Independent Polygenic Component Scores Link Multivariate Brain Imaging Genetics to Diverse Phenotypes

Abstract

Complex traits are influenced by many, often overlapping genetic influences, which makes it difficult to separate shared from trait-specific genetic mechanisms. We previously introduced Genomic Independent Component Analysis (genomICA), a novel, data-driven, multivariate method to disentangle genetic effects across thousands of univariate GWAS statistics into statistically independent genomic components (i.e. latent factors). Each component captures patterns of genetic covariance across traits. Our goals in this paper are threefold: To validate the capability of this new method to reliably recover genetic effects across thousands of complex phenotypes; to map shared and distinct genetic architectures across brain features and their relations with other complex traits; and to provide a new framework for polygenic scoring that improves out-of-sample prediction, interpretability, and stratification of individuals along independent genetic dimensions. Using the SNP-loadings from our previously derived genomICA components, derived from thousands of brain imaging GWASs, we generated polygenic component scores (PCS) in an independent cohort. We first evaluated mutual independence among PCS, as a condition for stratification utility. Next, we tested their out-of-sample predictive performance and specificity for 1269 neuroimaging and 858 behavioural, clinical, and lifestyle phenotypes. Additionally, we tested whether association patterns with brain traits recapitulated the original trait-loadings, as a cross-validation. Correlations between PCSs were low (|r| < 0.05), confirming mutual independence. PCSs explained substantial variance in brain traits (R2max =0.12). For behavioural/clinical/lifestyle phenotypes (not included in the discovery data) explained variance was also significant, but much lower (R2max =0.02). Individual PCSs were associated with distinct groups of neuroimaging categories and brain tissues. PCSs showed meaningful patterns of associations with clinical/behavioural/lifestyle phenotypes. For instance, PCS 8 captured associations with lifestyle, behavior, diet, and socioeconomic factors, while PCS 15 was linked to cardiovascular health outcomes. The dominant genomic influences captured here predicted physical, cognitive, lifestyle and environmental phenotypes, but surprisingly not mental health or neurological diagnoses. In the present study we present a novel polygenic scoring approach based on multivariate independent genomICA components. Each PCS reflects a reliable, distinct individual disposition towards a combination of brain-features. Individual PCSs align with distinct phenotypic domains. Overall, our results indicate that PCSs aid prediction and stratification utility of high-dimensional GWAS statistics.