Personalized brain circuit scores identify clinically distinct biotypes in depression and anxiety
个性化脑回路评分识别抑郁症和焦虑症中的临床不同生物型

Depression and associated anxiety disorders are an important global public health burden¹, the treatment of which has been hindered by etiological and phenotypic heterogeneity. The current psychiatric diagnostic system assigns a single label to syndromes that may involve the dysfunction of multiple and overlapping neurobiological processes which, in turn, would probably each require a different treatment. This is evident from the fact that more than a third of patients diagnosed with major depressive disorder, and approximately half of patients diagnosed with generalized anxiety disorder, do not respond to first-line treatment^2,3. Unlike the ‘one-size-fits-all’ approach, a precision medicine approach to care requires standardized metrics that are personalized for individual patients and are interpretable to clinicians. However, the promise of this approach is currently limited by a lack of personalized and interpretable measures for quantifying neurobiological dysfunctions in patients with depression and associated anxiety disorders. We believe that such measures should help to elucidate the underlying neurobiological dysfunctions within a neuroscientific theoretical framework, rather than remain an algorithmic black box. Using these measures, patients could be stratified prospectively into subgroups that share similar neurobiological dysfunctions, or ‘biotypes’, each of which would possibly implicate a different set of treatment approaches or a different treatment trajectory.
抑郁症及相关焦虑障碍是一个重要的全球公共卫生负担，其治疗受到病因和表型异质性的阻碍。目前的精神病诊断系统对可能涉及多个重叠神经生物过程功能障碍的综合症赋予单一标签，而这些过程可能各自需要不同的治疗。这一点从这样一个事实中可以看出：超过三分之一被诊断为重度抑郁症的患者，以及大约一半被诊断为广泛性焦虑障碍的患者，对一线治疗没有反应。与“一个适合所有人”的方法不同，精准医学的护理方法需要为个体患者量身定制的标准化指标，并且这些指标能够被临床医生解读。然而，这种方法的前景目前受到缺乏个性化和可解读的量化抑郁症及相关焦虑障碍患者神经生物功能障碍的测量工具的限制。 我们相信，这些措施应该有助于阐明神经科学理论框架内的潜在神经生物学功能障碍，而不是停留在算法黑箱中。利用这些措施，患者可以前瞻性地分层为具有相似神经生物学功能障碍的亚组或“生物型”，每种生物型可能涉及不同的治疗方法或不同的治疗轨迹。

Efforts to characterize biotypes of depressed and anxious patients with similar brain circuit dysfunctions have typically used task-free functional magnetic resonance imaging (fMRI)^4,5,6,7. For example, one pioneering study has found biotypes characterized by aberrant connectivity in frontostriatal and limbic networks that respond differently to repetitive transcranial magnetic stimulation (TMS)⁴. Other researchers have found biotypes characterized by hyper- and hypoconnectivity of the default mode network⁵, biotypes that distinguish comorbid anxiety within the context of depression⁶ and biotypes that are associated with a poorer response to standard antidepressants⁷.
对具有相似脑回路功能障碍的抑郁和焦虑患者进行生物类型特征描述的努力通常使用无任务功能性磁共振成像（fMRI）4,5,6,7。例如，一项开创性研究发现了以前额-纹状体和边缘网络中异常连接为特征的生物类型，这些生物类型对重复经颅磁刺激（TMS）反应不同 4。其他研究人员发现了以默认模式网络的超连接和低连接为特征的生物类型 5，区分抑郁背景下共病焦虑的生物类型 6，以及与标准抗抑郁药物反应较差相关的生物类型 7。

Nevertheless, we lack evidence about biotypes in depression and anxiety that are based on the participant-level quantification of measures derived from task-evoked imaging modalities. Patients with depression and anxiety exhibit dysfunction in the activity and connectivity of brain circuits in response to specific probes of general and emotional cognition. In other words, in depression and anxiety, the brain continually and flexibly engages different circuits under task-evoked and task-free conditions. Therefore, both sources of information may be useful in delineating biotypes and biotype-guided treatments. This is analogous to cardiac imaging being collected during both rest and task conditions in which the activity of the heart is elicited (for example, stress tests) to enable precise diagnoses and treatment plans, a necessity given the complexity of this organ and its functions⁸. Indeed, clinical trials have found that measures derived from task-based fMRI often predict response in depression treatment (for example, refs. ^9,10,11,12) and have recently been the biomarker of choice for new pharmacotherapy development (for example, ref. ¹³).
然而，我们缺乏基于参与者级别量化的任务诱发成像模式的抑郁和焦虑生物类型的证据。抑郁和焦虑患者在对特定的普遍和情感认知探针的反应中表现出大脑回路活动和连接的功能障碍。换句话说，在抑郁和焦虑中，大脑在任务诱发和无任务条件下不断灵活地参与不同的回路。因此，这两种信息来源可能对描绘生物类型和生物类型指导的治疗有用。这类似于在静息和任务条件下收集心脏成像数据，在这些条件下心脏的活动被引发（例如，压力测试），以便进行精确的诊断和治疗计划，这在考虑到这个器官及其功能的复杂性时是必要的。确实，临床试验发现，基于任务的 fMRI 所获得的测量常常能够预测抑郁治疗的反应（例如，参考文献 9,10,11,12），并且最近已成为新药物治疗开发的首选生物标志物（例如，参考文献 13）。

Foundational studies using whole-brain, task-free connectivity biomarkers have often taken an unsupervised whole-brain approach that uses thousands of features for biotyping. However, we posit that clinical translation requires a theoretically informed approach that relies on a well-defined, tractable set of inputs. Such an approach also addresses the potential for obtaining overly optimistic results (overfitting) when thousands of inputs are used in a fully unsupervised manner—an issue that has been raised in the field¹⁴ (but see ref. ¹⁵, which addresses overfitting¹¹).
基础研究使用全脑、无任务的连接生物标志物，通常采用无监督的全脑方法，利用数千个特征进行生物分类。然而，我们认为临床转化需要一种理论上有依据的方法，依赖于一组定义明确、可处理的输入。这种方法还解决了在完全无监督的方式下使用数千个输入时可能获得过于乐观结果（过拟合）的潜在问题——这一问题在该领域已被提出 14（但见参考文献 15，讨论了过拟合 11）。

Finally, previous studies have assessed the ability of biotypes to predict response to a single treatment (for example, TMS⁴ or antidepressants⁷), rather than comparing responses across different classes of treatments. To maximize the translational value of biotypes, the optimal treatment for each biotype should eventually be determined by comparing how different biotypes respond when receiving the same treatment.
最后，以前的研究评估了生物类型预测对单一治疗（例如，TMS4 或抗抑郁药 7）反应的能力，而不是比较不同治疗类别之间的反应。为了最大化生物类型的转化价值，最终应通过比较不同生物类型在接受相同治疗时的反应来确定每个生物类型的最佳治疗方案。

In the present study, we demonstrate a new approach to generating biotypes of depression and anxiety based on task-evoked and task-free imaging data, quantified at the individual patient level and evaluated in the context of transdiagnostic symptoms, behaviors and outcomes with multiple types of treatments. Our approach relies on a standardized circuit quantification system that enables us to compute a manageable number of task-evoked and task-free measures of circuit function on an individual participant basis. These measures are firmly grounded in a theoretical synthesis of functional brain imaging studies that implicate dysfunction across large-scale circuits in the clinical features of depression and anxiety^16,17. Thus, our theoretically driven approach provides unique insights that may have been missed by previous studies that either relied only on task-free data or mined large numbers of features using exploratory data analysis techniques. In our sample of 801 participants with depression and anxiety (95% of whom were unmedicated), the use of the same fMRI sequences, symptoms and behavioral measures enabled us to clinically validate theory-driven biotypes and demonstrate that they differ in symptom profiles and performance on general and emotional, cognitive, computerized behavioral tests. Furthermore, a substantial portion of the participants were enrolled into randomized clinical trials of antidepressants or behavioral therapy, which enabled us to demonstrate that our biotypes differ in their outcomes across multiple treatments.
在本研究中，我们展示了一种基于任务诱发和无任务成像数据生成抑郁和焦虑生物类型的新方法，这些数据在个体患者水平上进行量化，并在跨诊断症状、行为和多种治疗结果的背景下进行评估。我们的方法依赖于一个标准化的回路量化系统，使我们能够在个体参与者基础上计算出可管理数量的任务诱发和无任务的回路功能测量。这些测量牢牢基于功能性脑成像研究的理论综合，这些研究涉及抑郁和焦虑的临床特征中大规模回路的功能障碍。因此，我们的理论驱动方法提供了独特的见解，这些见解可能被之前仅依赖无任务数据或使用探索性数据分析技术挖掘大量特征的研究所忽视。 在我们 801 名抑郁和焦虑参与者的样本中（其中 95%未接受药物治疗），使用相同的 fMRI 序列、症状和行为测量使我们能够临床验证理论驱动的生物类型，并证明它们在症状特征和一般情感、认知、计算机化行为测试的表现上存在差异。此外，参与者中有相当一部分被纳入了抗抑郁药或行为疗法的随机临床试验，这使我们能够证明我们的生物类型在多种治疗中的结果存在差异。

