Which Cell Is Signaling? Attributing Pathway Activity Inside Spatial Transcriptomics Spots

Spatial transcriptomics measures gene expression at known positions in a tissue slice. The most widely used whole-transcriptome platforms do not measure single cells. They measure "spots" about 55 micrometers across, and each spot mixes the RNA of roughly one to ten cells. Current software can estimate which cell types sit in each spot. It does this by assuming each cell type has one fixed expression profile everywhere in the tissue. That assumption discards the question mechanism studies care most about. When a spot shows strong inflammatory or growth-factor signaling, which cell type inside it is responding, and which neighboring cell type is driving the response?

We propose to answer that question by changing what deconvolution estimates. Deconvolution here means the computational step that splits a spot's mixed signal back into the cell types that produced it. In our version, each cell type's expression is allowed to shift away from its reference profile, but only along the known downstream "footprints" of a small set of signaling pathways such as NF-κB, TGF-β, MAPK and PI3K-AKT. A footprint is the set of target genes a pathway switches on or off. The size of the shift is a shared, sparse function of which cell types are nearby. The output is not a list of proportions. It is a directed map that reads like "fibroblast abundance drives TGF-β footprint activity in tumor cells, not in macrophages." Because the model can be trained on ordinary archival Visium data, it could bring cell-type-resolved pathway mapping to clinical cohorts that in situ CRISPR screens cannot reach.

Core Claim

Whether a pathway response belongs to one cell type or another inside a mixed spot cannot be determined from any single spot. It becomes determinable across many spots if two things hold. The response must be restricted to a low-dimensional, biologically fixed set of gene directions, and its magnitude must depend on the neighborhood through a function shared across the slide. A footprint gene that a cell type barely expresses at baseline cannot carry that cell type's response, and composition varies from spot to spot. Together these separate "tumor cells in fibroblast-rich spots activate TGF-β targets" from "fibroblast-rich spots contain more fibroblasts, which express TGF-β targets." We claim this footprint constraint is the ingredient that makes attribution work. A model given the same freedom along unconstrained or random gene directions should fail at attribution even when it fits spot counts equally well.

Where This Sits in Prior Work

The motivating tension is spelled out in the literature. A benchmark of 18 deconvolution tools across 50 datasets ranked CARD, Cell2location and Tangram as the strongest methods. All of them estimate only proportions under a static profile for each cell type, as does the earlier negative-binomial model stereoscope. In contrast, Perturb-FISH combined in situ CRISPR barcoding with multiplexed imaging and showed that responses of TNF and IL1A reverse between low and high local cell density. CONCERT modeled perturbation spread across tissue with spatial Gaussian-process kernels and showed strong niche dependence. Together these papers motivate the proposal: the static-profile assumption is biologically wrong in precisely the way that hides signaling. CONCERT's authors also state that their model contains no regulatory or signaling-pathway structure. We add that structure, but we do not predict perturbation outcomes.

Two lines of work already recover more than proportions, and they set the bar we must clear. Count Bridges models integer counts with Poisson birth-death bridges and samples virtual single cells whose counts sum exactly to the spot (companion entry). It encodes no pathway knowledge and no neighborhood dependence. DeepSpot2Cell predicts virtual single-cell expression from H&E images using spot-level supervision. It can only recover signaling that leaves a visible morphological trace. Starfysh combines Visium with histology. It found glycolysis and PI3K-AKT signatures co-located with Treg- and MDSC-rich hypoxic hubs in metaplastic breast cancer, but it scores pathways at the level of the whole hub and cannot say which cell type executes them. That unresolved Starfysh finding is our main real-data target.

SubNetDL and POPPY supply a methodological lesson rather than a component. SubNetDL propagates subclone-specific mutations over a protein-interaction network, and POPPY attaches bulk samples to a single-cell atlas. Both show that decomposing a bulk signal before interpreting it beats interpreting the aggregate. Neither has spatial coordinates. We borrow only the principle that fixed biological network structure should constrain what a decomposed component is allowed to do.

