Sep 2026· Cancer research and treatment : official journal of Korean Cancer Association· 0 citations
Medicine
TL;DR
It is argued that the most credible translational goal is not to infer every transcript from H&E, but to identify biologically coherent spatial states with reproducible morphological correlates that can be validated across independent platforms and cohorts before being mapped across large archival collections.
Abstract
Histopathology preserves tissue architecture, whereas spatial transcriptomics measures gene expression within its native spatial context. Integrating these modalities with artificial intelligence (AI) can link morphology to spatial molecular states and scale spatial discoveries to routine pathology cohorts. Early hematoxylin and eosin (H&E)-to-spatial models demonstrated the feasibility of local gene-expression prediction. However, transcriptome-wide reconstruction remains constrained by gene-specific predictability, molecular sparsity, image quality, platform heterogeneity, and limited generalization. Xenium, CosMx, and related technologies are moving molecular supervision from multicellular spots toward cellular and subcellular units. In parallel, OmiCLIP exemplifies the shift from task-specific regression toward reusable visual-omics representations. We review these experimental and computational transitions and their applications in tumor microenvironment (TME) analysis, biomarker discovery, prognosis, and treatment-response prediction. We argue that the most credible translational goal is not to infer every transcript from H&E, but to identify biologically coherent spatial states with reproducible morphological correlates. Predictive uncertainty should be quantified, and these states should be validated across independent platforms and cohorts before being mapped across large archival collections. Future progress should extend this framework beyond static two-dimensional transcriptomic maps toward spatial multi-omics, longitudinal morphomolecular trajectory prediction, and three-dimensional tissue modeling.
Spatial transcriptomics (ST) maps gene expression within intact tissues, but its cost and technical complexity limit broad adoption. A growing body of deep-learning methods now seeks to infer ST directly from routine histology images, especially hematoxylin and eosin (H&E)-stained slides, with the aim of turning archiv...
Ke-Xin Xu, H. Koohy, A. Antanaviciute· EXO· 0 citations
Spatial transcriptomics (ST) enhances single-cell RNA sequencing by revealing transcript distribution, offering critical insights into heterogeneous diseases such as breast cancer. However, the high cost and lengthy processes of generating high-quality ST data limit clinical application. Recent deep learning methods pr...
STP-BENCH is presented, a standardized benchmark for virtual ST models and systematically examines which genes and gene sets are recoverable from histomorphology, and evaluates the downstream biological utility of predicted profiles through cell-type deconvolution and spatial domain identification.
Youngmin Chung, Hun-Ha Ji, Andrew H. Song et al.· 0 citations
A workflow for sequential 10x Genomics Xenium spatial transcriptomics, COMET cyclic immunofluorescence, and haematoxylin and eosin histological staining on the same formalin-fixed paraffin-embedded tissue section is presented, illustrating the widespread applicability of the workflow beyond a single tissue type.
Emily Duchini, Colista Tsao, J. Madore et al.· bioRxiv· 0 citations
MoSPR (Morpho-Spatial Program Regression), a linear framework that couples an adjacency-informed histology representation with a low-rank molecular basis, provides an interpretable link between spatially coherent macrostate regions and their associated molecular programs.
Dongmyung Shin, Geong-Ran Lee, Yesung Cho et al.· arXiv.org· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.