Density, not Triad Geometry, drives Allograft Rejection Signature in Reference Cervical Cancer Atera Slide

Summary

Figure 1. Whole-section combined cell-type and malignancy mask, 717,576 cells. A broad immune-infiltrated band (teal-blue) traverses the malignant epithelial lobules (red). Legend (right): top, malignancy scale on non-immune cells; bottom, top immune cell-type categories ordered by count.

Immune triads — a CD8⁺ cytotoxic T cell, a CD4⁺ helper T cell, and a dendritic cell co-localized within 20 µm — have been proposed as the in situ spatial signature of productive T-cell licensing, the antigen-presentation handshake that activates CD8 cells for effector function (Espinosa-Carrasco et al., Cancer Cell 2024). To test whether this geometric motif marks productive licensing in HPV-negative cervical carcinoma, we analysed a single 10x Atera WTA FFPE section (717,576 cells, 18,028 gene symbols). The section reproduced the predicted geometry: 406 triad-anchor CD8 cells (23.8% of CD8) concentrated along the invasive margin, and the in-triad versus out-of-triad transcriptional contrast enriched the canonical T-cell-activation programmes — allograft rejection, interferon-γ response, IL-6 / JAK / STAT3, and a CXCL9→CXCR3 axis matching dendritic-cell chemokines to CD8 receptors — consistent with the licensing readout reported in recent triad-adopting cohorts.

The lead readout we use throughout is the HALLMARK_ALLOGRAFT_REJECTION composite — a 200-gene MSigDB Hallmark set that has been cross-cohort validated (n = 6,245 breast cancer patients; Oshi et al., npj Breast Cancer 2022) as the strongest single-pathway proxy for productive anti-tumour immune activity, correlating with cytolytic activity, CD8⁺ infiltration, and improved survival in triple-negative disease. On this section it is elevated both at triad anchors and in dense immune neighbourhoods, alongside the interferon-γ response and inflammatory response programmes — the expected qualitative pattern that motivates the density-confounding tests in §3.

Composite activation scores in spatial transcriptomics rise with local cell density irrespective of mechanism, a confound recurrent in spatial-statistics analyses. To distinguish a triad-specific licensing signal from this density floor, we applied two controls. First, a “would-be-triad” comparator (CD8 cells with a CD4 neighbour but no dendritic-cell neighbour within 20 µm) — under the licensing hypothesis these should not score like triads, because the dendritic cell is the licensing source. Contrary to this prediction, they scored indistinguishably from genuine triads (Cliff’s δ = 0.09, a rank-based effect size where |δ| < 0.15 is negligible). Second, multivariable regression of each per-cell score against local cell, immune, and malignant density, and re-testing the contrast in the residuals. After this adjustment, the apparent dendritic-cell-side activation on HALLMARK_ALLOGRAFT_REJECTION collapsed from Cliff’s δ = +0.410 (n = 680 vs 26,908) to β = +0.004, p = 0.18, indicating that the unadjusted contrast reports neighbourhood density rather than triad geometry.

The companion tertiary-lymphoid-structure analysis identified a candidate dendritic-cell-side mechanism. Of 138 B/T-cell aggregates on the section, 11.6% met mature-TLS criteria. The LAMP3⁺ migratory-dendritic-cell compartment — LAMP3 marks the mature, antigen-presenting state of these cells — partitioned into a mature pole in the stromal interior and an arrested pole at the tumour border (median 7.6 vs 17.1 µm to the nearest malignant cell, about half the distance) with reduced CCR7 / CD83 maturation markers, an interferon-stimulated-gene-skewed transcriptome, density-independent reduction of the costimulatory receptor CD40 (β = −0.86, p = 1.5 × 10⁻⁵), and collapsed non-canonical NF-κB target output — the transcriptional programme that LTβR engagement normally drives during DC maturation (RELB / NFKB2 / BIRC3 / BAFF target score β = −0.87, p = 3.3 × 10⁻¹⁸) — despite intact upstream LTβR expression, suggesting that the receptor is present but downstream transduction is broken. The defect tracked proximity to malignant cells (β = −0.020, p = 1 × 10⁻⁶) rather than a measurable suppressive-cytokine field (β = −0.0008, p = 0.44), consistent with a contact- or short-range factor at the tumour interface as the proximate cause.

