Mutant analysis is the core of biological/pathological research, and measuring spatial gene expression can facilitate the understanding of the disorganised tissue phenotype1-5. The large numbers of mutants are worth investigating; however, the high cost and technically challenging nature of experiments to measure spatial transcriptomes may act as bottlenecks6. Spatial transcriptomes have been computationally predicted from single-cell RNA sequencing data based on teaching data of spatial gene expression of certain genes7; nonetheless, this process remains challenging because teaching data for most mutants are unavailable. In various machine-learning tasks, zero-shot learning offers the potential to tackle general prediction problems without using teaching data8. Here, we provide the first zero-shot framework for predicting mutant spatial transcriptomes from mutant single-cell RNA sequencing data without using teaching data, such as a mutant spatial reference atlas. We validated the zero-shot framework by accurately predicting the spatial transcriptomes of Alzheimers model mice3 and mutant zebrafish embryos with lost Nodal signaling9. We propose a spatially informed screening approach based on zero-shot framework prediction that identified novel Nodal-downregulated genes in zebrafish. We expect that the zero-shot framework will provide novel phenotypic insights by leveraging the enormous mutant/disease single-cell RNA sequencing data collected.