Personalized brain circuit scores define six biotypes
个性化脑回路评分定义六种生物类型

We began by implementing a new standardized image-processing procedure called ‘the Stanford Et Cere Image Processing System’ which quantified task-free and task-evoked brain circuit function at the level of the individual participants (Methods). We applied this procedure to a baseline dataset that consisted of brain scans acquired from both task-free and task conditions, utilizing identical scanning protocols, from 801 participants with depression and related anxiety disorders, as well as 137 healthy controls (Table 1 and Supplementary Table 1). At the time of baseline scanning, 95% of participants were not receiving any antidepressant treatments and none of the participants was diagnosed with a substance-dependent disorder. We used the same image-processing procedure in a treatment dataset consisting of 250 participants who were reassessed after completing treatment trials. During these trials, the participants were randomly assigned to receive one of three commonly prescribed antidepressant medications (escitalopram, sertraline or venlafaxine extended release (XR)¹⁸ (n = 164)) or an established behavioral intervention that integrated problem-solving with behavioral activation, compared with treatment as usual¹⁹ (n = 86) (Supplementary Tables 1 and 2).
我们首先实施了一种新的标准化图像处理程序，称为“斯坦福 Et Cere 图像处理系统”，该系统量化了个体参与者在无任务和任务诱发的脑回路功能（方法）。我们将此程序应用于一个基线数据集，该数据集由 801 名患有抑郁症和相关焦虑障碍的参与者以及 137 名健康对照者在无任务和任务条件下获取的脑扫描组成，使用相同的扫描协议（表 1 和补充表 1）。在基线扫描时，95%的参与者未接受任何抗抑郁治疗，且没有参与者被诊断为物质依赖障碍。我们在一个治疗数据集中使用了相同的图像处理程序，该数据集由 250 名参与者组成，他们在完成治疗试验后进行了重新评估。 在这些试验中，参与者被随机分配接受三种常用抗抑郁药物之一（艾司西酞普兰、舍曲林或延释氟伏沙明（XR）18（n = 164））或一种已建立的行为干预，该干预将问题解决与行为激活相结合，与常规治疗 19（n = 86）进行比较（补充表 1 和表 2）。

Using our image-processing system, we obtained 41 measures of activation and connectivity of 6 brain circuits of interest for each participant²⁰. We have previously shown that these circuit measures satisfy psychometric criteria for construct validation, internal consistency and generalizability²⁰. A unique feature of our image-processing system is that quantified circuit measures are expressed in terms of s.d. units from the mean of a healthy reference sample, and thus are interpretable for each individual. We refer to the resulting measures as ‘regional circuit scores’ (Fig. 1 and see Supplementary Methods for details).
通过我们的图像处理系统，我们为每个参与者获得了 6 个感兴趣的脑回路的 41 个激活和连接性测量值。我们之前已经证明，这些回路测量值满足构念验证、内部一致性和可推广性的心理测量标准。我们图像处理系统的一个独特特征是，量化的回路测量值以健康参考样本均值的标准差单位表示，因此可以对每个个体进行解释。我们将得到的测量值称为“区域回路评分”（图 1，详细信息见补充方法）。

a, Measures of task-based activation and functional connectivity and task-free connectivity derived from regions belonging to six circuits for which we have established relevance to depression and anxiety. (i) Default mode (D), salience (S) and attention (A) circuits were derived from the task-free periods of the fMRI. The Negative and Positive (P) circuits were engaged by a facial expressions task. In particular, the Negative circuit was engaged in Threat Conscious (NTC), Threat Non-conscious (NTN) and Sad (NS) conditions. The cognitive control circuit (C) was engaged by a Go–NoGo task. (ii) We defined the regions of interest comprising each circuit from the meta-analytic platform Neurosynth and refined them based on quality control, a set of psychometric criteria and whether they were implicated in depression and anxiety. (iii) We extracted functional connectivity between circuit regions for task-free circuits, and activation and connectivity of regions for task-engaged circuits (regions shown as sphere, connectivity shown as lines). b, We then expressed these measures as s.d. values compared with healthy participants to obtain personalized regional circuit scores for each individual. See Supplementary Table 18 for the full list of scores. c, We computed the distance between each pair of individuals as 1 − the correlation of their regional circuit scores. d, We show the distance matrix between the first 100 participants as a heatmap for illustrative purposes. e, We then used the distances obtained as input for a hierarchical clustering analysis. The individuals depicted have given permission to be included in published facial emotion stimulus sets^36,37. AG, angular gyrus; aI, anterior insula; aIPL, anterior inferior parietal lobule; amPFC, anterior medial prefrontal cortex; Amy, amygdala; dACC, dorsal anterior cingulate cortex; DLPFC, dorsolateral prefrontal cortex; LPFC, lateral prefrontal cortex; msPFC, medial superior prefrontal cortex; PCC, posterior cingulate cortex; PCU, precuneus; pgACC, pregenual anterior cingulate cortex; sgACC, subgenual anterior cingulate cortex; vmPFC, venteromedial prefrontal cortex.
a, 任务相关激活和功能连接以及来自六个回路的无任务连接的测量，这些回路与抑郁和焦虑的相关性已建立。(i) 默认模式(D)、显著性(S)和注意力(A)回路来自 fMRI 的无任务期。负性和正性(P)回路通过面部表情任务参与。特别是，负性回路在威胁意识(NTC)、威胁无意识(NTN)和悲伤(NS)条件下参与。认知控制回路(C)通过 Go–NoGo 任务参与。(ii) 我们从元分析平台 Neurosynth 定义了每个回路的感兴趣区域，并根据质量控制、一组心理测量标准以及它们是否与抑郁和焦虑相关进行了细化。(iii) 我们提取了无任务回路的回路区域之间的功能连接，以及任务参与回路的区域的激活和连接（区域以球体表示，连接以线条表示）。b, 然后我们将这些测量表示为标准差。 与健康参与者相比，获得每个个体的个性化区域回路评分。完整评分列表见补充表 18。c，我们计算了每对个体之间的距离，作为 1 减去它们的区域回路评分的相关性。d，我们将前 100 名参与者之间的距离矩阵以热图形式展示，以便于说明。e，然后我们将获得的距离作为层次聚类分析的输入。所描绘的个体已获得许可，包含在已发布的面部情感刺激集 36,37 中。AG，角回；aI，前岛；aIPL，前下顶叶；amPFC，前内侧前额叶皮层；Amy，杏仁体；dACC，背侧前扣带皮层；DLPFC，背外侧前额叶皮层；LPFC，外侧前额叶皮层；msPFC，内侧上前额叶皮层；PCC，后扣带皮层；PCU，前扣带；pgACC，前扣带皮层；sgACC，腹侧前扣带皮层；vmPFC，腹内侧前额叶皮层。

