Per-Nest Immune Coverage Distributions for Tumor Spatial Biology

Most spatial analyses of tumors reduce a tissue slide to one number per patient: the share of the slide that is stroma, the density of killer T cells, or an average "influence score" around tumor regions. Two patients can share that number and still look completely different under the microscope. In one, every island of cancer cells (a "tumor nest") is reached by some immune cells. In the other, the immune cells sit in a single pocket and most nests are untouched. Wireless network theory faced the same problem and solved it. Average coverage across a network hides which individual users actually get service, so the field moved to the meta distribution: the distribution, across individual links, of each link's own probability of success (On the Uplink SINR Meta Distribution of UAV-assisted Wireless Networks, On the Downlink SINR Meta Distribution of UAV-assisted Wireless Networks).

We propose to carry this object into tumor pathology. For every tumor cell we compute a local ratio of distance-weighted effector immune signal, such as CD8+ T cells, to distance-weighted suppressive signal, such as CD163+ macrophages. This is the biological analogue of a signal-to-interference ratio. For each nest we record what fraction of its cells clear a threshold. The patient descriptor is then the distribution of that fraction across nests, not its slide-wide average. A clustered point-process model, in which tumor nests are cluster centres and immune cells scatter around them, does two specific jobs. It estimates this distribution when the field of view is small and cuts nests at the image edge, as on tissue microarray cores and imaging mass cytometry regions. It also supplies a null model that says how much nest-to-nest heterogeneity to expect without any tumor–immune coupling. The claim we test is that per-nest heterogeneity of immune coverage carries prognostic and response information beyond slide-level averages and beyond standard multi-radius spatial statistics.

Core Claim and What Is New

The candidate contribution is the per-nest coverage distribution as a patient-level readout. The point-process machinery is an estimator and a null, not the headline. We separate these roles deliberately, because a reviewer could reasonably argue that empirical per-nest fractions computed straight from cell coordinates already capture the idea. Our experiments test that argument directly. If the empirical version matches the model-based one on large whole-slide fields, we will say so. The model's value would then be confined to small, truncated fields. We expect that regime to matter, since many available imaging mass cytometry and tissue microarray cohorts are exactly that.

Prior work connecting spatial statistics to the tumor microenvironment motivates this design and also shows where it falls short.

The closest work is A Novel Framework for Characterization of Tumor-Immune Spatial Relationships in Tumor Microenvironment. It counts immune cells inside a fixed 100 µm band around K17-defined tumor nests in pancreatic cancer and compares the counts against uniform random placement. It found that CD8+ and CD163+ cells deviate most from randomness (mean absolute deviation 1.36 and 0.99). That result motivates using exactly these two cell types as effector and suppressor. We differ in three ways. We replace the hard band with a distance-decaying weight. We replace the uniform null with a fitted clustered null. We keep nests separate instead of averaging over them.

The pancreatic whole-slide study Stroma and lymphocytes identified by deep learning are independent predictors for survival in pancreatic cancer shows that slide-level stroma and lymphocyte percentages already stratify survival. Our proposal must beat those averages, not ignore them.

Neighborhood-graph approaches are a different family. Deep-learning and transfer learning identify new breast cancer survival subtypes from single-cell imaging data builds 378 cell–cell interaction scores from neighborhood clustering coefficients and feeds them to a neural Cox survival model. Charting Cellular Networks in the Tumor Microenvironment: Graph Visualizations in Highly-Multiplexed Breast Cancer Tissue draws cell-niche graphs for exploration. Both pool interactions over the whole sample and do not ask how coverage varies nest by nest.

Two further methods take still other routes. Graph of graphs analysis for multiplexed data with application to imaging mass cytometry avoids cell segmentation by summarising marker co-expression in patches around tumor cells and embedding patches cohort-wide. That discards cell coordinates, which are exactly what we use. MONTAGE: A Computation Framework to Identify Spatially Resolved Functional Enrichment Gradients in the Tissue Microenvironment via Spatial Communities uses local Moran's I, a statistic for whether nearby grid cells have similar values, on gridded data. It is a lattice method rather than a point model.

