2026-09-28
SpaCEy models tumours as spatial cell graphs and predicts survival with an explainable GNN, reaching C-index 0.72 on breast cancer cohorts and beating all 24 non-spatial baselines.
Multiplexed imaging can now measure a few dozen proteins in every single cell of a tissue slice, along with its coordinates. The bottleneck is analysis: connecting how cells are arranged in space to how long a patient survives or whether the disease progresses. Classical survival models (Cox regression, random survival forests, survival SVMs) collapse an image into one feature vector, and spatial structure is lost at that step. Existing graph neural network approaches keep the geometry but take predefined cell-type labels as input, so the patterns they can find are framed by those annotations before learning starts; their interpretations rarely go beyond latent-space plots or per-cell importance scores. A pathologist needs a map overlaid on the tissue, not a vector.
SpaCEy (Spatial Clinical Explainability, from Heidelberg University, University of Zurich and collaborators) works in three stages:
The load-bearing choice is constraining explanations to connected subgraphs, so importance lands on tissue you can point at. Cell-type annotations never enter the model; they are only used for post hoc interpretation by compartment.
On lung adenocarcinoma (416 patients, imaging mass cytometry), progression classification over five-fold cross-validation: SpaCEy reaches 0.68 accuracy, 0.68 F1 and 0.62 AUC, against 0.61/0.59/0.57 for Ali et al. and 0.55/0.52/0.52 for SPACE-GM.
For censoring-aware survival prediction on two breast cancer cohorts, against 24 non-spatial baselines (four survival models x several feature aggregations) on identical splits:
| Cohort | SpaCEy C-index | Best non-spatial baseline | Relative gain |
| JacksonFischer (720 samples) | 0.720 ± 0.061 | 0.617 ± 0.092 | 16.5% |
| METABRIC (460 images) | 0.688 ± 0.038 | 0.570 ± 0.071 | 20.7% |
The explanations were validated three ways. The most direct is deletion: removing explainer-flagged nodes degrades the model far more than removing matched non-important nodes, with a mean output drop of 0.0818 versus 0.0281 at 10% removal, nearly 3x, and the pattern holds monotonically across removal levels from 1% to 80%. Synthetic data provided a known answer: on simulated tumour graphs, intratumoral CD8 T cells scored 0.0467 while stromal CD8 scored 0.0162 (p 8.3e-14), so the explainer picked up exactly the spatial motif it was built to contain. Biologically, the two breast cancer cohorts were trained independently and both enrich EMT and stromal markers (fibronectin, beta-catenin) in low-survival important regions while high-survival regions track hormone-receptor signals; Leiden clustering of the embeddings yields three subtype-agnostic groups with significantly separated survival curves (log-rank p = 0.00253 for JacksonFischer, p = 0.00371 for METABRIC), and the split holds within the HR+HER2- subtype alone (p = 0.031 and 0.015).
Cross-platform and scale: on a CODEX colorectal cohort, classification reaches AUPRC 0.846 and retains 0.801 when transferring across batches; a million-cell sample takes 7.42 s per forward pass, a 1.5-billion-cell epoch about 3 h, with peak GPU memory near 1.06 GB. Important regions are dominated by tumour (55.2% tumour, 29.3% stroma, 13.5% immune in JacksonFischer).
Anyone analysing spatial omics knows the situation: the model predicts well, nobody can say why, and the result stalls there. SpaCEy puts prediction, stratification and explanation in one pipeline, and the explanation is checkable, since deleting the cells it names really does move the prediction. For computational teams trying to convince biologist collaborators, that is a practical difference, not a rebranding of interpretability. The engineering bar is also low: a single RTX A6000 suffices, memory is bounded by tile size, and mid-size labs can work at whole-slide scale directly. Code and preprocessed data are open (GitHub saezlab/SpaCEy, archived on Zenodo).
The authors list several: only spatial proteomics is covered, and protein abundances give a partial view of cell state; explanation quality is capped by model performance, so at roughly 0.68 accuracy the interpretations deserve caution; importance scores are unsigned magnitudes that say which regions matter but not in which direction, and they are not causal evidence; crop-sensitivity tests show only moderate robustness to partial fields of view, and ROI restriction remains a real limitation.
Two more points stand out. Absolute performance is middling: AUC 0.62 on lung adenocarcinoma is far from decision-support grade, so the current use is hypothesis generation. And the survival gains are all measured against non-spatial methods; the two spatially aware baselines were compared only on the lung classification task, so a spatial-method-versus-spatial-method survival comparison is missing.