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

Depression and anxiety impose a major global public health burden, yet treatment is hindered by etiological and phenotypic heterogeneity. Current diagnostic labels aggregate diverse neurobiological dysfunctions, contributing to suboptimal outcomes, with over a third of major depressive disorder patients and about half with generalized anxiety disorder not responding to first-line treatment. Precision psychiatry requires standardized, personalized, and clinically interpretable biomarkers that map onto a theoretical neuroscience framework rather than black-box algorithms. Prior biotyping efforts have predominantly used task-free fMRI and large, exploratory feature sets, raising concerns about interpretability and overfitting. Moreover, most studies predict response to a single treatment rather than comparing treatment classes. This study asks whether individualized, theory-grounded brain circuit scores derived from both task-free and task-evoked fMRI can identify stable, clinically meaningful biotypes of depression and anxiety that differ in symptoms, behavior, and responses across multiple pharmacological and behavioral treatments.

Previous work has identified depression biotypes using resting-state fMRI, including frontostriatal and limbic connectivity patterns linked to differing TMS response, default mode network (DMN) hyper/hypoconnectivity, and biotypes associated with comorbid anxiety or poor antidepressant response. However, participant-level biotyping that leverages task-evoked activity/connectivity is lacking, despite evidence that task-based fMRI measures predict antidepressant outcomes and are used in biomarker development. Foundational whole-brain approaches using thousands of features can overfit and lack clinical interpretability. There is a need for a theoretically informed, tractable feature set grounded in large-scale circuit dysfunctions implicated in depression and anxiety (e.g., default mode, salience, attention, negative/positive affect, cognitive control). Additionally, prior studies often assess response to a single modality; translational value would be maximized by comparing responses across pharmacotherapy and behavioral therapy to identify optimal treatment by biotype.

Design and samples: Data were pooled from four studies (iSPOT-D, RAD, HCP-DES, ENGAGE). Baseline imaging was obtained from 801 clinical participants with depression/anxiety spectrum disorders and 137 healthy controls as a reference. At baseline, 95% of clinical participants were unmedicated and none had substance dependence. A treatment dataset included 250 participants reassessed after randomized treatments: escitalopram, sertraline, or venlafaxine XR (n=164) or an integrated behavioral intervention (I-CARE) versus usual care (U-CARE) (n=86). Diagnoses were established using MINI aligned to DSM-IV/DSM-5 or via PHQ-9 screening in ENGAGE.

Imaging acquisition and preprocessing: Participants completed a standardized protocol (Stanford Et Cere Image Processing System) probing six circuits: default mode (D), salience (S), attention (A), negative affect (sad, threat conscious, threat nonconscious), positive affect (P), and cognitive control (C). Task fMRI included Facial Expressions of Emotion tasks (conscious and nonconscious) and a Go–NoGo task; intrinsic (task-free) measures for D, S, A were derived from task data segments. Data were preprocessed with fMRIPrep; quality control excluded scans with artifacts or excessive motion. Multiple imputation (miceRanger) addressed missing circuit score values; scanner/site effects were harmonized with ComBat.

Feature derivation: Regions of interest (29 ROIs) were defined meta-analytically (Neurosynth) and by theory, with psychometric refinement. For each participant, 41 features were computed: task-evoked regional activation and within-circuit task-based connectivity (PPI) for affect and cognitive tasks, and intrinsic task-free within-circuit connectivity for D, S, A. All features were referenced to the healthy control mean and SD to yield individualized regional circuit scores (z-units), enabling interpretability.

Clustering and validation: Pairwise dissimilarity between participants was defined as 1 minus the Pearson correlation across the 41 scores. Hierarchical clustering (average linkage) was performed for 2–15 clusters. Candidate solutions were evaluated using: elbow method (within-cluster distances), simulation-based significance testing of the silhouette index versus a matched multinormal null (10,000 runs), permutation-based silhouette testing (10,000 permutations of circuit scores across participants), cluster stability via crossvalidation (leave-one-out and leave-20%-out ARI), split-half replication of cluster profiles, and concordance with a theoretical taxonomy (identifying circuit features deviating >0.5 SD from the healthy norm).

Clinical characterization: External validation used measures not involved in clustering. Symptoms: standardized self-reports assessed ruminative worry, ruminative brooding, anxious arousal, negative bias (depression), threat dysregulation (anxiety), anhedonia, cognitive dyscontrol, tension, insomnia, and suicidality; depression severity included HDRS (iSPOT-D) and SCL-20 (ENGAGE). Behavior: WebNeuro cognitive tests assessed sustained attention, executive function, cognitive control (Go–NoGo), explicit emotion identification, and implicit emotion priming. Treatment outcomes: severity scaled 0–1; response defined as ≥50% reduction from baseline; remission as HDRS ≤7 or SCL-20 ≤0.5. For each biotype, Wilcoxon tests compared in-biotype vs out-of-biotype medians (χ² for binary insomnia/suicidality). Stability was tested via split-half and leave-study-out procedures. Covariate checks assessed scanner, sex, age, and dataset distributions across biotypes.

Comparative biotyping inputs: Performance of the circuit score feature set (task + task-free) was compared against three alternatives based on resting-state connectivity (whole-brain connectome, DMN connectivity, angular gyrus-centered network), and against using task-free circuit scores alone, using identical validation criteria and resampling/permutation tests.

