Abstract 2439: Spatial transcriptomics-informed latent diffusion models generate histopathological images reflective of tumor microenvironment

Abstract While digitized whole slide images (WSIs) can improve cancer prognosis algorithms, limited data curation given high image complexity/size can hinder performance in underrepresented populations. Synthetic WSI generation offers a solution by simulating tissue heterogeneity and anonymizing data. However, current methods for histopathology cannot reproduce tumor microenvironment (TME) and other biomolecular features with high fidelity at gigapixel scale, limiting further application. Spatial transcriptomics (ST) paired with H&E-stained WSIs can provide this missing context. We present a novel method for using ST to guide image generation by conditioning latent diffusion models (LDM) on 55-micron resolution gene expression data from Visium ST. Preliminary evaluation demonstrates that LDMs can reproduce realistic TME histological features of primary colon tumors. We profiled 41 colon tumors with paired 40X H&E WSIs and Visium ST. A total of 255, 744 WSI patches (512-pixel) were matched with 55-micron ST spots for subsequent LDM training. Inspired by single-sample gene set enrichment analysis methods, spot-level ST data were encoded into pathway representations via a CLIP text encoder using the top 50 genes, ranked through gene-wise comparison to non-zero median expression. A modified UNET was fine-tuned for 100, 000 iterations (learning rate: 1e-5) to denoise CLIP-encoded ST data into histology images. Two pathologists evaluated 90 held-out randomly selected tissue patches, comparing original and ST-generated images at the same tissue locations. Pathologists- blinded to image origin (real/generated)- assessed the abundance of tumor, stromal, immune, and fat features on a four-point scale and determined whether the original and generated images contained similar histology. Hierarchical bivariate ordinal regression assessed correlations between original/generated features and calculated intraclass correlation coefficients (ICC) as a measure of interrater agreement. Hierarchical logistic regression determined proportion of images with similar histology, with image/pathologist random intercepts. Pathologic examination showed strong correlations for tumor (ρ=0.71; p<0.001) and immune (ρ=0.67; p<0.001) TME features. Excluding artifacts, stromal (ρ=0.68; p<0.001) and fat (ρ=0.57; p<0.001) features also correlated. Inter-rater agreement was high/comparable for most features (Real/Generated ICC- Tumor: 0.95/0.96, Stroma: 0.92/0.88, Immune: 0.63/0.69, Fat: 0.57/0.38; p>0.05). Pathologists found 86% of images reflected similar histologies (95% CI: [69%-95%]). This study is the first to generate realistic histology from spatially co-registered transcriptomic data. Further refinement is needed to scale the approach to whole tissue slides for prognostic modeling in cohorts with limited data. Citation Format: Gokul Srinivasan, Minh-Khang Le, Zarif Azher, Harsimran Kaur, Fred Kolling, Keluo Yao, Ken S. Lau, Xiaoying Liu, Louis J. Vaickus, Joshua J. Levy. Spatial transcriptomics-informed latent diffusion models generate histopathological images reflective of tumor microenvironment [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2439.