Figure 4. Construction and validation of neuron-clocks. (A) The training is done on the Jaffe et al. dataset encompassing 416 prefrontal cortex (PFC) samples. Fractions for 3 broad cell-types (neurons, glia and endothelial/stromal) are inferred using the HiBED algorithm. Subsequently, CellDMC is applied to infer DMCTs in each cell-type. The Fisher-test one-tailed P-value of overlap between Neu-DMCTs and Glia-DMCTs is given. Finally, elastic net predictors of chronological age are trained from the Neu-DMCTs parameterized by a penalty parameter. (B) Scatterplots of adjusted DNAm values (adjusted for variations in cell-type fractions) of one Neu-DMCT and one Glia-DMCT against the corresponding Neuron and Glia fraction, respectively. Samples have been colored according to age-group and regression lines for each age-group have been added. (C) Optimal model (parameter) selection is performed using a completely independent PFC dataset from Philstrom et al. Scatterplot displays the predicted vs. chronological age for the optimal model. (D) Barplot of Pearson Correlation Coefficients (PCC) for 4 clocks (Neu-In, Neu-Sin, Brain and Horvath) across sorted neuron and glia datasets. Number of samples in each dataset is given in brackets. (E) Barplot of Pearson Correlation Coefficients (PCC) for the same 4 clocks (Neu-In, Neu-Sin, Brain and Horvath) across brain and non-brain DNAm datasets. Each cohort is labeled with the number of samples in brackets. Abbreviations: WB: whole blood; PFC: prefrontal cortex; TL: temporal lobe; FL: frontal lobe; MTG/STG: middle/superior temporal gyrus; PL: parietal lobe; SC: sensory cortex; VC: visual cortex; MC: motor cortex; Hippo: hippocampus; EC: entorhinal cortex; Stria: striatum; CG: cingulate gyrus; NAc: nucleus accumbens; CRB: cerebellum; Mono: monocyte. (F) For each of the 4 clocks, boxplots of PCCs comparing the values obtained in brain vs. non-brain datasets. The number of datasets in each group is given below x-axis. P-value is from a one-tailed Wilcoxon rank sum test.