Presentation Information
[POS-63]A machine learning method to reconstruct mutant spatial transcriptomes
*Yasushi Okochi1, Takaaki Matsui2, Shunta Sakaguchi1, Takefumi Kondo3, Naoki Honda1 (1. Nagoya University (Japan), 2. NAIST (Japan), 3. RIKEN BDR (Japan))
Keywords:
Spatial transcriptomes,machine learning,mutant analysis
Mutant analysis plays a central role in biological and pathological research, and the spatial gene expression of mutants can greatly facilitate our understanding of phenotypic manifestations such as tissue disintegration. However, considering vast number of mutants that worth to investigate, experimental acquisition of spatial transcriptomes for mutants remains challenging with technically demanding and cost-prohibitive nature. Although several computational approaches have attempted to predict spatial transcriptomes from single-cell RNA sequencing (scRNA-seq) data by referencing the spatial expression patterns of a limited number of genes, these methods rely on the availability of spatial reference data, which is often lacking for most mutants. This scarcity of training data poses a significant obstacle to broader application. Zero-shot learning, which enables general prediction without the need for task-specific training data, has emerged as a promising paradigm across a range of machine learning domains. Building on this concept, we developed a novel zero-shot framework for predicting the spatial transcriptome of mutant organisms directly from their scRNA-seq data, without requiring any spatial reference atlas or supervised training specific to the mutant. To validate the effectiveness and generalizability of our method, we applied it to two distinct biological systems: a mouse model of Alzheimer’s disease and zebrafish embryos genetically deficient in Nodal signaling. In both cases, our method successfully reconstructed spatial gene expression patterns with high accuracy, despite the absence of mutant-specific spatial supervision. Furthermore, we introduced a spatially-informed screening strategy based on our method’s predictions, through which we identified previously unrecognized downregulated genes associated with the Nodal signaling pathway in zebrafish embryos. We expect that our method will serve as a powerful computational tool for extracting spatial insights from the large amount of scRNA-seq data of mutants and disease models, advancing our ability to decode complex spatial phenotypes in both basic and translational research contexts