Zero-shot reconstruction of mutant spatial transcriptomes.

Abstract

Mutant analysis is the core of biological/pathological research, and measuring spatial transcriptomes can facilitate the understanding of the disorganized tissue phenotype. However, the high cost and technical challenges of spatial transcriptome experiments hinder the investigation of large numbers of mutants. Spatial transcriptomes have also been computationally predicted from single-cell RNA sequencing data using teaching data of spatial expression of certain genes, but the lack of teaching data for most mutants remains challenging. In various machine-learning tasks, zero-shot learning offers potential for predictions without teaching data. Here, we provided ZENomix, the zero-shot framework for predicting mutant spatial transcriptomes without teaching data (e.g., mutant spatial atlases). ZENomix accurately predicted spatial transcriptomes in Alzheimer's model mice, Alzheimer's human brains, and Nodal-signaling-deficient mutant zebrafish embryos. We proposed a ZENomix-based screening approach, identifying Nodal-downregulated genes in zebrafish. We expect that ZENomix offers phenotypic insights by leveraging the enormous amount of mutant/disease single-cell RNA sequencing data.