Independent benchmark of H&E-based gene expression prediction in skin

Abstract

Given the widespread availability of H&E slides, there is considerable interest in determining whether molecular information can be inferred directly from tissue morphology, potentially reducing the need for costly spatial transcriptomic profiling. We assessed three state-of-the-art methods for predicting single-cell gene expression from H&E images across three skin disease contexts and two Xenium panels. As controls, we included simple linear regression models trained on embeddings from multiple foundation models, totalling 16 models evaluated in this study. We show that all models performed poorly: for most genes, prediction accuracy was near zero, and reliable predictions were largely restricted to keratinocyte-associated genes. Predicted expression failed to preserve cell-type identity and spatial organisation, with only keratinocytes forming coherent clusters, while immune, fibroblast, and other dermal populations were extensively mixed. Notably, simple ridge regression on pretrained embeddings matched or outperformed the more complex published architectures, indicating that the predictive signal originates primarily from image representations rather than model design. Our results demonstrate that current H&E-based gene expression prediction methods are not yet suitable for single-cell-level interpretation of spatial transcriptomics in skin tissue.