Independent Polygenic Component Scores of Brain Structure and Function

The polygenic and pleiotropic nature of behavioral, mental health and brain phenotypes complicates mechanistic interpretation of related genome-wide association study (GWAS) signals. We previously proposed genomic independent component analysis (genomICA) as a multivariate method to decompose a large set of univariate GWAS statistics of multimodal brain traits into statistically independent and potentially more interpretable genomic components. Here, we assess the out-of-sample predictive power and stratification potential of this method using polygenic scores derived from genomICA components. We posit that these neuroimaging-based polygenic component scores (PCS) capture pleiotropic polygenic patterns in a structured and interpretable way and explain phenotypic variance reliably, while simultaneously decorrelating individual genetic profiles to reveal independent dimensions of polygenic influence across brain imaging and other traits.

We used the SNP-loadings of our previously derived genomICA components, based on thousands of brain traits, as weights to compute PCS in an independent sample. We first assessed collinearity among PCS to confirm independence. Then, we tested the PCS predictive power in 1269 neuroimaging phenotypes and 858 non-neuroimaging phenotypes using general linear and logistic regression models, and quantified phenotypic variance jointly explained by all PCS. We then examined PCS correlations across individuals to assess whether they offer utility for stratifying phenotypic variation.

PCSs showed little collinearity (r_min= -0.04,r_mean= 4*10^-4,r_max= 0.05). They explained substantial variance in brain traits out-of-sample (R²_max=0.12,R²_mean=0.04). Variance explained in non-neuroimaging phenotypes was much lower (R²_max=0.02,R²_mean= 2*10^-4) but significant. Different individual PCS were associated with distinct groups of neuroimaging categories and brain tissues. PCSs explained meaningful patterns of phenotypic variance in non-neuroimaging phenotypes, showing a distinction along different categories of phenotypes. For instance, PCS 8 clearly captured associations with lifestyle, behavior, diet, and socioeconomic factors, while PCS 15 was linked to a wide range of cardiovascular health outcomes in both participants and their relatives.

These results demonstrate that genomICA-derived PCSs capture independent, uncorrelated patterns of associations across the brain and can predict substantial variance in a range of neuroimaging and other out-of-sample phenotypes. PCSs capture substantial variance in neuroimaging phenotypes and, to a lesser extent, non-neuroimaging traits, with individual PCS aligned to distinct phenotype domains. The inclusion of the brain phenotypes, but not other phenotypes, in the input for the decomposition likely accounts for this difference. GenomICA identifies mutually-independent dimensions in the space of genomic effects, which has utility for stratification of heterogeneous samples. Notably, the dominant genomic influences on the brain captured here reflect physical traits, lifestyle and environment but not mental health or neurological traits. Overall, this study highlights the potential of genomICA-based PCSs in modelling phenotypic variance across data domains. Future work will focus on refining these components to better capture variance in complex neuropsychiatric and behavioural traits by incorporating them in the decomposition process.