Pretrained gene representations transfer mean expression more broadly than spatial patterns in virtual spatial transcriptomics

Abstract

Virtual spatial transcriptomics models combine tissue images with pretrained gene representations to predict expression for unseen genes, but what these representations contribute is unclear. We separated their effects into overall gene-mean and spatial variation using held-out genes and individuals across four cohorts spanning three human brain regions and HER2-positive breast cancer. With fixed representations derived from regulatory DNA sequences and representations pretrained on single-cell transcriptomes, reductions in gene-mean error accounted for 91-96% of the total mean squared error reductions over matched random vectors. Mean-only models using gene representations without images retained 82-99% of the corresponding full-matrix correlation gain. Spatial gains were smaller, increased with expression variation in training tissue, and varied across cohorts and representations. Thus, pretrained gene representations broadly transfer mean expression but only selectively support spatial recovery, showing that gene-panel extension alone does not establish spatial recovery for new targets.