Developing Clinically Interpretable Neuroimaging Biotypes in Psychiatry

Abstract

Despite available treatments, major depressive disorder (MDD) remains one of the leading causes of disability across medical conditions. The current symptom-based diagnostic system groups patients with highly heterogeneous presentations, with no biomarkers to guide treatment, akin to diagnosing heart disease solely by chest pain without imaging to reveal the underlying pathology. Lacking biological guidance, clinicians rely on trial-and-error prescribing. Only 33% of individuals with MDD achieve remission in response to initial treatments, and most cycle through multiple treatments over an average of 7 years. The risk of relapse increases with each treatment failure, rising from 50% to 90%.

This critical review synthesizes studies showing how functional magnetic resonance imaging (fMRI) can predict treatment outcomes and identify which treatment is most effective for an individual based on their brain circuit profile. We illustrate one such method, a theoretically informed approach that quantifies dysfunction across 6 large-scale biotype circuits, relative to healthy reference norms. The resulting personalized circuit scores serve as predictors of response or failure and as moderators of differential treatment outcomes. Matching treatment to a patient’s biotype, defined by their personalized circuit scores, has the potential to double remission rates compared with unmatched treatment. We place this example in the broader context of precision imaging approaches to parsing MDD heterogeneity. We also discuss key challenges, limitations, and future directions for translating fMRI-based tools into clinical practice.

Major depressive disorder (MDD) affects 280 million people globally (1). An estimated 21 million adults in the United States have had at least 1 major depressive episode, representing 18.6% of young adults ages 18 to 25 years and 8.3% of all adults over the past year (2). Although MDD is currently diagnosed and treated as a single entity, accumulating evidence identifies brain-based subgroups that reflect its heterogeneity (3–5), highlighting the need for personalized treatment approaches tailored to individual needs and brain profiles.

In the absence of personalized approaches, only one-third of patients with MDD achieve remission (6), and it takes an average of 7 years to find an effective treatment (7). Each treatment failure increases the risk of recurrence, rising from 50% to 90% (8). This trial-and-error process contributes to the cumulative burden of MDD, which remains the leading global cause of disability (9–12).

This review focuses on functional magnetic resonance imaging (fMRI) as a direct measure of brain circuit function, capable of capturing individual differences that underlie the heterogeneity in MDD and informing more effective treatment selection. We illustrate how patient-level biotype circuit scores, generated using a standardized image processing system that incorporates methods developed in our laboratory (13), can accelerate and personalize treatment decisions by identifying neural targets linked to response. This approach is situated within the broader literature on fMRI studies of predictors of depression treatment outcomes. Consistent with the theme of this Biological Psychiatry special issue, we highlight how neuroimaging can support clinical decision making by outlining key opportunities and challenges for integrating fMRI-based tools into precision psychiatry.

ADVANCING PRECISION MEDICINE IN PSYCHIATRY THROUGH CIRCUIT-BASED fMRI

Current clinical approaches to diagnosing MDD rely on checklists of self-reported symptoms. However, 2 patients who both meet diagnostic criteria may present different symptom profiles (14). Furthermore, symptom reports do not distinguish among underlying brain dysfunctions that influence treatment response. Consequently, treatment decisions remain rooted in trial and error, guided by general trends rather than being informed by patients’ neurobiological characteristics. Reliance on symptoms alone does not improve remission rates beyond chance (15).

fMRI offers a noninvasive, scalable method for directly measuring brain circuit function. Over the past 2 decades, resting and task fMRI have identified large-scale brain circuits associated with the heterogeneity of MDD (16). The clinical utility of fMRI in psychiatry is gaining traction, following a trajectory similar to the evolution of imaging in cardiovascular medicine (17,18). The introduction of cardiac MRI, computed tomography, and echocardiography fundamentally transformed that field, shifting it from diagnosis by observable symptoms toward precision care guided by direct measures of cardiac structure and function (19,20).

Psychiatry is now at a similar inflection point. Clinical interviews remain central to diagnosis but offer limited insight into underlying biology. In contrast, fMRI enables the identification of neural circuit biotypes, distinct patterns of brain dysfunction that differentiate patients with similar symptoms but divergent underlying neurobiologies (Figure 1).

Figure 1.

The analogy between precision medicine in cardiology and psychiatry. An illustration comparing cardiology—in which imaging of the heart at rest and during stress tasks is the gold standard for informing diagnosis and treatment decisions—with the vision for precision medicine in psychiatry, in which brain imaging at rest and during tasks are used to guide diagnosis and treatment decisions for mental disorders.

In this review, we outline a road map for leveraging fMRI-derived measures to stratify patients and guide treatment selection. This approach supports not only selection among first-line treatments but also the early identification of patients unlikely to benefit from standard therapies, thereby enabling timely transitions to alternate treatments (Figure 2).

Figure 2.

Conceptual overview: from trial-and-error to circuit-guided precision psychiatry. A conceptual comparison of the current trial-and-error approach to treating major depressive disorder—which involves successive antidepressant trials, a base remission rate of 33%, and an average of 7 years to identify an effective treatment—with a circuit-guided precision medicine model. In this model, biotype circuit measures derived from functional magnetic resonance imaging (fMRI) are used to stratify patients based on specific circuit dysfunctions and guide treatment selection according to individualized profiles. This targeted approach aims to reduce prolonged trial-and-error prescribing and potentially double remission rates by addressing the underlying source of dysfunction.

A PLATFORM FOR PATIENT-LEVEL CIRCUIT-GUIDED TREATMENT SELECTION

To operationalize the vision for fMRI-guided care, we present a structured system for deriving patient-level circuit biotype scores that are reproducible, clinically interpretable, and benchmarked to healthy norms: the Stanford EtCere Image Processing System.

Deriving Personalized Circuit Scores

EtCere is designed to operate across both research and clinical scanners, using established task-based and resting-state fMRI sequences to assess 6 major neural circuits (Figure 3A and Supplemental Methods S1). Scan data are processed through a containerized, high-performance computing pipeline that automates functions and is comparable to those performed by MRIQC, SPM, and FSL, enabling rapid, standardized, and reproducible results. The framework applies rigorous quality control, standardized preprocessing, and validated patient-level quantification methods for deriving circuit measures (5,13).

Figure 3.

Framework and visualization of EtCere-derived circuit scores. (A) Visualization of the 6 large-scale biotype circuits quantified by the Stanford EtCere Image Processing System: default mode (blue), salience (green), attention (yellow), negative affect (orange), positive affect (purple), and cognitive control (dark red). A standardized procedure was used to define constituent regions and region-to-region connectivity for each circuit. See Table S1 for details. (B) Standardized personalized circuit scores for all 6 biotype circuits in 2 illustrative patients. Scores are referenced to a healthy normative dataset and expressed in units of standard deviations. Darkened bars denote the average personalized circuit score for each circuit. Shading indicates the primary dysfunctional circuit for each patient: patient 1 (blue) shows elevated default mode connectivity relative to the healthy reference, whereas patient 2 (red) shows reduced activity in the cognitive control circuit. Circuit scores can be converted to Standard Ten (STEN) values for clinical display using the formula z score × 2 + 5.5, yielding a mean of 5.5 and a range from 1 (≤ −2 SD) to 10 (≥ +2 SD). (C) Example STEN values for the average circuit score: 8.2 for the default mode circuit in patient 1 and 2.5 for the cognitive control circuit in patient 2. (D) Reproducibility of EtCere-derived circuit scores across 2 scanner sites on consecutive days in 3 healthy participants (2 sessions per site). Circuit scores, displayed in STEN units, show high within-subject stability across sessions and sites. (E) Visualization of the statistical independence of EtCere-derived personalized circuit scores. Lower overlap indicates greater circuit independence; overlap is quantified by shared variance [R² values in Figure S6 in Goldstein-Piekarski et al. (13)].

The 6 biotype circuits are defined based on a theoretical anatomical synthesis of functional brain imaging studies that link large-scale circuit dysfunction to the clinical features of depression and anxiety. These definitions are grounded in studies evaluating symptoms, behaviors, biotypes, and treatment outcomes across multiple modalities (5,13,21,22). Outputs include 41 primary circuit metrics: 14 intrinsic connectivity, 17 activation, and 10 task-based connectivity measures (5,13) (see Table S1 for region definitions and measures). Additional 69 quality-controlled regions (13) are available for future expansion; the standardized framework is designed to support the incorporation of new circuits as the evidence evolves.

