Connecting Gene Expression and Tissue Morphology with Conditional Generative Models

Frederieke Lohmann, Alberto Valdeolivas, Jelica Vasiljević

Keywords: spatial transcriptomics, digital pathology, gans, diffusion models, contrastive learning

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

Abstract: Inferring tissue morphology from gene expression remains widely unexplored. We present a two-stage conditional generative framework that leverages for the first time spatial transcriptomics data from the Visium HD platform to demonstrate this inference is feasible. Starting from near-whole-transcriptome profiles, the model synthesizes histology-like images that are plausible, as validated by FID scores and expert review. Model interpretation further reveals biologically meaningful links between specific genes and morphological patterns.

Submission Number: 78