Personalized brain circuit scores identify clinically distinct biotypes in depression and anxiety

Depression and anxiety are heterogeneous disorders with overlapping neurobiological processes, leading to suboptimal outcomes under current one-size-fits-all diagnostic and treatment approaches. A precision psychiatry framework requires standardized, individualized, and interpretable metrics of brain dysfunction to stratify patients into subgroups likely to benefit from different treatments. Prior work has largely focused on resting-state fMRI using thousands of features in unsupervised frameworks, which limits interpretability and risks overfitting, and often evaluates response to only a single treatment modality. This study tests the hypothesis that personalized, theory-driven scores of brain circuit function derived from both task-free and task-evoked fMRI can identify clinically meaningful and treatment-relevant biotypes of depression and anxiety. Using harmonized imaging and clinical measures across four studies, including unmedicated patients (n=801) and a treatment subset (n=250) randomized to pharmacotherapies or behavioral therapy, the authors aim to derive biotypes and assess their external validity against symptoms, cognitive-behavioral performance, and treatment outcomes.

Previous biotyping efforts in depression and anxiety predominantly used task-free (resting-state) fMRI and unsupervised whole-brain approaches, identifying subtypes with aberrant frontostriatal and limbic connectivity and default mode network hyper-/hypoconnectivity. Such approaches can be difficult to interpret clinically, may overfit when using thousands of features, and typically evaluate response to a single treatment (e.g., TMS or a specific antidepressant). Task-evoked fMRI measures, probing emotional and cognitive processes, have robust links to treatment response and have been used as biomarkers in recent pharmacotherapy development, yet participant-level, task-based biotyping evidence has been lacking. This work adopts a theoretically grounded, tractable feature set spanning both task-free and task-evoked circuits to enhance interpretability, mitigate overfitting, and compare responses across multiple treatment classes.

Design and samples: Pooled data from four studies (iSPOT-D, RAD, HCP-DES, ENGAGE). Clinical participants: n=801 with depression/anxiety spectrum; healthy controls: n=137 for reference norms. A treatment subset (n=250) completed randomized trials: escitalopram, sertraline, venlafaxine XR (n=164) or behavioral intervention I-CARE vs usual care U-CARE (n=86). At baseline, 95% were unmedicated; substance dependence excluded. MRI protocol and tasks: Standardized ‘Stanford Et Cere Image Processing System’ assessed six circuits: default mode, salience, attention (task-free); negative affect (sad, threat-conscious, threat-nonconscious), positive affect (happy), and cognitive control (Go-NoGo) under task conditions. Measures included task-based activation and PPI connectivity, and task-free connectivity. Feature derivation: ROIs defined via Neurosynth meta-analyses and theory-driven selection; 29 ROIs; 41 imaging features across activation and connectivity modalities. Clinical participants’ features were referenced to healthy controls as z-standardized ‘regional circuit scores’ (deviation in SD from healthy mean). Preprocessing and harmonization: fMRIPrep pipelines; exclusion for artifacts/incidental findings/excess motion; multiple imputation for missing features via miceRanger (random forests) per scanner; ComBat to remove scanner/site effects. Clustering: Computed pairwise dissimilarity as 1 − Pearson correlation between each participant’s 41-feature vector. Hierarchical clustering with average linkage. Candidate solutions from 2–15 clusters. Cluster validation and selection: Multiple convergent checks: elbow method; silhouette significance vs (a) simulated multinormal null with conserved covariance (10,000 runs) and (b) permutation null by shuffling participant labels (10,000 permutations); stability via cross-validation (leave-one-out, leave-20%-out); split-half replication of cluster circuit profiles; and match to a priori theoretical circuit taxonomy by identifying features deviating ≥0.5 SD from healthy. External clinical validation: Compared biotypes on (1) symptom severity across domains (e.g., anhedonia, anxious arousal, negative bias, threat dysregulation, ruminative worry/brooding, cognitive dyscontrol, tension; insomnia and suicidality via QIDS-SR items), (2) computerized cognitive/emotional performance (WebNeuro: sustained attention, executive function, Go-NoGo, explicit and implicit emotion tasks), and (3) treatment outcomes (scaled HDRS/SCL-20, response ≥50% reduction; remission thresholds HDRS≤7 or SCL-20≤0.5). Statistics: Wilcoxon rank-sum tests with effect size r; χ² tests for categorical outcomes. Replication via split-half and leave-study-out procedures. Assessed demographic and scanner balance; diagnosis frequencies across clusters; and benchmarking against alternative feature sets (whole-brain connectomes, default mode connectivity, angular gyrus network; and task-free-only subset). Code available on GitHub.

Biotype derivation and validation:

