Spatial transcriptomics (ST) offers unprecedented insight into how gene expression patterns are organized within the tissue microenvironment, but its adoption in both research and clinical practice has been slowed by the technical, financial, and logistical barriers of laboratory-based ST. A growing body of work demonstrates that artificial intelligence (AI) can bridge this gap by inferring ST directly from routine hematoxylin and eosin (H&E) histopathology slides, unlocking the potential of ST at scale.

In this joint session, we will present two complementary applications of AI-inferred ST in colon cancer. The first highlights how image-based AI models trained on paired H&E and Visium data can accurately reconstruct spatial gene expression profiles across tumors. Applied to large whole-slide image datasets, these inferred maps capture prognostically relevant tumor biology, yielding strong prediction of T-stage and uncovering upregulated genes and pathways linked to disease progression. This work demonstrates the fidelity of AI-inferred ST and its capacity to serve as a new class of biomarkers.

The second application focuses on tumor purity estimation (TPE), a critical parameter for genomic testing and treatment decision-making. Traditional visual purity estimates are subjective and prone to overestimation, while manual annotation or IHC-based methods lack precision and scalability. By integrating deep learning with ST-guided ground truth, we demonstrate a scalable framework for estimating tumor purity directly from H&E images. Our model achieves high concordance with both Visium-informed tumor cell proportions and independent HoVerNet-derived epithelial fractions, showing feasibility in external cohorts. This approach offers rapid, reproducible, and generalizable TPE, with potential to replace manual scoring in both research and clinical settings.

Together, these studies illustrate how AI-informed spatial transcriptomics can connect morphology and molecular state, expand the discovery of actionable biomarkers, improve tumor purity assessment for molecular testing, and ultimately enhance precision oncology. We will also discuss future directions including multi-site validation, 3D ST-H&E integration, and adaptation of these methods to other tumor types and problems in translational digital pathology.