Benchmarking the translational potential of spatial gene expression prediction from histology

Abstract

Spatial transcriptomics has enabled the quantification of gene expression at spatial coordinates, offering crucial insights into molecular underpinnings of diseases. In light of this, several methods predicting spatial gene expression from paired histology images have offered the opportunity of enhancing the utility of readily obtainable and cost-effective haematoxylin-and-eosin-stained histology images. To this end, we conducted a comprehensive benchmarking study encompassing six developed methods. These methods were reproduced and evaluated using HER2-positive breast tumour and human cutaneous squamous cell carcinoma datasets, followed by external validation using The Cancer Genome Atlas data. Our evaluation incorporates diverse metrics which capture the performance of predicted gene expression, model generalisability, translational potential, usability and computational efficiency of each method. Our findings demonstrate the capacity of methods to spatial gene expression from histology and highlight key areas that can be addressed to support the advancement of this emerging field.