These output metrics are expressed in raw form (beta estimates, connectivity coefficients) and in z score standard deviation units relative to a healthy reference sample, making the magnitude and direction of dysfunction intuitively interpretable for each patient (Figure 3B). The healthy referencing procedure is consistent with the principles of normative modeling (23). Following neuropsychological test conventions, EtCere provides both circuit-level composite scores and region-specific values. z Scores are optionally converted to Standard Ten Scores (STENS) units from 1 to 10 for clinician-facing reports (Figure 3C). Circuit scores have demonstrated internal consistency, test-retest reliability, construct validity, and reproducibility across scanners and sites and low shared variance (13,24) (Figure 3D, E).

Using Biotype Circuit Scores to Parse Depression Heterogeneity

Personalized circuit scores generated by EtCere have been validated across test, held out, and external generalization samples (13). Standardized against healthy norms, scores enable the classification of patients into circuit-based subgroups—or biotypes—based on the circuit location of their primary dysfunction (25). This approach relies on the established independence of the 6 circuits (Figure 3E) (13) and previous findings showing that 90% to 96.5% of individuals with MDD and comorbid anxiety have a primary dysfunction in one of these circuits, defined as a score exceeding at least 0.5 SDs from healthy reference norms (25).

These metrics can be prospectively applied to stratify new patients in precision medicine trials (26), similar to previous use of positron emission tomography (PET)–derived patient-level scores for treatment allocation (27–29). The 41 circuit scores can also support complementary machine learning clustering approaches to identify biotypes defined by both primary and co-occurring dysfunctions (5). Combining theory-driven regions with clustering algorithms builds on foundational work using whole-brain connectivity to define biotypes (3,30,31) and complements emerging frameworks that integrate clustering with dimensional analyses (4).

EtCere-Based Prediction of Treatment Outcomes in MDD

To illustrate how personalized biotype circuit scores can be used to guide treatment selection, we applied EtCere to model treatment prediction in 329 adults with depression from precision medicine trials that included randomized treatment to antidepressants (escitalopram, sertraline, venlafaxine-XR) and behavioral problem-solving therapy, as well as open-label transcranial magnetic stimulation (TMS). Circuit scores were prospectively generated for each patient and standardized against a healthy reference sample.

Stratifying patients by circuit dysfunction and matching them to specific treatments improves response and remission rates—more than doubling them in some cases—compared with unstratified or unmatched treatment (Figure 4). Pretreatment circuit scores serve as treatment-specific and differential predictors, outperforming symptom-based models in predicting outcomes. For brevity, we illustrate this with 3 of the 6 circuits: default mode, negative affect, and cognitive control.

Figure 4.

Matching by neural circuit biotypes improves depression treatment outcomes. Illustration of improved treatment outcomes when patients are matched to specific biotypes defined by dysfunction in 3 circuits. (A) Schematic of the 3 biotype circuits: default mode (top), negative affect (middle), and cognitive control (bottom). (B) Circuit dysfunction reflects both hyperfunction (high scores) and hypofunction (low scores) relative to a healthy reference dataset. (C) Our EtCere analysis identified the treatment predicted to be most effective based on circuit dysfunction. Patients were classified as “matched” when assigned to their circuit-predicted optimal treatment and “unmatched” when assigned to a nonpredicted treatment. (D) Bar charts show treatment success rates (response or remission), comparing biotype-matched (green), sample average (white), and biotype-unmatched (gray) groups. (E) To facilitate clinical interpretation, the improvement in success rate with biotype-matched treatment is quantified relative to the average success rate. As an example, patients with higher anterior medial default mode connectivity showed a 70% remission rate on sertraline compared with 16% when unmatched—representing a 106% improvement (or 2.1 times higher) over the average base remission rate of 34% for sertraline in this sample.

INTEGRATING PERSONALIZED CIRCUIT PREDICTIONS WITH THE CONTEXT OF FOUNDATIONAL fMRI TREATMENT PREDICTION FINDINGS

In this section, we place our illustrative examples within the broader fMRI literature, organized around 6 core circuits. We synthesize foundational studies that used pretreatment fMRI biomarkers to predict treatment outcomes in MDD, including those identifying differential moderators (Table 1), guided by classic definitions of predictors and moderators (32,33). Our focus is on studies of antidepressant medications, particularly studies that have used randomized or active control designs, but we also highlight key findings from behavioral therapy and TMS where fMRI has informed differential treatment prediction. While the synthesis is centered on the 6 core circuits emphasized in this review, we also note additional circuits implicated across the literature. Throughout, we draw connections between previous findings and our EtCere-based patient-level illustrations, supporting the clinical utility of personalized circuit scores in precision psychiatry.

Table 1. Synthesis of Circuit-Level Predictors of Treatment Outcomes in Major Depressive Disorder.

