70. MULTI-PGS BASED CHARACTERIZATION OF BIPOLAR DISORDER CLINICAL AND GENETIC SUBTYPES USING A NOVEL COMBINATION OF SEMI-SUPERVISED AND UNSUPERVISED LEARNING

Background Bipolar disorder (BIP) is a highly heritable and severe mental disorder characterized by depressive and manic (Type I) or hypomanic episodes (Type II). While GWAS of BIP have pointed to differences in the genetic architectures of the two types, such as higher genetic correlation with schizophrenia (SCZ) but a lower one with depression (DEP) for Type I, it remains an open question how distinguishable these subgroups are genetically at the individual level. Methods BIP patients with a matched healthy control (HC) group (N = 7.6k / 30k) were drawn from FinnGen biobank participants. Polygenic scores (PGS) based on psychiatry-related GWAS were calculated with LDpred2 and pruned for high genetic correlation (max rg =.95), leaving 197 PGS, which were then residualized on the first 20 population stratification PCs. Logistic LASSO prediction models for four endpoints (BIP / HC, Type I / HC, Type II / HC, Type I / II) using the PGS as predictors were trained and evaluated in a repeated (10x) tune-train-test scheme (3:5:2 split). Performances were evaluated with AUC in the test folds. For unsupervised grouping of BIP cases multi-PGS components were computed with partial least squares discriminant analysis (PLS-DA) in a repeated (1000x) train-predict procedure (1:1), followed by k-means clustering on the predicted components to split the BIP cases into two groups. The final grouping was derived by applying k-means (k=2) on the resulting co-clustering rate matrix, whose column variances also served as individual-level clustering stability indices (CSI). Finally, associations of the group assignment with the Types (I / II), psychiatric comorbidity, medication rates, hospitalizations and age of onset were analyzed with logistic and linear regressions. Results The prediction models produced AUCs of.68,.70,.67 and.60 for BIP / HC, Type I / HC, Type II / HC and Type I / II, respectively. The top predictors included the PGS for BIP, SCZ, DEP, various neuroticism items and educational attainment, but their relative contributions varied between the models. The group assignments were significantly associated with Type I / II (OR=0.62, p <.001), multiple psychiatric comorbidities (e.g. negatively for psychosis and positively for anxiety), lithium intake (OR=0.84, p <.001), age of onset (beta = 0.02, p=.016) and hospitalizations (log(beta) = -1.9, p <.001). The associations were stronger when restricted to individuals with high CSI, but they did not generally remain significant after adjusting for type I / II distinction. Furthermore, restricting the clustering to Type I or II cases only decreased the mean CSI (all =.31, Type I =.20, Type II =.14), an effect less pronounced for the unclassified cases (.31 vs.27). Discussion This study is the first to apply a multi-PGS strategy to characterize the genetic profiles of BIP I and II, which align with the corresponding summary level genetic differences reported in BIP GWAS. Notably, we show that unsupervised PGS-based classification aligns with the clinical classification phenotypically and genetically, supporting its biological validity. The analyses within the types also indicate that the applied algorithm can differentiate between diagnostic groups with a strong genetic substructure and those without one, enabling the investigation of other disorders for which subtypes have been hypothesized, such as unipolar depression. Replication analyses in an independent sample with clinical interview based type I/II classifications are pending.