Revealing multi-scale population structure in large cohorts

Abstract Genetic structure in large cohorts results from technical, sampling and demographic variation. Visualisation is therefore a first step in most genomic analyses. However, existing data exploration methods struggle with unbalanced sampling and the many scales of population structure. We investigate an approach to dimension reduction of genomic data that combines principal components analysis (PCA) with uniform manifold approximation and projection (UMAP) to succinctly illustrate population structure in large cohorts and capture their relationships on local and global scales. Using data from large-scale genomic datasets, we demonstrate that PCA-UMAP effectively clusters closely related individuals while placing them in a global continuum of genetic variation. This approach reveals previously overlooked subpopulations within the American Hispanic population and fine-scale relationships between geography, genotypes, and phenotypes in the UK population. This opens new lines of investigation for demographic research and statistical genetics. Given its small computational cost, PCA-UMAP also provides a general-purpose approach to exploratory analysis in population-scale datasets. Author summary Because of geographic isolation, individuals tend to be more genetically related to people living nearby than to people living far. This is an example of population structure, a situation where a large population contains subgroups that share more than the average amount of DNA. This structure can tell us about human history, and it can also have a large effect on medical studies. We use a newly developed method (UMAP) to visualize population structure from three genomic datasets. Using genotype data alone, we reveal numerous subgroups related to ancestry and correlated with traits such as white blood cell count, height, and FEV1, a measure used to detect airway obstruction. We demonstrate that UMAP reveals previously unobserved patterns and fine-scale structure. We show that visualizations work especially well in large datasets containing populations with diverse backgrounds, which are rapidly becoming more common, and that unlike other visualization methods, we can preserve intuitive connections between populations that reflect their shared ancestries. The combination of these results and the effectiveness of the strategy on large and diverse datasets make this an important approach for exploratory analysis for geneticists studying ancestral events and phenotype distributions.