Together these findings suggest a testable therapeutic stratification for HPV-negative cervical tumours that present triad geometry without dendritic-cell maturation: a CD40 agonist to bypass the downregulated receptor, paired with FLT3L to expand the conventional dendritic-cell compartment that the HPV-negative setting is depleted in. Because all findings derive from a single FFPE section, multi-patient validation with matched density strata and protein-level CD40 / CD40L measurement would be required before this stratification can be evaluated.

The transferable methodological observation is that composite activation scores in spatial-omics data report local neighbourhood density by default; multivariable density adjustment, dendritic-cell-subtype resolution, and a density-matched negative control are required before any cell-cell-signalling claim.

2. Productive-triad signature, baseline analysis

A triad anchor is a CD8⁺ T cell with at least one CD4⁺ T cell and one dendritic cell within 20 µm — the spatial signature predicted to mark sites of productive licensing (Espinosa-Carrasco et al., Cancer Cell 2024). With cell identity resolved by Cell-Ontology descent, 406 of 1,703 CD8 cells (23.8%) satisfied the triad criterion (275 retained after 80 µm greedy thinning for display). Triads were enriched for the licensing-competent dendritic-cell subtypes: conventional dendritic cells (cDCs) rose from 26% to 42% of the dendritic-cell pool inside triads, and plasmacytoid dendritic cells from 8.5% to 19.5%. Cytotoxic and memory CD8 cells were the most triad-engaged CD8 subsets, consistent with their role as the licensing-competent effector pool. The triads themselves concentrated along the invasive margin between malignant epithelial lobules and immune-infiltrated stroma.

Figure 2. 1,200 × 1,200 µm interior crop centred at (6,941, 2,734) µm. CD8 yellow, CD4 cyan, dendritic cell green, NK cell blue, macrophage tan, malignant (p ≥ 0.5) red. Magenta circles, 306 thinned triad anchors within the window (461 unthinned); orange circles, 77 tetrads (triads with an adjacent macrophage).

To test whether the triads carry the expected transcriptional signature of productive licensing, we contrasted in-triad versus out-of-triad cells on each of the three triad lineages by whole-transcriptome differential expression, followed by gene-set enrichment against the MSigDB Hallmark and Reactome collections of co-regulated programmes. The CD8 cells inside triads showed coordinated enrichment of the classical T-cell-activation programmes — allograft rejection, interferon-γ response, IL-6 / JAK / STAT3, and inflammatory response — together with an enriched chemokine-receptor axis whose receptors match the chemokines upregulated on the dendritic-cell side (most notably the CXCL9→CXCR3 axis that draws effector CD8 cells toward dendritic-cell-presented antigen). The CD4 cells showed recruitment and costimulation programmes; the dendritic cells showed a mature-migratory phenotype. This three-lineage pattern reproduces the licensing readout reported in recent triad-adopting cohorts (Damle 2026; Lopez-Janeiro 2026; Oh 2025), and on its own would support the licensing-handshake interpretation. Full enrichment statistics, the ligand-receptor concordance table, and per-lineage volcano plots are in the technical report §1.3.

Figure 3. Whole-section spatial distribution of triad anchors. Malignant cells (p ≥ 0.5) muted red; CD8 lineage yellow, CD4 lineage cyan-blue, dendritic-cell lineage green; magenta circles, 275 thinned triad-anchor CD8 cells.

4. Productive immunity in immune-dense regions and high-TLS aggregates

The §3 controls indicate that the composite Hallmark signal at current 20 µm triad anchors is fully explained by neighbourhood density. They do not, however, address a distinct question: whether productive immunity has occurred somewhere and the licensed effector cells are now present on the section, regardless of where the licensing event itself took place. To address this, we asked whether cell-intrinsic markers of effector function and proliferation — gene products expressed by an individual CD8 cell only after antigen engagement, and therefore independent of who its current neighbours are — are spatially structured.

Two readings are compatible with the §3 result. (A) Null reading: no productive licensing has occurred anywhere relevant to this section, and the composite-Hallmark elevation is a density artefact end to end. (B) Active-immunity reading: productive licensing has occurred and the licensed CD8⁺ T cells are present on the section, concentrated in the dense immune compartment (Figure 5). Within (B), the location of the licensing event admits two sub-readings that a single FFPE section cannot distinguish: (B-intratumoral) licensing happened on this section, at past triads or now-dissolved seeding events, and the productive niche is now self-sustained; (B-LN) licensing happened in the tumour-draining lymph node and the licensed CD8s circulated back and homed to the dense compartment of the tumour. Under (A) the cell-intrinsic markers should be flat across the section; under (B) they should be elevated in immune-dense regions and inside high-TLS aggregates, where the licensed CD8s would be concentrated (Figure 6).

