Reproducibility of Principal and Independent Genomic Components of Brain Structure and Function

Abstract

IntroductionThe highly polygenic and pleiotropic nature of behavioural traits, psychiatric disorders, and structural and functional brain phenotypes complicate mechanistic interpretation of related genome-wide association study (GWAS) signals, such that the underlying causal biological processes remain obscure. We propose the novel method of genomic principal and independent component analysis (PCA, ICA) to decompose a large set of univariate GWAS statistics of multimodal brain traits into more interpretable latent genomic components. Here we introduce this new method and evaluate its various analytic parameters and reproducibility across independent samples. MethodsTwo releases of GWAS summary statistics from the UK biobank (UKB), with 11,086 and 22,138 participants respectively, were retrieved from the Oxford BIG-40 server. GWAS summary statistics were clumped resulting in n=165,364 single nucleotide polymorphisms (SNP) and m=2,240 imaging derived phenotypes (IDPs). Both genome-wide beta-values and their corresponding, standard-error scaled z-values were decomposed using multivariate exploratory linear optimised decomposition into independent components (MELODIC). We evaluated variance explained at multiple dimensions up to 200. We tested the reproducibility of output of dimensions 5, 10, 25, and 50 by computing Pearsons correlation between component loadings, and Fisher Exact tests on overlap of the top SNP loadings across samples. Reproducibility statistics of the original raw and z-transformed univariate GWAS served as benchmarks. We also inspected the clustering of genomic components across neuroimaging modalities using t-SNE. ResultsThe first five PCs derived from z-transformed GWAS captured 31.9% of the variance across SNP effect sizes, while 200 PCs increased the variance explained to 79.6%. Reproducibility of 10-dimensional PCs and ICs showed the best balance between model complexity and robustness, and variance explained (PCs: |rz-max|=0.33, |rraw-max|=0.30; ICs: |rz-max|=0.23, |rraw-max|=0.19), with decreasing model stability and reproducibility at higher dimensions. Both genomic PC and IC reproducibilities improved substantially relative to mean univariate GWAS reproducibility up to a dimension of 10. Genomic components clustered along neuroimaging modalities. ConclusionOur results indicate that these novel methods of genomic ICA and PCA decompose genetic effects on IDPs from raw GWAS statistics with high reproducibility by taking advantage of the inherent pleiotropic patterns. These findings encourage further applications of genomic ICA and PCA as fully data-driven methods to effectively reduce the dimensionality, enhance the signal to noise ratio, and improve interpretability of high-dimensional multi-trait genome-wide analyses.