Preprint: sparse sampling and rare-variant depletion shown to distort PCA plots of population structure

Principal component analysis (PCA) is the standard first step for visualising genetic population structure in human and non-human genomic datasets. A preprint deposited on bioRxiv (not yet peer-reviewed) presents simulation evidence that two common features of real-world genomic studies — sparse geographic sampling and rare-variant depletion arising from minor-allele-frequency filters — interact in ways that systematically distort low-dimensional PCA plots. The authors, using spatially explicit simulations, show that these distortions can produce triangular or three-ray patterns, artificial outlier samples, and misleading gradient clines that do not reflect true population history.

To address this, the team develops an objective-guided manifold-learning framework that systematically explores genotype normalisation strategies, PCA dimensionality choices, distance metrics, and non-linear dimensionality-reduction algorithms including UMAP, densMAP, and PHATE. The framework selects parameter combinations according to a quantitative objective rather than by visual inspection, aiming to reduce subjectivity in the downstream interpretation of population-structure plots.

The work is of direct relevance to researchers conducting population genetics and archaeogenomics analyses, statistical geneticists evaluating or building analysis pipelines, and bioinformaticians working on genomic visualisation. It also has value for educators teaching population genomics methods, as the artefacts described are common sources of misinterpretation in both published literature and teaching datasets. As a preprint, these findings have not yet undergone formal peer review.