Figure 1

Overview of DeepSpaCE. Deep learning model for Spatial gene Clusters and Expression (DeepSpaCE) is a method for predicting gene-expression levels and transcriptomic cluster types from tissue spot images. DeepSpaCE is composed of two parts: the model training and gene-prediction parts. In the case of using semi-supervised learning as an option, unlabeled images are used to improve the prediction accuracy with predicted proxy labels. As practical applications of DeepSpaCE, we conducted super-resolution of spatial gene expression and tissue section imputation. (a) Super-resolution was used for predictions with unmeasured spot images (e.g., images among spots whose expression profiles were measured using the in situ capturing platform or images on spots with technical errors). Left spatial expression pattern shows that some spots lack expression value because of a technical problem such as potential permeabilization error (dotted circle). Right image shows an additional spatial expression pattern imputed by DeepSpaCE, and its highly expressed region in the center of the section (dotted line). It is challenging to infer a functional boundary such as cancer infiltration from spatial expression profiles of sparse spots (left). Spatial expression profiles of dense spots imputed by DeepSpaCE and their gene annotations enable delineating a functional boundary. (b) Tissue section imputation was used to predict gene-expression levels in one of the tissue sections within consecutive sections. By using DeepSpaCE, the unmeasured spatial expression profiles of the slide (red frame) can be imputed by at least one adjacent slide (black frame) whose expression profiles were measured using the in situ capturing platform.