Next-generation precision medicine for suicidality prevention

The study addresses the need for objective, quantitative tools to assess and prevent suicidality, a growing public health crisis. The research question is whether blood-based gene expression biomarkers, discovered and validated using multi-cohort, whole-genome approaches, can track suicidality state and predict future risk, and how these biomarkers integrate with psychosocial measures to enable precision medicine. Building on prior biomarker work and the CFI-S scale, the authors employ larger cohorts, RNA-seq, machine learning, and longer longitudinal follow-up to expand identification, prioritization, validation, and testing of suicidality biomarkers. The purpose is to develop clinically actionable panels and reports, understand underlying biology, and match therapies, thus improving prevention and treatment. The importance lies in providing objective, scalable, and personalized assessment to stem rising suicide rates.

The authors’ earlier work identified blood gene expression biomarkers for suicidality and developed the CFI-S score focusing on social determinants without asking about current ideation. Other groups have validated blood-based approaches. These expression studies complement genetic studies (GWAS, association, CNVs, linkage) that have identified suicidality risk loci and overlap with psychiatric disorders. The study integrates literature via a Convergent Functional Genomics (CFG) platform encompassing 551 suicidality papers (human genetics, brain, and peripheral studies), assigning evidence points to prioritize candidates. Prior evidence implicates serotonin (e.g., SLC6A4), MAOA, stress-related and apoptosis pathways, and comorbidity with depression, alcohol use, and stress. Previous results indicated that combining biomarkers with clinical measures improves prediction, and that biomarkers can inform pharmacogenomics and repurposing.

Design: Multi-step, multi-cohort study integrating discovery, prioritization (CFG), validation, and independent testing, with cross-sectional and longitudinal analyses; followed by pathway, pharmacogenomic, drug repurposing, reporting, subtype analyses, and machine learning. Ethics: Indiana University IRB-approved protocol (1011004024). Informed consent obtained; postmortem samples via coroner’s office with next-of-kin consent. Cohorts:

