Spatial Niche Bottlenecks: Tissue Models That Must Explain Themselves in Arrangements a Pathologist Can Check

When cells are separated from a tissue for sequencing, the record of which cells were next to which is lost. Across several diseases in the material reviewed here, that lost arrangement carries the clinical signal. In triple-negative breast cancer, chemotherapy responders and non-responders both contain cytotoxic T cells. What separates them is where those T cells sit: inside tumor nests, or held back in the surrounding stroma (Identifying tissue states by spatial protein patterns related to chemotherapy response in triple-negative breast cancer). The models that capture this signal are black boxes. Graph neural networks over cell neighborhoods, and large models pretrained on spatial gene expression, are explained after the fact, if at all.

We propose a model that has to route its prediction through a small set of named spatial arrangements, which we call niche concepts. An example is "fraction of CD8 T cells within 20 µm of tumor cells, relative to those in stroma." Each concept is computed explicitly from the cell map. The final prediction is a sparse, readable function of those concepts. This design makes the explanation faithful by construction: the concepts are the only path to the output. It also makes each explanation physically checkable, because every concept value corresponds to a region of the slide that a pathologist can look at. This second property matters because fieldwork on a deployed clinical model found that trust came from being able to verify the model's flags directly, not from seeing its internal weights (Katharine Henry et al.).

Core Claim

A concept bottleneck is a model forced to first predict human-named intermediate quantities and then predict the outcome only from those quantities. Our claim is that a concept bottleneck whose concepts are differentiable spatial relations over a measured cell map can do three things. First, it can match or beat a black-box graph network at predicting chemotherapy response in triple-negative breast cancer. Second, its explanations pass intervention tests that post-hoc graph explainers fail. Third, it outperforms simple hand-computed spatial statistics because it learns which relations and length scales matter.

The claim is falsifiable at each step. If hand-computed neighborhood statistics with logistic regression match the bottleneck, the learned part adds nothing. If shuffling cell coordinates leaves performance intact, the spatial part adds nothing.

How This Differs From the Closest Work

Graph models with post-hoc explanation. The TNBC study supplies both the motivation and the strongest baseline. It reached an area under the ROC curve (AUROC) of 0.71 with graph convolutional networks on imaging mass cytometry (a method that measures about 40 proteins per cell in place). It then used GNNExplainer, a post-hoc perturbation method, to name important markers. We keep the cohort and the endpoint but replace post-hoc attribution with an intrinsic bottleneck.

Concept bottlenecks over graphs. Graph concept bottlenecks (Haotian Xu et al.) pass messages between concepts. Relational bottlenecks (Relational Concept Bottleneck Models) produce logic rules over relational data. These works supply the method, but they were evaluated on synthetic logic tasks, citation networks, and flat image datasets. In our setting the relations are physical distances between cells.

Spatially resolved concepts. WHE-CBM (Ran Eisenberg & Ofir Lindenbaum) keeps two-dimensional concept maps so that edits stay local. It relies on bounding boxes around discrete objects. Tissue is a continuous carpet of cells with no natural boxes.

Concepts that filter slide features. CATE (The "Free Lunch" in Pathology) filters slide features using text-derived concepts. Our concepts are instead defined by measured molecular identity and position.

Spatially aware black boxes. STAMP (Fusing Pixels and Genes) and SToFM (SToFM) show that spatial molecular data improves representations. Both remain end-to-end black boxes with no human-readable intermediate layer.

Research Program

Input. The model starts from a segmented cell map. Each cell has coordinates and a measured protein profile. A small encoder assigns each cell soft memberships over $K$ cell states. Some states are anchored to known types (tumor, CD8 T cell, fibroblast, macrophage) by marker-based labels; a few are left free so that unlabeled populations can be found.

Niche concepts. Each concept is a relation between two states $a$ and $b$ at a learned length scale $\sigma$, optionally restricted to a tissue compartment such as stroma or tumor nest:

$$c_{ab}(\sigma) = \frac{1}{N_a}\sum_{i}\sum_{j \neq i} p_{ia}\, p_{jb}\, \exp\!\left(-\frac{\lVert x_i - x_j\rVert^2}{2\sigma^2}\right).$$

Here $p_{ia}$ is cell $i$'s membership in state $a$, $x_i$ is its position, and $N_a = \sum_i p_{ia}$. Each such concept can be rendered as a map over the slide, highlighting where the relation holds.