Full size image

To generate biotypes based on regional circuit scores of clinical participants, we used these scores as inputs for a hierarchical clustering algorithm (Fig. 1 and Methods). We generated solutions for 2–15 clusters and evaluated them as shown in Fig. 2.
为了基于临床参与者的区域回路评分生成生物类型，我们将这些评分作为层次聚类算法的输入（图 1 和方法）。我们生成了 2 到 15 个聚类的解决方案，并如图 2 所示进行了评估。

a, We selected candidate biotype solutions selected based on the sum of within-cluster distances. b, We evaluated the silhouette index of our solutions relative to a null multinormal distribution with conserved covariance between individuals¹⁴. c, We compared the silhouette index of our solutions relative to a solution using permuted participant labels, such that participant–brain correspondence was broken. d,e, We repeated our clustering approach leaving one participant out, 801× (d), as well as leaving out 20% of participants, 10,000× (e). In each iteration, we subsequently evaluated the overlap between participant biotype assignment in our original solution and each iterative solution by calculating the ARI. f, We evaluated the circuit measurements associated with each biotype across our original dataset and in two random halves of our original dataset separately. Circuit measurements that were consistently >0.5 s.d. from the mean across all these three samples were considered to be stable. g, We referenced the profile of circuit dysfunction to those found in the literature. h,i, To establish the clinical validity of our biotypes, we evaluated the cluster-specific differences in reported symptoms (h) and performances in a computerized cognitive battery (i). After establishing these differences in the full sample, we evaluated the stability of these symptom and behavioral profiles across two random half-splits of our data, deriving, each time, biotypes from the first half and assigning participants in the second half to a biotype derived from the first. We also followed the same procedure in a leave-study-out framework, leaving one of four of our studies out in each iteration. j,k, We subsequently evaluated the stability of biotype-specific symptom (j) and cognitive (k) differences relative to out-of-biotype participants in each iteration. We considered a difference to be stable when it was statistically significant in the whole sample and in each of the two random half-splits or in each of the two splits of a leave-study-out iteration. l, To evaluate the clinical utility of our cluster biotypes, we tested for differential symptom severity of each biotype to multiple depression treatments. Plots in this figure are only for illustrating the steps of our analysis.
a, 我们选择了基于簇内距离总和的候选生物类型解决方案。b, 我们评估了相对于具有保留个体协方差的零多元正态分布的解决方案的轮廓指数 14。c, 我们比较了相对于使用置换参与者标签的解决方案的轮廓指数，这样参与者与大脑的对应关系被打破。d,e, 我们重复了我们的聚类方法，排除一个参与者，801× (d)，以及排除 20%的参与者，10,000× (e)。在每次迭代中，我们随后通过计算 ARI 评估了我们原始解决方案与每个迭代解决方案中参与者生物类型分配的重叠。f, 我们评估了与每个生物类型相关的回路测量，在我们的原始数据集中以及原始数据集的两个随机部分中分别进行。所有这三个样本中，回路测量值始终大于均值的 0.5 标准差被视为稳定。g, 我们将回路功能障碍的特征与文献中发现的特征进行了对比。 为了建立我们生物类型的临床有效性，我们评估了报告症状（h）和计算机认知测试（i）中集群特异性差异。在全样本中建立这些差异后，我们评估了这些症状和行为特征在我们数据的两个随机半分中的稳定性，每次从第一半部分推导出生物类型，并将第二半部分的参与者分配到从第一部分推导出的生物类型。我们还在留出研究的框架中遵循相同的程序，在每次迭代中留出四项研究中的一项。随后，我们评估了每次迭代中生物类型特异性症状（j）和认知（k）差异相对于非生物类型参与者的稳定性。当差异在整个样本中以及在两个随机半分或在每次留出研究迭代的两个分割中都具有统计显著性时，我们认为该差异是稳定的。为了评估我们集群生物类型的临床效用，我们测试了每种生物类型对多种抑郁症治疗的症状严重性差异。 此图中的图表仅用于说明我们分析的步骤。

Full size image

Biotype validation 生物型验证

We validated our biotypes using six convergent sources of evidence: the elbow method (Fig. 2a); two procedures proposed by Dinga et al.¹⁴ to evaluate the evidence for biotypes of depression and anxiety (simulation-based significance testing of the silhouette index (Fig. 2b) and stability using leave-one-out, and leave-20%-out crossvalidation (Fig. 2d,e)); an additional permutation-based significance testing of the silhouette index (Fig. 2c); split-half reliability of the cluster profiles (Fig. 2f); and the match of the solution to a theoretical framework of circuit dysfunction in depression and anxiety supported by previous brain imaging research¹⁷ (Fig. 2g).
我们使用六个汇聚的证据来源验证了我们的生物类型：肘部法（图 2a）；Dinga 等人提出的两种程序 14，用于评估抑郁和焦虑的生物类型证据（基于模拟的轮廓指数显著性检验（图 2b）和使用留一法和留 20%法的交叉验证稳定性（图 2d,e））；额外的基于置换的轮廓指数显著性检验（图 2c）；聚类轮廓的分半信度（图 2f）；以及解决方案与抑郁和焦虑中回路功能障碍的理论框架的匹配，该框架得到了之前脑成像研究的支持 17（图 2g）。

The elbow method showed an elbow at five clusters and another, smaller elbow, at nine clusters, which suggested that the optimal solution lay between these two values (Supplementary Fig. 1). Simulation-based significance testing of the silhouette index showed that solutions with five or more clusters had a silhouette index that was significantly higher than that obtained by clustering data from a multivariate normal distribution (all P < 0.05; Supplementary Fig. 2) and significantly higher than that obtained by a permutation of the circuit scores across participants (P < 0.05; Supplementary Fig. 3). Assessment of cluster stability using crossvalidation showed that all solutions had good stability (adjusted Rand index (ARI) > 0.75 for leave-one-out and ARI > 0.28 for leave-20%-out) (Supplementary Fig. 4).
肘部法则显示在五个聚类处有一个肘部，在九个聚类处有另一个较小的肘部，这表明最佳解决方案位于这两个值之间（补充图 1）。基于模拟的轮廓指数显著性测试显示，五个或更多聚类的解决方案的轮廓指数显著高于从多元正态分布中聚类数据所获得的轮廓指数（所有 P < 0.05；补充图 2），并且显著高于在参与者之间对回路评分进行置换所获得的轮廓指数（P < 0.05；补充图 3）。使用交叉验证评估聚类稳定性显示所有解决方案具有良好的稳定性（调整兰德指数（ARI）> 0.75 对于留一法，ARI > 0.28 对于留 20%法）（补充图 4）。

Across all validation analyses, six emerged as a viable number of clusters. The silhouette index tests comparing the data with data from a multivariate normal distribution and with a permutation of the circuit scores across participants were significant for this solution (mean silhouette = 0.065, P = 0.016 and P < 0.0001, respectively) and crossvalidation showed that it had good stability (leave-study-out ARI = 0.80 and leave-20%-out ARI = 0.35). Also, in the six-cluster solution, a cluster emerged that was characterized by reduced task-evoked activation during cognitive control, which we had specifically hypothesized^16,17.
在所有验证分析中，六个被认为是可行的聚类数量。轮廓指数测试将数据与多元正态分布的数据以及参与者之间的回路评分置换进行比较，对于该解决方案是显著的（平均轮廓 = 0.065，P = 0.016 和 P < 0.0001），交叉验证显示其具有良好的稳定性（留出研究的 ARI = 0.80 和留出 20%的 ARI = 0.35）。此外，在六聚类解决方案中，出现了一个聚类，其特征是在认知控制期间任务诱发的激活减少，这是我们特意假设的 16,17。

The six resulting biotypes were distinguished by specific profiles of both task-free and task-evoked activity and/or connectivity, relative both to each other and to our healthy reference sample. To assign a name to these distinctive circuit profiles, we determined which circuit features, activity or connectivity were distinguished by a difference of at least 0.50 s.d. in magnitude away from the healthy reference sample. The distinct activity and connectivity profiles of each biotype are illustrated using a circuit schematic and numerical plot in Fig. 3 with further details illustrated in bar plots in Supplementary Fig. 5. We named each biotype according to the circuits and circuit features that specifically differentiated them at this threshold relative to each other and to the healthy reference sample. We used the following nomenclature (each circuit is indicated with a letter): D, default mode; S, salience; A, attention; NS, negative affect circuit evoked by sad stimuli; NTC, negative affect circuit evoked by conscious threat stimuli; NTN, negative affect circuit evoked by nonconscious threat stimuli; P, positive affect circuit; C, cognitive circuit. The distinguishing circuit feature is indicated as a subscript: C, connectivity; A, activity, and the direction of dysfunction is indicated by + or −. These distinct profiles were also replicated when conducting the clustering procedure on a random half of the data and assigning participants in the second independent half of the data to each cluster (Supplementary Fig. 6).
这六种生物类型通过特定的无任务和任务诱发活动和/或连接性特征相互区分，并与我们的健康参考样本进行比较。为了给这些独特的回路特征命名，我们确定了哪些回路特征、活动或连接性在与健康参考样本的比较中，至少有 0.50 标准差的显著差异。每种生物类型的独特活动和连接性特征在图 3 中通过回路示意图和数值图进行了说明，补充图 5 中的条形图提供了进一步的细节。我们根据这些生物类型在此阈值下相互之间以及与健康参考样本的具体区分的回路和回路特征为每种生物类型命名。我们使用了以下命名法（每个回路用字母表示）：D，默认模式；S，显著性；A，注意力；NS，由悲伤刺激诱发的负性情感回路；NTC，由意识威胁刺激诱发的负性情感回路；NTN，由非意识威胁刺激诱发的负性情感回路；P，积极情感回路；C，认知回路。 区分回路特征用下标表示：C，连接性；A，活动，功能障碍的方向用 + 或 − 表示。当对数据的随机一半进行聚类程序并将第二独立一半的数据中的参与者分配到每个聚类时，这些不同的特征也得到了复制（补充图 6）。

