Integrative Multi-omics Analysis of Childhood Aggressive Behavior

Abstract

This study introduces and illustrates the potential of an integrated multi-omics approach in investigating the underlying biology of complex traits such as childhood aggressive behavior. In 645 twins (cases = 42%), we trained single- and integrative multi-omics models to identify biomarkers for subclinical aggression and investigated the connections among these biomarkers. Our data comprised transmitted and two non-transmitted polygenic scores (PGSs) for 15 traits, 78,772 CpGs, and 90 metabolites. The single-omics models selected 31 PGSs, 1614 CpGs, and 90 metabolites, and the multi-omics model comprised 44 PGSs, 746 CpGs, and 90 metabolites. The predictive accuracy for these models in the test (N = 277, cases = 42%) and independent clinical data (N = 142, cases = 45%) ranged from 43 to 57%. We observed strong connections between DNA methylation, amino acids, and parental non-transmitted PGSs for ADHD, Autism Spectrum Disorder, intelligence, smoking initiation, and self-reported health. Aggression-related omics traits link to known and novel risk factors, including inflammation, carcinogens, and smoking.

Keywords: Childhood aggression; DNA methylation; Genetic nurturing; Metabolomics; Multi-omics; Polygenic scores.

Fig. 1

Overview of the biomarker identification approach for childhood aggression—details of statistical analyses and data included in each analysis. We employed an analytical design comprising three phases: (1) single-omics analyses; (2) pairwise cross-omics analyses; and (3) multi-omics analyses. First, we performed univariate polygenic score (PGS) analysis in 70% of the twin data and built multivariate single-omics biomarker panels in the twin cohort, with 70% of the twin data for model training (training data), 30% of the twin data and the clinical cohort for model testing (test data). Second, we examined the overall pairwise cross-omics correlations and the pairwise correlations of those omics variables selected by the single-omics models in the training data. Third, using the same data, we compared three multi-omics models, with different assumptions of the correlations among the omics blocks, and describe the multi-omics relationships of the selected omics variables. We offer the analytical details in the “Materials and methods” section

Fig. 2

Clustered heatmaps of the relationships obtained by the pairwise cross-omics Partial Least Squares (PLS) regression models including only the omics variables as selected by the single-omics sparse Partial Least Squares Discriminant Analyses (sPLS-DA). We generated the hierarchical clustering using the Ward linkage algorithm on Euclidean distances of the PLS variates. For each dendrogram we identified two clusters (cluster 1 = pink, cluster 2 = blue). We have depicted positive relationships among the omics variables in red and negative relationships in blue. For the polygenic scores (PGSs), the ‘_NTm’ suffix denotes non-transmitted maternal PGSs, the ‘_NTf’ suffix denotes the non-transmitted paternal PGSs, and childhood aggression is abbreviated as “aggression”, Attention-Deficit Hyperactivity Disorder as “ADHD”, Major Depressive Disorder as “MDD”, Autism Spectrum Disorder as “Autism”, Educational Attainment as “EA”, and wellbeing spectrum as “wellbeing”. For the metabolites, the ‘amines.’ prefix shows we measured these metabolites on the Liquid Chromatography Mass Spectrometry (LC–MS) amines platform, the ‘steroids.’ prefix shows we measured these metabolites on the LC–MS steroids platform, and the ‘OA.’ prefix shows we measured these metabolites on the Gas Chromatography (GC-) MS organic acids platform. a Relationships of the 1,614 CpGs and 90 metabolites included in the 3-component DNA methylation-metabolomics PLS model, where the selected CpGs are represented in the columns and the metabolomics traits in the rows. We included the cluster assignments and the full data matrix in Data S4-S5. b Relationships of the 36 PGSs and 90 metabolites included in the 2-component PGSs-metabolomics PLS model, where the selected PGSs are represented in the columns and the metabolomics traits in the rows. We included the cluster assignments and the full data matrix in Table S9 and Data S6, respectively. c Relationships of the 36 PGSs and 1,614 CpGs included in the 2-component PGSs-DNA methylation PLS model, where the selected PGSs are represented in the columns and the CpGs in the rows. We included the cluster assignments and the full data matrix in Data S7-S8 (Color figure online)

Fig. 3

Strong cross-omics connections of the multi-omics traits identified in the 5-component multi-block sparse Partial Least Squares Discriminant Analysis (MB-sPLS-DA) model including the empirical design matrix. The outer ring depicts the PGSs, CpGs, and metabolites in yellow, pink, and green, respectively. For the polygenic scores (PGSs), Attention-Deficit Hyperactivity Disorder is abbreviated as “ADHD”, and Educational Attainment as “EA”, and the ‘_NTf’ suffix denotes the non-transmitted paternal PGSs. For the metabolites, the ‘amines.’ prefix shows we measured these metabolites on the Liquid Chromatography Mass Spectrometry (LC–MS) amines platform, and the ‘OA.’ prefix shows we measured these metabolites on the Gas Chromatography (GC-) MS organic acids platform. The inner plot depicts the connetions among the omics variables. Here, we depict only high absolute correlations of the PLS variates (|r|≥ 0.60) between variables of at least two omics blocks, with blue lines reflecting negative correlations and red lines positive correlations. We averaged correlations of the PLS variates across all components in the MB-sPLS-DA model. We included the full data matrix in Data S10 and the patterns in Table S15 (Color figure online)