• Discovery (within-subject): 90 adults with psychiatric disorders (69 males, 21 females) with at least one diametric change in HAMD-SI from 0 to ≥2 or vice versa; 273 blood samples; multiple visits (2–6 per subject). Biomarker signals derived from longitudinal within-subject changes to factor out inter-individual variability.
• Validation (postmortem): 101 suicide completers (83 male, 18 female); last-seen-alive PMI ≤24 h; methods largely non-overdose (e.g., 59 gunshot, 27 hanging), minimizing expression confounds.
• Independent Testing (state): 406 subjects (330 males, 76 females), 820 samples spanning no, intermediate, high SI.
• Independent Testing (trait – 1-year hospitalizations): 326 subjects with ≥365 days follow-up (279 males, 47 females), 685 samples; outcomes classified by presence/absence of suicidality-related hospitalization in first year.
• Independent Testing (trait – all future hospitalizations): 365 subjects (310 males, 55 females), 745 samples; follow-up up to 17.2 years (mean 7.8 years). Medications: Heterogeneous; effects mitigated by within-subject discovery, gender-based Z-scoring in test cohorts, and CFG prioritization incorporating genetics independent of medication. Assays:
• Whole blood collected in PAXgene tubes; RNA extracted.
• Microarrays (Affymetrix): n=794 samples; RMA normalization, then gender-specific Z-scoring.
• RNA-seq: n=248 samples; transcript-level DE; zeros substituted with next lowest non-zero for fold-change stability. Integrated Data Processing:
• Discovery Step 1: Separate DE analyses for microarray probesets and RNA-seq transcripts to score tracking of suicidality across visits; map transcripts to probesets (Ensembl 110); integrate scores per probeset; assign percentiles and points (≥80%:6; ≥50%:4; ≥33.3%:2). Threshold ≥2 points to proceed. Initial yield: 9184 unique probesets.
• Prioritization Step 2 (CFG): Map probesets to genes via Ensembl/NetAffy/UCSC/GeneCards; score literature evidence from curated databases (551 suicidality papers): Brain expression 6, Peripheral expression 4, Genetics 2; add to discovery score; 2438 probesets (score ≥6) advanced.
• Validation Step 3: Integrate platforms using housekeeping normalization (ACTB-224594_x_at) and gender-specific scaling factors (microarray intensity/RNA-seq TPM averages) to create virtual probesets; ANOVA across Discovery No SI, Discovery High SI, and Suicide Completers to identify stepwise expression changes; Bonferroni significant=6 points, nominal=4, stepwise=2. Outcomes: 739 nominally significant; 382 Bonferroni significant.
• Candidate Set: Combine Steps 1–3 into CFE3 (max 24); 2340 top candidates with CFE3 ≥8 carried forward.
• Testing Step 4 (clinical utility): Independent cohorts; predictors computed cross-sectionally (levels) and longitudinally (levels, slopes, max levels, max slopes; minima for decreased markers). Metrics: ROC AUC for state and 1-year hospitalizations; t-tests; Pearson correlations; Cox regression and odds ratios for all-future hospitalizations. Scoring: significant in all=4 points; significant by gender=2 points (max 4 across cross-sectional/longitudinal). Final CFE4 max=36.
• Generalizability Step 5: Retest all nominally significant Step 4 biomarkers in entire database (n=1127; males=893, females=234). Build gender-specific panels: top 12 biomarkers each for state, 1-year, and all-future risk (36 per gender) to generate patient reports and treatment matching. Pathway and Network Analyses: DAVID (v2023q3), IPA (Qiagen v107193442) for canonical pathways and diseases; STRING for protein interaction networks. Therapeutics:
• Pharmacogenomics: Match biomarkers to drugs/nutraceuticals known to modulate expression opposite to disease direction using CFG databases.
• Drug repurposing: Connectivity Map (L1000) query on top biomarkers; analyze normalized connectivity scores (experimental drugs filtered). Reporting: Generate prototype physician reports with percentile risk scores (state, 1-year, future) and personalized treatment matches by gender using weighted biomarker panel scores combined with CFI-S and HAMD-SI. Subtypes: Unsupervised two-way hierarchical clustering on stress (SSS4), anxiety (SAS4), mood (SMS7), psychosis (PANSS Positive) in high SI discovery cohort (n=103) revealed 16 subtypes; relate to hospitalization frequencies. Machine Learning: Compare SVM, RF, XGBoost, Transformer, and DNN using biomarker panels + CFI-S items + HAMD-SI. Gender-specific DNN models trained (females: train 44, test 115; males: train 217, test 570) with specified architectures (batch norm, dropout 0.2, ReLU, sigmoid output, lr=0.001, BCE loss). Feature saliency derived via first-order Taylor approximation. ML results compared to simple additive bio-socio-psychological score integration.

• Biomarkers: The top increased biomarker was SLC6A4 (serotonin transporter); the top decreased was TINF2 (telomere shelterin complex). Other key markers: INSR (↑), CLN5 (↓), BCL2 (↓), APOE (↓), MAOA (↑). After full pipeline, 30 top biomarkers achieved high CFE4 scores, including SLC6A4, TINF2, INSR, CLN5, PKP4, SLC49A4, SKP1, ECHDC1, BCL2, SELENOF, SYNE2, NDFIP1, VTI1B, E2F1, CTIF, MTCH2, PRKAR2B, ANGPT1, KLF12, CDH4, APOE, MYH10, UBL3, CALD1, APC, MAP3K7, MAOA, LINC01432, S100A10, AGO2.
• Validation: 739 candidates nominally significant; 382 Bonferroni significant in stepwise change from No SI → High SI → Suicide completers.
• Predictive performance (examples): • APOE: State (All) AUC 0.74, p=1.30E-22; First-year hospitalizations (All) AUC 0.67, p=9.04E-09; All-future hospitalizations (All) OR 1.38, p=1.26E-05; stronger in females (e.g., state AUC up to 0.88). • SLC6A4: State (Females) AUC ~0.68, p=0.02; First-year hospitalizations (Females) AUC ~0.79, p=0.0007; All-future hospitalizations (Females) OR ~2.71, p=0.002. • TINF2: State (Females) AUC 0.72, p=0.005; First-year (Females) AUC 0.67, p=0.03; All-future (All) OR 1.2, p=0.005. • INSR: State (All) AUC ~0.60, p=0.01; All-future OR 1.38, p=0.002; stronger in females (OR ~1.86, p=0.02). • CLN5: State (All) AUC ~0.66, p=0.0002; All-future OR 1.85, p=0.004.
• Biology: Top pathways involve apoptosis, autophagy, insulin secretion signaling, synaptogenesis, neurotransmitter clearance. Upstream regulators include prednisolone (p=3.05E-07), 3-methyladenine, thyroid hormone, KLF3, simvastatin.
• Comorbidities (genomic overlap among top biomarkers): Alcohol (90%), depression (83.3%), stress (80%), dementia (76.7%), schizophrenia (66.7%), aging (50%), bipolar (50%), pain (46.7%), addictions (46.7%), anxiety (40%). Six of 30 genes (20%) relate to circadian mechanisms.
• Therapeutics: Overall top matches—lithium (26.7%), clozapine (23.3%), ketamine (20%); omega-3 fatty acids top nutraceutical (13.3%). Gender-specific: lithium stronger in females; clozapine stronger in males. Drug repurposing suggested renin-angiotensin modulators (e.g., lisinopril, losartan, ramipril) and cyclooxygenase inhibitors (celecoxib, pranoprofen, tenoxicam), and sertraline.
• Reporting: Prototype individualized reports show integrated bio-socio-psychological risk scores (state, 1-year, future), and ranked treatment matches by gender.
• Subtypes: 16 suicidality subtypes identified; those with high stress/anxiety had the highest subsequent hospitalization rates.
• Predictive integration and ML: Simple additive integration of biomarkers + CFI-S + HAMD-SI outperformed ML, with AUC 0.87 (females) and 0.78 (males) for predicting 1-year suicidality-related hospitalizations. DNN was the best ML approach but below the additive model. Top ML features for first-year hospitalization: females—HAMD-SI and JOSD1; males—JOSD1 and THY1; for imminence: females—“Feeling Useless” and HAMD-SI; males—Medical Problems and Age.

