Penetrance and pleiotropy of polygenic risk scores for schizophrenia in 90,000 patients across three healthcare systems

Abstract

BACKGROUND: One of the major rate-limiting features of advancing treatment and intervention for psychiatric disorders is the lack of robust risk stratification tools. Given the substantial heritability of many psychiatric disorders, quantitative measures of genetic risk may be useful towards this end. Polygenic risk scores (PRS), in particular, are easy and cheap to generate and can be calculated well before illness onset. However, their utility for predicting clinical diagnoses collected in a real-world clinical setting remains unclear as PRS have typically been applied only in research samples. Furthermore, the specificity of these risk scores, especially given the overlap of genetic risk across psychiatric disorders, remains unassessed in a clinical setting.

METHODS: As part of the PsycheMERGE Consortium, we generated PRS for several psychiatric disorders using summary statistics available from the Psychiatric Genomics Consortium and applied them to real-world clinical datasets from three large-scale, independent healthcare systems. PRS were calculated for individuals of European-American ancestry with genomic data from the Partners Healthcare Biobank (N=15,363), Vanderbilt University Medical Center biobank (BioVU) (N=11,647) and Geisinger Health System MyCode dataset (N=23,296). Associations between each PRS and all medical outcomes available in patient electronic health records were assessed using univariate logistic regression in a phenome-wide association study (PheWAS). Medical outcomes were defined using ‘phecodes’: a hierarchical grouping of ICD-9 diagnostic codes used to reduce tens of thousands of individual codes into fewer than 2000 disease categories. Cases for a given phecode had two or more relevant ICD-9 codes, and phecodes with 100 cases or fewer were excluded. All tests were corrected for the number of phecodes tested using a Bonferroni adjustment.

RESULTS: Patterns of results were remarkably similar across health systems and reflected many predicted associations. For example, schizophrenia PRS was associated with schizophrenia in all three health systems (p’s < 5.6 x 10-6), with bipolar disorder in two samples (p’s < 1.1 x 10-5), and with anxiety disorders in two samples, (p’s < 3.6 x 10-7).

DISCUSSION: We have provided an important proof-of-concept demonstration of the utility of PRS using real-world clinical diagnoses across multiple health care systems. Across three independent clinical samples, the associations were remarkably consistent and particularly strong for psychiatric disorders. These results suggest psychiatric PRS could be applied with relative success and specificity in hospital settings, elevating their potential for risk stratification efforts in the future.