SpatialFusion - preprocessed benchmark dataset

# Zenodo Repository Description ## Title SpatialFusion Benchmark Dataset: Preprocessed Spatial Transcriptomics and Histopathology Data Across Four Cancer Types SpatialFusion: A lightweight multimodal framework for pathway-informed spatial niche mapping Josephine Yates, Mitra Shavakhi, Toni K. Choueiri, Sabrina S. Camp, Eliezer M. Van Allen, Caroline Uhler bioRxiv 2026.03.16.712056; doi: https://doi.org/10.64898/2026.03.16.712056 ## Description This repository contains four preprocessed spatial transcriptomics datasets used to benchmark **SpatialFusion**, a multimodal spatial foundation-model-enabled framework that integrates paired histopathological and transcriptomic data to produce biologically grounded spatial neighborhood embeddings at single-cell resolution. The datasets span four cancer types profiled across two spatial transcriptomics platforms: 10x Xenium (ovarian, breast, and prostate cancer) and Visium HD (lung adenocarcinoma). Each file is stored in AnnData (`.h5ad`) format and contains raw gene expression, spatial coordinates in multiple reference spaces, cell-type annotations, pathologist-annotated spatial niche labels, and precomputed Leiden cluster assignments from SpatialFusion and all benchmarked baseline methods (NicheCompass, BANKSY, Nicheformer, scGPT-spatial, OmiCLIP, Novae, and Scanpy). These datasets were used to evaluate spatial niche rediscovery performance using clustering metrics (ARI, NMI, homogeneity, completeness) and spatial coherence metrics (PAS, CHAOS), as reported in the SpatialFusion manuscript. ## Files - `preprocessed_OVCA.h5ad` — Ovarian cancer, 10x Xenium, 338,695 cells × 5,101 genes - `preprocessed_BRCA.h5ad` — Breast cancer, 10x Xenium, 132,508 cells × 313 genes - `preprocessed_PRAD.h5ad` — Prostate cancer, 10x Xenium, 186,207 cells × 9,984 genes - `preprocessed_LUAD.h5ad` — Lung adenocarcinoma, Visium HD, 203,878 bins × 18,028 genes ## Keywords spatial transcriptomics, multimodal integration, spatial niches, tumor microenvironment, foundation models, H&E histopathology, Xenium, Visium HD, ovarian cancer, breast cancer, prostate cancer, lung adenocarcinoma, benchmarking