a–f, Schematic circuit images illustrating the profile of circuit dysfunction defining each biotype (biotypes are labeled a–f). Circuits are distinguished by colors that correspond to the circuit measure inputs (Fig. 1c). Spheres represent the regions within each biotype-defining circuit and the size of the spheres represents the magnitude of activation deviation from the healthy reference (small spheres, activation ≤0.5 s.d. below the healthy reference; large spheres, activation ≥0.5 s.d. above the healthy reference). The thickness of lines between the spheres denotes a connectivity deviation (dashed lines, decreased connectivity ≤0.5 s.d. below the healthy reference; thick lines, increased connectivity ≥0.5 s.d. above the healthy reference). Column plots display the average activity across regions that define each circuit or the average connectivity between regions that define each circuit. A visualization of each regional circuit score by biotype is in Supplementary Fig. 5. In bar plots, we highlight circuits that showed a mean difference of at least 0.50 s.d. below or above the healthy reference. We named each biotype according to the features that differentiated it from the healthy reference. Each circuit is indicated with a letter, the distinguishing circuit feature is indicated as a subscript and the direction of dysfunction is indicated by + or −. The subscript x indicates that the sixth biotype is not differentiated by a prominent circuit dysfunction. Besides this nomenclature, we suggest a short description for each biotype, which connects them with our theoretically synthesized biotypes: D_C+S_C+A_C+, default with salience and attention hyperconnectivity (n = 169 participants); A_C−, attention hypoconnectivity (n = 161 participants); NS_A+P_A+, sad-elicited negative affect with positive affect hyperactivation (n = 154 participants); C_A+, cognitive control hyperactivation (n = 258 participants); NTC_C−C_A−, cognitive control hypoactivation with conscious threat-elicited negative affect hypoconnectivity (n = 15 participants); and D_XS_XA_XN_XP_XC_X, intact activation and connectivity (n = 44 participants).
a–f，示意电路图展示了定义每种生物类型的电路功能障碍特征（生物类型标记为 a–f）。电路通过与电路测量输入相对应的颜色进行区分（图 1c）。球体代表每种生物类型定义电路中的区域，球体的大小表示与健康参考的激活偏差的大小（小球体，激活≤健康参考下方 0.5 标准差；大球体，激活≥健康参考上方 0.5 标准差）。球体之间线条的粗细表示连接性偏差（虚线，连接性降低≤健康参考下方 0.5 标准差；粗线，连接性增加≥健康参考上方 0.5 标准差）。列图显示定义每个电路的区域的平均活动或定义每个电路的区域之间的平均连接性。每种生物类型的区域电路评分可在补充图 5 中查看。在条形图中，我们突出显示了显示平均差异至少为健康参考下方或上方 0.50 标准差的电路。 我们根据每种生物型与健康参考的特征差异命名。每个回路用一个字母表示，区分回路特征用下标表示，功能障碍的方向用 + 或 − 表示。下标 x 表示第六种生物型没有明显的回路功能障碍。除了这种命名法，我们为每种生物型建议了一个简短的描述，将它们与我们理论上合成的生物型联系起来：DC+SC+AC+，默认状态下的显著性和注意力超连接（n = 169 名参与者）；AC−，注意力低连接（n = 161 名参与者）；NSA+PA+，悲伤引发的负面情感与正面情感超激活（n = 154 名参与者）；CA+，认知控制超激活（n = 258 名参与者）；NTCC−CA−，认知控制低激活与意识威胁引发的负面情感低连接（n = 15 名参与者）；以及 DXSXAXNXPXCX，完整的激活和连接（n = 44 名参与者）。

Full size image

Biotype D_C+S_C+A_C+ (n = 169) was distinguished by relative intrinsic hyperconnectivity within the default mode circuit, as well as in the task-free salience and attention circuits (Fig. 3a). In contrast, biotype A_C− (n = 161) was distinguished by a relative reduction in intrinsic connectivity specific to the attention circuit (Fig. 3b). Biotype NS_A+P_A+ (n = 154) was characterized by heightened activity during conscious emotion processing, specifically within the negative affect circuit evoked by sad stimuli and within the positive affect circuit evoked by happy stimuli (Fig. 3c). Biotype C_A+ (n = 258) was distinguished specifically by increased activity within the cognitive control circuit during the inhibition of NoGo stimuli (Fig. 3d). Biotype NTC_C-C_A− (n = 15) was a smaller cluster differentiated by a relative loss of functional connectivity within the negative affect circuit during the conscious processing of threat faces, as well as by reduced (rather than heightened) activity within the cognitive control circuit during the inhibition of NoGo stimuli (Fig. 3e). Biotype D_XS_XA_XN_XP_XC_X (n = 44) was not differentiated by a substantial circuit dysfunction relative to other biotypes or to the healthy norm; we indicated this by using the subscript x instead of + or − (Fig. 3f).
生物型 DC+SC+AC+ (n = 169) 通过默认模式回路内的相对内在超连接性以及在无任务显著性和注意力回路中的相对内在超连接性而被区分（图 3a）。相反，生物型 AC− (n = 161) 通过注意力回路内的相对内在连接性降低而被区分（图 3b）。生物型 NSA+PA+ (n = 154) 的特征是在意识情绪处理过程中，特别是在由悲伤刺激引发的负情感回路和由快乐刺激引发的正情感回路内的活动增强（图 3c）。生物型 CA+ (n = 258) 特别通过在抑制 NoGo 刺激时认知控制回路内的活动增加而被区分（图 3d）。生物型 NTCC-CA− (n = 15) 是一个较小的簇，通过在意识处理威胁面孔时负情感回路内的功能连接性相对丧失以及在抑制 NoGo 刺激时认知控制回路内的活动减少（而非增加）而被区分（图 3e）。 生物型 DXSXAXNXPXCX（n = 44）与其他生物型或健康标准相比，并未表现出显著的回路功能障碍；我们通过使用下标 x 而不是+或−来表示这一点（图 3f）。

These distinct biotype circuit profiles were not explained by differences in scanners, because we removed scanner effects from our data using ComBat (Methods) and verified that the distribution of biotypes did not differ across scanners (χ² = 12.773, two-sided P = 0.237).
这些不同的生物类型回路特征并不能通过扫描仪的差异来解释，因为我们使用 ComBat（方法）从数据中去除了扫描仪效应，并验证了生物类型在不同扫描仪之间的分布没有差异（χ2 = 12.773，双侧 P = 0.237）。

Biotypes differ on symptoms, behavior and treatment response
生物类型在症状、行为和治疗反应上存在差异

To further characterize the clinical phenotypes distinguished by each circuit biotype, we evaluated the biotype profiles on three different domains of clinically meaningful measures (Fig. 4): severity of symptoms, performance on general and emotional cognitive tests and differential treatment response. We highlight that the circuit biotypes derived from clustering were differentiated using only circuit inputs assessed independently from these domains of clinical information such that symptoms, performance and treatment response represented external validation measures.
为了进一步表征每种回路生物型所区分的临床表型，我们在三个不同的临床相关测量领域（图 4）上评估了生物型特征：症状严重程度、一般和情感认知测试的表现以及差异化治疗反应。我们强调，从聚类中得出的回路生物型仅使用独立于这些临床信息领域评估的回路输入进行区分，因此症状、表现和治疗反应代表了外部验证测量。

