The paper introduces iSCALE, an innovative machine learning framework designed to significantly enhance spatial transcriptomics (ST) analysis for large tissue samples, overcoming the size and resolution constraints of conventional platforms. iSCALE achieves this by integrating gene expression data from multiple small ST sections (daughter captures) with large-scale histology images (mother images) to predict super-resolution gene expression across the entire tissue. Through comprehensive benchmarking on gastric cancer and normal tissues, the authors demonstrate that iSCALE accurately reconstructs tissue architecture and performs robust cell type annotation, outperforming existing methods like iStar and RedeHist. Finally, applying iSCALE to large human multiple sclerosis (MS) brain samples reveals crucial, cellular-level characteristics associated with lesions, demonstrating the frameworkâs utility in complex disease research and its potential for cost-effective out-of-sample prediction using only histology.
References:
