Clustering approaches for patient stratification in psychiatry

Hagenberg, Jonas (2025): Clustering approaches for patient stratification in psychiatry. Dissertation, LMU München: Faculty of Biology

Abstract

Background Mental health is essential for thriving and actively participating in society. However, psychiatric disorders have a global lifetime prevalence of nearly 30% and significantly contribute to the global disease burden. Among the most prevalent disorders are anxiety disorders and depression as well as schizophrenia to a lesser extent. These disorders result from a complex interplay between environmental and biological factors. At the same time, the underlying biological mechanisms are poorly understood. It is known that genetics, a dysregulated stress response and inflammation play a role in their pathophysiology. Nevertheless, effective treatments for depression and other psychiatric disorders are lacking. Reasons for this are a categorical disorder classification based on symptoms and not pathophysiology; symptom heterogeneity within one disorder; and insufficient understanding of the underlying biological mechanisms. In the last years therefore patient stratification emerged as a promising tool to define subgroups beyond the classical disorder classification that share symptoms, pathophysiology and treatment targets. This thesis aims to improve patient stratification in two ways. First, I develop longmixr, a method for robust clustering of longitudinal mixed data consisting of both categorical and continuous variables, which is common in psychiatric research. Second, I use multi-omics data to investigate the association of depression symptoms with inflammation for possible patient stratification. Another goal of this thesis is to better understand the biology of the so called immune-related depression. Therefore, I lastly further examine the cell type specificity of the influence of immune markers on transcriptional changes in immune cells. Methodology For the method development, I combined consensus clustering with flexible mixture models to model longitudinal mixed data and to provide diagnostic metrics for a robust selection of the number of clusters. The method is implemented as an R package complete with extensive documentation. I examined its properties in a simulation study and applied it to a data set of patients with schizophrenia. For the inflammation-based stratification, I leveraged RNA-seq data of peripheral blood mononuclear cells (PBMCs) and a panel of 43 immune markers from a transdiagnostic cohort consisting of 237 individuals and an additional 36 controls without a psychiatric diagnosis. The data was clustered together with age and body mass index (BMI), and the resulting clusters were characterized with symptom data, polygenic risk scores, heart rate variability and structural magnetic resonance imaging data. For the analysis of cell type-specific influence of immune markers on immune cells, I used a single cell RNA-seq data set with approximately 38000 PBMCs from 13 individuals with depression and other psychiatric disorders to calculate differentially expressed genes (DEGs). Results The simulation study showed that the determination of the number of clusters is sample size-dependent. Applied to a data set of patients with schizophrenia, longmixr identified two clusters that differed in their disease severity as well as symptom variability over time. The patient stratification based on multi-omic measures of the inflammation revealed four clusters in the initial clustering: a mild depression symptoms cluster (MIDS), a low immune-related depression symptoms cluster (LIRDS) and two high immune-related depression symptoms clusters (HIRDS). The HIRDS clusters were characterized by elevated BMI, higher depression severity and elevated levels of immune markers such as interleukin-1 receptor antagonist (IL-1RA), C-reactive protein (CRP) and C-C motif chemokine ligand 2 (CCL2). BMI was also the most important variable in the clustering. On the other hand, the RNA-seq data mainly differentiated the MIDS cluster from the other clusters with high depression severity. The gene sets based on these genes were enriched in brain-related pathways. The heart rate variability showed a similar pattern. When I additionally included the predicted cell type proportions in the clustering, this resulted in a three cluster solution. One of these clusters exhibited elevated immune marker concentrations. In general, however, the RNA-seq data had a greater role in clustering than the immune marker data for this clustering. It was found that the cell type estimates were most pronounced in an intermediate depression symptoms cluster. This suggests that immune markers and RNA-seq capture different aspects of immune dysregulation in the context of immune-related depression. The further investigation how immune markers influence the transcriptional state in the single cell RNA-seq data set revealed 55 DEGs with regard to the CRP concentration, 47 for IL-6 and 9 for BMI across all cell types. Most of the DEGs were detected in CD4 TCM cells. The DEGs included genes related to inflammation, the brain and psychiatric disorders including two genes identified in a genome-wide association study for depression. Conclusion In summary, this thesis significantly contributes to patient stratification in psychiatry and the understanding of immune-related depression. Longmixr provides a robust clustering method for longitudinal mixed data and enables other researchers to use their data sets for patient stratification. Additionally, I contributed to the understanding of immune-related depression by assessing it on multiple molecular levels. The results underline the complementary information contained in immune marker measurements and PBMC RNA-seq data about immune dysregulation. Furthermore, they highlight the potential of other immune markers besides CRP for patient stratification, especially IL-1RA and chemokines. The importance of BMI in the clustering further underscores the intricate relationship between obesity and inflammation in the context of immune-related depression. Moreover, the single cell analysis unravels how immune markers interact with immune cells in patients with depression. Taken together, this thesis provides novel immune markers and genes as potential targets for clinical stratification and new therapeutic intervention for people with depression symptoms.

• Item Type: Theses (Dissertation, LMU Munich)
• Keywords: Psychiatry, Depression, Inflammation, Bioinformatics
• Subjects: 500 Natural sciences and mathematics; 500 Natural sciences and mathematics > 570 Life sciences
• Faculties: Faculty of Biology
• Language: English
• Date of oral examination: 7. May 2025
• 1. Referee: Schmidt, Mathias
• Signature of the printed copy: 0001/UMC 31955