a–f, Circuit biotypes are visualized using circuit schematics on the left (biotypes are labeled a–f). We first compared these circuit biotypes on symptoms of depression and related anxiety (column ‘Symptom severity’). Next, we compared biotypes on behavioral performance on general and emotional cognitive tests relevant to social and occupational function (column ‘Behavioral dysfunction’). We compared biotypes on severity after treatment with one of three antidepressant pharmacotherapies (escitalopram, sertraline or venlafaxine XR), a behavioral problem-solving therapy (I-CARE) or usual care (U-CARE) (column ‘Severity after treatment’). To facilitate comparison across units of analysis, all measures were scaled between 0 and 1 so that 0 would represent minimum severity/dysfunction and 1 maximum severity/dysfunction. The column ‘Severity after treatment’ shows differences in symptom severity posttreatment (that is, lower values correspond to better treatment response). Comparisons on severity after treatment were conducted only for biotype/treatment combinations having n ≥ 5, so only those are shown. We used the biotype nomenclature used previously. The subscript x indicates that the sixth biotype is not differentiated by a prominent circuit dysfunction relative to other biotypes. Besides this nomenclature, we suggest a short plain-English description for each biotype (in quotes), which connects them with our theoretically synthesized biotypes (as shown in Fig. 3).
a–f，回路生物型通过左侧的回路示意图进行可视化（生物型标记为 a–f）。我们首先比较了这些回路生物型在抑郁症和相关焦虑症状上的表现（列“症状严重性”）。接下来，我们比较了生物型在与社会和职业功能相关的一般和情感认知测试中的行为表现（列“行为功能障碍”）。我们比较了生物型在接受三种抗抑郁药物治疗（艾司西酞普兰、舍曲林或文拉法辛 XR）、一种行为问题解决疗法（I-CARE）或常规护理（U-CARE）后的严重性（列“治疗后的严重性”）。为了便于跨分析单位的比较，所有测量均缩放在 0 到 1 之间，以便 0 表示最低严重性/功能障碍，1 表示最高严重性/功能障碍。列“治疗后的严重性”显示了治疗后症状严重性的差异（即，较低的值对应于更好的治疗反应）。治疗后的严重性比较仅针对样本量 n ≥ 5 的生物型/治疗组合进行，因此仅显示这些。我们使用了之前使用的生物型命名法。 下标 x 表示第六种生物类型与其他生物类型相比，并没有通过显著的回路功能障碍来区分。除了这种命名法，我们建议为每种生物类型提供一个简短的通俗英语描述（用引号括起来），将它们与我们理论上合成的生物类型联系起来（如图 3 所示）。

Full size image

We first asked whether the biotypes were distinguished by the severity of symptoms of depression and anxiety. To address this question, we used Mann–Whitney U-tests to compare the symptom severity of each biotype to the median symptom severity of all clinical participants not in the biotype (Supplementary Fig. 10 and Supplementary Tables 3 and 4). For insomnia and suicidality, these comparisons were conducted using χ² tests instead (Supplementary Fig. 11 and Supplementary Table 5). We considered significant tests for which P < 0.05. We then replicated significant findings in split-half and leave-study-out analyses (Fig. 2h,j).
我们首先询问生物类型是否通过抑郁和焦虑症状的严重程度来区分。为了解决这个问题，我们使用 Mann-Whitney U 检验将每个生物类型的症状严重程度与所有不属于该生物类型的临床参与者的中位症状严重程度进行比较（补充图 10 和补充表 3 和 4）。对于失眠和自杀倾向，这些比较则使用χ2 检验进行（补充图 11 和补充表 5）。我们考虑 P < 0.05 的显著性检验。然后，我们在分半和留出研究分析中复制显著发现（图 2h,j）。

Second, we assessed whether biotypes are distinguished by performance on a computerized battery of general and emotional cognitive tests relevant to daily social and occupational function. We conducted these analyses as described above for symptoms (Supplementary Fig. 12 and Supplementary Tables 6 and 7). We then replicated significant findings in split-half and leave-study-out analyses (Fig. 2i,k).
其次，我们评估了生物类型是否通过在与日常社交和职业功能相关的计算机化一般和情感认知测试中的表现而有所区分。我们按照上述描述对症状进行了这些分析（补充图 12 和补充表 6 和 7）。然后，我们在分半和留出研究分析中复制了显著发现（图 2i,k）。

Third, we assessed whether the biotypes predicted differential treatment response to one of the three pharmacotherapies or to behavioral therapy versus usual care. We conducted these analyses as described above for symptoms and behavior (Fig. 2l, Supplementary Fig. 13 and Supplementary Tables 8–10).
第三，我们评估了生物类型是否预测了对三种药物治疗之一或对行为疗法与常规护理的不同治疗反应。我们按照上述描述对症状和行为进行了这些分析（图 2l，补充图 13 和补充表 8-10）。

Biotype D_C+S_C+A_C+, characterized by task-free circuit hyperconnectivity, had slowed behavioral responses in identifying sad faces (effect size (ES) = 0.289, P = 0.001, confidence interval (CI) = (−0.072, 0.289), replicated in leave-study-out), increased errors in an executive function task (ES = 0.175, P = 0.044, CI = (9−0.182, 0.166)), fewer commission errors in a cognitive control task (ES = −0.275, P = 0.002, CI = (−0.505, −0.217), replicated in leave-study-out) slowed responses to target stimuli in a sustained attention task (ES = 0.336, P = 0.0001, CI = (0.714, 1.099)) (see Fig. 4a and Supplementary Figs. 10–12 for detailed visualization and Supplementary Tables 3–7 for comparisons). The biotype D_C+S_C+A_C+ responded better to I-CARE compared with other biotypes (ES = −0.612, P = 0.037, CI = (0.137, 0.306), responders = 42%, remitters = 25%) (Fig. 4a, Supplementary Fig. 13 and Supplementary Tables 8–10).
生物型 DC+SC+AC+ 以无任务回路超连接为特征，在识别悲伤面孔时表现出减慢的行为反应（效应大小 (ES) = 0.289, P = 0.001, 置信区间 (CI) = (−0.072, 0.289)，在留出研究中重复），在执行功能任务中错误增加（ES = 0.175, P = 0.044, CI = (−0.182, 0.166)），在认知控制任务中较少的错误（ES = −0.275, P = 0.002, CI = (−0.505, −0.217)，在留出研究中重复），在持续注意任务中对目标刺激的反应减慢（ES = 0.336, P = 0.0001, CI = (0.714, 1.099)）（见图 4a 和补充图 10–12 以获取详细可视化，补充表 3–7 以进行比较）。生物型 DC+SC+AC+ 对 I-CARE 的反应优于其他生物型（ES = −0.612, P = 0.037, CI = (0.137, 0.306)，反应者 = 42%，缓解者 = 25%）（图 4a，补充图 13 和补充表 8–10）。

Biotype A_C-, characterized by task-free attention circuit hypoconnectivity, had relatively less severe tension (ES = −0.196, P = 0.049, CI = (11.5, 15)), but was also differentiated by relatively lower cognitive dyscontrol (ES = −0.305, P = 0.006, CI = (15.5; 17.5)). In computerized tests, A_C− was distinguished by faster responses to target Go stimuli on the Go–NoGo task, (ES = −0.383, P = 6.20 × 10⁻⁶, CI = (0.180, 0.510), replicated in split-half), more commission and omission errors on the sustained attention task (ES = 0.300, P = 0.0004, CI = (−0.302, −0.019); ES = 0.198, P = 0.020, CI = (−0.308, −0.010)) and faster responses to priming by implicit threat stimuli (ES = −0.256, P = 0.002, CI = (−0.111, 0.112)) (see Fig. 4b and Supplementary Figs. 10–12 for detailed visualization and Supplementary Tables 3–7 for comparisons). The A_C− biotype had comparatively worse response to I-CARE (ES = 0.593, P = 0.002, CI = (0.219; 0.350), responders = 26%, remitters = 22%) (Fig. 4a, Supplementary Fig. 13 and Supplementary Tables 8–10).
生物型 AC-，其特征为无任务注意回路低连接性，表现出相对较轻的紧张（效应量 = −0.196，P = 0.049，置信区间 = (11.5, 15)），但在认知失控方面也表现出相对较低（效应量 = −0.305，P = 0.006，置信区间 = (15.5; 17.5)）。在计算机化测试中，AC− 在 Go–NoGo 任务中对目标 Go 刺激的反应速度较快（效应量 = −0.383，P = 6.20 × 10−6，置信区间 = (0.180, 0.510)，在分半验证中重复），在持续注意任务中表现出更多的误报和漏报错误（效应量 = 0.300，P = 0.0004，置信区间 = (−0.302, −0.019)；效应量 = 0.198，P = 0.020，置信区间 = (−0.308, −0.010)），以及对隐性威胁刺激的启动反应更快（效应量 = −0.256，P = 0.002，置信区间 = (−0.111, 0.112)）（详见图 4b 和补充图 10–12 以获取详细可视化，补充表 3–7 以进行比较）。AC− 生物型对 I-CARE 的反应相对较差（效应量 = 0.593，P = 0.002，置信区间 = (0.219; 0.350)，反应者 = 26%，缓解者 = 22%）（图 4a，补充图 13 和补充表 8–10）。