• Six stable, theory-consistent biotypes emerged, each defined by distinct circuit dysfunction profiles relative to healthy norms and to each other (n=801 clinical participants):
  1. D_C+ S_C+ A_C+ (n=169): intrinsic hyperconnectivity in default mode, salience, and attention circuits.
  2. A_C− (n=161): intrinsic hypoconnectivity specifically within the attention circuit.
  3. NS_A+ P_A+ (n=154): hyperactivation during conscious emotion processing within sad-elicited negative affect and positive affect circuits.
  4. C_A+ (n=258): hyperactivation within the cognitive control circuit during NoGo inhibition.
  5. NTC_C−, C_A− (n=15): reduced connectivity in conscious threat negative affect circuit and hypoactivation in cognitive control.
  6. D_x S_x A_x N_x P_x C_x (n=44): no prominent circuit dysfunction.
• Cluster validation: The six-cluster solution showed beyond-chance clustering and stability (mean silhouette = 0.065; P=0.016 vs multinormal null; P<0.0001 vs permuted data; leave-one-out ARI ≈ 0.80; leave-20%-out ARI ≈ 0.35). Split-half analyses replicated circuit profiles.
• Symptom and behavioral distinctions (selected robust examples): • D_C+ S_C+ A_C+: slowed responses identifying sad faces; increased executive errors; fewer Go–NoGo commission errors; slowed sustained attention responses. This biotype responded better to behavioral therapy (I-CARE) than others (responders 42%, remitters 25%; significant Wilcoxon test). • A_C−: relatively less tension and lower cognitive dyscontrol; faster Go responses on Go–NoGo but more sustained attention commission/omission errors; faster responses to implicit threat primes; comparatively worse response to I-CARE (responders 26%, remitters 22%). • NS_A+ P_A+: greater anhedonia and ruminative brooding. • C_A+: greater anhedonia, anxious arousal, negative bias, and threat dysregulation; more executive errors and time, more Go–NoGo commission and sustained attention omission errors; better response to venlafaxine XR than other biotypes (responders 64%, remitters 40%; significant Wilcoxon test). • NTC_C−, C_A−: less ruminative brooding; faster reaction times to implicit sad faces. • D_x S_x A_x N_x P_x C_x: slower reaction times to implicit threat priming.
• Transdiagnostic nature: Biotypes cut across DSM diagnoses; only current major depressive disorder frequency differed across biotypes (χ² = 24.235, P = 0.0002), with the A_C− biotype having the highest and the intact-profile biotype the lowest proportion with MDD.
• Scanner/site and demographics: After ComBat harmonization, biotype distribution did not differ by scanner (χ² = 12.773, P = 0.237). Sex distribution did not differ (χ² = 12.643, P = 0.244). The A_C− biotype was slightly older on average yet within young to mid-adult range.
• Comparative feature performance: The combined task + task-free regional circuit scores outperformed whole-brain connectomes and DMN connectivity in clustering performance (significant silhouette differences) and were the only set to beat the multinormal null (P=0.016). They did not outperform an angular gyrus-centered network on silhouette alone, though that approach did not yield generalizable symptom differences. Using task-free circuit scores alone did not surpass null clustering and failed to generalize some symptom/behavior differences, underscoring the value of including task-evoked measures.

This study demonstrates that individualized, theory-grounded quantification of both task-free and task-evoked brain circuit function can parse the biological heterogeneity of depression and anxiety into clinically meaningful biotypes. The six biotypes align with a neural circuit taxonomy of mood and anxiety disorders and map to distinct symptom and behavioral profiles, supporting mechanistic interpretability. Crucially, biotypes also showed differential responses across pharmacotherapy and behavioral therapy, suggesting potential for guiding personalized treatment selection. The approach addresses limitations of prior whole-brain, unsupervised methods by using a tractable, hypothesis-driven feature set that reduces overfitting risk and enhances clinical interpretability. Including task-based fMRI features was essential for beyond-chance clustering and for identifying biotypes with distinct emotion and cognitive control dysfunctions, providing insights that resting-state-only approaches might miss. The results support precision psychiatry efforts to move beyond one-size-fits-all diagnoses toward biotype-informed care pathways.

Using standardized, personalized regional circuit scores grounded in a theoretical taxonomy, the authors identified six robust, clinically validated, and interpretable brain-derived biotypes in depression and anxiety. These biotypes differ in symptoms, cognitive and emotional task performance, and response to multiple treatments, providing a practical framework for stratifying patients in precision psychiatry. Future research should replicate these biotypes in independent datasets; prospectively test biotype-guided treatment assignment across modalities; expand task paradigms tapping similar domains; increase sample sizes for treatment comparisons; examine demographic and comorbidity influences; and develop streamlined, clinically accessible protocols to facilitate implementation.

• Task fMRI acquisition adds burden compared with task-free imaging; tasks used are not yet widely available.
• Treatment response analyses were limited by small cell sizes for many biotype-by-treatment combinations (n < 10 for 90% of comparisons), precluding split-half or leave-study-out validation; treatment findings require replication in larger trials.
• Effect sizes for symptom differences were generally small; many did not remain significant in split-half or leave-study-out analyses, highlighting the need for larger samples and finer-grained clinical measures.
• The diverse, transdiagnostic sample improves generalizability but may dilute detection of setting-specific biotypes; some biotypes may reflect comorbidity influences.
• Potential demographic and dataset composition differences could partially drive biotypes; although scanner effects were harmonized and sex distributions did not differ, age differed slightly for one biotype.
• Using a specific set of tasks and ROI definitions may limit generalizability; replication with alternative tasks probing similar domains is needed.