Cancer immunotherapy works only if killer T cells can physically reach tumor cells. In many tumors they cannot. Dense scar-like tissue (stroma) made by fibroblasts and stellate cells walls the tumor off, leaving T cells piled up at its border. Today this "exclusion" is usually measured by counting immune cells in two hand-drawn compartments, tumor and stroma. Wireless engineers face a problem with the same geometry. They need to know which locations a set of transmitters can reach when buildings and terrain block the signal, and they summarize the answer as a coverage manifold: a map of signal quality over space whose holes are dead zones. We propose to treat effector immune cells as transmitters, tumor cells as receivers, and stromal barrier cells as buildings. We then compute a blockage-aware coverage map of each tissue section and ask whether the fraction of tumor lying in "immune dead zones" predicts survival and immunotherapy response beyond what simple compartment counts already give.
The new ingredient is a density-normalized coverage deficit. It compares the coverage actually observed against the coverage expected if the same number of immune cells were scattered by a clustered, spacing-constrained random process of the kind telecom uses for drone swarms. This separates two situations that counts blur together: a tumor that is cold because it has few T cells (an "immune desert"), and a tumor that has enough T cells but where the geometry of its stroma keeps them out ("immune excluded"). If the deficit adds no prognostic value over intratumoral density and nearest-distance features, the idea is wrong, and the experiments are built to show that quickly.
Core Claim
Our claim is that the spatial arrangement of barrier cells between immune cells and tumor cells carries clinical information that is not captured by how many immune cells sit inside versus outside the tumor, nor by how far they are from tumor cells. The motivating evidence is that location, not abundance, matters. In bladder cancer, Microlocalization and clinical significance of stabilin-1+ macrophages found that stromal macrophage density (20 ± 17 cells/field) had no survival association, while the far sparser intratumoral population (4 ± 8 cells/field) was an independent hazard (HR 2.371). In pancreatic cancer, Integrating multi-modal transcriptomics identifies cellular subtypes with distinct roles in PDAC progression mapped a tumor-promoting niche of POSTN⁺ fibroblasts and SPP1⁺ macrophages sitting at tumor–stroma boundaries, opposed by an anti-tumor niche of CCL4⁺ CD8⁺ T cells and plasma cells. PLXDC1+ Tumor-Associated Pancreatic Stellate Cells showed that a specific stellate population builds a desmoplastic, immunosuppressive barrier. All three describe barriers through correlations or compartment counts. None models which tumor regions are geometrically shielded.
On the telecom side, three papers supply the method. Terrain-Based Coverage Manifold Estimation gives an accelerated recipe for blockage-aware coverage: it checks for obstructions exactly within a threshold distance and uses a probabilistic line-of-sight model beyond it. That recipe cuts brute-force cost from 60–90 minutes per location to 10–25 seconds. Outage Probability Analysis of Uplink Heterogeneous Non-terrestrial Networks introduces the Matérn hard-core cluster process (MHCCP). This is a random point model in which points form groups but no two points come closer than a minimum distance. The paper shows that plain Poisson placement misrepresents such layouts. We reuse that process as the null model for immune cells, which both cluster (for example, into tertiary lymphoid structures, the lymph-node-like immune aggregates that CONFIRM linked to nivolumab sensitivity) and cannot overlap. Deep Learning-Based Coverage and Rate Manifold Estimation mapped base-station layouts as images to coverage maps with an autoencoder, a network that compresses an input and reconstructs a target from the compressed code, and cut coverage error by 40% relative to stochastic geometry. That paper plays two roles here. It is a warning that analytic spatial averages miss local blind spots, and it is the template for an end-to-end competitor.
The research file reports no prior work linking these fields. The contribution is therefore not "spatial statistics for tissue," which already exists in descriptive forms such as nearest-neighbor distances and Moran's I. It is two specific things: blockage-aware reachability, and comparison against a matched clustered/hard-core null model.
