, and all of them were on the X or Y chromosome.
The same procedure was followed in mice. From 12 datasets combining published and unpublished data (101,161 cells at the time of integration), the authors obtained 6,592 tumor-infiltrating dendritic cells in 11 clusters. When the human gene signatures were converted to their mouse orthologs and applied, cDC1 and CCR7⁺ DC were well conserved across species, and ISG⁺ DC also showed a conserved pattern. The populations that earlier studies had called by separate names, LAMP3⁺ DC, mregDC and migratory DC, overlapped substantially in transcriptional profile with this study’s C4_CCR7⁺ DC.
Are the Two Activation States Connected?
When the CCR7⁺ DC were divided more finely, cells expressing CCR7 and ISG15 at the same time were found. There were two kinds: CCR7⁺ISG15hiCADM1⁺, which retain the cDC1 marker CADM1, and CCR7⁺ISG15hiCD1C⁺, which express the cDC2 marker CD1C.
A trajectory inference method called Monocle3 suggested that these may correspond to an early stage of the path along which cDC1 and cDC2 mature into CCR7⁺ DC. When origin was estimated with a method called SingleR, the most likely source of C5_ISG⁺ DC was the inflammatory DC2 of C0 and C7. In mice, too, a trajectory was drawn in which cDC2 pass through the ISG⁺ DC state on their way to CCR7⁺ DC.
The authors are cautious, however. ISG⁺ DC may be a waypoint on the way to CCR7⁺ DC, but given that they exist as a large population, some of them may remain within the tumor in the ISG state. They write that the data are compatible with either.
The Cytokine and Chemokine Genes They Express Differ
Comparing, in humans and mice, the genes that distinguish CCR7⁺ DC from ISG⁺ DC, many of the top differentially expressed genes moved in the same direction in both species (120 genes and 148 genes in the concordant quadrants, and 50 genes and 38 genes in the discordant quadrants).
Among them were genes for cytokines and chemokines (substances that attract cells). What follows is a comparison of RNA expression, not a measurement of the amount of protein secreted.
- C4_CCR7⁺ DC: IL15, IL32, CCL17, CCL19, CCL22
- C5_ISG⁺ DC: CXCL9, CXCL10
In addition, IL12B was expressed specifically in the CCR7⁺ISG15hiCADM1⁺ population, which appears to derive from cDC1, and IL27 was also highly expressed in the same population. There were further differences: CXCL9 was seen broadly in cDC1, CCR7⁺ DC and ISG⁺ DC, whereas CXCL10 was higher in ISG⁺ DC and in cDC2-derived CCR7⁺ DC.
👦 Student: If the substances they use to call cells are different, are the cells that gather also different?
🧬 Dr. Exotaro: That is the next question. If they are broadcasting on different frequencies, the audiences that gather should be different too. But from gene expression alone, all we can say is that they “seem to be calling.” To know who is actually next to them, we have to look on the tissue itself.
Method ②: Drawing the “Formation Diagram” with Spatial Transcriptomics
The data used here came from 3 cases of human colorectal cancer. The authors reanalyzed a previously published dataset in which both scRNA-seq and spatial transcriptomics (Visium HD) were obtained from serial sections of the same FFPE (formalin-fixed, paraffin-embedded) blocks.
Visium HD is a technology that reads, square by small square on a tissue section, “which genes are expressed at which location.” The analysis was performed with 8 µm bins. Because a single bin can contain mixed information from more than one cell, they used RCTD, a deconvolution method (separating mixed signals) that estimates the cell type of each bin using scRNA-seq from the same specimen as the reference. For the distance analysis they used bin2cell, which reassembles the 2 µm bins into single-cell units based on the nuclei in the H&E-stained image, and for inferring cell–cell communication they used CellChat.
Three things were found.
First, CCR7⁺ DC and ISG⁺ DC were densely clustered. This contrasts with cDC1 and cDC2, which were sparsely scattered. Moreover, the places where the two gathered were distinct, although they partly overlapped.
