QUANTIFYING INDIVIDUAL DEPRESSIVE SYMPTOM BURDEN OVER THE LIFE COURSE USING AREA UNDER THE CURVE (AUC): AN ALTERNATIVE PHENOTYPE FOR GENOME-WIDE ASSOCIATION STUDIES (GWAS)

Abstract

Depression can be a chronic, fluctuating condition, yet most research relies on cross-sectional estimates of symptom severity that fail to capture individual trajectories over time. This project used longitudinal population data and multilevel growth curve models to quantify individual-level depressive symptom burden using AUC methods. This approach provides a scalable, interpretable, and theoretically grounded alternative to the prior cross-sectional phenotype of depression.

This project used data from the UK Biobank of ∼490,000 adults with repeated measures of the 2-item Patient Health Questionnaire (PHQ-2) across eight occasions. A mixed-effect model was fitted via MLwiN in Stata 18.5. The models included fixed effects for age, age-squared and age-cubed, and random intercepts and slopes for age to capture individual deviations in symptom onset and linear change. Age was centred around the grand mean to improve interpretability and reduce collinearity between polynomial terms.

Person-specific predicted symptom trajectories were derived from the estimated model parameters, incorporating both fixed and random effects. The AUC was then calculated for each individual using the trapezoidal rule, integrating predicted PHQ-2 scores across observed time points. This resulted in a continuous measure of cumulative depressive symptom burden—scaled for comparability across individuals with varying follow-up durations. The resulting distribution of AUC values captures substantial inter-individual variability, with higher scores reflecting longer duration and greater intensity of symptoms over time. This phenotype was then used to run a GWAS using REGENIE, controlling for principal components, removing SNPs with genotype missingness > 10%, INFO < 0.1, and HWE p > 1e-15. REGENIE uses a two-step method that enables rapid whole-genome regression modelling. 2 lead SNPs were identified.

This method addresses three key limitations of traditional phenotyping approaches: (1) it preserves longitudinal structure and avoids arbitrary cutoffs, (2) it enables comparison of lifetime symptom burden across individuals, and (3) it accommodates varying observation schedules through flexible multilevel modelling.

We discuss the implications of using AUC as a phenotypic summary in psychiatric genomics. For example, how the cumulative burden derived from model-based predictions may serve as a more stable and interpretable outcome for a Genome-Wide Association Study compared to timepoint-specific symptom scores. Our work demonstrates how longitudinal modelling combined with AUC integration provides a powerful tool for capturing dynamic mental health processes over the life course.