Spatial transcriptomics machines such as 10X Visium do not measure single cells. They measure "spots": small circular patches of tissue, about 55 µm across, each of which pools the RNA of several cells into one read-out. Software then has to guess which cell types were in each spot, a step called deconvolution. Rare cells are the hardest case. A single therapy-resistant clone or a handful of infiltrating immune cells can be diluted below detection in one spot, or wrongly spread as a faint signal across many. Current methods mostly treat each spot as a bag of an unknown number of cells, or smooth estimates across neighboring spots with a hand-built graph. Neither approach says anything principled about how many cells of each type should physically fit in a disk of a given size, or how often a disk should contain none of a given type.
Wireless network theory has worked on exactly this question for base stations rather than cells. Stochastic geometry treats transmitters as random points scattered in the plane and derives, in closed form, how many land in a region and how often a region is empty. We propose to model tissue the same way: each cell type is one "tier" of a random point pattern, possibly clustered. Integrating that pattern over the spot disk gives an analytic prior on the per-type cell counts in every spot, including the probability that a rare type is absent and the way rare-type clusters spill across neighboring spots. Two things follow. The first is a deconvolution prior whose few parameters (densities and cluster radii per type) are shared across the whole slide and fitted from the data, in place of per-spot sparsity knobs. The second is a detectability law that predicts, from spot diameter and cell-type density, below what abundance a rare type stops being recoverable. That law is directly analogous to the outage-scaling results for dense networks.
Core Claim
A clustered marked point-process prior on cell positions, integrated analytically over the capture disk, improves rare-cell-type presence calls and proportion estimates in multi-cellular spots compared with flat or graph-smoothed priors. It also yields a testable scaling law for how rare-type detectability depends on spot size. The new ingredient is not "spatial information" in general, which CARD and FAST already use. It is a generative model of cell placement whose void probabilities and cross-spot correlations come from geometry and are not tuned per spot.
Grounding in Prior Work
The wireless side supplies the method. Coverage and Economy of Cellular Networks with Many Base Stations shows that ignoring the probability that a base station's cell is empty badly misestimates interference in dense networks. Once empty cells are silenced, outage scales as $\beta\lambda_u/\lambda_b$, which leads to a square-root law for optimal density. We reuse the void-probability-and-thinning logic for spots that contain no cells of a given type, and the scaling-law style of analysis for spot-size design. On Exact Distribution of Poisson-Voronoi Area in K-Tier HetNets derives exact moments of multi-tier cell areas using probability generating functionals (a tool for computing expectations of products over all points of a random pattern). It also reports that the common Gamma approximation is "~12% erroneous" in the second moment. That result motivates our use of exact functionals rather than moment-matched shortcuts. Most tractable network results assume independent, homogeneous Poisson patterns. Tissue is clustered, so we deliberately work with cluster processes, the regime the HetNet literature itself flags as hard.
The biology side supplies the gap and the inference engine. Count Bridges models integer gene counts with Poisson birth–death dynamics and deconvolves spots into single-cell count vectors by expectation-maximization with projection-guided sampling. However, it treats spots as isolated bins with no model of how many cells, of which types, physically occupy them. We plug our geometric prior into that inference as the distribution over the latent number and identities of cells per spot. FAST regularizes non-negative matrix factorization with a graph Laplacian built from spot distances and histology. The spatial deconvolution benchmark of 18 tools across 50 datasets found CARD, Cell2location and Tangram most robust. None of these methods models cell placement as a point process. Laplacian smoothing also pulls sharp rare-type niches toward their neighbors, which is precisely the failure we target. Non-spatial methods such as POPPY and SubNetDL recover heterogeneity from bulk data through reference atlases or subclone clustering. They motivate the problem but cannot place cells in space, so they serve as context rather than competitors.
The Model
Cells of type $k$ are placed as a Thomas cluster process. Parent centers fall as a Poisson pattern with density $\mu_k$. Each parent produces a Poisson number of cells with mean $c_k$, scattered around it with Gaussian spread $\sigma_k$. A plain Poisson pattern is the limit with no clustering. For a spot disk $B$, let $p_k(x)$ be the probability that a cell spawned by a parent at $x$ lands in $B$. The probability that the spot contains no type-$k$ cell is then
$$ P(N_k(B)=0)=\exp\!\Big(-\mu_k\!\int_{\mathbb{R}^2}\big(1-e^{-c_k\,p_k(x)}\big)\,dx\Big), $$
and the full count distribution and the joint law for two overlapping or adjacent spots follow from the same generating functional. In the Poisson limit this reduces to $\exp(-\lambda_k|B|)$ with $\lambda_k=\mu_k c_k$. Clustering raises the void probability at fixed mean density, which is the effect a flat sparsity prior cannot represent with a shared parameter. Adjacent spots share parents, so the prior couples them strongly when $\sigma_k$ is comparable to spot spacing and weakly otherwise. This gives data-fitted, type-specific spatial coupling in place of one global smoothing graph. A packing constraint (a total-cell cap from nucleus counts in the histology image, when available) thins the joint draw.
Parameters are fitted per slide by alternating two steps. The first step uses the current per-spot cell assignments from the expression likelihood, run through Count Bridges' expectation-maximization or, for a lighter variant, a negative-binomial mixture. The second step refits $(\mu_k, c_k, \sigma_k)$ by maximizing the point-process likelihood of the implied counts across all spots. Because only three parameters per type are shared across thousands of spots, the prior cannot overfit individual spots.
