Abstract
Plasma proteomic biomarkers hold significant promise for enhancing clinical assessments and improving early detection of psychosis conversion; however, concerns about their reproducibility and generalizability persist. Previous studies, largely conducted in Caucasian cohorts, have identified proteomic biomarkers predictive of psychosis conversion in individuals at ultra-high risk (UHR) of psychosis. In this study, we acquired plasma proteomics data from an Asian UHR cohort, the Longitudinal Youth at Risk Study (LYRIKS). We established a robust machine learning framework, through which we developed and evaluated prediction models for psychosis conversion. We showed that proteomic signatures previously identified in a predominantly Caucasian UHR cohort generalized to the LYRIKS cohort (best AUC = 0.81). Furthermore, we developed three prediction models using the LYRIKS dataset that demonstrated superior performance (best AUC = 0.96). Through these models, we identified novel proteomic signatures with high predictive performance. Despite the differences in individual protein composition between Asian- and Caucasian-derived signatures, functional convergence was observed across key pathways and protein families, namely the complement and coagulation cascade, apolipoproteins, inter-alpha-trypsin inhibitor heavy chain proteins, and serine protease inhibitors. Although current literature is divided on the utility of blood plasma biomarkers in psychiatric diagnosis, our study supports their use by demonstrating cross-population generalizability of Caucasian-derived signatures and deriving signatures with high predictive value from an Asian cohort that converge functionally.