We can also define ancestry specific partial PS (aspPSA) as a proxy for the total standardized PS that uses only the genomic portion pertaining to the ancestry A, losing therefore information as p decreases, but remaining virtually unbiased. A model including this metric (Supplementary Fig. 1, blue line) would outperform the traditional PS (PSA) in case of high directional bias and p and, with low transferability, even a model that takes into account p as a separate variable to explain the directional bias (Supplementary Fig. 1, red line). The limitation of not considering parts of the genome, makes the aspPS only partially useful in improving trait predictability. On the other hand, considering the availability of another PS (PSB) originating from a GWAS on a population closer to the other ancestry B, and therefore not suffering of transferability issues, one can imagine to combine the two aspPS, weighting for p, in order to obtain a combined ancestry specific PS (casPS). In our simulation, the casPS would outperform the traditional PS in all cases where PS transferability represents a problem and where a sufficiently predictive PS is present for the two ancestries independently (Supplementary Fig. 1, purple line).