Model

Let $y_{sg}$ be the count of gene $g$ in spot $s$. Let $w_{sk}$ be the abundance of cell type $k$ in the spot, $x_{kg}$ the reference expression of gene $g$ in type $k$, and $F_{pg}$ the signed footprint weight of pathway $p$ on gene $g$, fixed in advance. We model the expected count as

$$\mu_{sg} = \ell_s \sum_k w_{sk}\, x_{kg} \exp\Big(\sum_p a_{kp}(z_s)\, F_{pg}\Big),$$

where $\ell_s$ is a spot-level library-size factor. The observed count $y_{sg}$ is negative-binomial around $\mu_{sg}$. The term $a_{kp}(z_s)$ is the activity of pathway $p$ in cell type $k$. It depends on a neighborhood vector $z_s$ made of the estimated cell-type abundances in spot $s$ and its adjacent spots. We use the simplest form that stays interpretable:

$$a_{kp}(z_s) = b_{kp} + \sum_j \beta_{kpj}\, z_{sj},$$

with an L1 penalty on the coefficients $\beta$. A nonzero $\beta_{kpj}$ is the model's claim that the presence of cell type $j$ drives pathway $p$ in cell type $k$. The collection of these coefficients is the sender-to-receiver map. The term $b_{kp}$ is a baseline activity shared by all spots.

Abundances $w$ are initialized from Cell2location and refined jointly with the pathway parameters. The reference profiles $x$ come from a matched single-cell atlas, with per-gene platform-correction factors as in reference-based deconvolution. Footprints come from curated pathway target-gene sets of the PROGENy type. They can also be estimated from perturbation data where it exists. A nonlinear version of $a$ is a later extension, used only if the linear form clearly underfits.

Research Program

The work has three stages.

The first stage builds ground truth for attribution, which no current benchmark provides. We reuse the curated Xenium-derived pseudo-spot benchmark from DeepSpot2Cell: 20 lung, 7 breast and 2 pancreatic samples, with single cells aggregated into simulated 55 micrometer spots. Because the true single cells are known, we can compute each cell's footprint score on the genes in the panel. From those scores we derive the true niche dependence for each cell type. We restrict evaluation to pathways that have enough footprint genes in the Xenium panel, and we report that coverage explicitly. Panels of a few hundred genes are the main practical limitation, since they cap how many pathways can be evaluated. We add a planted-response benchmark on the same cells. In it, we artificially boost a pathway's footprint in one cell type as a known function of its neighbors, so the true sender-to-receiver edge is known exactly.

The second stage uses perturbation data as mechanistic ground truth. The Perturb-FISH melanoma xenografts come from an NF-κB pathway screen, and the dataset records which cells carry which guide RNAs. After aggregating those cells into pseudo-spots, the question is whether the model recovers the density-dependent NF-κB target responses and places them in the correct cells. The MERFISH mouse brain data used by Count Bridges serve as a second, non-tumor tissue.

The third stage applies the model to archival clinical data: the Starfysh breast cancer Visium cohort covering ER-positive, triple-negative and metaplastic subtypes. The question is which cell types execute the PI3K-AKT and glycolysis programs inside the hypoxic immunosuppressive hubs, and which neighboring types drive them. We treat this stage as producing hypotheses, with the internal checks described below, and we do not count it as validation.

Baselines

The naive reference point is spot-level pathway scoring. It computes a footprint score for each spot, as Starfysh-style hub analysis does, and then splits the score among cell types in proportion to their estimated abundance. This is what an analyst would do today, and it builds in the confound we are trying to remove.

The strongest proportion-only pipeline is Cell2location or CARD, the top performers in the 50-dataset benchmark, followed by a two-stage regression. Spot footprint scores are regressed on terms that multiply each cell type's abundance by each neighbor type's abundance, and the fitted interaction coefficients are read as the attribution. This baseline gets the same neighborhood features and the same footprints as our model, so it is the fairest test of whether fitting jointly at the gene level matters. That benchmark reports only rankings, so there is no reported number on Lacuna for Cell2location or CARD proportion error on our data. For reference, Starfysh reports proportion RMSE ≈ 0.09 on its breast cohort. We require that our refined abundances do not lose proportion accuracy relative to the Cell2location initialization.

