Spatial transcriptomics has enabled unprecedented insights into tissue organization by linking gene expression to histological context; however, the high cost, technical complexity, and limited coverage of sequencing-based technologies restrict their widespread application. Inferring biologically meaningful molecular states directly from routine histology remains a major challenge. Here, we introduce PATH (PAthway acTivation from Histology), a deep learning framework for predicting biologicalpathway activation from hematoxylin and eosin (H&E) stained images. PATH
is trained on paired spatial transcriptomics and histology data to learn pathway-level representations rather than gene-level expression. PATH
leverages a pre-trained Vision Transformer with Low-Rank Adaptation (LoRA) for efficient fine-tuning, combined with an adversarial learning
strategy to remove patient-, slide-, and dataset-specific confounding signals from the learned embeddings. By operating at the pathway level,PATH reduces noise and biological ambiguity inherent to gene-level prediction while improving generalization across datasets. Across multiplespatial transcriptomics datasets, PATH substantially outperforms gene-expression and pathway-based baselines, while exhibiting markedly reduced batch effects. We show that PATH captures biologically relevant processes, including immune signaling, cell cycle regulation, and oncogenic
pathways, and generalizes to high-resolution Visium HD slides containing tens of thousands of spatial locations. Overall, PATH demonstrates
that pathway-centric modeling enables robust, interpretable, and scalable inference of molecular states from histology, opening new avenues for
leveraging routine pathology images to study tissue biology and disease.

Co-authors: Asaf Madi, Roded Sharan