Cellular transcriptional states are shaped by local tissue context, yet quantifying how gene expression varies with proximity to different cell types is challenging in cell-resolved spatial transcriptomics. These assays are often sparse and vulnerable to contamination from neighbouring cells via diffusion, imperfect segmentation and cell overlap. Such technical factors complicate distinguishing true cell-state changes from artefacts. The authors present PACE (Proximity-Associated Changes in Expression), a modelling framework designed to quantify cell-type-resolved proximity effects on gene expression while addressing these challenges.
PACE is built as a hierarchical empirical Bayes model that uses partial pooling to stabilise inference across genes and cell types. Partial pooling shares information across related parameters, which helps regularise estimates when individual gene or cell-type counts are sparse. The empirical Bayes hierarchy facilitates data-driven estimation of prior distributions, enabling more robust identification of proximity-associated expression changes than methods that analyse each gene independently.
A core feature of PACE is explicit separation of contamination from biologically meaningful spatial associations. The framework recognises that measured transcripts in a cell can reflect both true endogenous expression and contributions from neighbouring cells due to technical spillover. By modelling these components, PACE aims to reduce false inference driven by contamination and better isolate genuine proximity effects. The approach therefore improves interpretability and reliability of spatially resolved differential expression analyses in the presence of common spatial assay artefacts.
Beyond gene-level effects, PACE identifies coordinated transcriptional programs underlying each proximity effect. Rather than treating genes as independent signals, the framework seeks groups of co-regulated genes that respond together to proximity, which can highlight biologically meaningful pathways and cellular programs implicated by cell–cell spatial relationships. This program-focused output supports mechanistic interpretation of how local tissue organisation shapes cellular states.
The authors applied PACE to breast cancer tissue profiled with the Xenium platform. In this dataset, PACE revealed tumour-associated reprogramming of stromal cells and macrophages located at tumour interfaces. These findings indicate that proximity to tumour cells is associated with distinct transcriptional states in adjacent stromal and immune populations. The source abstract reports these qualitative results but does not provide detailed quantitative metrics or lists of specific genes or pathways in the previewed content.
PACE was also applied to melanoma samples profiled with the CosMx platform. In these data, the method identified fibroblast responses to tumour proximity, including extracellular matrix transcriptional programs. Importantly, these extracellular matrix programs differed between tumours from patients with progressive disease and those with stable disease following immunotherapy, suggesting that proximity-associated fibroblast programs may relate to clinical response. The abstract summarises these comparative observations without reporting detailed program composition or statistical summary in the preview.
Across applications, PACE produced interpretable results linking tissue architecture to cellular reprogramming at tumour interfaces. In breast cancer tissue, adjacency to tumour regions corresponded with altered stromal and macrophage states. In melanoma, tumour-proximal fibroblasts upregulated matrix-associated programs that varied with patient outcome after immunotherapy. These examples illustrate how a proximity-aware model can reveal biologically and clinically relevant spatial patterns that might be obscured by technical contamination or sparse measurements.
PACE provides a principled modelling approach for investigators using cell-resolved spatial transcriptomics who need to disentangle contamination from true local cell–cell effects. By combining hierarchical empirical Bayes estimation, partial pooling and program-level inference, PACE aims to improve robustness and interpretability of proximity analyses. The abstract emphasises the framework’s applicability to different spatial platforms (Xenium, CosMx) and highlights its potential to reveal context-dependent reprogramming relevant to tumour biology and treatment response. The source preview does not report implementation details, benchmarking statistics or full program gene lists; readers should consult the full preprint for methodological specifics, performance evaluation and supplementary results.