Computational framework for multi-organ cell type annotation of HuBMAP single-cell spatial-omics data

Introduction: : Spatial omics technologies have transformed our ability to study tissue organization by capturing molecular and spatial information at single-cell resolution. As these technologies advance, an increasing number of datasets are being generated across a wide range of tissues and biological contexts.
One major initiative advancing this field is the Human BioMolecular Atlas Program (HuBMAP), which aims to map the structure and organization of healthy human tissues at single-cell resolution. To date, HuBMAP has produced over 4,300 datasets across 31 organs, all publicly accessible through its data portal. While these datasets provide rich spatial and molecular information, they currently lack cell type labels—a critical component for downstream biological interpretation. This gap significantly limits the broader utility of the datasets by the scientific community and constrains efforts to perform comparative, multi-organ analysis of cellular organization and tissue architecture.
To address this challenge, we developed GARDEN (Generalized Annotation pipeline for Resolving DiversE cell types in multi-orgaN spatial omics datasets) (Fig. 1). GARDEN provides a modular framework that integrates data standardization, marker-guided clustering, spatial data visualization, and cell type annotation. It is designed for flexibility across imaging platforms, tissue types, and biological contexts, and applies broadly beyond HuBMAP to other spatial omics datasets from both healthy and diseased tissues.
We aim (1) to provide an open-source, user-friendly pipeline that the broader scientific community can adopt for their own spatial datasets, and (2) to establish foundational, harmonized annotations across multiple organs, enabling integrative, multi-organ spatial analyses, including hierarchical cellular neighborhood analysis and cell-cell interaction studies.

Materials and

Methods: : CODEX datasets from HuBMAP were pre-processed by the HuBMAP HIVE team using Cytokit, a spatial omics analysis toolkit. The processed outputs, including single-cell level marker quantification and spatial coordinates, are available through the HuBMAP data portal and extracted for downstream analysis.
Marker names were first standardized across datasets, and the intersection of shared antibody markers was identified to ensure consistency. Cell-level marker intensities were then Z-normalized, performed within each donor for intestine datasets and within each dataset for other organs, to account for variability in staining intensity and acquisition conditions. Following normalization, cells with abnormally high total z-scores were removed as noise, and a nuclear intensity-based filter was used to retain cells with valid nuclear staining. Harmony integration was used to correct for batch effects where applicable.
We performed unsupervised Leiden clustering using a GPU-accelerated workflow to identify phenotypically distinct cell populations. Preliminary cluster annotation was based on canonical marker expression patterns, using both mean expression values and the proportion of cells expressing each marker. Annotated clusters were evaluated using Vitessce, an interactive spatial-omics visualization tool that overlays marker expression and cluster assignments on the raw images. Clusters containing mixed cell types were further refined by K-means sub-clustering or manual gating, then final cell type identity for each cluster were assigned.
Multi-cellular neighborhood analysis was performed as previously described. Briefly, Local cell type composition profiles, including each cell’s 10 nearest neighbors, were clustered, and distinct tissue structures were defined based on differences in cell type composition.

Results, Conclusions, and Discussions:: We first adapted the GARDEN cell type annotation pipeline into the HuBMAP Workspace, a cloud platform supporting direct dataset access (Fig. 2A). We then applied GARDEN to HuBMAP CODEX intestine datasets, covering over two million cells across 64 samples. After standardizing marker names, we identified 47 antibody markers shared across all datasets (Fig. 2B).
Following Z-normalization, imaging artifacts and noisy cells were filtered out based on total z-score thresholds and nuclear intensity (Fig. 2C, D). Batch effects were then corrected using Harmony integration (Fig. 2E) and Leiden clustering identified 73 clusters (Fig. 2F). To further demonstrate GARDEN’s generalizability, we applied the pipeline also to HuBMAP lymph node datasets (Fig. 2G).
After cluster splitting, merging, and validation back on raw CODEX images, we annotated 27 distinct cell types in intestine dataset (Fig. 2H). To evaluate GARDEN’s reliability, we quantitatively assessed the cell type annotations. Major immune, stromal, and epithelial populations were recovered, and the relative abundances of cell types matched known intestinal tissue composition (Fig. 3A). Several cell types, including Paneth cell, CD8+ T cell, etc., showed significant abundance differences between small and large intestine (Fig. 3B).
Spatial patterns were consistent across the 64 samples and reflected known intestinal architecture (Fig. 3C). In addition to recovering the expected epithelial crypt-villus zonation, our pipeline also captured immune cell enrichment within follicle-associated zones (Fig. 3D). To further investigate higher-order tissue organization, we performed multicellular neighborhood analysis and identified 22 distinct neighborhoods characterized by unique cell type compositions (Fig. 3E). Notably, B cell-enriched inner follicles and CD4+ T cell-enriched outer follicle structures were clearly delineated, consistent with known intestinal follicle organization (Fig. 3F).
These results validate the accuracy of our cell type annotations and demonstrate GARDEN’s ability to recover both known tissue structures and biological differences across tissues. Future developments will extend GARDEN to additional organs beyond the intestine, including lung, heart, kidney, etc., enabling broader exploration of multi-organ cellular organization and tissue architecture.
In conclusion, by providing standardized, scalable annotations across diverse tissues, GARDEN lays the foundation for comparative spatial biology and accelerates the broader utility of public datasets like those from HuBMAP.

Acknowledgements and/or References (Optional)::