Exploring Experimental and Computational Strategies for Efficient and Cost-effective High-Definition Spatial Transcriptomics through Super-Resolution Methods based on Histological Images

High-resolution (HR) spatially resolved transcriptomics (SRT) technologies have emerged as powerful but expensive tools for spatially mapping gene expression at single-cell resolution. To achieve HR SRT computationally, bioinformaticians have developed methods to transform low-resolution (LR) SRT data into HR data. This transformation is possible because high-resolution histological images are often available along with LR SRT and can be modeled to bridge the LR and HR worlds. However, this field lacks standardized benchmarking and comprehensive evaluation of these super-resolution methods. This registered report outlines a study to systematically evaluate and compare super-resolution methods for SRT data. We will assess these methods based on accuracy, scalability and usability using a diverse set of metrics. High-resolution SRT datasets generated by different technologies will be collected and intentionally lowered in resolution to serve as input for the benchmarked methods, with the original data serving as the ground truth. Our evaluation will include both in-sample prediction (enhancing existing low-resolution data) and out-of-sample prediction (inferring gene expression in regions with only histological data available). Importantly, this study aims to address three critical, yet often overlooked, aspects of SRT experimental design: 1, The influence of gene characteristics, such as overall expression, expression variability, and spatial variability, on prediction performance. This will enable the selection and application of super-resolution methods according to the genes of interest. 2, The impact of SRT-measured region selection on out-of-sample prediction performance, guiding optimal tissue sampling strategies. 3, The relationship between initial spatial resolution, spot spatial layout, and super-resolution performance across different tissue types, identifying tissue-specific saturation points where further increases in resolution yield diminishing returns. By addressing these questions, our study will provide valuable insights for achieving efficient and cost-effective spatial transcriptomics. We anticipate that our findings will guide researchers in selecting appropriate super-resolution methods, designing more efficient SRT experiments both experimentally and computationally, and making informed decisions about resource allocation in spatial transcriptomics studies.