Profiling of blood-based transcriptomics and multimodal modelling of clinical outcome in ultra-high-risk individuals from the LYRIKS cohort

• Main Author: Tan, Samuel Ming Xuan
• Other Authors: Goh Wen Bin Wilson
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Medicine, Health and Life Sciences; Mental health; Psychosis; RNA-seq; Ultra-high-risk; Schizophrenia; Clinical outcome prediction
• Online Access: https://hdl.handle.net/10356/180555
• Institution: Nanyang Technological University

Psychosis is the defining characteristic of schizophrenia spectrum and psychotic disorders (American Psychiatric Association, 2013). The onset of psychosis is often preceded by a period of subthreshold symptoms commonly referred to as the prodrome (Yung et al., 1996). Individuals identified to be in the prodrome are designated as Ultra-High-Risk (UHR) (Yung et al., 2005). Accurate Identification of individuals at high risk of developing psychosis is a key component in clinical intervention strategies (Fusar-Poli et al., 2017a). However, existing UHR instruments such the Comprehensive Assessment of At-Risk Mental States (CAARMS) and the Structured interview for Prodromal Syndrome (SIPS) are prone to false positives (Oliver et al., 2022). Over the years, substantial research effort has been invested into developing improved prognostic markers across numerous fronts. Notable fronts include the use of neuroimaging markers (Smieskova et al., 2010), genetic and molecular markers from Genome Wide Association Studies (GWAS) (Singh et al., 2022; Trubetskoy et al., 2022) or transcriptomic sequencing methods (Chaumette et al., 2019; Kuzman et al., 2009; Mongan et al., 2020; Sanders et al., 2017), and the use of machine learning methods to develop outcome prediction models (Cannon et al., 2008; Fernandes et al., 2020; Koutsouleris et al., 2018). However, at the time of writing, the prospect of clinical deployment of these methods remains distant (Meehan et al., 2022). This is largely due to the inability of candidate biomarkers and prediction models to generalize across different study cohorts (Chekroud et al., 2024; Farrell et al., 2015). This is a problem that likely can only be solved through the combined efforts an entire generation of investigators, involving significant advancement in our understanding on the biological basis of psychosis; better technology; and substantial improvement in study (and predictive model) designs. We seek to contribute towards this goal in two ways: 1) contribute to the understanding of peripheral blood transcriptomic changes associated with psychosis conversion and UHR remission using longitudinal cohort (as opposed to cross-sectional cohorts more commonly seen in this space). And 2) The development of a framework for a more reliable evaluation of outcome prediction models. The Longitudinal Youth at Risk Study (LYRIKS) is a Singapore-based longitudinal observational cohort study that tracks the outcome of participants assessed to be at Ultra-High-Risk of developing psychosis (Lee et al., 2013). Prior studies on the LYRIKS cohort have revealed that cognitive deficits are strong predictors of psychosis conversion (Lam et al., 2018) and that Polygenic Risk Scores (PRS) can be used to differentiate UHR participants from healthy control participants (Lim et al., 2020). The peripheral whole blood samples collected in the LYRIKS cohort study provide us with the opportunity to examine transcriptomic changes associated with UHR outcomes using longitudinal cohorts and contribute to the limited number of studies in this area. In Part 1, we profiled the transcriptomic changes of psychosis conversion in the peripheral blood by RNA-sequencing. We identified 7 genes differentially expressed between both the comparisons involving samples collected before and after psychosis conversion, and the comparison involving samples from UHR individuals who do not convert. These genes are theoretical candidate biomarkers as they are associated with psychosis onset and are differentially expressed between patients with different outcomes. These genes are CYP51A1P1, UBE3AP2, CA12, COL9A2, ENSG00000285966, ENSG00000248936, ENSG00000278112. Our gene ontology enrichment analysis findings on the genes differentially expressed in psychosis conversion are consistent with the patterns of immune dysregulation. Additionally, we found that the differential genes identified in a longitudinal study design will not be the same as those identified in a cross-sectional study design. In Part 2 we examined the transcriptomic changes associated with non-conversion outcomes namely remission and maintaining UHR status. We performed Weighted Gene Correlation Network Analysis (WGCNA) and found that the transcriptomic patterns of UHR individuals who maintain UHR status deviated from that of healthy controls. This deviation was not observed in UHR individuals who remitted. In Part 3, we demonstrated a framework for comparing different combinations of data modalities in a development of multimodal outcome prediction models. We paid particular attention to the control of potential background associations in the data. In our framework, the performances of the models were compared against models trained on noise data (null models). We found that even models with seemingly good performance often fail to outperform null models. Our findings in this section suggest that the apparent performances of outcome prediction models should be examine in greater depth and the use of null models warrants greater exploration. In its sum, we believe that the findings in this series of work can contribute materially towards improved understanding of the biological basis of psychosis conversion and the development of clinically viable outcome prediction models. Our findings in Part 1 revealed that longitudinal study designs are paramount when studying the transcriptomic changes of psychosis conversion. Our findings in Part 2 revealed for the first time the gene expression patterns of UHR individuals who maintained UHR status deviate from that of control individuals. Lastly, we hope that the framework presented in Part 3 would be useful to investigators seeking to evaluate their outcome prediction models while controlling for background associations.