Research Program
From segmented multiplex immunofluorescence or Xenium sections, we assign each cell a type: tumor, effector (CD8⁺ T cell, and as a secondary analysis pro-inflammatory macrophages), or barrier (fibroblast or stellate cell, plus collagen-dense regions where a stain exists). For a tumor cell at location $x$, we call it covered at range $d$ if some effector lies within distance $d$ of $x$ and the straight segment between them crosses fewer than $k$ barrier cells. The section's coverage is
$$C(d,k) = \frac{1}{|T|}\sum_{x \in T} \mathbf{1}\left[\exists\, e \in E:\ \|x-e\| \le d,\ b(x,e) < k\right],$$
where $T$ is the set of tumor cells, $E$ the set of effectors, and $b(x,e)$ the number of barrier cells the segment crosses. The straight segment is a deliberate simplification taken from radio line-of-sight. A sensitivity analysis replaces it with shortest paths that are cheaper through non-barrier tissue, to check that conclusions do not hinge on it. Following the accelerated-simulation recipe, barrier crossings are counted exactly within a threshold distance, and beyond it a fitted crossing probability that depends on local barrier density is used. Computation stays tractable on whole slides with millions of cells. A learned surrogate in the style of the coverage-manifold autoencoder is built only if exact computation proves too slow; it is not part of the claim.
The deficit is then
$$\Delta = C_{\text{obs}}(d,k) - \mathbb{E}_{\text{null}}\left[C(d,k)\right].$$
The expectation is taken over effector placements drawn from an MHCCP. Its cluster intensity, cluster spread, and minimum spacing are fitted to the section's own effector pattern, and the tumor and barrier cells are held fixed. A strongly negative $\Delta$ means the effectors present would cover far more tumor if the barrier geometry did not shield it. Repeating this over a grid of tiles yields a coverage map with explicit dead zones, which pathologists can inspect.
Baselines
The naive reference point is the clinical default: PD-L1 status where available, or nothing beyond stage and grade. The CONFIRM analysis found that tumor mutational burden did not separate responders from non-responders, so we expect this reference to be weak in mesothelioma. We include it so the absolute size of any gain is visible.
The direct spatial competitor is compartment density: intratumoral and stromal CD8⁺ (or Stabilin-1⁺ macrophage) densities, the measure behind the HR of 2.371 in the bladder cohort. It is strongest because it already captures "inside versus outside." The deficit beats it only if it significantly improves a Cox model that already contains both densities, judged by a likelihood-ratio test and by gain in the concordance index under nested cross-validation.
The strongest descriptive spatial baseline is a feature set of nearest effector-to-tumor distances, the fraction of tumor cells with an effector within $d$, Ripley's K cross-functions, and Moran's I. These are fit with the same penalized Cox or logistic model and the same tuning budget as the deficit. This set contains everything in our metric except blockage and null normalization. It is therefore the critical comparison.
The strongest existing clinical predictors in the file are expression-based. The pan-cancer macrophage census in Tumor-associated macrophage clusters linked to immunotherapy in a pan-cancer census reports a random-forest response model with test AUC 0.73 across immunotherapy cohorts that include IMvigor210. CONFIRM reports an IL24-to-EMT expression ratio reaching AUROC 0.889 for nivolumab sensitivity, but only on its 29-patient multi-omic subcohort. Where expression and tissue come from the same patients, we compute these signatures directly and test whether the deficit adds to them. We will not claim to beat 0.889 on 29 patients; with that sample size, confidence intervals are too wide for any such claim to mean much.
Two ablations isolate the key ingredients. The first sets $k=\infty$, removing blockage, so coverage reduces to pure distance. The second uses $C_{\text{obs}}$ without subtracting the null expectation. The proposal stands only if the full metric beats both ablations. An end-to-end competitor is also included: a convolutional network on rasterized cell-type maps, following the coverage-manifold autoencoder design, trained under identical splits and with a matched hyperparameter search budget.
Why not just…?
Why not just count intratumoral immune cells, as the bladder study did? This is the most serious threat. Intratumoral density already encodes much of the exclusion story. Our view is that density cannot separate a tumor with few T cells from one where many T cells are held at a barrier. The deficit is designed to make exactly that separation. The nested Cox comparison settles the question. If $\Delta$ adds nothing once both densities are in the model, we drop the claim.
