Multi-omic biomarker identification and validation for diagnosing warzone-related post-traumatic stress disorder

Abstract

Post-traumatic stress disorder (PTSD) impacts many veterans and active duty soldiers, but diagnosis can be problematic due to biases in self-disclosure of symptoms, stigma within military populations, and limitations identifying those at risk. Prior studies suggest that PTSD may be a systemic illness, affecting not just the brain, but the entire body. Therefore, disease signals likely span multiple biological domains, including genes, proteins, cells, tissues, and organism-level physiological changes. Identification of these signals could aid in diagnostics, treatment decision-making, and risk evaluation. In the search for PTSD diagnostic biomarkers, we ascertained over one million molecular, cellular, physiological, and clinical features from three cohorts of male veterans. In a discovery cohort of 83 warzone-related PTSD cases and 82 warzone-exposed controls, we identified a set of 343 candidate biomarkers. These candidate biomarkers were selected from an integrated approach using (1) data-driven methods, including Support Vector Machine with Recursive Feature Elimination and other standard or published methodologies, and (2) hypothesis-driven approaches, using previous genetic studies for polygenic risk, or other PTSD-related literature. After reassessment of ~30% of these participants, we refined this set of markers from 343 to 28, based on their performance and ability to track changes in phenotype over time. The final diagnostic panel of 28 features was validated in an independent cohort (26 cases, 26 controls) with good performance (AUC = 0.80, 81% accuracy, 85% sensitivity, and 77% specificity). The identification and validation of this diverse diagnostic panel represents a powerful and novel approach to improve accuracy and reduce bias in diagnosing combat-related PTSD.

Fig. 1

Overview of PTSD biomarker identification approach—details of cohort recruitment, and biomarker identification, down-selection, and validation

Fig. 2

Overview of molecular datasets and cohort symptom severity. a Flow diagram for participant recruitment and enrollment. Participant eligibility was determined through a phone pre-screen and a baseline diagnostic clinical interview. Eligible participants completed fasting blood draws for multi-omic molecular assays. Participants in the initial discovery cohort were invited to return for follow-up in the recall cohort. Some participants returned with symptom changes, including “subthreshold” PTSD symptoms (below original study inclusion criteria). b Trajectory of PTSD symptoms in recalled participants. CAPS total for current symptoms at baseline (T0) and follow-up (T1) for each participant are connected. Participants who remained in the PTSD + group at both time points are shown in red. Participants who remained in the PTSD- group are shown blue. Participants with PTSD status changes are shown in gray, including participants who became “subthreshold” PTSD cases. c Distribution of molecular data types at three stages of biomarker identification: full exploratory dataset (All Data), reduced set of 343 potential biomarkers (candidate set) and the final panel of 28 biomarker (final set). Methylation and GWAS data represents 99% of initial data screen due to high-throughput arrays. Other molecular data types are well represented in the second and final stages of biomarker identification and selection

Fig. 3

Validation of biomarker panels. a ROC curve for identified biomarker panel (28 markers), illustrating good performance in an independent validation dataset (26 cases, 26 controls). Shaded region indicates 95% confidence interval, determined by 2000 bootstrapping iterations. Operating point closest to (0,1) on ROC curve used for calculating sensitivity, specificity, and accuracy. b Predicted probability of PTSD based on trained random forest model using a biomarker panel of 28 features. In PTSD participants, predicted PTSD probability is correlated with PTSD symptom severity, measured by CAPS (r = 0.59, p < 0.01). c Random forest variable importance of the final 28 biomarkers. Variable importance was determined using biomarker model training data (cohorts 1 and 2). The top 10 biomarkers, based on random forest variable importance, contain multiple data types, including methylation markers (cg01208318, cg20578780, and cg15687973), physiological features (heart rate), miRNAs (miR-133a-1-3p, miR-192-5p, and miR-9-1-5p), clinical lab measurements (insulin and mean platelet volume), and metabolites (gammaglutamyltyrosine). d Correlation between PTSD biomarkers. Pearson correlation coefficients were computed in the combined set of all three cohorts. The final set of identified biomarkers show small clusters of moderately correlated features, primarily grouped by molecular data type (proteins, miRNAs, and methylation markers). e Biomarker panel performance evaluation during panel refinement, across molecular data types, and in nonlinear features. The validation AUC improves after biomarker down-selection and model refinement. The final biomarker panel validates with greater AUC over the initial biomarker candidate pool (343 markers, AUC = 0.74), and stage one refined panel (77 markers, AUC = 0.75). The final multi-omic panel also outperforms each individual molecular data type. Performance metrics for nonlinear feature combinations, Global Arginine Bioavailability Ratio (GABR) and lactate/citrate. Both nonlinear combinations outperform their individual components in AUC (0.60 vs. 0.51 and 0.55 vs. 0.52 in GABR and lactate/citrate, respectively). Error bars indicate 95% confidence interval, determined by 2000 bootstrapping iterations. f Validation performance by ethnicity, and in the presence of major depressive disorder (MDD). Validation performance in Hispanic participants was higher than other ethnicities (non-Hispanic White, non-Hispanic Black, non-Hispanic Asian). PTSD cases with comorbid MDD (n = 9) are easily distinguishable from all combat-exposed controls (n = 26), with AUC = 0.92, while PTSD cases without comorbid MDD (n = 17) are only moderately distinguishable from controls (n = 26), with AUC = 0.73