Second, each hub sat in a different signaling environment. CCR7⁺ DC were located in regions rich in CCL19 and CCL21, the ligands of CCR7. Cancer-associated fibroblasts and endothelial cells were found broadly throughout the tumor, but when proximal and distal zones were compared in the specimen from patient 2 using 50 µm as the boundary, the cells expressing CCL19 and CCL21 were skewed toward the vicinity of CCR7⁺ DC. ISG⁺ DC, on the other hand, were in regions where interferon signaling was active, and the surrounding cells also showed features of an interferon response. The CellChat analysis suggests that the source of interferon is not a single cell type but is spread across both immune and non-immune cells.
Third, the neighboring T cells were different. Both hubs are T cell-rich regions, but the cast differs.
- Around CCR7⁺ DC: a variety of CD4 T cells, including naive-like CD4 T cells, follicular helper T cells (Tfh), Th17 cells and regulatory T cells (Treg). The analysis pointed to the possibility that CCL22 from CCR7⁺ DC attracts them via CCR4
- Around ISG⁺ DC: CD8 T cells expressing exhaustion-associated transcripts (CD8.EX). The accumulation of CXCL9 and CXCL10 overlapped with this region, suggesting communication via CXCR3
Proliferating immune cells were abundant in both hubs.
Method ③: Watching the Changes Caused by Treatment with Multiplexed Imaging
Last come the data newly generated for this paper. CODEX (also called PhenoCycler) is a technology that stains dozens of proteins simultaneously on a single tissue section. Antibodies are tagged with DNA barcodes, and the corresponding fluorescent labels are repeatedly “attached and imaged, then stripped off so the next can be attached.”
The material is a mouse melanoma model called Y1.7LI. It is a cell line in which the BrafV600E Pten−/− Cdkn2a−/− YUMM1.7 line was made to express neoantigens for MHC class I and for class II, and it responds to anti-PD-1 antibody, anti-CTLA-4 antibody, their combination, and a neoantigen vaccine alike. This makes it possible to compare treatments with one another without the comparison being confounded by a difference between “works” and “does not work.”
The experimental conditions were as follows.
- 5 × 10⁵ cells were implanted subcutaneously into the flank of male C57BL/6J mice aged 8–12 weeks
- Anti-PD-1 antibody (clone RMP1-14) and anti-CTLA-4 antibody (clone 9D9) were given intraperitoneally at 200 μg on days 7, 10 and 13 after implantation
- The neoantigen vaccine (NeoVAX) consisted of 10 μg of the mLama4 synthetic long peptide and 50 μg of polyI:C, given subcutaneously on days 7 and 13
- There were two kinds of control: control antibody, and a control vaccine using an irrelevant peptide (mAlg8)
- Tumors were harvested on day 15, cut into frozen sections (8 µm thick), and imaged
6 treatment groups, 3 independent experiments, and 20 tumor sections in total. More than 1.4 million cells were analyzed with more than 30 markers.
The method of assigning cell types was also carefully devised. Because scRNA-seq data from the same model under the same treatment conditions already existed, the authors used an algorithm called MaxFuse to match the “RNA roster” to the “protein images” and transferred the roster’s labels to the cells on the imaging side. Only cells with a matching score of 0.4 or higher were labeled, and cells that fell short were left as “unassigned.” In this way, 10 major immune cell types and fine subtypes of dendritic cells and T cells were identified.
A permutation test was used to evaluate neighbor relationships. For each cell, the cells within a radius of 20 µm are counted; then, keeping the cell positions fixed, only the labels are shuffled 1,000 times to build a distribution of “how often they would be neighbors by chance,” and the ratio to the observed value is calculated.
First, in untreated (control antibody) tumors, the same picture as in human colorectal cancer was confirmed. ISG⁺ DC and CCR7⁺ DC occupied separate places while partly overlapping, and their neighboring T cells differed. CCR7⁺ DC were near Treg.
Then come the changes caused by treatment.
With anti-PD-1 antibody, the tendency of CCR7⁺ DC to gather with one another became markedly stronger. The CCR7⁺ DC aggregates defined by density-based clustering (DBSCAN) were larger than in controls, and were closer to effector/exhausted CD8 T cells (EffExh_CD8) and proliferating CD8 T cells. ISG⁺ DC showed a similar tendency, with increased proximity to EffExh_CD8 and effector CD4 T cells (Eff_CD4).
This is the crux. Even when the scRNA-seq of the same model, and the integrated human and mouse data, were compared across treatment conditions, the changes in dendritic cell composition and transcriptional state were very slight. Within the detectable range, the transcriptional states of the cells hardly changed, and yet their arrangement had changed. It is a change you would miss if you looked only at the roster.
