Contextualizing postmortem bias for single-nuclei transcriptomic studies of human brain

The molecular underpinnings of brain disorders remain unknown. This is in part due to the inability to study brain tissue from living people. Instead, the field has relied on postmortem human brain tissue, which may not be an accurate representation of living human brain tissue. The Living Brain Project (LBP) has allowed researchers to ethically biopsy living cortical tissue to study multi-omic differences between living and postmortem brains. Single-nuclei RNAseq was generated from 31 living (LIV) and 21 postmortem (PM) cortical samples and clustered/annotated into neural cell types. Differentially expressed genes (DEGs) were identified across cell types using a linear mixed model. Gene regulatory networks (regulons) were defined using SCENIC and then tested for differentially active regulons (DARs) between groups. The utility of the LIV vs. PM expression signature was explored via elastic net to create a postmortem linear predictor score (PMlink) for postmortem bulk samples. After QC, the 52 samples yielded 362,390 quality cells (61% postmortem and 39% living) which were clustered into 10 cell types. Differential expression revealed massive signals across all cell types, with 64% of genes differentially expressed in at least one cell type (n = 16,759 genes). Neuronal markers are disproportionately upregulated in postmortem neurons and oligodendrocyte markers are upregulated in living across all cell types. In examining DARs, living cells show increased regulon activity associated with RNA processing, whereas postmortem cells display increased activity of regulons associated with neuronal signaling. Finally, the calculated PMlink had perfect predictive strength in determining LV vs PM classification across all pseudobulk testing permutations (ROC AUC = 1). When PMlink is incorporated as the dependent variable in a linear mixed model containing only postmortem bulk samples, the LV vs PM pseudobulk DE signal is replicated (ρ=0.65). All key findings using pseudobulk were replicated in an independent bulk dataset. The combined DEG and DAR results reveal ubiquitous dysregulation of key biological systems throughout the postmortem brain that are not reflective of the living expression profile. These findings provide necessary context for interpreting postmortem gene expression as a snapshot of biological processes at death rather than a proxy for living brain function. Machine learning methods provide utility for creating correction algorithms that can help address this problem in past and future postmortem transcriptomic studies of the brain.