Biotype NS_A+P_A+, distinguished by circuit hyperactivation during conscious emotion processing, was distinguished by more severe anhedonia (ES = 0.343, P = 0.014, CI = (2, 4.5)) and ruminative brooding (ES = 0.294, P = 0.036, CI = (55.5, 63)) (Fig. 4c; see Supplementary Figs. 10–12 for detailed visualization and Supplementary Tables 3–7 for comparisons).
生物型 NSA+PA+在意识情感处理过程中以回路过度激活为特征，其表现为更严重的快感缺失（效应量 = 0.343，P = 0.014，置信区间 = (2, 4.5)）和反刍性沉思（效应量 = 0.294，P = 0.036，置信区间 = (55.5, 63)）（图 4c；详见补充图 10-12 以获取详细可视化，补充表 3-7 以进行比较）。

Biotype C_A+, distinguished by heightened activity within the cognitive control circuit, had more severe anhedonia than other biotypes (ES = 0.295, P = 0.015, CI = (2, 3.5)), more anxious arousal (ES = 0.218, P = 0.003, CI = (15.5, 17.5)), more negative bias (ES = 0.188, P = 0.003, CI = (15, 18.5), replicated in split-half) and more threat dysregulation (ES = 0.317, P = 5.07 × 10⁻⁷, CI = [7.5, 9], replicated in split-half and leave-study-out). Behaviorally, C_A+ had more errors and completion time in the executive function task (ES = 0.164, P = 0.017, CI = (−0.268, −0.027) and ES = 0.152, P = 0.027, CI = (−0.164, 0.090)), more commission errors in the Go–NoGo task (ES = 0.158, P = 0.022, CI = (−0.201, 0.035), replicated in split-half) and more omission errors to target stimuli on the sustained attention task (ES = 0.275, P = 6.46 × 10⁻⁵, CI = (−0.045, 0.170), replicated in split-half and leave-study-out) (Fig. 4c; see Supplementary Figs. 10–12 for detailed visualization and Supplementary Tables 3–7 for comparisons). This biotype showed a better response to venlafaxine compared with the others (ES = −0.426, P = 0.034, CI = (0.132, 0.226), responders = 64%, remitters = 40%) (Fig. 4c, Supplementary Fig. 13 and Supplementary Tables 8–10).
生物型 CA+，以认知控制回路的活动增强为特征，表现出比其他生物型更严重的快感缺失（ES = 0.295，P = 0.015，CI = (2, 3.5)），更高的焦虑唤醒（ES = 0.218，P = 0.003，CI = (15.5, 17.5)），更强的负面偏见（ES = 0.188，P = 0.003，CI = (15, 18.5)，在分半中重复）以及更多的威胁失调（ES = 0.317，P = 5.07 × 10−7，CI = [7.5, 9]，在分半和留出研究中重复）。在行为上，CA+在执行功能任务中表现出更多的错误和完成时间（ES = 0.164，P = 0.017，CI = (−0.268, −0.027) 和 ES = 0.152，P = 0.027，CI = (−0.164, 0.090)），在 Go–NoGo 任务中表现出更多的错误反应（ES = 0.158，P = 0.022，CI = (−0.201, 0.035)，在分半中重复）以及在持续注意任务中对目标刺激的遗漏错误更多（ES = 0.275，P = 6.46 × 10−5，CI = (−0.045, 0.170)，在分半和留出研究中重复）（图 4c；详见补充图 10–12 的可视化和补充表 3–7 的比较）。与其他生物型相比，该生物型对文拉法辛的反应更好（ES = −0.426，P = 0.034，CI = (0.132, 0.226)，反应者 = 64%，缓解者 = 40%）（图）。 4c，补充图 13 和补充表 8-10）。

Biotype NTC_C-C_A-, differentiated by loss of functional connectivity within the negative affect circuit during the conscious processing of threat faces, as well as reduced activity within the cognitive control circuit, had less ruminative brooding compared with the other biotypes (ES = −0.902, P = 0.036, CI = (46, 5)), as well as faster reaction times to implicit sad faces (ES = −0.669, P = 0.024, CI = (−1.316, −0.315)) (Fig. 4d; see Supplementary Figs. 10–12 for detailed visualization and Supplementary Tables 3–7 for comparisons).
生物型 NTCC-CA-，通过在意识处理威胁面孔时负面情感回路内功能连接的丧失以及认知控制回路内活动的减少而区分，与其他生物型相比，具有较少的反刍沉思（ES = −0.902，P = 0.036，CI = (46, 5)），以及对隐含悲伤面孔的反应时间更快（ES = −0.669，P = 0.024，CI = (−1.316, −0.315)）（图 4d；详见补充图 10–12 进行详细可视化，补充表 3–7 进行比较）。

Biotype D_XS_XA_XN_XP_XC_X was not differentiated by a prominent circuit dysfunction relative to other biotypes or the healthy norm; however, it was distinguished by slower reaction times to implicit threat priming (ES = 0.516, P = 0.001, CI = (0.254, 0.611)) (Fig. 4e; see Supplementary Figs. 10–12 for detailed visualization and Supplementary Tables 3–7 for comparisons).
生物型 DXSXAXNXPXCX 与其他生物型或健康标准相比，并没有通过显著的回路功能障碍来区分；然而，它的隐性威胁启动反应时间较慢（效应量=0.516，P=0.001，置信区间=(0.254, 0.611)）（图 4e；详见补充图 10-12 以获取详细可视化，补充表 3-7 以进行比较）。

Finally, we also considered the demographic factors of age and biological sex. The biotypes did not differ in sex distribution (χ² = 12.643, P = 0.244) and only the A_C− biotype was, on average, slightly older than the other biotypes; importantly, however, participants in this biotype were still within the young to mid-adult age range (mean age: 39.69 years, s.d. = 15.739, F = 8.761, P = 4.21 × 10⁻⁸). Biotypes were also represented differently between datasets, which we expected given the clinical differences between the participants enrolled into each study (χ² = 161.37, P = 2.2 × 10⁻¹⁶) (Supplementary Table 11).
最后，我们还考虑了年龄和生物性别的人口统计因素。生物类型在性别分布上没有差异（χ2 = 12.643，P = 0.244），而且只有 AC−生物类型的平均年龄略高于其他生物类型；然而，重要的是，这一生物类型的参与者仍然处于年轻到中年成人的年龄范围内（平均年龄：39.69 岁，标准差 = 15.739，F = 8.761，P = 4.21 × 10−8）。不同数据集之间的生物类型表现也有所不同，这一点是我们所预期的，因为参与每项研究的参与者之间存在临床差异（χ2 = 161.37，P = 2.2 × 10−16）（补充表 11）。

As a context for the above evaluation of how biotypes were distinguished by symptoms, performance and treatment response, we evaluated the correlations between circuit scores and these external measures in the full sample across clusters combined (Supplementary Figs. 7–9). When thresholded with the false discovery rate correction for all pairwise correlations, we observed significant associations between circuit scores and 21% of the symptom measures, 10% of the performance measures and 31% of the treatment response measures.
作为上述评估生物类型如何通过症状、表现和治疗反应进行区分的背景，我们评估了在合并的各个簇中，回路评分与这些外部测量之间的相关性（补充图 7-9）。当对所有成对相关性进行假发现率校正时，我们观察到回路评分与 21%的症状测量、10%的表现测量和 31%的治疗反应测量之间存在显著关联。

Biotypes are transdiagnostic
生物类型是跨诊断的

The distinct clinical and treatment profiles that distinguish the six biotypes indicate that these circuit-derived biotypes dissect the heterogeneity of the traditional diagnostic classification of depression. We next asked whether biotypes transcend diagnostic classifications across the diagnoses that are related to and comorbid with depression. Our sample was composed of participants who met traditional diagnostic criteria for major depressive disorder (n = 375), generalized anxiety disorder (n = 192), panic disorder (n = 75), social anxiety disorder (n = 179), obsessive–compulsive disorder (n = 47) and post-traumatic stress disorder (n = 37). Several participants also met criteria for more than one diagnosis (n = 221) (Table 1).
这六种生物类型所区分的独特临床和治疗特征表明，这些源自回路的生物类型剖析了传统抑郁症诊断分类的异质性。接下来，我们询问生物类型是否超越了与抑郁症相关和共病的诊断分类。我们的样本由符合传统重度抑郁症诊断标准的参与者组成（n = 375）、广泛性焦虑症（n = 192）、惊恐障碍（n = 75）、社交焦虑症（n = 179）、强迫症（n = 47）和创伤后应激障碍（n = 37）。一些参与者还符合多个诊断的标准（n = 221）（表 1）。

