Fig. 2

Visualization of the mediating role of the early-PN-induced alterations in DNA methylation in explaining impaired long-term emotional/behavioral outcomes with early-PN. This figure summarizes the results of the mediation analyses that were performed to investigate any role of early-PN-induced adversely altered leukocyte DNA methylation of 37 CpG-sites [5] in statistically explaining the harmful impact of early-PN on internalizing, externalizing and total behavioral problems identified at 4-year follow-up. All multivariable models resulted from random forest machine learning, with covariate significance levels obtained via permutation importance [11, 12], including the baseline risk factors age, center, race, sex, geographic origin, language, history of malignancy, a predefined “syndrome” (Additional file 2), diagnosis and severity of illness (PIM3 and PeLOD score), risk of malnutrition (STRONGkids score) and the randomized intervention, without (panel A) or with (panel B) the 37 differentially methylated CpG-sites evoked by early-PN [5]. The robustness of these results was evaluated in 100 bootstrapped replicates [13, 14]. Each row corresponds to the 100 bootstrap replicates of the multivariable non-linear models for the internalizing, externalizing and total behavioral problems. Columns correspond to the 37 CpG-sites that were differentially methylated by early-PN. Color intensity of the boxes reflects the frequency with which a CpG-site was found to be independently and significantly (P < 0.05) associated with the behavioral outcomes in the 100 bootstrapped replicated analyses, with darker orange colors corresponding to a higher frequency. Outcomes and CpG-sites were clustered based on these frequencies, with the clustering hierarchy shown in the dendrograms. The base of the column dendrogram is color-coded according to CpG-site functional classes “Cerebral/neuronal”, “Growth/Development/Locomotion”, “Metabolism”, “Gene expression/DNA regulation/Epigenetic regulation”, “Intra/intercellular Signaling/Transport”, “Non-coding RNA/pseudogene”, “Oncogenic/Apoptosis”, or “Intergenic”, as previously described [5]