Background/Purpose: Autoimmune diseases such as rheumatoid arthritis (RA) involve complex spatial organization of immune and stromal cells within inflamed tissues. As spatial molecular profiling methods grow in scale and resolution, there is increasing opportunity to identify clinically and pathophysiologically informative tissue structures through case-control comparisons. Existing approaches to define key pathological structures rely either on bespoke manual scoring systems such as the Krenn inflammation score or on clustering spatial neighborhoods of segmented and typed cells based on cell-type abundances. These approaches are limited by confirmation bias, segmentation error, loss of local spatial relationships among cells, and inability to model spatial overlap of multiple biological processes.

Methods: We developed variational inference-based microniche analysis (VIMA), an artificial intelligence-driven method for case-control analysis of spatial data. VIMA uses AI to learn numerical “fingerprints” of small tissue patches that capture their biological content, groups similar patches into thousands of overlapping “microniches” across samples, and identifies microniches whose abundance differs by condition. VIMA requires no cell segmentation or typing and can be used on any high-resolution spatial modality.

Results: We first applied VIMA to a 7-marker N&#3f22 immunohistochemistry (IHC) dataset in RA. Without prior labels, VIMA stratified samples into an “immune” and a “stromal” subtype (P=6.5e-4). The stromal subtype contained swathes of CLIC5+ staining and CD34+ regions; the immune subtype contained perivascular infiltrates and macrophages contacting sublining fibroblasts. These subtypes matched cell-type abundance phenotypes (CTAPs) defined previously in a much larger (N&#3f79) non-spatial single-cell dataset, showing that VIMA can extract information from IHC that previously required state-of-the-art sequencing at large samle sizes to uncover. VIMA also revealed striking within-sample spatial heterogeneity, with many RA samples containing patches from both the stromal and immune subtypes, challenging the notion of one subtype per patient. We confirmed similar heterogeneity in subsequently collected synovial samples profiled with spatial transcriptomics as well.We also applied VIMA to a 54-marker N&#3f34 CODEX dataset in UC. Here, VIMA was able to powerfully distinguish UC samples from healthy samples (P=6.5e-4), and then to identify a dose-response relationship between duration of TNF inhibition and absence of lymphoid aggregates (P=1.4e-3). VIMA can therefore detect clinically relevant spatial signals — including signatures of therapeutic immune modulation — across diseases, technologies, and tissue types.

Conclusion: VIMA enables flexible, high-resolution spatial analysis of autoimmune disease. It identifies biologically meaningful signals missed by existing tools and reveals within-sample heterogeneity, demonstrating the promise of AI-based methods for advancing our understanding of autoimmune disease pathogenesis.

Examples of VIMA’s microniches in the RA IHC dataset. We show a UMAP of all the patches in the dataset generated using the patch fingerprints, together with randomly selected patches drawn from 7 example microniches; these microniches are characterized by (clockwise from bottom-left): high CD34 intensity with interspersed CLIC5 signal, high CLIC5 intensity with minimal CD34 signal, high CLIC5 intensity but with a blurrier texture and interspersed DAPI signal, a vertical boundary between CLIC5-high/DAPI-low pixels and CLIC5-low/DAPI-high pixels, a DAPI staining artifact, DAPI-high infiltrate surrounding CD34-high regions that are likely blood vessels, and perivascular CD90-high pixels. All scale bars represent 100um.

VIMA analysis of the RA IHC dataset. A) plot of the first two VIMA-determined principal components, with each dot representing a sample and samples colored by RA subtype (CTAP) as determined using a larger scRNA-seq dataset. F: fibroblast, TF: T-cell/fibroblast, EFM: endothelial/fibroblast/myeloid, M: myeloid, TM: T-cell/myeloid, and TB: T-cell/B-cell. B) Results of VIMA case-control analysis for the “stromal” (F/TF/EFM) versus “immune” (M/TM/TB) subtypes: a UMAP of all the patches in the dataset is generated using the patch fingerprints, and the anchor patches for microniches passing the 10% FDR threshold are colored in proportion to the correlation between the abundance of each microniche and sample-level stromal vs immune status, with red denoting stromal-enriched and blue denoting immune-enriched. The centroids of the two subclusters of each set of patches, respectively, are indicated with numbers. The global P-value for the association test is overlaid. C) A subset of 6 samples from the dataset exemplifying both spatial homogeneity in local stromal vs immune status within a sample (the leftmost and rightmost samples) and spatial heterogeneity in stromal vs immune status within a sample (middle samples), with disease subtype indicated above each sample. Each patch in each sample is colored red if it is stromal-associated by VIMA or blue if it is immune-associated. D) Randomly selected patches from each of the two clusters of immune-associated patches. E) Log-odds ratios among the immune-associated patches for belonging to cluster 1 (positive log-odds) vs cluster 2 (negative log-odds) as a function of disease subtype, with 95% confidence intervals. F) Randomly selected patches from each of the two clusters of stromal-associated patches. G) Log-odds ratios among the stromal-associated patches for belonging to cluster 3 (positive log-odds) vs cluster 4 (negative log-odds) as a function of disease subtype. H) A volcano plot showing that a cluster-based rather than microniche-based case-control analysis using the patch fingerprints learned by VIMA does not identify as robust of a signal. The Bonferroni significance threshold is shown with a dotted line. Scale bars represent 100um throughout.

VIMA analysis of UC CODEX dataset. A) Results of VIMA case-control analysis for UC versus healthy: a UMAP of all the patches in the dataset is generated using the patch fingerprints, and the anchor patches for microniches passing the 10% FDR threshold are colored in proportion to the correlation between the abundance of each microniche and UC status, with red denoting to positive correlations and blue denoting negative correlations. The global P-value for the association test is overlaid. B) Distributions of average intensity per patch of selected markers among UC-associated and healthy-associated patches. The selected markers are grouped using dotted lines according to the relationship of their directions of effect to those in panel (F). C) Selected samples shown with cytokeratin staining in blue and contiguous areas of UC-associated patches outlined in red. D) Results of VIMA analysis of UC-associated patches only, comparing the samples with longstanding TNF inhibition (longTNFi) to those with no TNF inhibition, prior TNF inhibition only, or recent TNF inhibition only (little/noTNFi), with gold denoting positive correlation to longTNFi status and green denoting negative correlation. The global P-value for the association test is overlaid. E) Comparison of fraction of longTNFi-associated vs little/noTNFi-associated patches in each sample. For each sample, the percent of patches in that sample that anchor microniches significantly associated with longTNFi is plotted against the percent of patches in that sample significantly associated with little/noTNFi. Each sample is then colored by its detailed TNFi status. F) Distributions of average intensity per patch of selected markers among TNFi-associated and non-TNFi-associated patches. The selected markers are grouped using dotted lines as follows: markers that are higher in UC compared to control and in little/noTNFi compared to longTNFi (left), markers that are higher in UC compared to control but lower in little/noTNFi compared to longTNFi (middle), and markers that are lower in UC compared to control and lower in little/noTNFi compared to longTNFi (right). G) Randomly selected patches from the little/noTNFi-associated patches (left) and the longTNFi-associated patches (right).

« Back to ACR Convergence 2025

ACR Meeting Abstracts - https://acrabstracts.org/abstract/powerful-and-accurate-case-control-analysis-of-spatially-resolved-molecular-data-in-autoimmune-disease/