We must consider several important limitations of our work, as our results are a consequence of the analyzed GWAS populations, polygenic traits and diseases, and available experimental data to create functional annotations. First, our functional insights are limited by biases in public TF ChIP–seq data: preference to cell lines over primary cells, rare or difficult-to-assay cell types, and preference to TFs with known regulatory roles and specific antibodies. As experimental strategies are developed to map regulatory elements, such as high-throughput CRISPR screens paired with assays for open chromatin, the IMPACT framework may need to be adapted to incorporate different types of training data. Second, the robustness of multi-ancestry comparisons rely on properties surrounding the recruitment of individuals or the exact genotyping platform used in various biobanks, which may result in cohort bias that inflates within-population PRS prediction accuracy. For example, BBJ is a disease ascertainment cohort, in which each individual has any one of 47 common diseases58,59; therefore, BBJ control samples are not comparable to healthy controls of UKBB. Other biases may arise from clinical differences in phenotyping. Third, we