With the neoantigen vaccine, the tendency of cDC2 and ISG⁺ DC to sit next to each other appeared even more strongly than under any of the ICT conditions. This zone of adjacency lies in places rich in EffExh_CD8. The authors raise the possibility that IFN-γ produced by neoantigen-reactive CD8 T cells is pushing the cDC2 on the spot into the ISG⁺ DC state.
The combination of anti-CTLA-4 and anti-PD-1 antibodies did not turn out to be the sum of the “promising-looking features” of each single agent.
- The enlargement of CCR7⁺ DC aggregates seen with anti-PD-1 alone was not seen with the combination
- Anti-CTLA-4 alone reduced Treg both overall and around dendritic cells, but with the combination there were more Treg around both kinds of dendritic cells than with anti-CTLA-4 alone, most notably around CCR7⁺ DC
- On the other hand, Eff_CD4 (a population that includes Th1 cells) did not increase near dendritic cells with anti-CTLA-4 alone, but did increase near CCR7⁺ DC with the combination
On the first point, the authors draw on a recent report that niches formed by Treg around lymphatic vessels impede the migration of CCR7⁺ DC to the lymph nodes, and speculate that anti-CTLA-4 antibody may lift that constraint and send CCR7⁺ DC out to the draining lymph nodes, making large aggregates less likely to remain within the tumor. They state explicitly, however, that what this remodeling means for antitumor immunity “remains to be determined.”
How Will the Future Change? (The Road to the Clinic)
I need to say first that this study is not at a stage where it connects directly to the clinic. The human data are a reanalysis of published work, and the treatment-induced changes were examined in a single mouse model. With that said, there are three implications at the level of how we think.
The first is the viewpoint of reading how immunotherapy works through “arrangement.” When counting the types and amounts of cells in a tumor before and after treatment shows no difference, looking at the arrangement may reveal one. There is room for multiplexed staining of pathology specimens and spatial analysis to be added to response prediction in the future.
The second is the possibility that the two hubs have separate roles. Multiple CD4 T cell populations and Treg were abundant around CCR7⁺ DC, and CD8 T cells carrying exhaustion-associated transcripts were abundant around ISG⁺ DC. If each dendritic cell state is in fact attracting them, the design of a treatment could change depending on which hub one strengthens. Earlier work reported that ISG⁺ DC promoted CD8 T cell-mediated antitumor immunity even in Batf3−/− mice, which lack cDC1, and the authors attach importance to this point.
The third is a caution that “combining treatments” does not necessarily amount to addition. The observation that there were more Treg around dendritic cells in the combination group than in the anti-CTLA-4 single-agent group is a point of concern when thinking about combination therapy. But this is an observation in one mouse model at one time point, and its clinical meaning is unknown.
👦 Student: So it does not mean that treatment for patients will change right away.
🧬 Dr. Exotaro: No, it will not. This is a study that drew a map. A map is not proof that you have reached the destination. But without a good map, you cannot even decide where to dig next. The authors, too, position this atlas as a “resource” for future research.
How to Read This Study Critically (Limitations, and Routes to Higher Quality)
First, the strengths. Integration on the scale of 28 human and 12 mouse datasets; confirmation of robustness with different correction methods; a design that matched RNA and images in the same model under the same treatment conditions; and a policy of releasing data and code (although, as noted above, access to the Zenodo files was restricted when I checked). All of these are careful for a study of this kind. The authors themselves list three limitations: that much of the analysis is based on transcripts, that the trajectory analysis is a computational prediction and not lineage tracing, and that the scale of the human spatial analysis is small.
With that in mind, here are the points to watch when reading this as a paper that has not been peer reviewed.
① The human spatial data come from 3 cases, and from colorectal cancer only. The atlas covers 12 cancer types, but the human data supporting the spatial conclusion that “the two hubs are skewed to different places while partly overlapping” are 3 cases of colorectal cancer. Whether the arrangement is the same in other cancer types is not yet known.
② Causality has not been shown. What was observed are associations: “they are neighbors” and “they increase together.” There are no experiments that selectively remove CCR7⁺ DC or ISG⁺ DC, or block the CCL22–CCR4 or CXCL9/10–CXCR3 pathways, to check whether the hubs fall apart or the therapeutic effect declines. The ligand–receptor interactions are inferences made by CellChat.