This work demonstrates that blood gene expression biomarkers can objectively track suicidality state and predict future suicidality-related hospitalizations, addressing the need for quantitative measures in clinical practice. The multi-step convergent approach (within-subject discovery, literature-based prioritization, validation in suicide completers, and independent testing) produced robust biomarkers with mechanistic insight—highlighting apoptosis, stress biology, and neurotransmitter clearance. Integrating biomarker panels with psychosocial measures (CFI-S) and current ideation (HAMD-SI) yields synergistic, clinically relevant predictions that surpass ML models trained on the same features. Gender-stratified analyses improved accuracy, with stronger predictive performance in females and differential therapeutic matches. The findings support a model of suicidality as an extreme, stress-driven active aging/apoptotic process with actionable biological targets, enabling pharmacogenomic matching (e.g., lithium, clozapine, ketamine; omega-3) and informing repurposing candidates (RAS and COX inhibitors). Subtyping by stress/anxiety/mood/psychosis further refines risk stratification and potential personalized interventions.

The study advances next-generation precision medicine for suicidality by delivering a validated pipeline of blood biomarkers, integrated clinical reports, and pharmacogenomic treatment matching. Key contributions include: robust, transdiagnostic biomarkers (e.g., SLC6A4, TINF2) predictive of state and future risk; biological insights implicating apoptosis and stress-related pathways; and practical, gender-personalized clinical reporting that synergizes biomarkers with psychosocial measures. Top therapeutic matches align with approved anti-suicidality treatments (lithium, clozapine) and identify repurposing opportunities. Future research should: expand to larger, diverse populations; refine gender- and diagnosis-specific panels; develop point-of-care assays (including protein-level markers); prospectively test biomarker-guided interventions; and further evaluate repurposed agents in clinical trials.

• Medication and comorbidity heterogeneity: Participants were on varied psychiatric and medical treatments, with potential gene expression effects; mitigated by within-subject design and gender Z-scoring, but residual confounding is possible.
• Platform integration: Combining microarray and RNA-seq required housekeeping normalization and scaling; although empirically validated, platform-related artifacts or biases may remain.
• Postmortem validation: Despite PMI controls and normalization, postmortem changes could affect expression.
• Generalizability: Cohorts are psychiatric and VA/academic center-based with male predominance; external validity to general population needs further testing.
• Follow-up variability: Differential lengths of follow-up and potential loss to follow-up could understate predictive power in ROC analyses; Cox models addressed time-to-event but limitations persist.
• ML training/testing not fully independent: DNN hyperparameter tuning used test cohort performance, possibly inflating ML estimates; nevertheless, simple additive integration outperformed ML.
• Peripheral proxy: Blood expression may not fully reflect brain changes; convergence with brain literature and within-subject discovery partially addresses this.