Prediction. The outcome is predicted by sparse logistic regression on a few dozen candidate concepts, with an $L_1$ penalty that selects a handful. There is no residual path around the bottleneck. This rules out by design the leakage problem common to concept models, where information slips past the named concepts. Removing the residual path risks the "distraction" loss that Emmeke Veltmeijer et al. observed when clinical criteria were forced on a melanoma model. We therefore measure the cost directly against a variant that does have a residual path, rather than assume it away.

Clinical interaction. A pathologist can edit an intermediate concept, for example by marking a region the model called "tumor nest" as stroma. The prediction is then recomputed. This is the kind of override that Susanne Gaube et al. suggest is needed: 27.54% of radiologists in their study followed incorrect advice every time.

Baselines

Naive reference: cell abundance. This is a logistic regression on per-region cell-type fractions, which is the information a dissociated or deconvolved profile would provide. The TNBC study reports that such counts did not separate responders from non-responders. It gives no numeric score for them, so we will compute one ourselves. Any useful model must clearly exceed it.

Strongest existing method: graph convolutional network. We use the GCN from Identifying tissue states, which reported AUROC 0.71 on the 63-patient, 813-region imaging mass cytometry cohort. We will re-run it on our exact patient-level folds with the same hyperparameter search budget as our model, and report both its published and re-run numbers. Beating it means a higher patient-level AUROC with overlapping or better confidence intervals. Matching it within the interval while passing the faithfulness tests below also counts as success, because the claim is equal accuracy with intrinsic explanation, not raw accuracy.

Hand-computed spatial statistics. This baseline computes fixed neighborhood indices on the same cell calls: infiltration ratios, nearest-neighbor distances, and co-localization at fixed radii. It then fits $L_1$-penalized logistic regression with the same feature budget. This is the most dangerous competitor, and it also serves as the ablation for the learned states and scales.

Coordinate-shuffle ablation. Our model is trained on regions in which cell positions are permuted within each region. This keeps composition but destroys arrangement, isolating the spatial ingredient.

Flat concept model. A concept bottleneck over per-region marker averages tests whether relations matter beyond composition.

Why not just…?

Why not just compute standard spatial statistics and fit a linear model? This is a real threat. On measured cell maps it may match our model. No result on Lacuna reports such a baseline against the 0.71 GCN. Our design makes the comparison exact: same cells, same feature count, same penalty. The only differences are that our states and length scales are learned. If this baseline ties, we will report that the contribution reduces to an interpretable pipeline that matches graph networks. That is still a useful negative finding about the necessity of deep graph models.

Why not just keep the GCN and explain it with GNNExplainer? Post-hoc explanations can look plausible without reflecting what the model actually computed. We test this directly with tissue interventions. We relocate T cells in silico from stroma into tumor nests, or the reverse, and check whether each model's output moves in the direction its explanation implies. Our bottleneck model satisfies this test by construction. The GCN-plus-GNNExplainer pair may not. We measure how often its attributions predict its own response to the intervention.

Why not just use a pathology vision foundation model with saliency maps? In a mammography reader study, heatmaps served only as sanity checks (Friederike Jungmann et al.). Post-hoc attributions in WHE-CBM's localization tests placed only about 8% of their mass on the target. More importantly, H&E images lack the molecular identity that defines our concepts. The secondary experiment below tests whether niche concepts can be inferred from H&E at all.

Proposed Experiments

Primary experiment. We use the TNBC imaging mass cytometry cohort with nested, patient-level cross-validation. The cohort has only 63 patients, so we report bootstrap confidence intervals and pre-register the comparison against the GCN. We do not chase small differences.

Faithfulness checks. These are the in-silico relocation tests described above. In addition, we measure whether pathologists who are shown concept maps can identify which predictions the model got wrong. As a control, we compare against the same pathologists shown GCN attributions.

Secondary experiment: transfer to routine slides. We train an H&E-to-concept predictor on paired spot data from HEST. Its concepts are spot-level, because Visium spots span about 55 µm and contain several cells. We then test the bottleneck on TCGA-LUAD EGFR/KRAS prediction against STAMP. No reported result on Lacuna gives STAMP's numeric AUROC for these mutations, so both models will be run on identical splits.

Decisive checks. The proposal fails if any of three things happens:

  • The coordinate-shuffle ablation matches the full model.
  • The hand-computed statistics match it and the faithfulness advantage over GNNExplainer is absent.
  • Accuracy drops below the GCN's confidence interval.