• Six-cluster solution supported across validation procedures. Average silhouette = 0.065; significant vs multinormal null (P=0.016) and vs permutation (P<0.0001). Stability: leave-study-out ARI = 0.80; leave-20%-out ARI = 0.35 (and generally good stability across other CV checks). Split-half profiles replicated. Scanner effects removed and not driving biotype distribution (χ²=12.773, P=0.237).
• Six biotypes (n values):
  1. D_C S_C A_C (n=169): task-free hyperconnectivity within default mode, salience, and attention circuits.
  2. A_C (n=161): task-free hypoconnectivity within attention circuit.
  3. N_S_A P_A (n=154): hyperactivation during conscious emotion processing in negative (sad) and positive (happy) affect circuits.
  4. C_A (n=258): hyperactivation in cognitive control circuit during NoGo inhibition.
  5. NTC_C C_A− (n=15): reduced connectivity in negative affect circuit (conscious threat) and hypoactivation in cognitive control circuit during NoGo.
  6. D_X S_X A_X N_X P_X C_X (n=44): no prominent circuit dysfunction relative to healthy or other biotypes. Clinical distinctions (selected examples; see paper for full set):
• D_C S_C A_C: Slower responses identifying sad faces (ES=0.289, P=0.001); increased executive function errors (ES=0.175, P=0.044); fewer commission errors on Go-NoGo (ES=−0.275, P=0.002); slowed sustained attention response times (ES=0.336, P=0.0001). Better response to behavioral I-CARE vs others (ES=−0.612, P=0.037; responders 42%, remitters 25%).
• A_C: Lower tension (ES=−0.196, P=0.049) and lower cognitive dyscontrol (ES=−0.305, P=0.006); faster Go responses (ES=−0.383, P≈6.2×10−?; split-half replicated); more commission and omission errors in sustained attention (ES=0.300, P=0.0004; ES=0.198, P=0.020); faster responses to implicit threat priming (ES=−0.256, P=0.002). Worse response to I-CARE (ES=0.593, P=0.002; responders 26%, remitters 22%).
• N_S_A P_A: More severe anhedonia (ES=0.343, P=0.014; CI (2, 4.5)) and greater ruminative brooding (ES=0.294, P=0.036; CI (55.5, 63)).
• C_A: Greater anhedonia (ES=0.295, P=0.015), anxious arousal (ES=0.218, P=0.003), negative bias (ES=0.188, P=0.003; split-half), and threat dysregulation (ES=0.317, P=5.07×10−7; split-half and leave-study-out). Behavioral deficits: more executive errors/time (ES=0.164, P=0.017; ES=0.152, P=0.027), more Go-NoGo commission errors (ES=0.158, P=0.022; split-half), and more sustained attention omission errors (ES=0.275, P=6.46×10−5; split-half and leave-study-out). Better response to venlafaxine (ES=0.426, P=0.034; responders 64%, remitters 40%).
• NTC_C C_A−: Less ruminative brooding (ES=−0.902, P=0.036) and faster reaction times to implicit sad faces (ES=−0.669, P=0.024).
• D_X S_X A_X N_X P_X C_X: Slower reaction times to implicit threat priming (ES=0.516, P=0.001). Transdiagnostic nature: Biotypes cut across DSM diagnoses; only current major depressive disorder differed in frequency across biotypes (χ²=24.235, P=0.0002); A_C had highest proportion with current MDD; D_X…C_X had lowest. Benchmarking features: The combined task-evoked + task-free regional circuit scores outperformed whole-brain connectomes and default mode resting connectivity on clustering performance (silhouette comparisons), and task-based measures were necessary for beyond-chance clustering and generalizable clinical associations. Task-free-only features did not outperform null clustering under simulation. Angular gyrus network connectivity performed similarly on silhouette but did not yield generalizable symptom differences. Correlative context: Across all participants, significant associations were observed between circuit scores and 21% of symptom measures, 10% of performance measures, and 31% of treatment response measures after FDR correction.

Personalized, interpretable brain circuit scores spanning task-free and task-evoked conditions identified six biotypes that reconcile biological heterogeneity in depression and anxiety with clinically meaningful symptom, behavioral, and treatment-response profiles. This approach addresses limitations of prior unsupervised, high-dimensional resting-state methods by using a theory-driven, tractable feature set and rigorous validation. The biotypes generalized across datasets, were largely independent of scanner effects, transcended DSM categories, and, importantly, showed differential responses to both pharmacotherapies and behavioral therapy. These findings support a precision psychiatry model in which patient stratification by circuit dysfunction can guide selection between behavioral interventions and specific antidepressants (e.g., venlafaxine for cognitive control hyperactivation). Including task fMRI was pivotal, as task-evoked measures captured dysfunctions (e.g., cognitive control hyper-/hypoactivation; emotion-processing hyperactivation) that were critical for biotype definition and clinical prediction. Overall, the results demonstrate a clinically interpretable, biologically grounded biotyping pipeline that can inform prospective treatment assignment and trial design.

This study introduces and validates a theory-driven, participant-level biotyping method using standardized regional circuit scores from both task-free and task-evoked fMRI. Six biotypes were identified with distinct neural profiles that aligned with symptom patterns, cognitive-behavioral performance, and differential responses to multiple treatments. The approach outperformed several alternative feature sets and highlights the value of integrating task-based measures for precision psychiatry. Future work should prospectively test biotype-guided treatment allocation, replicate findings in independent cohorts and clinical trials, evaluate generalization across alternative tasks probing similar domains, and refine clinical phenotyping to detect smaller effects more robustly.

• Effect sizes for many symptom associations were small, and some did not survive split-half or leave-study-out validations.
• Treatment-response analyses were limited by small sample sizes within some biotype-by-treatment cells (n<10 in many comparisons), precluding robust replication analyses; findings should be interpreted cautiously until validated in larger trials.
• Task fMRI acquisition is more burdensome than resting-state only; tasks used here may not be widely available, potentially limiting immediate generalizability.
• The diverse, transdiagnostic sample (a strength for external validity) may dilute detection of additional setting-specific biotypes and could reflect influences from comorbidities or demographic differences between datasets.
• Potential age effects across biotypes (A_C slightly older on average) and dataset composition could contribute to biotype differences despite harmonization.
• While scanner effects were harmonized via ComBat and not associated with cluster distribution, residual site-specific factors cannot be fully excluded.