22. STABLE SYMPTOM DEPRESSION SUBTYPES WITH DISTINCT GENETIC AND HEALTH SIGNATURES

Background Depression is a heterogeneous disorder, with individuals often showing diverse combinations of symptoms that may reflect distinct underlying biological mechanisms. Data driven approaches offer an opportunity to examine this heterogeneity and connect symptom profiles to genetic, health, and lifestyle factors. Methods We applied unsupervised machine learning to UK Biobank data, examining independent samples using the Mental Health Questionnaire from 2017 (Q1) and the Mental Wellbeing Questionnaire from 2023 (Q2). Symptoms for individual's worst ever depressive episodes (N:Q1 = 17,435 and N:Q2 = 14,301) or current depressive episodes (N:Q1 = 3,096 and N:Q2 = 3,240) were obtained. Two machine learning clustering approaches, multivariate Bernoulli mixture models and agglomerative hierarchical clustering, were used to identify co occurring symptoms and subgroup individuals. Symptom cluster stability was evaluated across Q1 and Q2 and across approaches. Associations between clusters and sociodemographic and lifestyle variables, eight health conditions, and polygenic scores for bipolar disorder, schizophrenia, and attention deficit/hyperactivity disorder were tested. Results Based on the symptoms experienced during individuals’ worst ever depressive episodes, there were 14 and 13 clusters identified in Q1 and Q2, respectively. There were 11 clusters identified in both Q1 and Q2 for current depressive episode symptoms. Symptom clusters were highly stable across time (mean correlation > 0.81) and across clustering approaches (Rand Index > 0.83). Several clusters aligned with known clinical subtypes, such as atypical and melancholic depression, while additional clusters reflected novel sets of symptoms. Atypical presentations (e.g., hypersomnia and weight gain) emerged consistently at both time points and were associated with younger age (P < 3.10e-12) and higher BMI (P < 2.00e-16). Distinct clusters combining insomnia, weight gain, and thoughts of death showed strong associations with asthma (P < 7.44e-7), suggesting potential inflammatory mechanisms. Clusters characterised by psychomotor disturbance were robustly associated with Parkinson’s disease (P < 1.00e-6), both before and after mental health assessment. Discussion These findings identify reproducible and clinically meaningful depressive symptom subtypes and demonstrate the value of integrating machine learning with genetic and health data to refine phenotypes relevant to psychiatric aetiology.