Circuit	Study	Parent Study/Study Type	(Es) Citalopram	Sertraline	Venlafaxine	Duloxetine	Bupropion	Behavior Therapy	Placebo/Control	Other	Analytic Method and Sample Size	Predictor Type	Circuit Predictor	Results	Summary
Default Mode	Chin Fatt et al., 2020 (36)	EMBARC/RCT	–	✔	–	–	–	–	✔	–	ROI-based, connectivity (N = 279)	Pretreatment predictor (treatment-specific)	Connectivity within DMN, and between DMN and FPN/CEN	Responders vs. nonrespondersa: effect size (moderator) = 0.4 Sertraline: Accuracy = 47.3% Sensitivity = 81.3% Specificity = 27.7% Placebo: Accuracy = 74.2% Sensitivity = 56.1% Specificity = 82.4%	Response to sertraline was predicted by higher within-DMN connectivity and between-network connectivity of DMN and FPN/CEN, while response to placebo was predicted by hippocampal connectivity
	Andreescu et al., 2013 (40)	NA/OL	✔	–	✔	✔	✔	–	–	–	ROI-based, connectivity (n = 21)	Pretreatment predictor	Connectivity of amPFC with PCC and PCUN with PCC	Responders vs. nonrespondersa: amPFC–PCC (t = 2.52) PCUN–PCC (t = 2.97)	Treatment-responsive patients had higher baseline amPFC–PCC and PCUN–PCC connectivity
	Goldstein-Piekarski et al., 2022 (13)	iSPOT-D/RCT	✔	✔	✔	–	–	✔	–	–	ROI-based, connectivity, patient-level referenced to norms (n = 137)	Pretreatment predictor (differential treatment moderator)	Connectivity between PCC and AG	Responders vs. nonrespondersb: effect size (moderator): β = 3.53, z = 1.99	Response to venlafaxine, compared with SSRIs, was differentially predicted by lower connectivity between PCC–AG of posterior DMN subdivision.
	Korgaonkar et al., 2020 (38)	iSPOT-D/RCT	✔	✔	✔	–	–	–	–	–	Whole-brain, connectivity (n = 163)	Pretreatment predictor	Connectivity within DMN and between DMN and FPN	Remitters vs. nonremittersa: Accuracy = 68.8% Sensitivity = 63.1% Specificity = 72.4%	Remission to antidepressants was predicted by higher connectivity between DMN and FPN, while nonremission was predicted by lower connectivity.
	Goldstein-Piekarski et al., 2018 (39)	iSPOT-D/RCT	✔	✔	✔	–	–	–	–	–	ROI-based, connectivity (n = 75)	Pretreatment predictor	Connectivity of amPFC and PCC	Remitters vs. nonremittersa: Accuracy = 81.9% Sensitivity = 78.4% Specificity = 82.0%	Remission to antidepressants was predicted by higher amPFC–PCC anterior subdivision connectivity, while nonremission was predicted by lower connectivity.
	Ju et al., 2023 (43)	NA/RCT	–	–	–	–	–	–	–	✔ mixed	Whole-brain, connectivity (n = 110)	Pre-post treatment response biomarker	Connectivity in the DMN	Remitters vs. nonremittersa: Accuracy = 80.0% Sensitivity (nonremitters) = 46.2% Specificity (remitters) = 90.5%	Remission was classified by changes in DMN connectivity over the 6-mo treatment period.
	Klöbl et al., 2020 (42)	NA/RCT	✔	–	–	–	–	–	–	–	Whole-brain, connectivity (n = 29)	Early change predictor (treatment-specific)	Connectivity involving PFC, ACC, PCC, PCUN, aMCC, STC, SMG, and Ins.	Remitters vs. nonremittersa: Accuracy = 0.73 Responders vs. nonrespondersa: Accuracy = 0.68 Symptom improvement (correlation)a r = 0.45–0.55	Response and remission to escitalopram vs. placebo were predicted by acute changes in connectivity.
	Cheng et al., 2017 (99)	NA/OL	✔	–	–	–	–	–	–	–	Whole-brain, activity (fALFF) (n = 38)	Early change predictor	Activity in right MTG, CdN, temporal, and occipital cortices	Remitters vs. nonremittersa: AUC = 0.80–0.89 Sensitivity = 56.5–95.7% Specificity = 66.7–100%	Remission on escitalopram was predicted by acute first-dose changes in functional activity.
	Komulainen et al., 2018 (41)	NA/RCT	✔	–	–	–	–	–	–	–	Whole-brain and ROI-based, activation (n = 32)	Pre-post treatment response biomarker	Whole-brain: activation in SFG and PCUN/PCC ROI: activation in mPFC and ACC	Escitalopram vs. placebob,c: Decreased to self-referential words: z = 3.60–4.63	Response to escitalopram was associated with decreased activation during self-referential processing after 1 week.
	Lai and Wu, 2012 (49)	NA/OL	–	–	–	✔	–	–	–	–	Whole-brain, homogeneity (ReHo) (n = 15)	Pre-post treatment response biomarker	Homogeneity in the right SFC, MFC, and STC	Symptom improvement (correlation)a: r = 0.68	Symptom improvement on duloxetine was associated with homogeneity changes.
Salience	Geugies, et al., 2019 52)	NESDA/OL	–	–	–	–	–	–	–	✔ mixed	Whole-brain, connectivity, activation (n = 49)	Pretreatment predictor	Connectivity between the right AI and SN	Treatment-resistant vs. treatment-responsive groupd: z = 3.86	Treatment resistance was predicted by lower AI connectivity at 2 years.
	Kaiser et al., 2022 (53)	EMBARC/RCT	–	✔	–	–	–	–	✔		Whole-brain, coactivation pattern (n = 259)	Early change predictor	Expression of brain states consistent with the SN and DMN	Symptom improvementa: mediated effects = 0.06; 95% CI, 0.02 to 0.15	Symptom improvement was mediated by early treatment-related increases in time spent in SN and related networks.
	Crowther et al., 2015 (54)	NA/RCT	–	–	–	–	–	✔ BA	–	–	ROI-based, connectivity (n = 23)	Pretreatment predictor	Connectivity of the right-sided AI and MTG	Symptom improvementc (connectivity × time interaction): BDI anhedonia subscale t = −2.19 BDI somatic subscale: t = −1.88 BDI cognitive subscale: t = 2.55 BDI total: t = −1.80	Symptom improvement on BA therapy was predicted by higher connectivity between AI and MTG.
Attention	Goldstein-Piekarski et al., 2022 (13)	ENGAGE/RCT	–	–	–	–	–	✔ PST BA	✔		ROI-based, connectivity, activation, patient-level referenced to norms (n = 68)	Pretreatment predictor (differential treatment moderator)	Connectivity of LPFC and aIPL	Responders vs. nonresponderse: effect size (moderator): β = 5.14, z = 3.09	Response to PST, compared with usual care, was differentially predicted by reduced LPFC–aIPL connectivity.
Negative Affect	Vai et al., 2016 (73)	NA/OL	✔	–	–	–	–	–	–	–	ROI-based, connectivity (n = 33)	Pretreatment predictor	Connectivity between Amy ACC and vLPFC	Responders vs. nonrespondersa: Task connectivity ACC to Amy: F = 6.83 Intrinsic connectivity Amy to vLPFC: F = 4.84 Resting connectivity Amy to ACC: F = 5.91	Nonresponse was predicted by reduced intrinsic connectivity from Amy to ACC and vLPFC, but increased connectivity from ACC to Amy during threat processing.
	Goldstein-Piekarski et al., 2022 (13)	iSPOT-D, ENGAGE/RCT	✔	✔	✔	–		✔ PST-BA	–	–	ROI-based, connectivity and activation, patient-level referenced to norms (n = 205)	Pretreatment predictor (differential treatment moderator)	SSRI/SNRI: connectivity between Amy with sgACC and dACC evoked by threat PST-BA: left Amy activation evoked by threat	Responders vs. nonresponders: SSRI vs. SNRI (nonconscious threat) Right Amy activationa: effect size (moderator): β = 2.82 Left Amy–sgACC connectivity: effect size (moderator) β = −2.74, z = −2.28 SSRI vs. SNRI (conscious threat) Left Amy–dACC connectivitya: effect size (moderator) β = 3.93, z = 3.0 Right Amy–dACC connectivity: effect size (moderator) β = −3.5, z = −2.83 PST (conscious threat) Left Amy activatione: effect size (moderator) β = 10.41, z = 2.11	Response to antidepressants was predicted by generally lower circuit function during nonconscious threat processing. Response to the SNRI venlafaxine, compared with SSRIs, was predicted by lower Amy activation and higher Amy–sgACC connectivity during nonconscious threat processing, and lower Amy–dACC connectivity during conscious threat processing. Response to PST, compared with usual care, was predicted by lower Amy activation during conscious threat processing.
	Godlewska et al., 2016 (67)	NA/RCT	✔	–	–	–	–	–	✔	–	Whole-brain, activation (n = 35)	Early change predictor	Activation in the ACC, Ins, Amy, STC, and THAL in response to fearful vs. happy stimuli	Responders vs. nonrespondersa: Ins/Amy z = 3.76 ACC z = 3.96 STC z = 3.56 THAL z = 3.18	Response was predicted by decrease in activation during fear processing at week 1.
	Williams et al., 2015 (61)	iSPOT-D/RCT	✔	✔	✔	–	–	–	–	–	ROI-based, activation (n = 80)	Pretreatment predictor (predictor, differential treatment moderator)	General response: Amy activation to subliminal happy and threat Treatment-specific response: Amy activation to subliminal sadness	Predictora: effect size (predictor): Cohen’s d = 0.63–0.77 Accuracy = 75% Sensitivity = 77% Specificity = 72% Venlafaxine vs. SSRIa: effect size (moderator): Cohen’s d = 1.5 Accuracy = 81% Sensitivity = 87% Specificity = 73%	Response across antidepressants was predicted by lower Amy activation during nonconscious threat processing, while nonresponse was predicted by higher activation. Response to SNRI vs. SSRI was predicted by lower Amy activation during nonconscious threat processing, while nonresponse was predicted by higher activation.
	Fu et al., 2008 (62)	NA/OL	–	–	–	–	–	✔ CBT	–	–	Whole-brain, activation (n = 19)	Pretreatment predictor	dACC activity during nonconscious load-dependent processing of sad faces	Remitters vs. nonremittersa: Accuracy (remission) = 85.7% Accuracy (residual symptoms) = 88.9% (χ2 = 4.19)	Remission on CBT was predicted by dACC activation during load-related nonconscious sad processing.
	Dunlop et al., 2017 (64)	PReDICT/RCT	✔	–	–	✔	–	✔ CBT	–	–	ROI-based, connectivity (n = 122)	Pretreatment predictor (treatment-specific)	Connectivity between the SCC (sgACC) and left vLPFC, left AI, left vmPFC, and dorsal midbrain	Remission vs. failurea: AUC (remission) = 72%–78% AUC (failure) = 75%–89%	Remission on escitalopram was predicted by higher sgACC-based connectivity with vLPFC, AI, vmPFC, and dorsal midbrain, while remission on CBT was predicted by lower connectivity.
	Szczepanik et al., 2016 (115)	NA/RCT	–	–	–	–	–	–	✔	✔ scopolamine	Whole-brain, activation (n = 11)	Pretreatment predictor	Activation in Amy during conscious processing of sad faces	Symptom improvement (correlation)f: r = 20.72 to 20.55	Symptom improvement on scopolamine was predicted by lower Amy activation during conscious sad processing.
	Goldstein-Piekarski et al., 2021 (72)	ENGAGE/RCT	–	–	–	–	–	✔ PST-BA	–	–	ROI-based, activation (n = 108)	Early change predictor (treatment-specific mediator)	Amy activation evoked by threat	Symptom improvemente: Right Amy: effect size (differential mediator) = 0.53; 95% CI, 0.14 to 0.91 Left Amy: effect size (differential mediator) = 0.36; 95% CI, 0.08 to 0.65	Symptom improvement on PST vs. usual care was differentially mediated by early reductions in Amy activation during nonconscious threat processing.
Positive Affect	Goldstein-Piekarski et al., 2022 (13)	ENGAGE/RCT	✔	✔	✔	–		✔ PST-BA	–	–	ROI-based, activation, patient-level referenced to norms (n = 68)	Pretreatment predictor (differential treatment moderator)	vmPFC activation to happy faces	Responders vs. nonresponderse: PST vs. usual care: effect size (moderator): β = 4.57, z = 2.42	Response to PST vs. usual care was differentially predicted by lower vmPFC activation during happy processing.
	Nguyen et al., 2022 (80)	EMBARC/RCT	–	✔	–	–	✔	–	✔	–	Whole-brain, activation (n = 222)	Pretreatment predictor (treatment-specific)	Sertraline: activation during reward processing in the PFC and cerebellar crus 1 Buproprion: activation during reward processing in the CC, CdN, OFC, and crus 1	Symptom improvementa: Sertraline: effect size: R2 = 48%; 95% CI, 33% to 61%; NNT = 4.86 Bupropion: effect size: R2 = 34%; 95% CI, 10% to 59%; NNT = 1.68	Greater symptom improvement on sertraline was predicted by higher PFC and visual/temporal activation during reward processing, while less improvement was predicted by higher cerebellar crus 1 and parietal activation. Greater symptom improvement on bupropion was predicted by higher CC, CdN, and hippocampal activation during reward processing, while less improvement was predicted by higher OFC and cerebellar crus 1 activation.
	Greenberg et al., 2020 (76)	EMBARC/RCT	–	✔	–	–	–	–	✔	–	Whole-brain, activation (n = 222)	Pretreatment predictor (differential treatment moderator)	Striatum activation during reward processing	Symptom improvementa: Sertraline vs. placebo (vs. reward index , z = 20.21): effect size (moderator): d = 0.32; 95% CI, 0.06 to 0.58; t193 = 2.38	Response to sertraline vs. placebo was differentially predicted by a lower VS index during reward processing.
	Dunlop et al., 2020 (78)	CAN-BIND/OL	✔	–	–	–	–	–	–	–	ROI-based, connectivity (n = 87)	Early change predictor	Connectivity between the VS and rACC	Symptom improvement (correlation)f: r = 0.39 6 0.09 SE; 95% CI, 0.20 to 0.55	Symptom improvement on escitalopram was predicted by early (2 weeks) increase in frontostriatal connectivity during reward anticipation.
	Ventorp et al., 2022 (83)	NA/OL	–	–	–	–	–	–	–	✔ pramipexole	ROI-based, activation (n = 8)	Pre-post treatment response biomarker	Striatum activation during reward processing	Symptom improvementf: increases in reward-related activation in right NAc/VS, bilateral putamen and CdN (t = 2.74 to 4.15)	Symptom improvement on pramipexole was associated with treatment-related increases in NAc/VS, CdN and putamen during reward processing.
Cognitive Control	Tozzi et al., 2020 (92)	iSPOT-D/RCT	✔	✔	✔	–	–	–	–	–	ROI-based, connectivity and activity (n = 124)	Pretreatment predictor (differential treatment moderator)	Connectivity between dLPFC–IPL and IPL–MTG during response inhibition with Go/No-Go task	Responders vs. nonrespondersb: Sertraline: Accuracy = 83% Balanced accuracy = 84% Sensitivity = 95% Specificity = 74% Venlafaxine: Accuracy = 77% Balanced accuracy = 81% Sensitivity = 67% Specificity = 94%	Response to SSRI sertraline was predicted by higher baseline connectivity during No-Go response inhibition. Response to SNRI venlafaxine was predicted by lower baseline connectivity during No-Go response inhibition.
	Zhang et al., 2024 (24)	ENGAGE/RCT	–	–	–	–	–	✔ PST-BA	✔	–	Whole-brain, activation (N = 108)	Early change predictor	Activation of the left dLPFC, right IPL, PCUN and FFG (interaction) and IPL and SPL (prediction) during response inhibition	Responders vs. nonrespondersf: Interaction: effect size: z = 4.15–4.65 Prediction: r = 0.51 Sensitivity = 90% Specificity = 40% PR-AUC = 46% AUROC = 0.65% Predictive improvement = 104%	Symptom improvement with PST vs. usual care at 6 mo was predicted by earlier 2 mo changes in control circuit activation during response inhibition. Symptom improvement with PST over 24 months was predicted by cognitive control circuit activation during response inhibition.
	Zhukovsky et al., 2025 (91)	EMBARC, CANBIND/RCT	✔	✔	–	–	–	–	✔	–	ROI-based, connectivity(N = 363)	Pretreatment predictor	Connectivity in the dACC	Responders vs. nonrespondersa,f: Balanced accuracy = 63%–65% AUC = 0.62–0.67	Response to escitalopram and sertraline, generalized across 2 trials, was predicted by a model utilizing pretreatment clinical features and dACC connectivity.
	Gyurak et al., 2016 (85)	iSPOT-D/RCT	✔	✔	✔	–	–	–	–	–	Whole-brain, activation (N = 80)	Pretreatment predictor (predictor, differential treatment moderator)	Activation of the right dLPFC and IPL during response inhibition	Remitters vs. nonremittersa: Prediction (dLPFC): effect size: η2 = 0.17; 95% CI, 0.04 to 0.31 Sensitivity = 48.9% Specificity = 64.3% Interaction (IPL): effect size: d = 0.67, z = 4.60 Sensitivity = 69.0% Specificity = 45.9%	Remission across antidepressants was predicted by higher activation in the right dLPFC during response inhibition and nonremission was predicted by lower activation. Remission on SSRIs was differentially predicted by activation in the IPL during response inhibition, while remission on the SNRI venlafaxine was predicted by lower activation.
	Crane et al., 2017 (90)	NA/OL	✔	–	–	✔	–	–	–	–	Whole-brain, ICA and HRF modeling (Go/No-Go n = 36; ICA n = 29)	Pretreatment predictor	Activation during Go/No-Go commission errors (HRF): dACC, rACC, MCC, dmPFC, right vLPFC, lOFC Activation during Go/No-Go commission errors (ICA): dACC, rACC, MCC, mPFC, FP	Remitters vs. nonremittersa: Accuracy = 90% Sensitivity = 90% Specificity = 88.9% Symptom improvement: R2 = 0.65	Remission on escitalopram and duloxetine was predicted by a model combining behavioral commission errors with lower error-related activation of dACC and right vLPFC during response inhibition.