The method itself comes from wireless stochastic geometry. A Thomas cluster process places "child" points around each "parent" with Gaussian spread. A Matérn cluster process places them uniformly within a fixed radius. Modeling and Analysis of D2D Millimeter-Wave Networks with Poisson Cluster Processes used the Thomas model and showed that performance there is limited by interference within clusters, with signals attenuated by distance and blocking obstacles. Performance Analysis of Clustered LoRa Networks used the Matérn model for bounded clusters. The two-moment Beta approximation of the meta distribution comes from the UAV papers. That work also warns that the Beta fit breaks down in some regimes, so we always check it against the empirical distribution. We borrow these as tools. We do not claim that tissue obeys wireless physics.

Research Program

The input is a list of cell centroids with type labels: tumor, CD8+ effector, and CD163+ or other suppressive myeloid cells. Using centroids rather than full cell outlines matters, because boundary segmentation is the step that fails in crowded tissue. Tumor nests are defined as connected components of tumor cells within a linking distance. For a tumor cell at location $x$ we compute

$$ S(x) = \frac{\sum_{e \in E} g(\lVert x-e \rVert)}{\sum_{s \in M} g(\lVert x-s \rVert) + \eta}. $$

Here $E$ is the set of effector cells and $M$ the set of suppressive cells. The weight $g(r)=\exp(-r/\ell)$ decays with distance $r$ over a length scale $\ell$, and $\eta$ is a small constant playing the role of noise. We fix $\ell$ by inner cross-validation over a grid of 10–100 µm, so that the hard 100 µm band is one point in the comparison.

For nest $c$, let $P_c(\theta)$ be the fraction of its tumor cells with $S(x) > \theta$. The patient-level meta distribution is

$$ \bar F(\theta, t) = \frac{1}{N}\sum_{c=1}^{N} \mathbf{1}\big[P_c(\theta) > t\big], $$

the share of the patient's $N$ nests in which more than a fraction $t$ of cells are "covered." We summarise $\bar F$ by its first two moments, by a Beta fit, and by low and high quantiles. The lower tail captures the "untouched nests" phenotype that averages hide.

The model-based estimator fits a marked cluster process to each sample. Nest centroids are parents, drawn from an inhomogeneous Poisson process. Effector and suppressor cells are children, with type-specific Thomas spread parameters and cluster sizes that vary from nest to nest. The model is fitted by minimum contrast on cross-type pair correlation functions, which measure how much more often two cell types appear at a given distance than chance predicts. With the fitted model we simulate complete, untruncated tissue and compute $\bar F$ on the simulation. This corrects for nests cut by the image edge and for nests containing only a handful of cells. The same fit with immune children placed independently of nest identity supplies the null. "Excess heterogeneity" is the gap between the observed and null meta distributions.

Before touching patient data, we validate on synthetic tissue. We simulate Thomas and Matérn configurations with known parameters at whole-slide scale. We then crop to tissue-microarray-sized windows and measure how far the empirical and model-based $\bar F$ drift from the full-slide truth.

Baselines

The naive reference is slide-level composition: CD8+ density, CD163+ density, and their ratio. This is the cell-level analogue of the lymphocyte and stroma percentages that Stroma and lymphocytes identified by deep learning are independent predictors for survival in pancreatic cancer showed are prognostic. We enter it in a Cox survival model with the same clinical covariates. On the breast cancer imaging mass cytometry cohort, clinical molecular subtyping alone gives a C-index of 0.63 (Deep-learning and transfer learning identify new breast cancer survival subtypes from single-cell imaging data). C-index is concordance, the probability that the model ranks two patients' survival times correctly.

A second simple reference is the fixed-band influence score of A Novel Framework for Characterization of Tumor-Immune Spatial Relationships in Tumor Microenvironment, reimplemented on the same centroids.

The strongest published spatial competitors are two. For survival, it is the 378 interaction scores with a neural Cox model from that same breast paper, which reports a test C-index of 0.80. For immunotherapy response, it is the graph-of-graphs method, which reports ROC AUC 0.945 on the 55-patient non-small cell lung cancer cohort and 1.0 on the 29-patient cohort (Graph of graphs analysis for multiplexed data with application to imaging mass cytometry). Because those cohorts are tiny, we rerun this method ourselves under our own repeated nested cross-validation and compare against the rerun rather than the published figure.

The strongest classical spatial-statistics baseline is multi-radius cross-type Ripley K and pair-correlation summaries between tumor and each immune type. These measure clustering at several distances, pooled over the slide. No reported result was found on Lacuna for cross-K features on these cohorts.

