EPCO-16. METHOD TO MODEL AND CORRECT NOISE IN SINGLE-NUCLEI RNA-SEQUENCING DATA FROM POST-MORTEM BRAIN

Abstract Post-mortem brain samples are crucial for neuropathological research. For brain cancers, they provide insights about disease progression, microenvironmental factors, and treatment resistance. In some brain diseases, brain tissue may only be available post-mortem. Establishing a reference atlas of healthy brain also requires post-mortem donations. Beyond logistical and ethical challenges of post-mortem samples, there are also post-mortem factors that mask biological signal. In this study, we specifically aim to model post-mortem gene expression changes in single-nuclei RNA-sequencing (snRNA-seq) data. We analyze a public snRNA-seq dataset of pediatric high-grade glioma (pHGG) that has matched samples at initial resection and autopsy. We find a clear gene signature, composed of mitochondrial genes, stress-related genes, and metabolism-related genes, that is upregulated post-mortem. This signature also appears in post-mortem snRNA-seq data from other brain diseases and healthy donors. We classify it as “post-mortem noise”. We model the post-mortem noise as a combination of technical and biological artifact. The technical artifact is correlated with the percentage of mitochondrial reads and is a general measure of sample quality. The biological artifact is correlated with post-mortem interval and represents gene expression changes due to post-mortem stress. We apply our model to the pHGG data by regressing out the technical and biological artifact. Removing the post-mortem noise results in an improvement in malignant versus normal cell prediction, cell type classification, and the uncovering of pathways related to glioma progression. We confidently identify several normal cell types and predict their interactions with cancer cells, highlighting the utility of wider possible margins post-mortem. Therefore, after removing post-mortem noise, we analyze glioma progression at an unparalleled timepoint and characterize late-stage changes in the tumor microenvironment. In conclusion, we develop a method that models post-mortem noise in snRNA-seq brain data. Our method can be broadly applied to future brain cancer studies using post-mortem tissue.