n denotes the number of depressed participants included in the final analysis for the result shown, and N denotes the total number of participants matching the analytic sample.

ACC, anterior cingulate cortex; AG, angular gyrus; AI, anterior insula; aIPL, anterior inferior parietal lobule; aMCC, anterior mid-cingulate cortex; amPFC, anterior medial prefrontal cortex; Amy, amygdala; AUROC, area under the receiver-operating characteristic curve; BA, behavioral activation; BDI-II, Beck Depression Inventory II; CAN-BIND, Canadian Biomarker Integration Network in Depression; CBT, cognitive behavior therapy; CC, cingulate cortex; CdN, caudate nucleus; dACC, dorsal anterior cingulate cortex; dLPFC, dorsolateral prefrontal cortex; dmPFC, dorsomedial prefrontal cortex; EMBARC, Establishing Moderators and Biosignatures of Antidepressant Response for Clinical Care; fALFF, fractional amplitude of low-frequency fluctuations; FFG, fusiform gyrus; FP, frontal pole; HDRS, Hamilton Depression Rating Scale; HRF, hemodynamic response function; ICA, independent component analysis; IDS, Inventory of Depressive Symptomatology (self-report); Ins, Insula; IPC, inferior parietal cortex; iSPOT-D, International Study to Predict Optimized Treatment for Depression; lOFC, lateral orbitofrontal cortex; LPFC, lateral prefrontal cortex; MADRS, Montgomery–Åsberg Depression Rating Scale; MFC, medial frontal cortex; MPC, medial parietal cortex; mPFC, medial prefrontal cortex; MTG, middle temporal gyrus; NA, not available; NAc, nucleus accumbens; NESDA, Netherlands Study of Depression and Anxiety; NNT, number needed to treat; OFC, orbitofrontal cortex; OL, open-label; PCC, posterior cingulate cortex; PCUN, precuneus; PFC, prefrontal cortex; PPV, positive predictive value; PReDICT, Predictors of Remission in Depression to Individual and Combined Treatments; PST-BA, problem-solving therapy with behavioral activation; QIDS, Quick Inventory of Depressive Symptomatology (self-report); rACC, rostral anterior cingulate cortex; RCT, randomized controlled trial; ReHo, regional homogeneity; ROI, region of interest; SCC, subcallosal cingulate cortex; SCL-20, Hopkins Symptom Checklist-20; SFC, superior frontal cortex; sgACC, subgenual anterior cingulate cortex; SNRI, serotonin norepinephrine reuptake inhibitor; SPL, superior parietal lobule; SSRI, selective serotonin reuptake inhibitor; STC, superior temporal cortex; THAL, thalamus; vLPFC, ventrolateral prefrontal cortex; vmPFC, ventromedial prefrontal cortex; VS, ventral striatum.

