Finding differentially expressed genes between cell fates predicted by image-based deep learning

Abstract

Finding early trigger genes involved in cell-fate-determining processes is important for understanding molecular mechanisms of, e.g., differentiation and disease progression. One of the powerful tools for the finding is hypothesis-free omics measurements, e.g., gene expression analysis (transcriptome analysis) by RNA sequencing (RNA-seq). However, because whole single-cell RNA-seq requires cell disruption and the fate of the disrupted cell is generally unknown, it is difficult to find fate-related genes by single-cell RNA-seq profiles, especially in the early stages of cell-fate determination. Meanwhile, deep learning has successfully predicted cell fates using individual cell images. Here, we developed an approach by integrating image-based cell-fate prediction using deep learning and single-cell whole-transcriptome analysis to find differentially expressed genes (DEGs) between different predicted fates. As a proof of principle, we applied this approach to cells fated to die and survive. First, we applied temporary heat stress to a mammalian cell line to induce a certain fraction of cells to die, and performed time-lapse imaging to observe this process. Second, we made image-based deep learning models trained with our dataset for the cell fate prediction (survival and death). Third, we picked the cells after another time-lapse imaging and performed single-cell RNA-seq. Finally, we compared the transcriptomes between cells predicted to die and survive. We successfully detected the DEGs when the transcriptomic profiles did not show clear multiple clusters that may correspond to the heat-induced different fates in a dimension-reduced plane. Our approach may contribute to a deeper understanding of cell-fate regulation and new molecular marker detection.

Caption of Graphical Abstract

We developed a platform to find candidate genes associated with early-stage cell-fate determination by combining label-free image-based deep learning, picking of microscopically observed single cells, and RNA sequencing. We applied temporary heat stress to cells and identified cell fates (survival and death) using time-lapse imaging. Then, we made deep learning models to predict these fates using cell images at multiple timepoints. Finally, we picked the observed cells, performed RNA-sequencing, and compared the transcriptomes between cells predicted to survive or die. This approach enabled the detection of differentially expressed genes between cells predicted to have different fates.