Replication and transdiagnostic validation of genetic risk profiles in schizophrenia

Subtyping is essential to address the marked clinical heterogeneity observed in psychiatric disorders. Transdiagnostic validation further evaluates whether such subtypes generalize across diagnostic categories, thereby strengthening their clinical relevance. Previous analyses identified two genetic risk profiles (GRPs) in schizophrenia (SCZ) patients compared with healthy controls (HC). Subgroups were derived via k-means clustering and subsequent random forest (RF) modelling, predicting SCZ [1]. We tested whether these GRPs can be replicated in patients with recent-onset psychosis (ROP), and whether they extend transdiagnostically, although in decreasing associative strength, to individuals at clinical high-risk for psychosis (CHR), and patients with recent-onset depression (ROD).

We applied the multi-polygenic score (multi-PGS) pipeline PRScope [1] to compute 45 psychiatry-related PGS in the PRONIA cohort (n = 992, including ROP, CHR, ROD, and HC). The previously trained models [1] were then applied across 13 feature sizes (k ∈ {1, 2, 3, 4, 5, 6, 12, 16, 24, 28, 33, 40, 45}, ordered by genetic correlation with SCZ), defined by the number of PGS used for k-means clustering during model training. RF model application resulted in one prediction per GRP and feature size (n = 26). Individuals were assigned to a GRP only if a single model correctly classified their diagnostic status; otherwise, they remained unassigned. Associations between GRP membership and diagnosis were quantified using odds ratios (OR), with significance assessed via Fisher’s exact test and Benjamini-Hochberg correction within each diagnostic group across feature sizes and models. Clinical profiles of GRPs were examined using linear or logistic regression models, with phenotypic variables (z-scored if continuous) as outcomes and GRP as predictor, adjusting for age, sex, and SCZ PGS. Multiple testing correction was applied within each feature size and model type.

In ROP, both GRP’s models significantly distinguished patients from controls from feature size 12 onward (for feature size 45: OR_GRP1=1.59, OR_GRP2=2.63, p-FDR_GRP1=0.017, p-FDR_GRP2<0.0001). In CHR, only the GRP1 model showed significant discrimination, whereas in ROD, neither model performed above chance level (for feature size 45: CHR: OR_GRP1=1.08, OR_GRP2=2.40, p-FDR_GRP1=0.71, p-FDR_GRP2<0.0001; ROD: OR_GRP1=0.94, OR_GRP2=1.74, p-FDR_GRP1=1.00, p-FDR_GRP2=0.11). Clinical characterization based on feature size 45 for simplicity indicated that GRP1 was associated with a more severe phenotype, including higher symptom burden (social anxiety and depressive symptoms), reduced functioning and well-being, and lower perceived social support and self-efficacy (all p-FDR<0.041).

As expected, predictive accuracy decreased with increasing diagnostic distance from SCZ. Consistent with previous findings, one GRP reflected a more severe clinical profile. While the association with depressive symptoms was replicated, the previously reported increase in medication use was not observed, potentially reflecting the partially subclinical nature of the cohort. In addition, novel associations emerged with measures of perceived social support and resilience. In summary, we replicated prior GRP findings in an independent cohort and demonstrated their partial transdiagnostic applicability. Ongoing analyses will assess whether these profiles are linked to imaging-based signatures and whether they generalize to population-based samples.

References

1. Kocak, E., et al., 2025. Development and validation of polygenic risk profiles of schizophrenia. medRxiv (preprint). https://doi.org/10.1101/2025.04.29.25326656