^aHDRS.

^bQIDS.

^cBDI.

^dIDS.

^eSCL-20.

^fMADRS.

Default Mode Circuit

The default mode circuit (Figure 3A and Table S1), particularly its anterior and posterior subdivisions, has consistently been identified as a significant predictor of antidepressant treatment outcomes. Meta-analyses have highlighted the anterior subdivision, centered in the anterior medial prefrontal cortex (amPFC), as a key region associated with remission (34). In the EMBARC (Establishing Moderators and Biosignatures of Antidepressant Response for Clinical Care) trial (35), group-level analyses showed that baseline connectivity between the default mode and frontoparietal attention/central executive networks predicted differential outcomes to sertraline versus placebo (36). The iSPOT-D (International Study to Predict Optimized Treatment for Depression) trial found similar significant predictions (37); higher baseline connectivity between the default mode and frontoparietal attention network was associated with remission (38), and amPFC–posterior cingulate cortex (PCC) connectivity predicted antidepressant response (39). Consistent with these findings, higher amPFC–PCC connectivity also predicted antidepressant response in late-life depression (40) (Table 1).

Early changes in default mode connectivity have emerged as potential biomarkers of treatment response. In a small randomized trial, escitalopram altered anterior circuit connectivity and reduced rumination within hours of administration (41). Similarly, citalopram-induced early shifts in default mode and central executive connectivity predicted symptom reductions (42). Furthermore, treatment-related changes in default mode connectivity distinguished remitters from nonremitters after antidepressant treatment with escitalopram, sertraline, or venlafaxine-XR (38). Increases in default mode network activity—first anterior, then posterior—also tracked response to treatments such as deep brain stimulation (44).

Building on these foundational findings, our EtCere-based patient-level illustration demonstrates that connectivity patterns within subdivisions of the default mode biotype circuit serve as treatment-specific predictors of response to the selective serotonin reuptake inhibitors (SSRIs) escitalopram and sertraline and the serotonin and norepinephrine reuptake inhibitor (SNRI) venlafaxine-XR (Box 1, Figure 4, and Supplemental Results S1). Compared with unstratified, unmatched treatment, stratifying patients based on personalized circuit scores and matching them to the predicted treatment improved response rates from approximately 16% to 22% (unmatched) to 53% to 70% (matched), representing a 2.4- to 4.5-times-higher treatment response rate.

Box 1. Clinical Use Case 1: Personalized Default Mode Circuit Scores Can Inform Choice of Antidepressants.

To illustrate how fMRI-derived circuit features can guide specific antidepressant choices, we analyzed data from the iSPOT-D antidepressant biomarker trial (n = 198; escitalopram n = 65, sertraline n = 68, venlafaxine-XR n = 65) (37). We derived personalized biotype circuit scores using the Stanford EtCere Image Processing System by calculating intrinsic connectivity between key default mode regions, standardized to the healthy reference dataset: the amPFC and PCC representing the anteromedial subdivision, and bilateral angular gyrus representing the posterior subdivision (13,21,112). We note that both high and low circuit scores indicate distinct forms of dysfunction rather than better or worse functioning, similar to the way that both high and low blood pressure are considered pathological in cardiology because both represent deviations from a healthy range.

Treatment-Specific Predictors

Personalized circuit scores for default mode subdivisions significantly predicted differential treatment outcomes among the SSRIs escitalopram and sertraline and the SNRI venlafaxine-XR (Figure 4; see Supplemental Results S1 for details).

• Sertraline: High connectivity between the amPFC and PCC involving the anterior default mode subdivision predicted remission (β = 0.76; sensitivity = 73.9%, specificity = 73.3%, AUC = 0.75 [95% CI: 0.60, 0.87]), outperforming the symptom-only baseline model (Δχ² = 5.85, p = .02).

• Escitalopram: Low connectivity between the amPFC and angular gyrus of the posterior default mode subdivision predicted response (β = −0.57; sensitivity = 77.4%, specificity = 61.8%, AUC = 0.74 [95% CI: 0.61, 0.86]), outperforming the symptom-only model (Δχ² = 3.85, p = .05).

• Venlafaxine-XR: Low connectivity between the amPFC and angular gyrus of the posterior default mode subdivision predicted remission (β = −0.49; sensitivity = 52.6%, specificity = 63.0%, AUC = 0.67 [95% CI: 0.52, 0.81]), outperforming the symptom-only model (Δχ² = 4.61, p = .03).

Circuit-Treatment Matching

To evaluate the clinical utility of predictors based on default mode connectivity, we compared outcomes between patients predicted to respond from their assigned treatment (matched) and those predicted not to respond (unmatched) (Supplemental Methods S2 for details). Matched patients consistently demonstrated significantly higher response rates across all treatment arms, supporting the potential of circuit-guided treatment selection to improve outcomes (Figure 4).

• Sertraline: The remission rate among predictor-matched patients was 70% compared with just 16% when unmatched (p < .001, χ² test; OR = 12.41 [95% CI: 3.74, 41.18]; Cohen’s h = 1.16). Matched patients achieved a 106% improvement (or 2.1 times higher) over the average base remission rate of 34%.

• Escitalopram: The response rate among predictor-matched patients was 68% compared with just 16% when unmatched (p < .001, χ² test; OR = 10.90 [95% CI: 3.10, 38.34]; Cohen’s h = 1.11). Matched patients achieved a 42% improvement (or 1.4 times higher) over the average response rate of 48%.

• Venlafaxine-XR: The remission rate among predictor-matched patients was 53% compared with 22% when unmatched (p = .04, χ² test; OR = 4.05 [95% CI: 1.20, 13.66]; Cohen’s h = 0.66). Matched patients achieved an 83% improvement (or 1.8 times higher) over the average remission rate of 29%.

Clinical Implications

Personalized circuit scores representing connectivity within default mode subdivisions offer clinically meaningful and neurobiologically grounded guidance for antidepressant selection. Approximately 17% of patients with major depressive disorder show primary dysfunction in the default mode circuit, defined by biotype circuit scores that fall at the extremes relative to healthy reference data (25). For these individuals, integrating personalized default mode profiles into treatment planning may enhance the likelihood of achieving treatment success.

amPFC, anterior medial prefrontal cortex; AUC, area under the curve; fMRI, functional magnetic resonance imaging; iSPOT-D, International Study to Predict Optimized Treatment for Depression; OR, odds ratio; PCC, posterior cingulate cortex; SNRI, serotonin norepinephrine reuptake inhibitor; SSRI, selective serotonin reuptake inhibitor.

Specifically, high amPFC–PCC connectivity predicted remission with sertraline (70% matched vs. 16% unmatched), whereas low amPFC–angular gyrus (AG) connectivity predicted response to escitalopram (68% matched vs. 16% unmatched), and low AG–PCC connectivity predicted remission with venlafaxine-XR (53% matched vs. 22% unmatched). See Supplemental Results S1 for more details.

These patient-level predictions for sertraline response, based on the anterior subdivision of the default mode circuit, converge with group-level findings from EMBARC and iSPOT-D. In contrast, involvement of the posterior subdivision in predicting remission with escitalopram and venlafaxine-XR generalizes earlier patient-level results from the first-wave iSPOT-D cohort (13).

The specificity of these predictions—both in terms of default mode subdivisions and antidepressant class—highlights the importance of investigating default mode variation further as a contributor to depression heterogeneity and outcome prediction across different pharmacological treatments. MDD heterogeneity encompasses profiles of both hyper- and hypoconnectivity (31,45–47), and distinct default mode subdivisions have been associated with specific depressive features; for example, amPFC–PCC connectivity has been linked to self-reflective rumination (48). Notably, duloxetine, another SNRI, has also been shown to normalize default mode connectivity (49), emphasizing the need for studies that broaden circuit-based predictors across a wider range of commonly prescribed antidepressants.

Salience Circuit

The salience circuit, also referred to as the midcingulo-insular network, involves core regions in the anterior insula, cingulate, and extended amygdala (50) (Figure 3A and Table S1). It functions to identify and prioritize both internal and external stimuli, directing other networks, such as the default mode and frontoparietal attention/central executive networks, accordingly.