Why not just use nearest-neighbor distance or Ripley's K features? These capture proximity but are blind to what lies between cells. A T cell 30 µm from a tumor cell across open tissue and one 30 µm away behind a fibroblast wall look identical to them. The $k=\infty$ ablation turns our metric into exactly such a distance measure. If it performs as well as the full metric, blockage is irrelevant and this baseline wins.
Why not just train a deep network end-to-end on the tissue maps? With enough data, a convolutional network could learn blockage implicitly. That is the lesson of the coverage-manifold autoencoder outperforming stochastic geometry. But the spatial cohorts here are small (16 gastric patients, 29 CONFIRM patients, a few hundred bladder patients), and that regime favors a metric with a handful of parameters. The end-to-end baseline is run anyway. If it wins at the largest available cohort size, we report that the mechanism-based metric is mainly an interpretability tool.
Why not just use bulk expression signatures, which are cheaper? The spatially informed mFISHseq assay showed that bulk profiles are diluted by non-tumor tissue, and that spatially informed profiling reclassified 30% of patients. Even so, bulk signatures are cheap and clinically closer to deployment, so we test incremental value over them rather than replacement.
Proposed Experiments
The first check is synthetic and fast. We generate tissues from an MHCCP with planted barrier walls of varying thickness, holding effector counts fixed. We verify that $\Delta$ tracks the planted exclusion while compartment densities and distance features do not. If $\Delta$ fails to track the planted exclusion, we stop.
The mechanistic check uses the multi-modal PDAC atlas with Xenium and multiplex IHC sections. Samples and regions dominated by the POSTN⁺/SPP1⁺ stromal-remodeling niche should show more negative deficits than regions dominated by the CCL4⁺ CD8⁺ immune-activation niche, at matched CD8⁺ density. Within the PDAC data, we also test whether PLXDC1⁺ stellate cells add more blockage per cell than other fibroblasts when barriers are weighted by type.
The clinical tests use three cohorts. For prognosis, we use the 283-patient treatment-naïve bladder cohort, re-digitized and segmented, with overall and recurrence-free survival. The reference to beat there is intratumoral Stabilin-1⁺ density, HR 2.371, plus intratumoral CD8⁺ density. For immunotherapy response, we use the CONFIRM multiplex immunofluorescence subcohort and the 16-patient gastric cohort. Because these are small, they are reported as pre-registered effect-direction checks with confidence intervals rather than headline AUCs. For every outcome analysis, $d$ and $k$ are chosen inside the training folds only.
The decisive result is narrow. A negative $\Delta$ must predict worse survival in the bladder cohort after adjusting for intratumoral and stromal densities and for the distance feature set. It must also beat both ablations. If it does, coverage geometry carries information that counts do not, and dead-zone maps are worth pursuing as a biomarker. If it does not, the correct conclusion is that compartment counting already captures what blockage modeling would add.
References
- Microlocalization and clinical significance of stabilin-1+ macrophages in treatment-naïve patients with urothelial carcinoma of the bladder
- Integrating multi-modal transcriptomics identifies cellular subtypes with distinct roles in PDAC progression
- PLXDC1+ Tumor‐Associated Pancreatic Stellate Cells Promote Desmoplastic and Immunosuppressive Niche in Pancreatic Ductal Adenocarcinoma
- Terrain-Based Coverage Manifold Estimation: Machine Learning, Stochastic Geometry, or Simulation?
- Outage Probability Analysis of Uplink Heterogeneous Non-terrestrial Networks: A Novel Stochastic Geometry Model
- Constitutive inflammation and epithelial-mesenchymal transition dictate sensitivity to nivolumab in CONFIRM: a placebo-controlled, randomised phase III trial
- Deep Learning-Based Coverage and Rate Manifold Estimation in Cellular Networks
- Tumor-associated macrophage clusters linked to immunotherapy in a pan-cancer census
- The spatially informed mFISHseq assay resolves biomarker discordance and predicts treatment response in breast cancer