Figure 5. Conceptual model for interpretation B-intratumoral. An infrequent CD8 / CD4 / DC interaction seeds local activation; cytokine and chemokine amplification recruit further immune cells; the niche matures into a TLS-like structure that maintains antigen presentation and T-cell help independent of any single current triad. The composite Hallmark readout around current triads therefore reports the density of this self-sustained environment rather than the presence of an instantaneously productive triad. Under interpretation B-LN, the same end-state is reached by lymph-node licensing followed by tumour homing — the two routes produce the same per-cell pattern on a single FFPE section.

We used two conventional marker sets as proxies for licensed-and-active CD8 cells: proliferation markers (MKI67, TOP2A, PCNA, TYMS, MCM6) for cells in cell cycle — the strongest single signal of antigen-engaged clonal expansion — and effector-cytotoxicity markers (IFNG, GZMA, GZMB, PRF1, NKG7) for cells actively secreting effector molecules. Per-cell binary positivity is used in place of Tirosh-style composite scores because composite scores on sparsely detected functional markers are dominated by their density-tracking random-control baseline (technical report §2.5).

Figure 6. Per-cell positivity rate of CD8⁺ T cells (n = 1,847) for proliferation markers (any of MKI67, TOP2A, PCNA, TYMS, MCM6) and effector-cytotoxicity markers (any of IFNG, GZMA, GZMB, PRF1, NKG7). Top row: rate by local immune-density quartile (Q1 sparse → Q4 dense). Bottom row: rate by aggregate context (high-TLS = top quartile of aggregates). Bars show Wilson 95% confidence intervals. The proliferating-CD8 rate is about 2.5× higher in dense than sparse regions and about 1.5× higher inside high-TLS aggregates than outside; the effector-positive rate shows a weaker density gradient and no aggregate-context effect.

About one in eight CD8 cells on the section was in cell cycle, and about one in four was effector-positive — both substantial fractions for an HPV-negative tumour. Crucially, these populations were spatially structured rather than uniform: the proliferating-CD8 rate was about 2.5× higher in dense immune neighbourhoods than in sparse ones, and about 1.5× higher inside high-TLS aggregates than outside, consistent with active clonal expansion of licensed CD8s in those niches. The effector-positive rate showed a weaker density gradient and no aggregate-context effect.

Together this pattern rules out the strict null reading (A): CD8 cells in cell cycle have been licensed somewhere, and they are not distributed uniformly. The pattern does not, however, locate the licensing event — the same per-cell enrichment is consistent with both B-intratumoral and B-LN, and a single FFPE section without TCR clonality data, lineage tracing, or paired tumour / lymph-node biopsies cannot distinguish the two sub-readings. Several section-level observations weakly favour B-intratumoral — the in-situ mature-TLS aggregates and the licensing-competent mature LAMP3⁺ dendritic-cell pole inside the tumour (§5), plus a directionally consistent proliferation signal in the original triad-versus-would-be-triad CD8 contrast — but no one observation is decisive (technical report §2.5).

6. Discussion

Triad geometry as a spatial landmark, not a biomarker of productive immunity

Geometry is necessary but not sufficient: triads form without producing the licensing handshake. The closest published comparator (Syding 2025; Xenium on cervical and HPV⁺ HNSCC) reached the same conclusion from the suppressive-cell side: TLS quantity did not predict survival, but unfavourable TLS composition (high Treg / PMN-MDSC, high Arg1, reduced TCRζ) did. Cell-cell-signalling claims from spatial-omics data should be reported with multivariable density adjustment, dendritic-cell-subtype resolution, and a density-matched negative control.