Evidence from NESDA (Netherlands Study of Depression and Anxiety) (51) demonstrated that decreased connectivity of the right insula within the salience network distinguished patients with insufficient response to 2 or more antidepressants over 2 years from patients who received a single antidepressant (52). In the EMBARC trial (35), dynamic network analysis indicated that more time spent in brain states consistent with the salience and default mode networks was an early biomarker of response to sertraline (53). The first psychotherapy prediction study observed an opposite direction of effect, with higher pretreatment insula connectivity predicting better symptom outcomes (54). Further evidence for the differential prediction of antidepressants versus cognitive behavioral therapy (CBT) response involving insula connectivity is outlined in the Negative Affect Circuit section.

Using the original research pipeline of the patient-level circuit method underpinning EtCere, lower intrinsic salience network connectivity in MDD emerged as a general predictor of better treatment responses to escitalopram, sertraline, and venlafaxine-XR above and beyond symptom-based models (13). Given the variability in effect directions across studies, further research is needed to examine patient-level stratification within the salience network for predicting treatment response and remission outcomes across a range of treatments.

Attention Circuit

The frontoparietal attention circuit, anchored in the medial superior frontal cortices, anterior inferior parietal lobule, and precuneus (Figure 3A and Table S1), is broadly implicated in alertness, sustained attention, and recollection (21,55). Often referred to as the external attention system (55), it overlaps with the central executive network (56), and together with the default mode and salience networks, it forms the triple-network model of psychopathology (56). In depression, greater coupling among these networks has been linked to better outcomes with antidepressants, especially sertraline (36,38) (Table 1).

Group-level analyses using personalized circuit scores generated with EtCere identified a biotype characterized by low attention connectivity—more than 0.5 SDs below the healthy norm—which was associated with reduced benefit from behavioral therapy (5). Earlier findings using the method underpinning EtCere showed that this circuit moderates response to behavioral problem-solving therapy versus usual care (13) (Table 1).

This differentiation highlights the clinical relevance of attention-related circuit disruptions in MDD (57). Future studies are needed to examine how stratifying attention circuit connectivity—both within the circuit and across broader networks—can guide personalized treatment strategies and improve outcomes across a wider range of treatments.

Negative Affect Circuit

The negative affect circuit—comprising the amygdala, insula, and medial prefrontal regions, including the subgenual anterior cingulate cortex (sgACC) and dorsal anterior cingulate cortex (dACC) (Figure 3A and Table S1)—is among the most extensively studied networks in MDD. Dysfunction of this circuit has been studied using task-based fMRI (21,22,58) and PET imaging during mood induction (59) and deep brain stimulation (60).

Region-of-interest (ROI) analyses from the first wave of iSPOT-D (37) showed that elevated nonconscious amygdala activity predicted nonresponse to antidepressants, while lower activity was associated with better treatment response (61). Relatedly, dACC activity elicited by negative affect stimuli has been linked to treatment response with CBT (62).

Using resting-state fMRI, the PReDICT (Predictors of Remission in Depression to Individual and Combined Treatments) trial of never-treated MDD (63) identified lower affective circuit connectivity (left sgACC to insula, periaqueductal gray, and ventral PFC) as a differential predictor of remission with citalopram or duloxetine but failure with CBT, whereas higher connectivity predicted the opposite (64). PET evidence further supports differential prediction within negative affect and salience circuits; in recurrent MDD, right anterior insula metabolism predicted differential outcomes with escitalopram versus CBT (27), whereas sgACC metabolism predicted failure for both (65). These findings culminated in a clinically translatable resting-state connectivity biomarker tested in a prospective trial that successfully allocated biomarker-matched patients to escitalopram or CBT (28).

Posttreatment normalization of negative affect circuit activity has also been documented. In previously unmedicated patients, escitalopram normalized fear- and sadness-related activity in the amygdala, insula, and ACC (66,67). Attenuation of amygdala reactivity has also been observed following sertraline (68,69) and paroxetine (70). In contrast, with citalopram, amygdala reactivity to sad stimuli normalizes, but reactivity to threat-related stimuli persists (71).

Consistent with previous work, illustrative analyses using EtCere’s personalized negative affect circuit scores moderated treatment response (Box 2 and Figure 4). Patients with low scores responded better to antidepressants, while patients with high scores were more likely to remit with behavioral therapy, even after accounting for baseline symptom severity. Matching treatment to circuit profile nearly doubled remission rates (32% matched vs. 15% unmatched) (see Supplemental Results S2 for more details). These findings extend previous results from the iSPOT-D and ENGAGE trials (13,72) and are consistent with region-based predictors from the first-wave iSPOT-D cohort (61). A key driver of the moderation effect in our illustrative analyses was sgACC–amygdala connectivity during nonconscious threat processing. This circuit feature optimized predictions for remission with venlafaxine-XR versus behavioral therapy and is consistent with previous studies linking amygdala–ACC connectivity—both task-evoked and intrinsic—to SSRI outcomes in treatment-resistant depression (73). However, the direction of effects varies across paradigms and antidepressant classes, highlighting the need for more work in this area.

Box 2. Clinical Use Case 2: Personalized Negative Affect Circuit Scores Can Inform Early Treatment Decisions Between Behavioral Therapy and Antidepressants.

To illustrate how the negative affect biotype circuit may guide treatment modality selection, we combined RCT data from the iSPOT-D antidepressant biomarker trial (n = 198; escitalopram n = 65, sertraline n = 68, venlafaxine-XR n = 65) (37) and the ENGAGE precision medicine behavioral therapy trial (n = 50; problem-solving cognitive behavioral therapy) (113). Using the Stanford EtCere Image Processing System, we derived a personalized negative affect circuit score based on nonconscious processing of threat-related face stimuli (13,21). This biotype circuit score reflects a composite of relatively greater amygdala activation, reduced sgACC activation, and reduced sgACC–amygdala connectivity (13,21).

Differential Treatment Moderator

Personalized negative affect circuit scores significantly predicted differential outcomes between antidepressants and behavioral therapy, with opposing predictive effects across modalities (χ² = 9.43, p = .009). The interaction between circuit scores and treatment type was large (β = 2.34, p < .001, OR = 10.37 [95% CI: 3.04, 35.35]), outperforming the symptom-only model (Δχ² = 11.49, p = .009) (see Supplemental Results S2 for details). Connectivity between the sgACC and amygdala, a central component of the negative affect circuit score, emerged as a key driver of differential prediction, particularly for venlafaxine versus behavioral therapy.

• Behavioral therapy: High negative affect circuit scores predicted remission (β = 1.83; sensitivity = 66.7%, specificity = 63.4%, AUC = 0.75 [95% CI: 0.50, 0.98]).

• Venlafaxine-XR: Low negative affect circuit scores were most predictive of remission on venlafaxine-XR (β = −1.47; sensitivity = 57.9%, specificity = 58.7%, AUC = 0.66 [95% CI: 0.53, 0.80]).

Circuit-Treatment Matching

To evaluate the clinical utility of personalized negative affect circuit scores, we compared outcomes between patients predicted to respond to their assigned treatment (matched) and those predicted not to respond (unmatched) (see Supplemental Methods S2 for details). For both the negative affect circuit score and the specific sgACC–amygdala connectivity predictor, matched patients consistently demonstrated significantly higher response rates, supporting the potential of circuit-guided treatment selection to improve outcomes (Figure 4).

Matching Treatment With the Negative Affect Circuit Predictor

The remission rate among predictor-matched patients was 32% compared with 15% in unmatched patients, an improvement of 112% that doubles remission rates (see Supplemental Results S2 for details).

• Behavioral therapy: The remission rate among predictor-matched patients was 63% compared with 10% when unmatched (p = .002, χ² test; OR = 15.83 [95% CI: 2.71, 92.36]; Cohen’s h = 1.20). Matched patients achieved a 247% improvement (or 3.5 times higher) over the average remission rate of 18%.

• Venlafaxine-XR: The remission rate for predictor-matched patients was 41% compared with 14% when unmatched (p = .04, χ² test; OR = 4.09 [95% CI: 1.18, 14.21]; Cohen’s h = 0.61). Matched patients achieved a 39% improvement (or 1.4 times higher) over the average remission rate of 29%.

Matching Treatment With the sgACC–Amygdala Connectivity Predictor

• Behavioral therapy: The remission rate was 47% for matched patients compared with just 3% for unmatched patients (p < .001, χ² test; OR = 28.44 [95% CI: 3.13, 258.38]; Cohen’s h = 1.16). Matched patients achieved a 162% improvement (or 2.6 times higher) over the average remission rate of 18%.

