The Garrison Signature

A patient with colorectal cancer is given a checkpoint inhibitor. The drug, anti-PD-1, takes a brake off the patient's own T cells so they can attack the tumour, and in melanoma and a handful of other cancers it has changed what the disease means. In this patient's cancer, mismatch-repair-proficient colorectal, fewer than one in twenty respond. The pathology read on the tumour is cold, the immune system absent.

Non-response is the common case. Close to half of cancer patients in the US are now eligible for one of these drugs, and by one estimate about an eighth of patients respond (Haslam and Prasad, 2019). Most people who are treated do not benefit, and the field has no structural account of who they are or why.

The cells are in the next room

So look at the immune system in these tumours directly. In the Pelka colorectal single-cell dataset there are 23,486 CD8 T cells to read, and the cells are there. The build is complete. They carry effector machinery and receptors that recognise the tumour. The reserve marker IL7R is held. The stem marker TCF7 is held. The exhaustion markers PD-1 and TOX are low. And CXCL13, the signal a CD8 cell fires only once it has reached the tumour and engaged a target, is absent.

That last absence places them in time. These are finished effector cells caught before engagement, sitting in the tumour-draining lymph node. A built T cell does not leave the node on its own. It waits for a second instruction, a licensing signal handed to it by a dendritic cell, before it is released to travel to the tumour. In these patients that signal does not arrive, and the army stays in the barracks.

The count inside the tumour was low, and a low count was read as no response. The response had already happened, one step back, in the draining lymph node where the biopsy was not looking.

The gate the diagram drew as a line

There is a reason this state has no name in the clinic. The standard map of the anti-tumour response, the Chen and Mellman cancer-immunity cycle, draws a single arrow from priming in the lymph node to trafficking into the tumour. We rebuilt that one arrow into the steps it compresses, the sixteen-position cycle set out in the cancer-immunity cycle we completed. Four positions sit inside that arrow: clonal expansion, lineage commitment, the licensing test against the dendritic cell, and the selection of which built cells are released to ship out. The last is a gate. The diagram drew it as a line.

The positions numbered in that figure are coordinates on a cycle that did not begin in oncology. It comes from Persistence Dynamics (van der Klein, 2026), a framework on Encounter for systems that persist by running. Two operations are enough to describe such a system. It holds what it has built, and it crosses to engage what is beyond it. Each operation is either active or idle, which gives four phases, and the phases run in a fixed order: gather, build, engage, conserve. The order is forced, because each phase produces what the next one needs, and because only one operation changes at each step.

A T cell mounting a response is one of these systems. Its build phase makes the effector cells. Its engaged phase is the killing at the synapse. Between them sits a forced crossing, the moment a built cell is either released to ship out or held back. That crossing is the gate, and it was in the framework before any colorectal data was opened, which is why the phenotype could be written down first and tested second. The Egress barrier is that crossing in immune terms, and a Garrison is what a patient looks like when the response is stranded there.

Past the gate

Anti-PD-1 acts at position 10, the synapse, two positions past the gate. It releases a brake on a CD8 cell that has reached the tumour, formed a contact with a cancer cell, and had its signalling shut down through PD-L1. That is the entire mechanism. It has nothing to act on in a cell that has not reached the synapse. The parked cells are upstream of where the drug works, behind the gate, still in the node. The under-five-percent response in this group is that fact read at population scale.

The numbers

The phenotype is measurable, and it was written down before the data was opened. In the Pelka colorectal set it separates the checkpoint-failing mismatch-repair-proficient tumours from the checkpoint-responding mismatch-repair-deficient ones: 32 percent against 9, odds ratio 4.64, p = 0.0028. Take the whole built-and-stalled region, positions 7 and 8 together, and it widens to 59 percent against 15, odds ratio 8.25. The same definition, applied without a single change to the Bassez breast cohort, marks 12 of 20 non-responders and 0 of 9 responders, p = 0.0024. The framework predicted the phenotype should be rare in cancers that reach the engaged phase by default, and in pretreatment melanoma it is absent, 0 of 11. One definition, three cancers, three datasets.

