Multi-Omic Biomarker Identification and Characterization for Posttraumatic Stress Disorder

Abstract

Posttraumatic Stress Disorder (PTSD) is an anxiety disorder affecting 5-15% of the population that develops after exposure to a traumatic event. Diagnosis and treatment of PTSD is limited by the lack of understanding of the disease mechanism, inability to identify early cases of disease development, failure to determine prognosis, and poor determination of successful personalized treatment. Integrating multi-scale molecular data from blood can provide new insights into the disorder, and generate hypotheses regarding candidate diagnostic biomarkers, novel therapeutic targets, and disease subgroups. Utilizing multiple cohorts of male veterans with and without PTSD, we integrated molecular datasets including genetics, epigenetics, proteomics, metabolomics, and others to address some of these core challenges in PTSD diagnosis, prognosis, and treatment. Primarily, our efforts were focused in identification of blood-based biomarkers for diagnosing PTSD, and characterization of disease biology. Using an integrated methodology, we consolidated candidate biomarkers across all molecular data types, and performed two stages of biomarker panel refinement to identify a final candidate biomarker panel for diagnosing PTSD. We evaluated the performance of this final set of 28 biomarkers in an independent validation cohort, and achieved accuracy of 81%, sensitivity of 85% and specificity of 77% in diagnosing warzone-related PTSD in a cohort of combat-exposed male veterans. This biomarker panel consisted of a heterogeneous set of molecular data types, including DNA methylation, miRNAs, proteins, metabolites, and small molecules. These features include some with direct links to known PTSD biology, including methylation of the PDE9A and CPT1B genes, which have previously been shown to have altered expression in PTSD, as well as novel biology, including miRNAs associated with obesity, diabetes, and inflammation. Additionally, we have developed and implemented multiple novel methodologies for: (1) characterizing diseases from high-throughput datasets, (2) integrating incomplete multi-omic data, and (3) identifying biomarkers in clinical settings where false positive and false negative errors have unequal weights. We quantified changes in expression variance from high throughput datasets, and incorporated the differential expression variance genes into a network-based biomarker identification framework. These network-based biomarkers from gene expression variance resulted in larger disease-related subnetworks that can be used for classification. Next, we proposed a maximum probability integration strategy that allows for integration of incomplete multi-omic datasets. Using individual data type classifiers, we incorporated additional incomplete samples in the training dataset, and evaluated performance based on maximum probability in the test set, resulting in improvements in AUC. Finally, we used expression variance to generate a feature selection approach that improves either sensitivity or specificity of the prediction, without loss of accuracy. Overall, we have applied multiple novel approaches to identifying biomarkers and characterizing disease signals from multi-omic PTSD datasets. These algorithms have generated candidate biomarkers for further evaluation, identified disease signals for future studies, and provided tools for analyses of new datasets, in PTSD or any disease of interest.