• Venlafaxine-XR: The remission rate for predictor-matched patients was 73% compared with 20% for unmatched patients (p = .002, χ² test; OR = 10.42 [95% CI: 2.37, 45.93]; Cohen’s h = 1.11). Matched patients achieved a 149% improvement (or 2.5 times higher) over the average remission rate of 29%.

Clinical Implications

These findings demonstrate that personalized negative affect circuit scores—whether derived from composite circuit-level function or specific sgACC–amygdala connectivity—provide clinically meaningful, neurobiologically grounded guidance for patient stratification and treatment selection. At least 12% of patients with major depressive disorder show primary dysfunction in the negative affect biotype circuit during nonconscious threat processing, defined as circuit scores falling at the extremes relative to healthy reference data (25). For these individuals, incorporating personalized negative affect profiles into treatment planning could help guide the choice between pharmacological and behavioral treatment modalities, potentially improving treatment success.

iSPOT-D, International Study to Predict Optimized Treatment for Depression; OR, odds ratio; RCT, randomized controlled trial; sgACC, subgenual anterior cingulate cortex.

Positive Affect Circuit

Dysfunction within the corticostriatal reward system (Figure 3A and Table S1), particularly the ventral striatum and medial prefrontal regions, is linked to anhedonia, characterized by the diminished ability to experience pleasure, loss of motivation, and decreased willingness to expend effort for reward (21,22,74).

In the EMBARC trial (35), a reward circuit index derived from ventral striatal activity during a monetary incentive delay task (75) predicted better response to sertraline than placebo (76) (Table 1). Similarly, the CAN-BIND (Canadian Biomarker Integration Network in Depression) study (77) found that early increases in ventral striatal–ACC connectivity were associated with escitalopram response (78).

Reward circuit dysfunction may also underlie nonresponse to serotonergic antidepressants, which often fail to address the core features of anhedonia (79). This has fueled interest in dopaminergic agents. In the EMBARC trial (35), orbitofrontal, cingulate, and caudate activation predicted response to bupropion, while distinct prefrontal activity predicted response to sertraline (80). Complementing these findings, higher baseline frontostriatal connectivity predicted clinical benefit from bupropion but not sertraline in the stage 2 EMBARC sample (81).

Dopaminergic therapies such as pramipexole, a dopamine D₂/D₃ agonist with selective affinity for the ventral striatum (82), has shown promise in patients with reward deficits. Using a stratified approach, individuals with low baseline ventral striatal activity showed marked improvement in both global depression and anhedonia (83). Individuals with strong baseline striatal activation during reward learning tasks also showed greater benefit (84), further supporting the role of reward circuit engagement as a treatment-matching biomarker. Behavioral approaches designed to increase engagement in meaningful activities likewise show promise, with opposing profiles of ventral prefrontal activity during positive emotion processing serving as a differential moderator of response to active behavioral therapy versus usual care (13).

Cognitive Control Circuit

The cognitive control circuit includes the dorsolateral PFC (dLPFC), dACC (21) (Figure 3A and Table S1), and dorsal parietal regions. The dLPFC is engaged during tasks that require goal selection and response inhibition, such as the Go/No-Go task, and is frequently identified as a key region in resting-state fMRI studies of treatment prediction.

Findings from the first-wave iSPOT-D cohort (37) indicated that lower dLPFC activation during a Go/No-Go task predicted poorer treatment outcomes (85). A specific cognitive biotype characterized by low dLPFC and dACC activity and impaired cognitive performance relative to healthy control participants was observed in 27% of participants with MDD and was associated with poor remission rates (86). Group-level analyses in the ENGAGE-2 cohort of patients with depression and cognitive impairment (87) also indicated that reduced Go/No-Go cognitive control activity in the dLPFC and dACC predicted 6-month anxiety outcomes following behavioral therapy (88).

Task-based fMRI findings of the cognitive control circuit converge with resting-state results showing that dLPFC–dACC functional connectivity predicts remission in older adults with MDD treated with escitalopram (89). However, the direction of effects varies. In one study combining behavioral Go/No-Go commission errors with fMRI, higher error-related dACC and ventrolateral PFC activity predicted poorer response to both escitalopram and duloxetine (90). A large-scale analysis across the EMBARC (35) and CAN-BIND (77) cohorts demonstrated that lower dLPFC–dACC connectivity in early treatment (week 2) predicted better response to sertraline and escitalopram (91). These findings highlight the need to parse depression heterogeneity based on behavioral performance measures and to integrate task and resting fMRI predictors in the same patients.

In our illustrative analyses, personalized cognitive control circuit scores served as differential moderators of response to antidepressants and TMS (Box 3, Figure 4, and Supplemental Results S3). Specifically, low cognitive control circuit scores during a Go/No-Go task identified patients more likely to fail on standard antidepressants such as escitalopram and sertraline. These results are consistent with iSPOT-D findings (85) and with evidence showing that low cognitive control circuit function and corresponding cognitive behavioral deficits, a cognitive biotype, are associated with poor response to standard antidepressants (86). Furthermore, previous iSPOT-D findings showed that lower pretreatment functional connectivity during the Go/No-Go task predicted poor response to sertraline, while the opposite pattern predicted response to venlafaxine-XR (92).

Box 3. Clinical Use Case 3: Personalized Cognitive Control Circuit Scores Can Inform When to Select Antidepressants and When to Expedite to Other Treatments.

To illustrate the clinical utility of incorporating fMRI-derived biotype circuit scores in guiding treatment selection, we combined data from 3 trials: the iSPOT-D antidepressant biomarker trial (n = 198; escitalopram n = 65, sertraline n = 68, venlafaxine-XR n = 65) (37), the ENGAGE precision medicine behavioral therapy trial (n = 50) (113), and the B-SMART-fMRI trial of TMS (n = 81) (114). Using the Stanford EtCere Image Processing System, we computed a personalized cognitive control circuit score, defined as the composite average of activation in the dorsal anterior cingulate cortex and bilateral dorsolateral prefrontal cortex, along with their functional connectivity during a Go/No-Go task, standardized against a healthy reference dataset (13).

Differential Treatment Moderator

Personalized cognitive control circuit scores significantly predicted differential treatment outcomes across treatment modalities (χ² = 15.86, p = .003), outperforming symptom-only models (Δχ² = 16.42, p = .006). The strongest interaction effect was between antidepressants and TMS (β = −1.81, p = .001, OR = 0.16 [95% CI: 0.05, 0.50]) (see Supplemental Results S3 for details).

• Antidepressants: High cognitive control circuit scores predicted response (β = 0.54; sensitivity = 61.9%, specificity = 54.4%, AUC = 0.60 [95% CI: 0.52, 0.69]). Low scores identified patients likely to fail on standard antidepressants, indicating the potential need to expedite alternative treatments.

• TMS: Low cognitive control circuit scores predicted response (β = −1.1; sensitivity = 64.7%, specificity = 59.4%; AUC = 0.65 [95% CI: 0.50, 0.79]) and remission (β = −1.04; sensitivity = 75.0%, specificity = 69.6%, AUC = 0.78 [95% CI: 0.66, 0.89]).

Circuit-Treatment Matching

The response rate among predictor-matched patients was 43% compared with 24% when unmatched, nearly doubling the potential rate of response (p , .001, χ² test; OR = 2.37 [95% CI: 1.47, 3.83]; Cohen’s h = 0.41) (Figure 4; see Supplemental Results S3 for details).

• Antidepressants: The response rate in predictor-matched patients was 52% compared with 31% when unmatched (p = .004, χ² test; OR = 2.43 [95% CI: 1.35, 4.37]; Cohen’s h = 0.43). Matched patients achieved a 22% improvement (or 1.2 times higher) over the average response rate of 42%.

• TMS: The response rate in predictor-matched patients was 41% compared with 16% when unmatched (p = .049, χ² test; OR = 3.78 [95% CI: 1.16, 12.28]; Cohen’s h = 0.58), Matched patients achieved a 96% improvement (or 2.0 times higher) over the average response rate of 21%.

Clinical Implications

These findings show that personalized cognitive control circuit scores provide clinically meaningful, neurobiologically grounded guidance for patient stratification and treatment selection. In major depressive disorder, up to 25% of patients exhibit primary dysfunction in the cognitive control biotype circuit, defined as circuit scores falling at the extremes relative to healthy reference data (86). For these individuals, cognitive control circuit profiles may help identify patients unlikely to benefit from antidepressants, supporting expedited progression to alternative treatments such as TMS, potentially reducing time to effective care and improving overall outcomes.