Therapeutic implications

TLS-based stratification has not been tested as an immune-checkpoint-blockade predictor in cervical cancer (no lymphoid-aggregate endpoint in KEYNOTE-A18 or innovaTV 301), and TLS-induction therapy in invasive cervical disease is an open area — published TLS-induction work to date is in premalignant CIN2/3 via therapeutic HPV vaccination. The maturation-arrested LAMP3⁺ dendritic-cell phenotype identified here suggests a stratification that addresses each layer of the defect: a CD40 agonist to bypass the downregulated costimulatory receptor, paired with FLT3L to expand the cDC compartment that the HPV-negative setting is depleted in by background. The rationale is strongest under §4 interpretation B-intratumoral, where the licensing bottleneck is in the tumour and the arrested LAMP3⁺ pole is the proximate cause; it is weaker under B-LN, where the bottleneck is in the lymph node and the arrested-DC pole is downstream. Distinguishing B-intratumoral from B-LN with paired tumour / lymph-node TCR clonality and lineage tracing is a prerequisite for translation of this stratification.

Clinical context for the allograft-rejection readout

The composite used as the density-control proxy in §3 has been cross-cohort validated as the strongest single-pathway Hallmark proxy for anti-tumour immune activity in breast cancer (Oshi et al., npj Breast Cancer 2022; n = 6,245 across METABRIC, GSE96058, TCGA, GSE75688). It tracks cytolytic activity and broad immune-cell infiltration, but also tracks aggressive cancer biology and pro-cancer infiltration (Tregs, M2 macrophages). Like the spatial readout here, the bulk-cohort version of this composite reports immune-compartment engagement, not productivity-specific anti-tumour immunity.

The prognostic direction is subtype-conditional (Figure 10). In TNBC, a high allograft-rejection score is associated with significantly better disease-free, disease-specific, and overall survival under multivariate Cox adjustment. In ER+/HER2− breast cancer the same gene-set enrichment yields no survival separation despite the same enrichment magnitude, which Oshi et al. interpret as TNBC reaching a sufficient absolute level of productive infiltration to outweigh aggressive biology while ER+/HER2− does not — the cohort-level analog of the case study’s finding that productive immunity must occur within the engaged compartment for outcome to follow.

Figure 10. Reproduced from Oshi et al., npj Breast Cancer 8:92 (2022), Fig 6. Kaplan-Meier curves of disease-free, disease-specific, and overall survival by allograft-rejection score (high red, low blue) in METABRIC and GSE96058, stratified into whole cohort, ER+/HER2−, and TNBC. The high-score group has significantly better survival in TNBC but no separation in ER+/HER2− — consistent with the model that productive infiltration must exceed a subtype-specific threshold for outcome benefit. © 2022 the Authors. Open access under CC-BY; doi:10.1038/s41523-022-00466-2.

HPV-negative cervical SCC has not been analysed in this framework. The closest cervical-specific spatial comparator (Syding 2025, Xenium on cervical and HPV⁺ HNSCC) independently concludes that TLS composition — not quantity — determines outcome. The case study’s per-cell-positivity refinement (§4) is the spatial-resolution analog of the same principle: bulk-cohort or composite-Hallmark scoring cannot separate density-driven engagement from per-cell productive immunity. Multivariable density adjustment for cell-cell-signalling claims and per-cell positivity for sparse functional markers would resolve that ambiguity prospectively in spatial-transcriptomic cohorts. Full cross-cohort correlations, hazard ratios, and discussion are in Oshi et al. and in the technical report §4.1.

Limitations

Our study has several limitations. First, all findings derive from a single FFPE section, so the negative result on current 20 µm triads is necessarily narrow and section-specific; it reconciles with the published positive cohorts under the hypothesis that productive triads occur within mature TLS or cDC1-rich niches, a context the HPV-negative cervical section lacks. Second, the §4 active-immunity reading rules out a strict null but does not locate the licensing event — the proliferating-CD8 enrichment pattern is consistent with both intratumoral past-triad licensing and lymph-node licensing followed by selective tumour homing, and a single FFPE section cannot distinguish the two without paired TCR clonality, lineage tracing, or temporal data. The transferable element of this case study is therefore the methodological recommendation, not the section-specific negative outcome.

Methods and source artifacts

Single 10x Atera WTA Preview FFPE cervical squamous cell carcinoma section; ~18,000 gene symbols; processed with miratyper --panel all --run-malignant, tissue prior UBERON:0000995 α 1/10; MiraTyper proprietary gene-embedding model, 98.7% panel coverage. All effect sizes are Cliff’s δ (rank-biserial); |δ| < 0.15 negligible, 0.15–0.33 small, 0.33–0.47 medium, > 0.47 large.