2. EVALUATING THE TRANSFERABILITY OF POLYGENIC SCORE METHODS FOR SCHIZOPHRENIA IN AN ADMIXED POPULATION

Background Schizophrenia (SCZ) is a multifactorial disorder with a global prevalence of ∼1% and is a major contributor to disability. SCZ has a strong genetic component with a heritability of ∼80%. The expansion of consortia in psychiatric genetics has enabled the use of Polygenic Scores (PGS) for SCZ based on Genome-Wide Association Study (GWAS) results. However, PGS transferability remains limited for admixed populations, as available GWAS disproportionately based on individuals of European ancestry. Given Brazil's complex geographical history, the Brazilian population's genome is a mosaic of ancestries, making it unclear how the current PGS methods perform in this population. Some phenotypes, such as height (HGT), have already reached a substantial portion of SNP heritability capture in the GWAS, making PGS highly accurate. Therefore, HGT was included as a quantitative benchmark, providing a reference for comparison with SCZ. This study aims to benchmark the performance of five PGS construction methods in a Brazilian population, using schizophrenia (dichotomous) and height (quantitative) as primary phenotypes. Methods Global ancestry proportions were estimated using ADMIXTURE. The target sample for the PGS analysis comprised individuals from three Brazilian cohorts: the First Episode Psychosis Cohort (FEPC), the Brazilian High Risk Cohort Study (BHRCS), and the Schizophrenia Cohort (SCZC). As discovery samples, we used summary statistics from the Psychiatric Genomics Consortium (PGC, 2022) for SCZ and from Yengo et al. (2022) for HGT. We analyzed SCZ PGS in 2,193 individuals (1,585 controls and 608 SCZ cases) using PRSice2, PRS-CS, PRS-CSx, SBayesRC and DiscoDivas. HGT PGS was included as a comparison phenotype in 2,119 adults. We evaluated the predictions using the variance explained (R²) for HGT and the Nagelkerke R² for SCZ. We adjusted the models for 10 principal components, age, and sex, and we considered p < 0.05 after Bonferroni correction as statistically significant. Results Global ancestry confirmed the admixed composition, with European (67.9 ± 4.9%), African (21.6 ± 3.9%), and Native American (10.6 ± 2.6%) ancestries, indicating a higher European contribution. All PGS methods significantly predicted both traits (p-values < 0.001). For SCZ, PRSice2 achieved the highest performance (R² = 0.13, p = 1.4E-35), followed by PRS-CS (R² = 0.10, p = 5.9E-30), DiscoDivas (R² = 0.10, p = 3.2E-29), PRS-CSx (R² = 0.10, p = 6.2E-29), and SBayesRC (R² = 0.06, p = 7.7E-18). For HGT, PRS-CS performed best (R² = 0.24, p = 6.6E-129), followed by PRS-CSx (R² = 0.24, p = 1.4E-125), DiscoDivas (R² = 0.22, p = 2.6E-112), SBayesRC (R² = 0.21, p = 5.3E-107), and PRSice2 (R² = 0.18, p = 3.7E-91). Discussion We identified that PGS methods performed differently across the analyzed traits; PRSice2 performed better for SCZ prediction, and PRS-CS performed better for HGT. This discrepancy may be attributable to lower ancestral diversity and smaller sample sizes in available SCZ GWAS datasets relative to HGT GWAS, as well as to limitations in the LD reference panels underlying Bayesian approaches, which cannot fully reflect the complex admixture architecture of the Brazilian population. Future analyses will incorporate Local Ancestry-Aware PGS methods to better capture genetic architecture in this population.