We then tested the phenotype predictivity of aspPS on admixed genomes extracted from UKBB. We fitted a model on a joint, balanced set of admixed and unadmixed European individuals, and adopting as predictor a) the traditional PS or b) aspPSs, in addition to non-genetic covariates and the global ancestry proportion. This last predictor was introduced to account for predictivity gained simply scaling the total PS with the ancestry fraction: according to our simulations (Supplementary Fig. 1), this approach would be enough to recover predictivity loss due to PS directional bias alone. As shown in Fig. 5 (blue and yellow lines) aspPS computed on the European ancestral component, although comparable, never outperform the predictivity of the total PS for the admixed individuals. As anticipated by our simulations (Supplementary Fig. 1), this is expected when (A) transferability to the non-European ancestry of the trait-SNP associations discovered in Europeans is greater than zero and (B) the directional bias in PS introduced by the non-European genetic component is moderate. Nonetheless, this result confirms that the genomic portion discarded by the local ancestry deconvolution only
