Abstract
BACKGROUND: The predictive power of polygenic scores (PGSs) for lithium treatment response in bipolar disorder (BD) remains limited. AIM: To enhance prediction of lithium responsiveness by developing a multi-trait PGS (mt-PGS) combining genetic information from multiple phenotypes implicated in lithium response and/or BD aetiology. METHODS: We analysed data collected from BD patients who had received lithium treatment for at least six months and participated in the International Consortium on Lithium Genetics (ConLi(+)Gen, N=2,367) study. The ALDA scale was used to assess lithium responsiveness, and treatment outcome was defined as continuous ALDA score (0-10) and categorical outcome (favourable ≥7 vs unfavourable response). PGSs were calculated for 59 phenotypes grouped into five clinical-biological clusters: clinical lithium exemplar (#22 phenotypes), cardiometabolic (#17), autoimmune/inflammatory (#5), neurocognitive (#8) and renal function (#7). We applied cross-validated machine learning regression approaches in both outcomes within each cluster, and the selected features from each cluster were subsequently combined to construct the final mt-PGS models. Model performance was assessed using explained variance (R(2)) for the continuous outcome, and McFadden's pseudo-R(2) as well as standard classification model parameters for the categorical outcomes. RESULTS: The mt-PGS explained 5.07% (continuous outcome) to 9.02% (categorical outcome) of the interindividual variability in lithium responsiveness. Classification accuracy (AUC) for the categorical outcome was 68.13% (95% CI: 64.86, 71.77). Of the five clusters, the PGSs for clinical lithium exemplar phenotypes were most strongly associated with lithium responsiveness, accounting for 2.97%-6.20% of its variability. CONCLUSIONS: By integrating polygenic scores for multiple relevant phenotypes, predictive accuracy for lithium response improved up to nine-fold compared to single-trait methods. Future research incorporating larger, more diverse populations and combining genetic scores with clinical data holds promise for further enhancing prediction and advancing clinical implementation.