High-resolution virtual spatial transcriptomics from real-world H&E images in NSCLC

Background
Spatial transcriptomics (ST) offers critical value for mapping the tumor microenvironment (TME) and discovering immuno-oncology biomarkers, but high cost and operational complexity limit adoption in routine practice. Virtual ST offers a scalable alternative by utilizing deep learning to predict spatially resolved gene expression directly from standard, cost-effective Hematoxylin and Eosin (H&E) slides. However, many models are trained on legacy, low-resolution platforms with limited gene panels that fail to represent real-world variability, limiting clinical deployment. Here, we present a highly scalable, high resolution virtual ST pipeline that overcomes these limitations by leveraging a pathology foundation model and training on one of the largest real-world Visium HD NSCLC datasets to predict whole-transcriptome gene expression.

Methods
We developed a computationally efficient pipeline to predict spatial gene expression directly from H&E images. Morphological features were extracted from H&E tiles using the Virchow2 pathology foundation model. A dual-stream fully connected multi-layer perceptron (MLP) with shared backbone was trained to map high-dimensional morphological embeddings with ST profiles. The model was trained and evaluated on 34 primary and 12 metastatic NSCLC samples, comprising 53k tiles (~3.7 million cells) across 6 tissue sites. Model performance was evaluated using Spearman correlation (ρ).

Results
The model was applied to the validation set (10 samples, 13k tiles), achieving a mean correlation of ρ= 0.45 (IQR = 0.24-0.62) across all 18132 genes tested, and ρ= 0.61 (IQR = 0.57-0.71) across 500 highly variable genes. Enrichment analysis revealed top performing genes aligned with markers of distinct cellular lineages including plasma cells and myofibroblasts, while simultaneously capturing critical functional pathways including MYC signaling, interferon-gamma response and epithelial-mesenchymal transition. The model showed high predictive accuracy for key immune-related transcripts like CD3E (ρ= 0.67) and immune checkpoint markers including TIGIT (ρ= 0.69), as well as structural markers of tumor epithelial cells includingTACSTD2 (ρ= 0.80) and KRT18 (ρ= 0.79), and cancer-associated fibroblasts such asCOL3A1 (ρ= 0.83).

Conclusions
We report on a virtual ST model trained to predict gene expression from H&E slides, using a large real-world NSCLC dataset. Strong predictive performance was achieved using a simple architecture, underscoring the dense biological signal captured by modern pathology foundation models when paired with high-definition spatial omics. This highly scalable pipeline enables the cost-effective profiling of the TME directly from standard-of-care H&E slides, offering a powerful tool for characterizing spatial architectures in NSCLC.