W125. using physical and mental health polygenic risk scores to identify subtypes of schizophrenia and related diagnoses

Background Within current severe mental illness (SMI) diagnoses there is a large amount of phenotypic and genetic heterogeneity, as well as overlap in both symptoms and genetic liability between distinct diagnoses. This increases misdiagnosis and stalls treatment development, by making it difficult to identify the underlying biological targets. One approach to improve diagnostic classification is to cluster individuals based on a wide range of clinical, phenotypic and/or genetic data, not including clinical diagnosis. This can identify biologically homogeneous groups (“biotypes”) that can span traditional diagnosis boundaries or reflect within-disorder subtypes. An existing study (Kocak et al. 2026 [medRxiv preprint]) identified distinct schizophrenia (SZ) biotypes with homogeneous genetic risk profiles for certain psychiatric-related phenotypes by clustering individuals with SZ (n=7,486) on polygenic risk scores (PRS) from 413 relevant genome wide association studies. Here, similar methodology was applied to a cohort of individuals with SZ-related diagnoses, clustering on a more focussed set of PRS for 10 physical and mental health traits or disorders. Methods In 693 individuals with schizophrenia-related diagnoses (SZ, schizophreniform disorder, schizoaffective disorder or delusional disorder), PRS were calculated for 10 physical and mental health disorders or traits (Kappel et al. 2025, Lancet Psychiatry). These traits represent a range of psychiatric conditions and common physical comorbidities in individuals with SMI (SZ, bipolar disorder (BD), attention deficit hyperactivity disorder, major depressive disorder, type 2 diabetes, systolic blood pressure (SBP), coronary artery disease (CAD), asthma, low-density lipoprotein cholesterol (LDL-C) and rheumatoid arthritis). PRS were clustered via k-means clustering, using both two and three clusters. Quality was assessed using the silhouette score to measure cluster separation. Variable importance was investigated via a “leave-one-out” procedure, removing each PRS in turn from the input to assess the impact on the results. Another cluster quality metric, the Ratkowsky-Lance criterion, was also used to quantify per variable contribution. Results For k=2 means clustering, a mean silhouette value of 0.090 was achieved. Visual cluster separation predominantly occurred on low vs. high SZ and BD PRS. Both leave-one-out clustering and the Ratkowsky-Lance criterion indicated that SZ and BD PRS were the most important in the clustering solution. For three clusters, a mean silhouette value of 0.081 was obtained. Visually, cluster differentiation was again broadly based on SZ and BD PRS, but a third cluster captured differentiation based on certain physical health-related PRS (LDL-C, CAD, SBP). Leave-one-out clustering identified no major contributing single PRS, while BD, LDL-C, and SZ PRS were the most important according to the Ratkowsky-Lance criterion. Discussion The small silhouette values across clustering solutions highlight that only very minimal cluster separation is occurring. Any cluster separation that is achieved is predominantly driven by SZ and BD PRS. For three clusters, additional differentiation can be seen based on the PRS for certain physical health traits and disorders. The observed clustering could result from fine-scale population structure and/or local ancestry rather than genuine differences in PRS profiles, and this will be investigated further.