The strongest method that recovers per-cell expression is Count Bridges. We use it to sample virtual single cells consistent with each spot, then score footprints on those cells and fit the same sparse neighborhood regression. The file reports that it surpasses STDeconvolve on MERFISH but gives no number; there is no reported result found on Lacuna for Count Bridges' single-cell accuracy. The image-based alternative is DeepSpot2Cell with the same downstream scoring. Its reported strength is relative: a 46% improvement in cell-expression correlation for the top 100 genes over scstGCN in the lung out-of-sample setting. We rerun it on the same splits instead of quoting that figure.

Three ablations isolate the key ingredient. The first gives the model the same number of free directions but replaces the footprints with random gene sets of matched size and expression level. The second learns the directions freely for each cell type. The third removes the neighborhood input so that activity is constant for each cell type. All methods use the same reference atlas, the same splits and the same hyperparameter-search budget measured in GPU-hours, with footprints fixed before any test data is seen.

Beating the baselines means the following. On held-out samples, our model must show higher Spearman correlation between inferred and true activity per cell type and per spot. It must also show higher area under the precision-recall curve for recovering true sender-to-receiver edges. Both gains must hold across cancer types and must exceed the random-footprint ablation by a margin outside the bootstrap interval across samples.

Why not just…?

Why not just run Cell2location and regress spot pathway scores on abundance interactions? This is the real threat, and we treat it as one. The regression sees only a spot-level score, so it cannot use the fact that a given footprint gene is expressed by fibroblasts but not by tumor cells. Our model uses that fact because the reference expression $x_{kg}$ weights each cell type's contribution to each gene. That is the mechanism we expect to make attribution possible, but it is a hypothesis. If the two-stage regression matches our model on attribution accuracy in the planted and Xenium benchmarks, we will report that. The contribution would then shrink to the attribution benchmark plus a recommendation for the simpler pipeline.

Why not just deconvolve into virtual single cells with Count Bridges and score pathways on them? Count Bridges enforces exact count sums and realistic integer distributions. Nothing in it, however, ties how the spot's extra footprint signal is divided among cells to the neighborhood. We expect its samples to drift back toward reference cell states and to spread niche-driven responses across whichever cells are abundant. The Perturb-FISH pseudo-spots settle this directly, because the true responding cells are known.

Why not just predict single cells from H&E with DeepSpot2Cell? The research file notes that DeepSpot2Cell cannot infer signaling switches that have no morphological correlate. Many transcriptional responses downstream of NF-κB or TGF-β fall into that category. We test this by comparing its footprint-score recovery against its recovery of cell-type marker genes on the same benchmark. A large gap between the two would show the limitation directly.

Why not just use Xenium or Perturb-FISH directly? Where those assays exist, they should be used, and that is why we treat them as ground truth. But their panels hold at most a few thousand genes, and Perturb-FISH requires prospective lentiviral engineering. Neither can be applied retroactively to archival whole-transcriptome clinical cohorts, and those cohorts are the regime this model is built for.

Proposed Experiments and Decisive Checks

The decisive experiment is the random-footprint ablation on the Xenium pseudo-spot benchmark. Real and random footprints give the model the same number of parameters. If the real ones do not clearly beat the random ones at attribution, even when spot reconstruction likelihood is equal, the core claim is false.

The second check tests identifiability directly. On planted responses, we vary how strongly two cell types' abundances are correlated across spots and measure the point at which attribution collapses, for our model and for the two-stage baseline. This shows when attribution from spots should not be trusted at all.

The third check uses the Perturb-FISH melanoma pseudo-spots. We ask whether the model assigns the NF-κB footprint response to the guide-carrying cells and reproduces the reversal of TNF and IL1A responses with density.

Finally, on the Starfysh cohort, we test whether the sender-to-receiver edges inferred from metaplastic hubs replicate across patients held out from fitting. We also test whether they disappear when the neighborhood vectors are shuffled across spots. Edges that survive both checks become concrete, testable predictions about which cells run PI3K-AKT and glycolysis programs in those hubs.