Abstract 2085: Visium HD combined with deep-learning-based cell segmentation on H&E images yield accurate cell annotation at single-cell resolution

Abstract Background: Bulk and single-cell next-generation sequencing (NGS) have been instrumental tools for characterizing gene expression profiles of tumor samples. However, the lack of spatial and cellular context limits their utility in investigating tissue architecture and cellular interactions in the tumor microenvironment (TME). NGS-based Spatial Transcriptomics (ST) technologies have gained increasing attention for their ability to provide spatial context of gene expression, but they have been constrained by low resolutions until the recent launch of 10X Genomics Visium HD platform. This platform achieves whole-transcriptome profiling at 2 μm resolution. However, there are complications in downstream biological interpretations with the default binning of Visium HD at 8 μm resolution. Methods: Primary tumor samples were collected from 2 patients with non-small cell lung cancer (NSCLC). The 10X Genomics Visium HD platform was used to generate high-resolution ST data. To study spatial expression at the single-cell level, we trained a cell-segmentation neural network and applied it to H&E images, generating cell segmentation masks. These masks were used to summarize raw Visium HD data at 2 μm resolution into single-cell-level gene counts. Clustering was performed on the generated single-cell level data and a large language model-based (LLM-based) cell type annotator was used to infer the cell types of each cluster. Results: The LLM-annotated cell clusters from Visium HD were highly consistent with pathologist’s annotations based on H&E images. Specifically, in the immune-active sample, 93.1% of cells in annotated lymphocyte regions were classified as lymphocytes based on cell-level ST counts, and 88.8% of cells identified in annotated tumor regions were classified as cancer cells. Cell-level ST clustering also identified a benign epithelium cell cluster that is confirmed by pathologist’s annotations. Similarly, in the immune-inactive sample, 72.8% of cells in annotated tumor regions and 95.8% of cells in annotated stroma regions were classified as cancer cells and stromal cells respectively. Conclusions: Our findings showcase the feasibility and advantages of analyzing Visium HD data at single-cell resolution using deep-learning-based segmentation models applied to H&E images, empowering clinical biomarker discovery and new mechanistic insights. Citation Format: Tianyou Luo, Mario G. Rosasco, Christine M. Hoeman, Sonal Khare, Daniel J. Gorski, Tim A. Rand, Richard A. Klinghoffer, Justin Guinney, Chi-Sing Ho. Visium HD combined with deep-learning-based cell segmentation on H&E images yield accurate cell annotation at single-cell resolution [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 2085.