Article: Accurate and well-powered case–control analysis of spatial molecular data – (October 15, 2026)

Spatial profiling of tissue samples has long been a foundational tool in biology^1,2 and is now a pillar of biomedical research^3,4,5,6. In recent years, the scale and richness of spatial molecular modalities have dramatically increased, and modern modalities now profile scores of proteins and thousands of genes at high spatial resolution^{7,8,9,10,11,12,13,14}; however, this richness complicates systematic analysis, and traditional approaches in cancer^15,16 and elsewhere¹⁷ have relied on manually constructed scoring systems. This limits discoveries to those that fit existing paradigms and scales poorly as more genes and proteins are assayed.

Thus, the development of new computational methods for high-resolution annotation of tissue structures in spatial molecular data has become an active area of research. Modern methods based on local abundances of different cell types^16,18,19, spatially constrained clustering²⁰ and aggregation^21,22, spatial factorization^23,24, graph neural networks^25,26,27 and others^28,29,30 have enabled identification of granular tissue structures. More recent work has begun to explore the intuition that neural network architectures from image analysis can be built upon to provide powerful and flexible tissue representations^31,32,33.

Here, we focus specifically on the question of statistical case−control analysis in spatial data: given a spatial dataset with multiple case and control samples, are there patterns that are statistically significantly different between cases and controls, after multiple-testing correction? One simple strategy is to annotate the samples in such a dataset with a modern tissue-annotation method, use the annotations to cluster regions of tissue into ‘niches’ with similar biology, and then to test for differential abundance of the niches in cases versus controls; however, while this approach has been used successfully to find case−control associations^{16,19,21,22,32}, it may be under-powered if the tissue annotation is not sufficiently expressive, if the case−control differences are obfuscated by the hard clustering step, or if the tissue annotation is too sensitive to sample-specific idiosyncrasies.

To address this challenge, we introduce variational inference-based microniche analysis (VIMA), a method for statistical case−control analysis of spatial molecular data. VIMA uses neural network-based techniques from image processing to learn compact, informative ‘fingerprints’ for each of many small tissue patches in a spatial dataset, similarly to some recent methods^32,33; however, VIMA introduces three innovations. First, it uses a conditional variational autoencoder (cVAE) architecture designed to learn fingerprints that are minimally influenced by sample- or batch-specific effects, thereby enhancing sensitivity to true case−control signals. Second, VIMA leverages the stochasticity of neural network optimization by training an ensemble of cVAEs rather than just one, resulting in multiple representations that can capture complementary aspects of the same dataset while also increasing replicability. Third, rather than using the fingerprints to cluster the tissue patches into discrete groups, VIMA instead defines many small, potentially overlapping groups of highly similar tissue patches, which we term microniches, and quantifies the abundance of each microniche in each sample. VIMA then performs statistical tests to identify nontrivial correlations between the abundances of the microniches and case−control status, meta-analyzing across the ensemble of cVAEs.

VIMA produces both a global statistical test for the presence of any case−control difference in the spatial data as well as a local statistical test identifying the set of specific patches explaining that difference at a given false discovery rate (FDR) threshold. VIMA can be applied to any spatially resolved molecular technology, is well powered even at the modest sample sizes typical of research cohorts, and avoids traditional, parameter-intensive preprocessing steps such as cell segmentation, clustering of cells into discrete cell types, or clustering of tissue regions into discrete niches. VIMA produces properly calibrated P values and so can be used for statistical hypothesis testing, a fact that we confirm in simulations.

We apply VIMA to three spatial molecular case−control datasets containing four case−control signals and generated with three different technologies in rheumatoid arthritis (RA) (seven-marker immunofluorescence microscopy), ulcerative colitis (UC) (52-marker CODEX) and dementia (140-gene MERFISH). In each case, VIMA identifies spatial features that differ significantly between cases and controls and that both recapitulate known biology and provide new insights. We also benchmark VIMA against seven state-of-the-art methods and find that most of VIMA’s signals are not detectable with these methods. Finally, we show, using an ablation study, that each new component of VIMA is necessary for its performance.

…

Continue to read the full article here