Exploring the contribution of polygenic risk scores to explaining variance in symptom level after internet-delivered cognitive behavioral therapy for depression and anxiety disorders

While internet-delivered cognitive behavioral therapy (ICBT) has been proven effective in treating depression and anxiety disorders, many patients will continue to have a high symptom level following treatment. Establishing factors associated with post-treatment symptom severity is a critical first step in devising personalized interventions. The literature on predictors of treatment outcome is abundant, yet mostly based on a limited number of clinically available predictors. Genetic differences may account for part of the variance in the treatment outcome, similar to the role they play in any complex behavior. However, it is unclear to what extent the predictive power of PRS for different traits is attenuated when used in conjunction with rich phenotypic predictors. In this study, we evaluated the added benefit of utilizing polygenic risk scores (PRS) to predict post-ICBT symptom level in addition to clinical and registry-based socioeconomic predictors. 2668 patients with major depressive disorder (n=1330), panic disorder (n=727), and social anxiety disorder (n=641) were treated with ICBT at the Internet psychiatry clinic, Stockholm, Sweden. Prior to treatment, blood samples were collected, genotyped, and aggregated into PRS for eight traits (MDD, ASD, ADHD, Bipolar Disorder, Schizophrenia (SZ), Educational Attainment, IQ, and Cross Disorder). Clinical data consist of eight predictors such as self-reported pre-treatment symptom severity, comorbidities, and symptom duration. Socioeconomic predictors were derived from high-quality national registries (National Patient Register, Prescribed Drug Register, Longitudinal integrated database for health insurance and labor market studies, Total Population Register, and Social Insurance Agency) and comprise 30 predictors, including prior psychiatric diagnoses, medication use, education, employment, financial aid, and income. Three ordinary least squares regression models were developed: A baseline model using only PRS as predictors, a model with all clinical and socioeconomic predictors but excluding PRS, and the full model. Goodness of fit of the three models was compared based on their respective adjusted coefficients of determination (adjusted R2). The model including only PRS as predictors yielded an adjusted R2 of 0.00405, thus explaining 0.41% of the variance in post-treatment symptom level. The model with all the predictors besides PRS explained 33.80%, and the full model explained 34.11% of the variance. No PRS was statistically significantly associated with the post-treatment symptom level in bivariate models. We found that the addition of PRS provides only marginal additional value for predicting post-ICBT symptom level when an extensive set of phenotypic predictors are used (< 1%). A potential future avenue is development of a machine learning model that can handle a large number of features for individual-level predictions with unaggregated SNPs as opposed to PRS. In addition, there is preliminary evidence suggesting that PRS may play a bigger role in predicting relative symptom change.