Nucleus feature profiles from GTEx histological images across 12 tissues

This dataset contains nucleus features extracted from GTEx histological images using QuPath across 12 human tissues. It also includes summary statistics of these features, which were applied in image quantitative trait loci (imageQTL) analyses and differential expression (DE) analyses. The dataset enables the exploration of relationships between nuclear morphology, genetic variation, and gene expression. By providing feature measurements across multiple tissues, it serves as a resource for integrative studies in computational pathology, imaging genomics, and functional genomics. The histological image data used in these analyses were obtained from the GTEx Portal on 10/01/2022, and we acknowledge the GTEx Consortium for providing these publicly available resources.

Methods
QuPath v0.4.3 was used to extract nucleus features from GTEx WSIs. QuPath is publicly available (https://github.com/qupath/qupath/releases). Default settings were followed as outlined in the QuPath tutorial: https://qupath.readthedocs.io/en/stable/docs/tutorials/index.html. The image type was set to 'brightfield H&E,' and stain vectors were optimized. Subsequently, we ran a pixel classifier to identify relevant tissue areas. Cell and nucleus detection was performed using QuPath’s cell detection tool. Finally, the extracted cell and nucleus detection measurements for each WSI resulted in a structured dataset for further analysis. We configured the pipeline for each tissue type to accommodate different tissue morphologies. This automated feature extraction was crucial for ensuring consistent and objective quantification of tissue characteristics across all slides.

Four key nucleus features were included in the summary statistics: (1) nucleus area, (2) nucleus circularity, (3) nucleus eccentricity, and (4) nucleus-to-cell area ratio. We employed several parameters to characterize the distribution within each WSI. Outliers were filtered using a 1.5 interquartile range. We then applied kernel density estimation, partitioning the data into 500 bins, and selected the value corresponding to the highest density, representing the mode of the entire distribution. Additionally, we included the first quartile (Q1), second quartile (Q2), third quartile (Q3), mean (here we used a 90% trimmed mean), and standard deviation to describe the distribution. These six representative values for each nucleus feature were then utilized in subsequent analyses, resulting in 24 nucleus features in total.