AUC, area under the curve; fMRI, functional magnetic resonance imaging; iSPOT-D, International Study to Predict Optimized Treatment for Depression; OR, odds ratio; TMS, transcranial magnetic stimulation.

Low cognitive control circuit scores instead identified patients more likely to benefit from TMS (response rates: TMS, 41% matched vs. 16% unmatched) (Box 3, Figure 4, and Supplemental Results S3). Although TMS is currently reserved for treatment-resistant depression, stratifying patients by circuit function may expand its use to earlier stages of care. The involvement of the dLPFC and dACC as key regions in response to TMS is supported in an independent precision medicine trial showing that a cognitive biotype with low dLPFC–dACC connectivity mediated earlier behavioral improvement with TMS delivered in a clinical setting (93). Resting-state fMRI data also showed that stimulation with the SAINT (Stanford Accelerated Intelligent Neuromodulation Therapy) TMS protocol induced signaling shifts in both dLPFC and dACC, with dACC changes predicting response outcomes (94).

Together, these results support the potential utility of personalized cognitive control circuit scores for expediting patients to alternative treatments and minimizing trial-and-error approaches. Consistent with this, a recent prospective clinical trial demonstrated that patients who met criteria for a low cognitive control biotype achieved high remission rates (86.4%) when treated with guanfacine, an α_2A receptor agonist that enhances dLPFC neuroplasticity, activation, and connectivity (26,95).

Additional studies support the utility of cognitive control circuit scores in predicting treatment response. Cognitive training in middle-age and older adults with MDD enhanced circuit activation and improved mood and cognition (96). These findings are consistent with the PReDICT trial (63), in which remitters to behavioral therapy showed greater resting-state connectivity in networks tied to cognitive control (97). Posterior regions in the cognitive control circuit may also serve as response markers; in the ENGAGE trial, early attenuation of inferior parietal activity during a Go/No-Go task predicted long-term symptom and problem-solving improvements (24). Furthermore, fractional amplitude of low-frequency fluctuations studies in first-episode depression showed that escitalopram reduced posterior activity within hours and restored dLPFC and ACC activity at 4 weeks (98,99).

BRINGING fMRI TO THE CLINIC: CHALLENGES AND OPPORTUNITIES

In this section, we outline some key areas critical for accelerating the translation of fMRI-based approaches into real-world clinical settings. Box 4 highlights specific opportunities to drive progress.

Box 4. Opportunities for Driving Progress in Circuit-Guided Precision Psychiatric Care.

Expanding the Treatment Menu

Evidence directly comparing 2 or more active treatments is essential for guiding treatment selection (110). This review focused on circuit-based predictors for commonly prescribed antidepressants and illustrated how the EtCere system provides personalized circuit scores that differentially predict response to antidepressants, behavioral therapy, and TMS.

The same approach is scalable to a broader range of treatments and could be deployed earlier in the care pathway. Personalized fMRI-based circuit scores can help rapidly identify patients unlikely to respond to standard antidepressants, reducing trial-and-error delays and expediting access to more effective alternatives. For example, patients with persistent negative affect circuit hyperactivity—signaling nonresponse to standard antidepressants—could be expedited to alternative interventions such as the muscarinic antagonist scopolamine (115) or real-time fMRI neurofeedback targeting amygdala reactivity (116).

There is an urgent need for large-scale clinical efforts to implement these predictors across diverse interventions and control groups, including currently approved antidepressants, a wider array of behavioral therapies, neuromodulation (117,118), deep brain stimulation (119,120), and rapid-acting therapeutics such as ketamine (121) and psychedelics (122), within a standardized framework. Such efforts would also support monitoring of treatment trajectories and enable early detection of relapse or recurrence (120,123,124).

Reference Norms

In both cardiology and neuropsychology, the clinical utility of objective measures often relies on comparison to healthy and/or clinical reference norms. Similarly, advancing individualized functional brain imaging in psychiatry will require reference norms to support standardized and clinically meaningful interpretation of circuit metrics. Such norms will enable clinicians to assess the magnitude and specificity of circuit dysfunctions at the individual level. Normative modeling has been used to dissect individual differences in fMRI (125). In the use cases described above, personalized circuit scores calculated using the Stanford EtCere Image Processing System were standardized against healthy norms. This approach also enables classification of patients into circuit-based subgroups—or biotypes—based on the primary location of dysfunction (25).

Integration Into Clinical Infrastructure

Translation requires standardized fMRI sequences that align with existing clinical workflows and payer-recognized reimbursement structures. Patient-level tools such as EtCere can interface with broader precision imaging infrastructure, including motion monitoring systems (e.g., FIRMM) (126) and data platforms (e.g., Flywheel). Developing shared standards with input from clinicians, engineers, and researchers will be essential, paralleling protocols in other image-guided medical specialties.

fMRI, functional magnetic resonance imaging; TMS, transcranial magnetic stimulation.

Standardization, Reproducibility, Reliability, Validity, and Feasibility

For fMRI-based biomarkers to be adopted in practice, they must meet benchmarks for standardization, reproducibility, reliability, and validity. Both internal consistency and test-retest reliability, across short intervals and treatment-relevant timeframes, are important for using circuit scores as treatment predictors. While between-subject test-retest reliability is often modest, within-subject reliability is typically strong for both resting and task fMRI using anatomically defined ROIs or peak activations (100–102). Future work could prioritize within-subject reliability most relevant for clinical decision making (103).

Reproducibility across sites is also essential. The ABCD (Adolescent Cognitive Brain Development) Study demonstrates cross-scanner reproducibility using harmonized protocols across multiple academic sites (104). EtCere-derived personalized circuit scores show similar consistency across sites for 3 traveling participants, with strong 1-day test-retest reliability (Figure 3D). Visualization in STENS illustrates the stability of scores within individuals across sites and time.

Validity—how accurately imaging reflects neural processes—is equally critical. While strategies such as increasing scan duration may enhance reliability, they can compromise clinical interpretability due to participant fatigue or increased motion-related noise. Optimized task designs (e.g., block paradigms, active baselines) and individualized processing pipelines can improve both reliability and validity (105,106). Converging evidence from multiple treatment trials, summarized here, shows that fMRI-based circuit metrics predict treatment response more accurately than clinical features alone, a key requirement for translation. Long-term studies such as NESDA (101) further highlight the added value of imaging measures beyond symptom ratings.

The translational impact of circuit biotypes and predictors also depends on their generalizability across samples and settings. Both data- and theory-driven approaches support biotype generalizability (4,5,107). Patient-level metrics are central to enabling predictors to transfer across independent datasets and prospective designs. For example, an insula–sgACC connectivity–based treatment moderator identified using PET (27) was later operationalized as a patient-level metric and prospectively applied in a separate cohort (28).

The clinical feasibility of fMRI will require workflows analogous to those used for other high-burden illnesses where imaging is routinely ordered (16). Leveraging pretreatment circuit predictors to increase the proportion of patients who reach remission earlier could mitigate disability and lower long-term costs associated with depression. Provider and patient surveys indicate strong receptivity to using brain scans to personalize care and reduce stigma and self-blame (108).

Defining Meaningful Outcomes

Outcome definitions vary widely across studies; harmonizing remission and response criteria—tailored to treatment stage—would improve translation. For example, full remission may be the priority in first-episode depression, whereas partial response can be an appropriate goal in treatment-resistant cases (109). Trials should also broaden to younger ages and include features that have often been excluded but are central to burden, such as suicidality and comorbid substance use, while tracking functional outcomes and disability alongside symptoms (87,110,111).

CONCLUSIONS

The identification of clinically actionable, personalized circuit metrics marks a critical advance toward precision psychiatry. In this review, we presented evidence that patient-level fMRI measures improve prediction of treatment outcomes beyond symptom-based models, offering a path to faster and more effective care. The consistency of our norm-referenced EtCere findings with the broader literature underscores their potential to support timely, interpretable treatment decisions and, in some cases, nearly double response rates. Realizing this promise will require coordinated efforts to ensure methodological rigor, standardization, and integration into clinical practice. With these foundations in place, neuroimaging-based treatment selection has the potential to transform psychiatric care, bringing us closer to a model in which brain-based evidence routinely guides treatment decisions.

Supplementary Material

Supplementary material cited in this article is available online at https://doi.org/10.1016/j.biopsych.2025.08.019.