Superresolved Spatial Transcriptomics Transferred from a Histological Context

Shu Wang,Xiaocheng Zhou,Yan Kong,Hui Lu

DOI: https://doi.org/10.1007/s10489-023-05190-3

IF: 5.3

2023-01-01

Applied Intelligence

Abstract:Spatially resolved transcriptomics (SRT) is a vital technique in biology that allows for gene expression measurement at the resolution of individual spots while preserving spatial information. However, owing to technical limitations, single-spot resolution often includes data from multiple cells, leading to suboptimal results and opportunities for improvement. In this study, we propose a deep learning-based, plug-and-play method for enhancing spot resolution to obtain higher-resolution SRT data. Our approach involves training a convolutional neural network (CNN) model and introducing a shift-predict operation to obtain superresolution spots. Using a human breast cancer SRT dataset, we demonstrate that our method achieves 9 × superresolution, outperforming traditional superresolution techniques. Crucially, our method decreased the mean squared error (MSE) to 1.379 for all genes, 2.287 for tumor-related genes at 4 × superresolution, 1.866 for all genes, and 3.371 for tumor-related genes at 9 × superresolution, reflecting substantial improvements compared to the traditional approaches, including Gaussian RBF, multiquadric RBF, linear RBF, resize-predict, bilinear, and bicubic methods. Furthermore, we verify our method's effectiveness using external and simulated datasets. Our proposed method offers a substantial advancement in SRT by enabling higher-resolution gene expression data generation. By providing a deeper understanding of gene expression patterns and their underlying biological significance, this method contributes to progress in biology and medicine.