The only diagnosis with a different frequency across biotypes was current major depressive disorder (χ² = 24.235, two-sided P = 0.0002). In particular, the A_C− biotype had the highest proportion of participants with current major depressive disorder and the D_XS_XA_XN_XP_XC_X cluster had the lowest proportion (Fig. 5 and Supplementary Table 12).
在不同生物类型中，唯一一个频率不同的诊断是当前的重度抑郁症（χ2 = 24.235，双侧 P = 0.0002）。特别是，AC−生物类型中当前重度抑郁症参与者的比例最高，而 DXSXAXNXPXCX 簇的比例最低（图 5 和补充表 12）。

We show the proportion of participants in each biotype who meet diagnostic criteria for major depressive disorder, generalized anxiety disorder, panic disorder, social anxiety disorder, obsessive–compulsive disorder and post-traumatic stress disorder (biotypes are labeled a–f). χ² tests revealed that the frequency of major depressive disorder was significantly different across biotypes (two-sided χ² = 24.235, P = 0.0002). We used the same biotype nomenclature as previously. The subscript x indicates that the sixth biotype is not differentiated by a prominent circuit dysfunction relative to other biotypes. Besides this nomenclature, we suggest a short plain-English description for each biotype (in quotes), which connects them with our theoretically synthesized biotypes, again as expressed in the legend to Fig. 3.
我们展示了每种生物类型中符合重度抑郁症、广泛性焦虑症、惊恐障碍、社交焦虑症、强迫症和创伤后应激障碍诊断标准的参与者比例（生物类型标记为 a-f）。χ2 检验显示，重度抑郁症在不同生物类型之间的发生频率显著不同（双侧χ2 = 24.235，P = 0.0002）。我们使用了与之前相同的生物类型命名法。下标 x 表示第六种生物类型与其他生物类型相比没有明显的回路功能障碍。除了这种命名法，我们还为每种生物类型建议了一个简短的通俗英语描述（用引号标出），将它们与我们理论上合成的生物类型联系起来，再次如图 3 的图例所示。

Full size image

Brain circuit scores outperform other features for biotyping
脑回路评分优于其他特征用于生物类型鉴定

To compare prior approaches for biotyping with ours, we repeated our analysis using three competing alternative feature sets, each used in a recent paper reporting the identification of biotypes of depression using resting state fMRI. We then evaluated the results with the same criteria that we used for our own features (Fig. 2). Our findings show that our feature set is the only one that outperforms the null hypothesis of no clusters based on simulating data from a multinormal distribution with the same covariance as the original data (P = 0.016). In direct statistical comparisons of clustering performance between feature sets used as inputs, our combination of task and task-free regional circuit scores outperformed whole-brain connectomes (silhouette difference = −0.026, P_resample = 0.049, P_permute < 0.0001) and default mode network resting state connectivity (silhouette difference = −0.012, P_resample = 0.256, P_permute < 0.0001), but not connectivity of a network centered on the angular gyrus (silhouette difference = 0.155, P_resample = 1, P_permute = 1). The other feature sets also yielded associations among various metrics of biotypes, symptoms, behavioral performance and treatment response (Supplementary Tables 13 and 14).
为了将我们的方法与先前的生物类型识别方法进行比较，我们使用了三组竞争性的替代特征集重复了我们的分析，这些特征集均在最近的一篇论文中被使用，该论文报告了使用静息态 fMRI 识别抑郁症的生物类型。然后，我们使用与我们自己特征相同的标准评估结果（图 2）。我们的发现表明，我们的特征集是唯一一个在基于从具有与原始数据相同协方差的多元正态分布模拟数据的情况下，超越无聚类零假设的特征集（P = 0.016）。在对作为输入的特征集之间的聚类性能进行直接统计比较时，我们的任务和无任务区域回路评分的组合优于全脑连接组（轮廓差 = −0.026，Presample = 0.049，Ppermute < 0.0001）和默认模式网络静息态连接性（轮廓差 = −0.012，Presample = 0.256，Ppermute < 0.0001），但不优于以角回为中心的网络连接性（轮廓差 = 0.155，Presample = 1，Ppermute = 1）。 其他特征集也在生物类型、症状、行为表现和治疗反应的各种指标之间产生了关联（补充表 13 和 14）。

To assess the impact of including task fMRI measures in addition to task-free brain circuit scores only, we also evaluated, in the same way, the results obtained using only our task-free brain circuit scores as input. To do so we showed that limiting the analysis to task-free brain circuit scores generated results that did not outperform the null hypothesis of no clusters based on simulating data from a multinormal distribution with the same covariance as the original data. Task-based brain circuit scores were also necessary to obtain symptom differences that generalize across random split-halves and behavior differences that generalize across the leave-study-out splits, depending on the number of clusters chosen (Supplementary Table 15).
为了评估在仅使用无任务脑回路评分的基础上加入任务 fMRI 测量的影响，我们同样以相同的方式评估了仅使用我们的无任务脑回路评分作为输入所获得的结果。为此，我们展示了将分析限制在无任务脑回路评分上所生成的结果未能超越基于从具有相同协方差的多元正态分布模拟数据得出的无聚类假设。基于任务的脑回路评分也是获得在随机分割的半数中普遍适用的症状差异和在留出研究分割中普遍适用的行为差异所必需的，这取决于所选择的聚类数量（补充表 15）。

To enable more precise diagnosis and selection of the best treatment for each individual, we need to dissect the heterogeneity of depression and anxiety. The dominant ‘one-size-fits-all’ diagnostic approach in psychiatry leads to cycling through treatment options by trial and error, which is lengthy, expensive and frustrating, with 30–40% of patients not achieving remission after trying one treatment²¹.
为了实现更精确的诊断和为每个个体选择最佳治疗方案，我们需要剖析抑郁症和焦虑症的异质性。精神病学中主导的“一个尺寸适合所有”的诊断方法导致通过试错循环治疗选项，这个过程漫长、昂贵且令人沮丧，30-40%的患者在尝试一种治疗后未能达到缓解。

In the present study, we focus on the conceptualization of depression and anxiety as disorders of brain circuit function²². Using clustering and a new imaging system for the standardized quantification of circuit dysfunction at the level of the individual, we characterized six biotypes of depression and anxiety defined by specific profiles of dysfunction within both task-free and task-evoked brain circuits. These biotypes were validated using several procedures including simulations, crossvalidation and replication in held-out data. We found that the biotypes were distinguished by symptoms and behavioral performance on general and emotional cognitive tests that were not used as inputs in the clustering procedure. Importantly, some of these associations were replicated in split-half and leave-study-out procedures. We also showed that the six biotypes cut across the diagnostic boundaries of depression, anxiety and related comorbid disorders. Importantly for clinical translation, these biotypes predict response to different pharmacological and behavioral interventions.
在本研究中，我们将抑郁症和焦虑症概念化为脑回路功能障碍的疾病。通过聚类分析和一种新的成像系统，对个体层面的回路功能障碍进行标准化量化，我们描述了六种抑郁症和焦虑症的生物类型，这些生物类型由任务无关和任务诱发的脑回路中的特定功能障碍特征定义。我们使用多种程序对这些生物类型进行了验证，包括模拟、交叉验证和在保留数据中的复制。我们发现，这些生物类型通过症状和在一般和情感认知测试中的行为表现进行区分，而这些测试并未作为聚类程序的输入。重要的是，这些关联在分半和留出研究程序中得到了复制。我们还表明，这六种生物类型跨越了抑郁症、焦虑症及相关共病障碍的诊断边界。对于临床转化而言，这些生物类型能够预测对不同药物和行为干预的反应。

We believe that this is the first identification of brain-derived biotypes that uses standardized personalized quantification of both task-free and task-evoked brain circuit dysfunctions and assesses response of the biotypes across different types of treatment. Rather than pursuing a fully data-driven approach, we integrated an unsupervised clustering analysis with a theoretical framework suitable for interpretability (Supplementary Table 16). We did this to minimize the possibility of overfitting and to generate solutions suited to the prospective selection of patients by biotype for future precision psychiatry trials. In this hybrid approach, each biotype was typified by a specific circuit dysfunction relative to a healthy norm, which mapped on to a unique transdiagnostic clinical phenotype.
我们相信这是首次识别脑源生物型，使用标准化的个性化量化方法评估无任务和任务诱发的脑回路功能障碍，并评估生物型在不同治疗类型中的反应。我们并没有追求完全数据驱动的方法，而是将无监督聚类分析与适合可解释性的理论框架相结合（补充表 16）。这样做是为了最小化过拟合的可能性，并生成适合未来精准精神病学试验中按生物型选择患者的解决方案。在这种混合方法中，每个生物型通过相对于健康标准的特定回路功能障碍进行表征，这与独特的跨诊断临床表型相对应。