A natural experiment sits inside the colorectal numbers. Mismatch-repair-proficient colorectal, the group that piles at the parked position, responds to anti-PD-1 at under five percent. Mismatch-repair-deficient colorectal, where the cells have shipped and reached the synapse, responds near fifty. That ten-fold gap is the distance between an army parked and an army deployed, sitting in the clinical record already.

Garrison

The state needs a name, because a phenotype described in a sentence gets read once and lost. A garrison is a force that is built, armed, and stationed in place, held back from the front. That is the cell state, and the fingerprint that identifies it is the Garrison signature.

It comes in two forms. The first is the single-cell signature locked above, exact and already validated on public data. The second is a bulk-deconvolution and antibody-stain version that reads the same markers in ordinary tissue, the kind sitting in hospital and trial freezers. The second form is what makes the signature usable on samples that already exist, and it is the part that still needs building and checking.

Already in the literature

The pieces of this position are in the literature already, each described on its own. The tumour-draining lymph node holds a stable reservoir of stem-like TCF1+ CD8 cells that migrate into the tumour and only there differentiate and exhaust (Connolly et al., 2021). Conventional type-1 dendritic cells maintain that reservoir in the node (Schenkel et al., 2021), and they are the licensing partner. The draining-node cells turn out to be the true responders to PD-1 blockade (Huang et al., 2022), and the checkpoint acts in the node as well as in the tumour (Dammeijer et al., 2020). Reliance on these stem-like cells for response is already known to vary with the cancer (Escobar et al., 2023), which is the cross-cancer pattern above seen from another direction.

What the field has not done is treat the egress step as a place a patient is stuck. In those papers the reservoir is a phenotype and a niche. Here it is a position, with a fingerprint that names who is parked there and a lever that follows from where they are parked.

Releasing it

If the missing event is the licensing signal, the intervention is to supply it. The most direct lever is CD40 agonism, which hands the cell the dendritic-cell help it is waiting for. CD40 agonist antibodies are already in clinical use, sotigalimab and selicrelumab among them, with signals that have been real and mixed. The framework reads the mixed part as a selection problem. The agonist should help the Garrison-positive patient and do little for the rest, and a trial that does not stratify averages the two together into a muted result.

Sequence is part of the mechanism. Given before the checkpoint, with roughly the cycle's own traversal time of about two weeks between them, the agonist ships the army to the synapse first, and the checkpoint then arrives to find cells it can release. Given together, the checkpoint arrives at an empty front. Other levers act at the same position when the dendritic cell itself is unfit: STING agonists, type-I interferon and IL-12 to mature it, IL-2 variants aimed at the stem-like compartment, adoptive transfer of the lymph-node stem-like cells. Cell therapy skips the position altogether with an effector built outside the patient, which is consistent with cell therapy working in settings where checkpoint blockade does not.

The test that needs no new patient

CD40 agonists have already been given with checkpoint inhibitors, selicrelumab with atezolizumab, sotigalimab with nivolumab. The baseline biopsies from those trials exist. Score them for the Garrison signature and ask a single question: did the benefit of adding the agonist sit with the Garrison-positive patients? A mixed trial result turns into a stratified subgroup analysis once the signature is in hand, with no new patient enrolled and no new drug given. The prospective single arm, Garrison-positive patients given the agonist before the checkpoint, is the step that re-analysis earns.

Limits

Two limits, said plainly. In a tumour biopsy the fingerprint cannot yet separate the built-and-parked cell from a cell that failed the licensing test one position earlier. Both carry TCF7, and the marker that would split them sits below the resolution of these cohorts. A clean separation needs lymph-node tissue. And the signature is locked at single-cell resolution, while most stored samples are bulk or fixed tissue, so the retrospective test waits on the bulk and stain version being built and validated.

The cells in the draining lymph node have been described for close to a decade, mapped as a stem-like reservoir, named in paper after paper, catalogued as a kind of cell and never as a place a patient is stuck.

---

The full paper, with the signature definition, statistics, and references: The Garrison Signature (van der Klein, 2026), 10.5281/zenodo.20479100.