From Gene Expression to Tissue Morphology: Can Generative Models Uncover the Link?

Frederieke Lohmann, Alberto Valdeolivas, Jelica Vasiljević

Keywords: generative models, spatial transcriptomics, contrastive learning, digital pathology

TL;DR: Two-stage conditional generative framework that leverages spatial transcriptomics to infer tissue morphology from gene expression

Abstract: Spatial Transcriptomics technologies enable capturing gene expression within the native tissue context. Platforms such as 10x Visium and Visium HD integrate gene expression with histological imaging, providing a multi-dimensional view of tissue organisation. Motivated by the success of generative models in computer vision and natural language processing, we investigate the largely unexplored task of synthesising histological images directly from gene expression profiles. Leveraging recent advancements in Spatial Transcriptomics, particularly the 10X Visium HD platform, we introduce the first two-stage conditional generative framework to infer tissue morphology from near-whole transcriptome profiles. Competitive FID scores and a study involving multiple pathologists confirm that the synthesised images are plausible and that our framework generalises well to unseen standard Visium samples. Furthermore, model interpretation reveals connections between structurally relevant gene sets and specific morphological patterns, opening new avenues for studying the relationship between gene expression and tissue morphology.

Submission Number: 37