Fig. 7: Machine learning prediction of epigenetic drivers of gene expression in the aging mouse brain.

A Heatmap showing the balanced classification accuracy of random forest models for cohort 1 (top) and cohort 2 (bottom) over three classes (up/down/constant) across seven brain regions. The final prediction was defined as up or down if at least 75% of the models agreed, as previously described³³. For cross-region validation, the entire data of the tested region was used. The color scale shows the accuracy values. B Scatterplot of top 50 features obtained from machine learning on cohort 1 (left) or cohort 2 (right) and Lisa ranks of predicted TFs from upregulated DARs in pseudobulk data. C/EBPβ and YY1, labeled in red, were chosen for testing. C Bar plot of Cebpb mRNA in fiber tracts of young, middle, and old mice. Each dot shows the average expression of the two technical replicates for each mouse relative to mean of young group, equivalent to n = 4 biological replicates per group. Data were summarized as mean ± standard error of the relative mean counts. **p < 0.01, from an unpaired two-tailed t-test with Welch’s correction comparing middle vs. young and old vs. young (left). Schematic showing the study design of CUT&RUN experiments in primary microglia isolated from young and old mouse brains (right). D Metaplot showing binding of C/EBPβ near TSSs of 73 genes upregulated during aging (Fig. 2D, G) and shared across five regions (top) or a background set of 91 genes that are downregulated with age also from Fig. 2D (bottom). E Same as (D), except for YY1. For (D, E), ***p < 0.001 from an unpaired two-tailed t-test with Welch’s correction comparing AUC values across the gene body in old vs young. F Representative genome browser tracks showing binding of C/EBPβ (left) and YY1 (right) at key upregulated genes. G Comprehensive model derived from this study showing BBB damage and demyelination at fiber tracts with consequential microglia activation (WAM/DAM2 signature), astrogliosis, and neutrophil infiltration. Illustration credit for parts of (C) goes to Endosymbiont GmbH, and BioRender. Sen, P. (2025) https://BioRender.com/m70w321. Source data are provided as a Source Data file.