Project Summary/Abstract Single cell RNA-seq (scRNA-seq) data have revolutionized our understanding of biology at cell level. Spatial transcriptomics further moves the field forward by providing spatial context of gene expression. These exciting techniques have been applied in many basic science or clinical research projects to understand living systems or the biological basis for disease diagnosis, treatment, and prevention. The data generated by scRNA-seq or spatial transcriptomics typically has high dimension (the number of genes) and large sample size (the number of cells or spatial spots). Many biological processes underlying the observed gene expression data are likely non-linear functions of high dimensional gene expression data. Large sample size combined with non-linear signals of high dimensional data makes deep learning an appropriate tool to analyze scRNA-seq or spatial transcriptomics data. Earlier deep learning works on scRNA-seq or spatial transcriptomics focus on un- supervised tasks, such as de-noising or clustering. For many biomedical applications, a natural next step is supervised analysis, e.g., comparing scRNA-seq or spatial transcriptomics between two conditions. There are much fewer works in this direction where deep learning methods face two general challenges: interpretability and noisy labels of single cells. In this project, we aim to address the interpretability challenge by a flexible method to incorporate gene annotation into deep learning. To work with single cells with noisy labels, we propose a mixture model that iteratively refines cell labels and the neural network that predicts cell labels. Our work on spatial transcriptomics focuses on using these data to train deep learning models to interpret histological images, particularly H&E stained histological images. Our method provides spatial annotation of histological images in terms of cell type proportions and interactions between any two cell types. Histological images are universally available in many clinical settings. In contrast, spatial transcriptomics is harder to scale due to cost and logistic challenges. Our method enables the transfer of knowledge from spatial transcriptomics to histological images. Once trained by an appropriate training dataset with both spatial transcriptomics and histological images, our method can be applied to analyze datasets with only histological images and assess their associations with phenotypic or clinical outcomes. In summary, our computation methods address fundamental questions on scRNA-seq or spatial transcriptomics data analysis and they are applicable for most basic science or clinical research projects that produce relevant data.