Statistical method to rank spatially variable genes adjusted for mean-variance relationship

Author(s): Kinnary Shah,Boyi Guo,Stephanie Hicks

Affiliation(s): Johns Hopkins Bloomberg School of Public Health

Social media: https://twitter.com/kinnaryhshah

Recently, spatially resolved transcriptomics technologies have emerged that allow us to measure full transcriptome-wide expression in two-dimensional space. Current approaches rank spatially variable genes based on either p-values or some effect size, such as the proportion of spatially variable genes. However, previous work in RNA-seq has shown that a technical bias, referred to as the 'mean-variance relationship”, exists in these data in that the gene-level variance is correlated with mean RNA expression. We found that there is a mean-variance relationship in spatial transcriptomics data, and so we propose a statistical framework to prioritize spatially variable genes that is not confounded by this relationship. We fit a loess curve to estimate the mean-variance relationship. Then, similar to using weights in a weighted least squares model, we used weights that we plugged into a Gaussian Process Regression model fit with a nearest-neighbor Gaussian process model to the preprocessed expression measurements for each gene, i.e. one model per gene. This framework removes the bias and leads to a more informative set of spatially variable genes. We are currently working on implementing this method as a package, which we plan to submit to Bioconductor.