Pix2Path: Integrating Spatial Transcriptomics and Digital Pathology with Deep Learning to Score Pathological Risk and Link Gene Expression to Disease Mechanisms

Abstract

Spatial transcriptomics (ST) provides high-resolution mapping of gene expression within tissues, and integrating ST with digital pathology can offer unprecedented insights into the molecular mechanisms underlying various diseases. However, existing methods primarily focus on aligning these two distinct datasets, often neglecting the causal connections between spatial gene activity and pathological phenotype. We introduce Pix2Path, a deep learning-based approach utilizing conditional generative adversarial networks (cGANs), to bridge the gap between spatial transcriptomics and digital pathology. Pix2Path can process data from various spatial transcriptomics (ST) technologies, assess pathological risk scores across different conditions, and supports a leave-one-out spatial in silico gene perturbation strategy. As demonstrated in AD A{beta} plaques pathology, this approach allows to link gene expression changes to tissue morphology and pathology without relying on predefined conditions, providing a new perspective on understanding disease mechanisms.