Two ablations isolate the key ingredient. The slide-level mean of $P_c(\theta)$ tests whether heterogeneity matters beyond the average of the very same ratio. The empirical, model-free $\bar F$ tests whether the cluster-process fit adds anything.

Every method gets the same patient-level splits, the same cell calls, and the same hyperparameter search budget of 50 configurations per method in the inner loop. Each also gets an identical downstream learner: a penalised Cox model for survival and logistic regression for response. Neural Cox is reported in addition for the interaction-score baseline. Beating a baseline means a C-index gain of at least 0.02 whose 95% bootstrap interval excludes zero when the meta-distribution features are added on top of it. For response, it means a corresponding gain in repeated-cross-validation AUC. Matching 0.80 with five interpretable numbers instead of 378 would be a secondary, efficiency-style win.

Why not just…?

Why not just use cross-type Ripley K or pair correlation functions? These are the standard way to measure spatial clustering at several scales, and they already improve on a uniform-randomness test. They are pooled over the whole slide, though, so a patient with uniformly moderate infiltration and a patient with one heavily infiltrated nest among many empty ones can produce similar curves. This is a real threat. If the meta-distribution features add nothing in a Cox model that already contains cross-K summaries, the claim fails, and the experiment is designed to show that.

Why not just compute per-nest infiltration fractions directly, with no point-process model? On large whole-slide fields this may work just as well, and that would be a legitimate finding. It would mean the per-nest distribution, not the point-process fit, is what carries the signal. Our expectation that the model helps rests on the truncation problem: tissue-microarray cores and imaging regions cut nests at the edge and leave many nests with few cells. The synthetic cropping study and the real crop-and-compare experiment below decide this. If the empirical estimator is stable under cropping, we drop the model from the recommended pipeline.

Why not just train a graph neural network or use neighborhood interaction scores? The 0.80 C-index of Deep-learning and transfer learning identify new breast cancer survival subtypes from single-cell imaging data shows these are strong. They aggregate interactions over the sample, however, and their features have no direct reading as "fraction of nests left uncovered." We test complementarity rather than replacement by asking whether our five to ten features add to the interaction scores in a joint model.

Why not just use segmentation-free patch signatures, given their near-perfect AUC? The graph-of-graphs results come from 55 and 29 patients, where an AUC of 1.0 is fragile. The method also deliberately discards coordinates, so it cannot say how far immune cells reach into nests. We rerun it under identical cross-validation. Separately, we check our own reliance on cell detection by jittering centroids by 2–5 µm and randomly dropping 10–20% of cells, and we report how much $\bar F$ moves.

Proposed Experiments and Decisive Checks

The primary survival benchmark is the breast cancer imaging mass cytometry cohort, using the 27 phenotypes from Deep-learning and transfer learning identify new breast cancer survival subtypes from single-cell imaging data mapped onto effector and suppressor groups. The response benchmark is the two non-small cell lung cancer anti-PD-1 imaging mass cytometry cohorts. The 118 glioblastoma and 46 metastasis brain tumor samples from Spatial Coordinates as a Cell Language: A Multi-Sentence Framework for Imaging Mass Cytometry Analysis serve for disease-type discrimination. The 23-slide pancreatic multiplex immunohistochemistry cohort serves as a whole-slide field where truncation is minimal. This plan assumes cell-level coordinates for these cohorts can be obtained from their authors or public releases; we have not confirmed that.

Three checks decide the proposal.

The first is incremental value. Do the meta-distribution features improve C-index or AUC over a model that already contains the slide-level mean of the same ratio plus cross-K summaries, judged by likelihood-ratio test and bootstrap intervals?

The second is truncation robustness. We crop the pancreatic whole slides into core-sized windows and compare the per-window empirical $\bar F$ and model-based $\bar F$ against the full-slide value. The model-based estimator should show smaller bias and variance. If it does not, the point-process component is unjustified.

The third is null calibration. On synthetic tissue with no tumor–immune coupling, the excess-heterogeneity statistic should be centred at zero. On real tissue, its lower tail should mark nests that pathologists review blind and call immune-excluded.

A negative result on the first check, combined with a positive result on the second, would still yield a useful tool: an edge-corrected estimator of nest-level coverage for small-field imaging.