③ The newly performed imaging experiments are narrow in scope. The mouse single-cell atlas integrates 12 datasets, but the CODEX experiments that examined treatment-induced changes in arrangement used male mice only, a single tumor model, and a single harvest time point on day 15. Precisely because the authors went to the trouble of examining sex differences in the human data, the fact that the mouse side is male only stands out.
④ The unit of statistics. With 20 sections across 6 groups, there are only a few tumor sections per group. In the figure comparing the size of CCR7⁺ DC aggregates, each dot represents one “aggregate,” and the comparison is tested by one-way analysis of variance with Tukey’s multiple comparisons. Aggregates taken from the same tumor are not independent of one another, so using the aggregate as the unit risks making the sample size look larger than it really is and making differences easier to find.
⑤ How to read “the changes were minimal.” The conclusion that the changes in transcriptional state are small applies within the range detectable by scRNA-seq. The abstract, too, uses the wording that the “detectable changes” were minimal. Whether there were enough cells and enough statistical power to pick up small changes cannot be judged from what is written in the paper alone.
⑥ The breadth of the “exhaustion” label. Human CD8.EX is defined by the expression of 5 markers: TOX, PDCD1, LAG3, TIGIT and HAVCR2. The authors themselves note that activated cells and cells that are still functional may be mixed in.
⑦ Conflicts of interest. For the core technology, CODEX, it is disclosed that the co-authors include a co-founder of the company that developed it. It has been disclosed, and it serves as background information for readers when they evaluate the results.
As a minor point, there is a place where the main text refers to “Supplemental Figure 17A,” but there are only 11 supplemental figures. It is presumably the kind of error that will be corrected during peer review.
What would make this an even higher-quality study? Specifically, the following.
- Extend the human spatial analysis to multiple cancer types and to paired specimens from before and after immunotherapy
- Align the unit of statistics with the tumor (the individual mouse), or use a mixed-effects model with tumor as a random effect
- Reproduce the findings in female mice and in at least one more tumor model
- Confirm causality with experiments that selectively deplete particular dendritic cell states and by blocking chemokine pathways
- Back up the computational trajectory inference with genetic lineage tracing
- Harvest at multiple time points and follow when the hubs form and when they break down
- Link the size and arrangement of the hubs, animal by animal, to tumor shrinkage and survival
Dr. Exotaro’s Perspective
To be honest, this paper is some distance from my own field of research. What I work on is the treatment of neurological diseases, beginning with spinal cord injury, using mesenchymal stem cells (MSC) and their extracellular vesicles (EV). Neither MSCs nor EVs appear in this paper, and cancer immunology is not my specialty. I have never run CODEX myself, either. So when it comes to the fine details of cancer immunology, I am reading from the same side as my readers.
I chose to cover it nonetheless because the claim at the center of this study resonates across fields. Even when the detectable changes in gene expression are small, the arrangement of cells can change. What that change in arrangement means functionally has not yet been shown, even in this paper. Even so, I think the point carries weight: there are changes that cannot be seen if you measure only the “contents.”
A similar question arises at sites of neural injury. We, too, often examine what the administered cells and EVs have delivered. But where macrophages and microglia gather at the injury site, who they sit next to, and how treatment changes that arrangement has long remained hard to see. Methods that dissociate the tissue and count cell types will, in principle, miss the kind of change this paper showed. I believe that technologies which read many molecules while preserving space will become indispensable from here on for investigating the mechanisms behind the effects of regenerative medicine as well. This, however, is my own reading and not something the paper showed.
One more thing I want to learn from this paper is how it is written. On whether ISG⁺ DC are a waypoint or a stable state, the authors wrote that the data are “compatible with either.” On why the enlargement of the hubs disappeared under combination therapy, too, they stated a hypothesis and then concluded that the functional significance “remains to be determined.” Faced with large-scale data, one is tempted to tie everything into a single neat story. That they held back at that point is where I place my trust.
To repeat, this is a paper that has not yet been peer reviewed. Analyses may be added during review, and the wording of the conclusions may change. Once it is formally published, I will cover it again in detail, including what changed and what remained.
Note: this article was written with the assistance of AI.