Principal and Independent Genomic Components of Brain Structure and Function

Abstract

IntroductionHuman brain structure and function, as measured using magnetic resonance imaging (MRI), is heritable. The genetic associations are highly polygenic and pleiotropic, complicating translation of genetic association studies to concrete biological processes. We recently proposed genomic independent component analysis (ICA) to reduce a large array of genome-wide association study (GWAS) statistics into a number of more interpretable multivariate components. Here, we extend this work with genomic principal component analysis (PCA). We also estimate the inter-trait stability and test-retest reliability of genomic PCA and ICA. As a benchmark, we provide a head-to-head comparison with the reproducibility of original (i.e., raw univariate) GWAS variant effect sizes from which the components were derived.MethodsWe considered summary statistics of two rounds of brain MRI GWAS in UK biobank encompassing 22,138 and 11,086 subjects and thousands of imaging-derived phenotypes (IDPs) retrieved from Oxford BIG-40 server. We decomposed the clumped brain-wide genome-wide matrix (n=157,893 variants and m=2,240 brain IDPs) by PCA and ICA to obtain up to 50 principal and independent genome-wide components. We tested the reproducibility of these genomic components using tail overlap coefficient and Pearson’s correlation. Decompositions were repeated 20 times on randomly permuted split-half matrices to arrive at an empirical distribution of the above reproducibility metrics. As a benchmark, we also provide comparisons with the reproducibility of the original univariate GWAS effect sizes from which the components were derived.ResultsThe first ten principal genomic components were highly stable across trait-subsets within samples (r=0.82±0.15; mean ± standard deviation). PCA cross-sample reproducibility dropped to r=0.16±0.07 (r-max=0.30) was still a substantial improvement over univariate GWAS reproducibility r=0.09±0.05. ICA components showed higher within-sample reproducibility at lower dimensions (ICA dimension-5: r=0.81±0.16 vs. ICA dimension-50: r=0.41±0.13). The same trend was observed for their cross-sample reproducibility, achieving an r-max of 0.27 at the lowest ICA dimension of five (average r=0.20±0.3) vs. r-max of 0.11 at dimension 50 (average r=0.05±0.02).ConclusionOverall, the cross-sample reproducibility of PCA and ICA components are moderate, but considerably better than the reproducibility of the raw univariate GWAS effect sizes. Genomic PCA identified highly reproducible and pleiotropic genomic effects across many brain traits in just 1-3 factors. Genomic ICA captured mutually independent genome-wide effects that may be more sensitive to distinct biological pathways. As GWAS samples continue to increase and more (endo)phenotypes become available in biobanks, multi-trait genome-wide associations of the brain decomposed by ICA and PCA will likely become more reproducible and interpretable.

Distinct impact modes of polygenic disposition to dyslexia in the adult brain

“…These voxel-or fixel-wise t-score maps were then concatenated across all 13,766 variants and decomposed by MELODIC into ten independent components, separately per each imaging modality. The default MELODIC data transformations, including variance normalization and mean signal removal, were not applied as these momentums reflect meaningful signals in t-score maps 85 . The extracted independent components, henceforth referred to as genomic impact modes , capture the hidden sources that shape the brain-wide influence of dyslexia-disposing variants and approximate their heterogeneous spatial profiles through a limited number of features.…”

GUIDE deconstructs genetic architectures using association studies

“…Independent Component Analysis (ICA) is a statistical method for decomposing a mixed signal into a set of statistically independent ones ( e.g ., separating individual conversations from a recording of multiple conversations in the famous cocktail party problem) [8, 9, 10, 11]. It has been successfully applied in many applications in genomics as a denoising or dimensionality reduction alternative to principle components analysis (PCA) [12], most recently to analyze the GWAS summary statistics of brain image-derived phenotypes [13]. Our method, Genetic Unmixing by Independent Decomposition (GUIDE), uses TSVD and a two-sided application of ICA—simultaneously for both the SNPs-to-latents and latents-to-traits weights—to determine a set of sparse and independent latent factors given GWAS summary statistics.…”