The detectability law comes from two competing effects. The chance that a spot contains a type-$k$ cell grows roughly like $\lambda_k|B|$. The share of the spot's signal contributed by one such cell shrinks roughly like $1/(\lambda_{\text{tot}}|B|)$, where $\lambda_{\text{tot}}$ is the density of all cells. We hypothesize an optimal spot area for each rare type, and a minimum detectable density that depends on clustering. The precise form will be derived from the generating functional and checked empirically.
Research Program
The work has three parts. The first is analytic: closed-form or one-dimensional-integral expressions for per-spot count distributions and two-spot correlations under Thomas and Matérn cluster processes on disks, plus the detectability scaling. The second is the inference method: the geometric prior placed inside Count Bridges deconvolution and inside a simple negative-binomial proportion model, which shows the prior is not tied to one backbone. The third is an empirical study on single-cell-resolution data where ground truth is known, using pseudo-spots.
Baselines
The naive reference is a flat prior. It fixes the number of cells per spot at the dataset mean, places a uniform Dirichlet prior on proportions, and uses the same expression likelihood as our method. This shows how much any prior on placement matters.
The strongest existing methods are CARD, Cell2location and Tangram, identified by the benchmark as the most robust and accurate. CARD is the closest in spirit because it already uses spatial correlation, through a conditional autoregressive model. We run all three with their published defaults plus an equal hyperparameter search budget on a held-out slide, using the same reference atlases and the same pseudo-spots. The benchmark reports no single summary scalar for these tools, so we re-run them rather than quote numbers.
FAST is the representative graph-Laplacian method. Its reported figures on a simulated astrocyte–neuron mixture of 260 spots are an average Pearson correlation of 0.93 and an RMSE of 0.15 for proportions. We reproduce that setting as a sanity check. Matching it on common types while beating it on rare-type recall would count as success.
Count Bridges without our prior is the most direct comparison, because it shares our inference engine. It reportedly outperformed CIBERSORTx, MuSiC and STdeconvolve on MERFISH pseudo-spots, but no reported result found on Lacuna for its exact scalar accuracy, so we re-run it.
Two ablations isolate the key ingredient. The first replaces the cluster process with a plain Poisson pattern, which removes clustering. The second keeps per-spot zero-inflation but tunes it freely, with no geometric sharing. Beating Count Bridges alone but not these ablations would mean the gain is generic sparsity, not geometry.
Why not just…?
Why not just add a tuned sparsity or spike-and-slab prior on proportions? This is the real threat. On isolated spots, a well-tuned zero-inflation prior may match our void probabilities. Our bet is that two things set the geometric prior apart. Its void rates are tied to density and clustering through shared parameters, so they transfer across spot sizes without retuning. It also couples neighboring spots in a type-specific way. The second ablation tests this directly. We additionally evaluate transfer: fit on 55 µm pseudo-spots, test on 30 µm and 100 µm pseudo-spots without refitting. If tuned sparsity matches us there too, the geometric contribution reduces to the detectability law alone, and we will report it that way.
Why not just smooth with a graph Laplacian or a conditional autoregressive model, as FAST and CARD do? These smooth all types with one spatial kernel, and their reported accuracy is measured on common types. A rare type in a 1–2 spot niche is exactly what smoothing pulls toward its neighbors. We test this by stratifying all metrics by type abundance and niche size. If CARD matches us on rare types in small niches, the claim fails.
Why not just count nuclei in the histology image and use that as the per-spot cell number? Nucleus counts fix the total but not the per-type split or the void probability for a rare type. We already use them as a packing cap when available. The informative comparison is Count Bridges with nucleus-count totals versus Count Bridges with our full prior. If the totals alone close most of the gap, the per-type geometry adds little.
Why not just use single-cell imaging such as MERFISH? It is costly and limited to targeted gene panels, and spot-based and bulk assays still dominate clinical cohorts. That is also why the detectability law has practical value: it tells experiment designers what spot size can see a given rare type.
Proposed Experiments
Ground truth comes from single-cell-resolution datasets in the 50-dataset benchmark collection: seqFISH+, MERFISH, Slide-seqV2 and Stereo-seq. On these we tile synthetic circular spots at diameters from roughly 20 to 100 µm and aggregate true counts. We then apply zero-shot transfer to real 10X Visium slides, such as the mouse olfactory bulb slide used by FAST, where layer markers provide qualitative checks.
The primary metric is the area under the precision-recall curve for rare-type presence per spot, restricted to types under 2% abundance. Secondary metrics are proportion RMSE and Pearson correlation, the Energy Score on reconstructed single-cell counts, and calibration of presence probabilities.
Three checks are decisive. The first is whether the cluster prior beats both ablations on rare-type precision-recall at 55 µm. The second is whether it transfers across spot diameters without refitting while tuned sparsity does not. The third is whether the empirical detectability curve, meaning recall versus spot diameter for each rare type, shows the interior optimum and density-dependent threshold predicted analytically, within bootstrap error. A predicted optimum that does not appear in pseudo-spot data would falsify the scaling analysis even if the deconvolution gains hold.
References
- Coverage and Economy of Cellular Networks with Many Base Stations
- On Exact Distribution of Poisson-Voronoi Area in K-Tier HetNets With Generalized Association Rule
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Flexible Analysis of Spatial Transcriptomics Data (FAST): A Deconvolution Approach
- A comprehensive benchmarking with practical guidelines for cellular deconvolution of spatial transcriptomics
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer