Differences in polygenic risk score for major depressive disorder depending on case definition in an electronic health records study

Major depressive disorder (MDD) is a polygenic trait, and in recent years polygenic risk scores (PRS) have been used to estimate individuals’ overall genetic risk of the disease. Large research cohorts such as the Psychiatric Genomics Consortium and biobanks such as UK Biobank are typically used to estimate genetic variant effects that are subsequently used to compute PRS in other cohorts. However, approaches used to classify individuals as MDD cases vs. controls vary among studies (e.g., using self-reporting vs diagnosis codes in electronic health records [EHRs] research), and often characteristics of the discovery GWAS cohorts differ from cohorts where PRS are applied. This variation has implications for PRS-phenotype associations in the target population. In this study, we calculated a PRS for MDD in the Mayo Clinic Biobank (MCB), a relatively older cohort, and examined differences in MDD PRS with different case/control definitions. Using data from 44,103 MCB participants of European, we identified those with a history of MDD prior to enrollment (“existing” MDD) using ICD codes or self-report of a diagnosis. Participants that had their first EHR record for MDD after enrollment (“new” MDD) were also identified. PRS were estimated using LDpred2 based on estimates from a GWAS of PGC and UK Biobank data. Participants with existing MDD were classified into 3 groups: those who self-reported MDD only without ICD codes in EHR; those who had ≥1 ICD codes for MDD but did not self-report a MDD diagnosis; and those who had both self-report and ICD codes for MDD. PRS were compared among 3 groups with existing MDD using linear regression models. Similar analysis was conducted to compare MDD PRS among new and existing MDD groups and those with no evidence of MDD (“controls”). In this cohort (median age at enrollment = 63 [25th-75th percentile: 52-72]), 29% had a history of MDD at enrollment (existing MDD) and 5.3% had the first diagnosis in the Mayo EHR after enrollment (new MDD). Among those with existing MDD, MDD-PRS was higher in the self-reported only group than the ICD-only group (mean PRS: 0.187 vs 0.061; p =3.0 × 10^-7). The PRS in the group with both self-reports and ICD codes was similar to the self-report only group (mean PRS: 0.163 vs. 0.187; p =0.20). When comparing existing MDD, new MDD, and controls, although MDD-PRS was higher in both MDD case groups than in controls, MDD PRS was higher for participants with existing MDD than those with new MDD (mean PRS: 0.155 vs -0.031; p =1.4 × 10^-17). This study demonstrates how the MDD PRS can vary considerably between different groups of patients with MDD depending on how the patients are identified, particularly in the context of identifying older patients with MDD using ICD codes in EHRs vs self-reported diagnosis. It suggests MDD PRS may be less predictive in the context of new MDD in older adults. More generally, it emphasizes how the utility of PRS may differ for different patient groups. While this concept is already well recognized in the contexts of reduced PRS transferability across different ancestral groups, the same concept applied to other differences in study cohorts.