Although our identification of six biotypes is one of many possible solutions to disentangling heterogeneity, these biotypes indicate that there may be multiple neural pathways that result in the clinical manifestation of depression and anxiety. By combining imaging data with clinical symptoms and behavior, we delineated clinical patterns that are consistent with the putative function of the circuits underlying each biotype. Importantly, although some biotypes were characterized exclusively by alterations in task-free intrinsic connectivity, others were characterized by alterations in task-evoked changes in activity and connectivity.
尽管我们识别出的六种生物类型是解开异质性众多可能解决方案之一，但这些生物类型表明可能存在多条神经通路导致抑郁和焦虑的临床表现。通过将成像数据与临床症状和行为相结合，我们描绘了与每种生物类型所依据的回路假定功能一致的临床模式。重要的是，尽管一些生物类型仅通过任务无关的内在连接变化来表征，但其他生物类型则通过任务诱发的活动和连接变化来表征。

In the task-free state, D_C+S_C+A_C+ was distinguished by hyperconnectivity of the default mode circuit, coupled with hyperconnectivity of both salience and attention circuits, correlating clinically with slowed emotional and attentional responses, replicated in split-half analyses. Although previous studies have reported circuit alterations in each of these circuits in depression and anxiety, our findings indicate that the D_C+S_C+A_C+ biotype exhibits a combination of these alterations. In line with our theoretical taxonomy, the A_C+ biotype demonstrated hypoconnectivity rather than hyperconnectivity within the frontoparietal attention circuit. This pattern corresponded to a clinical profile of lapses in concentration and impulsivity, replicated in split-half analyses.
在无任务状态下，DC+SC+AC+通过默认模式回路的超连接性与显著性和注意力回路的超连接性相结合而被区分，临床上与情感和注意反应的减缓相关，并在分半分析中得到了重复。尽管之前的研究报告了抑郁和焦虑中这些回路的变化，但我们的发现表明，DC+SC+AC+生物型表现出这些变化的组合。与我们的理论分类相一致，AC+生物型在额顶注意力回路中表现出低连接性而非超连接性。这一模式对应于注意力缺失和冲动性的临床特征，并在分半分析中得到了重复。

Under task conditions, the NS_A+P_A+ biotype displayed heightened activation within subcortical and cortical brain regions associated with processing both sad and positive emotions. Clinically, this biotype also exhibited prominent anhedonia. This profile corresponds with previous findings of heightened activity in the medial prefrontal cortex in response to happy faces, which has been linked to levels of anhedonia^23,24 and is consistent with our theoretical taxonomy. Increased activation of the amygdala is a common observation in depression in response to negative emotion^25,26. Notably, biotype NS_A+P_A+ exhibits concurrent hyperactivation of the ventral striatum, which may indicate a negative bias alongside anhedonia¹⁷.
在任务条件下，NSA+PA+生物型在与处理悲伤和积极情绪相关的皮层和皮层下脑区显示出增强的激活。临床上，这种生物型也表现出显著的快感缺失。该特征与之前的研究结果相符，即在对快乐面孔的反应中，内侧前额叶皮层的活动增强，这与快感缺失的水平有关，并与我们的理论分类一致。杏仁核的激活增加是抑郁症中对负面情绪的常见观察。值得注意的是，NSA+PA+生物型表现出腹侧纹状体的同时过度激活，这可能表明在快感缺失的同时存在负面偏见。

Two additional biotypes displayed contrasting dysfunctions within the cognitive control circuit. Biotype NTC_C−C_A− exhibited reduced activation during a cognitive control task and decreased connectivity in processing threat consciously. These characteristics suggest impaired cognitive control which is also crucial for regulating emotions. In contrast, C_A+ showed increased activation of the cognitive control circuit. This was associated with threat-related symptoms, negative bias and poorer cognitive control, as well as working memory performance, confirmed by both split-half and leave-study-out analyses. The replication of biotype C_A+ reinforces its inclusion as an exploratory biotype in our theoretical taxonomy. Although early evidence suggested that heightened cognitive control activity might be compensatory and not necessarily linked to behavioral deficits²⁷, our findings indicate that it is associated with specific cognitive–behavioral impairments. These findings highlight the importance of including task fMRI measures in future precision psychiatry studies and the value of using multimodal approaches to achieve more precise diagnoses in depression²⁸.
两个额外的生物类型在认知控制回路中表现出对比的功能障碍。生物类型 NTCC−CA− 在认知控制任务中表现出激活减少，并且在有意识地处理威胁时连接性降低。这些特征表明认知控制受损，而认知控制对于调节情绪也至关重要。相反，CA+ 显示出认知控制回路的激活增加。这与威胁相关症状、负面偏见和较差的认知控制以及工作记忆表现相关，这一点通过分半和留出研究分析得到了证实。生物类型 CA+ 的复制强化了其作为我们理论分类中探索性生物类型的纳入。尽管早期证据表明，增强的认知控制活动可能是补偿性的，并不一定与行为缺陷相关，但我们的发现表明它与特定的认知-行为障碍相关。这些发现强调了在未来精准精神病学研究中纳入任务 fMRI 测量的重要性，以及使用多模态方法以实现更精确的抑郁症诊断的价值。

Our approach enabled us to compare the efficacy of different treatments for each biotype to advance neurobiologically informed precision psychiatry. Collecting identical imaging and clinical measures across patients and treatments enabled us to compare the response of each biotype for three antidepressants, a behavioral intervention and treatment as usual. By doing so, we found that the D_C+S_C+A_C+ biotype, characterized by hyperconnectivity of the default mode and other task-free circuits, was associated with a better response to behavioral treatment compared with the other biotypes. On the other hand, the biotype characterized by reduced attention circuit connectivity (A_C−), had a worse response to behavioral treatment. Finally, biotype C_A+, characterized by hyperactivation of the cognitive control circuit, had a better response to venlafaxine.
我们的方法使我们能够比较不同生物类型的治疗效果，以推进神经生物学知情的精准精神病学。通过在患者和治疗中收集相同的影像学和临床指标，使我们能够比较每种生物类型对三种抗抑郁药、一种行为干预和常规治疗的反应。通过这样做，我们发现 DC+SC+AC+生物类型，其特征是默认模式和其他无任务回路的超连接性，与行为治疗的反应相比其他生物类型更好。另一方面，特征为注意回路连接性降低（AC−）的生物类型，对行为治疗的反应较差。最后，特征为认知控制回路超激活的 CA+生物类型，对文拉法辛的反应更好。

We delineated and validated biotypes using a small number of theoretically motivated features. By integrating theoretically grounded, task-evoked and task-free measures, our analysis provides unique insights that are complementary to those of foundational large studies that have analyzed task-free data using whole-brain techniques^4,15. Nevertheless, as this is the first demonstration, to our knowledge, of a participant-level approach to cluster-derived biotyping using a small number of task-evoked and task-free features, our results should be interpreted with caution. Future studies are needed to investigate these biotypes in new datasets and to prospectively assign participants to treatment based on their biotypes. Also, we acknowledge that obtaining task fMRI measures can be more burdensome than collecting task-free measures only. We compared our results with results obtained using task-free data only and found that including both task and task-free data provided the best validation results, especially in beyond-chance clustering of subjects in feature space. In direct statistical comparisons of clustering performance, our combination of task and task-free regional circuit scores outperformed whole-brain connectomes, default mode network task-free connectivity and task-free regional circuit scores alone, but not connectivity of a network centered on the angular gyrus; however, the last approach did not provide generalizable symptom differences between clusters. Alternative feature sets also yielded several reproducible associations among clusters, symptoms and behavioral performance, consistent with the previous literature. This demonstrates that our approach, although potentially advantageous, does not negate the potential of other feature selection processes for depression biotyping. Future biotyping studies with both task-based and task-free data should consider comparing the performance of each.
我们使用少量理论驱动的特征 delineated 和验证了生物类型。通过整合理论基础、任务诱发和无任务测量，我们的分析提供了独特的见解，这些见解与基础大型研究的结果互为补充，这些研究使用全脑技术分析了无任务数据 4,15。然而，鉴于这是我们所知的首次展示，采用参与者级别的方法通过少量任务诱发和无任务特征进行聚类衍生的生物类型，我们的结果应谨慎解读。未来的研究需要在新的数据集中调查这些生物类型，并根据其生物类型前瞻性地将参与者分配到治疗中。此外，我们承认获取任务 fMRI 测量可能比仅收集无任务测量更具负担。我们将我们的结果与仅使用无任务数据获得的结果进行了比较，发现同时包含任务和无任务数据提供了最佳的验证结果，特别是在特征空间中超越偶然的聚类。 在对聚类性能的直接统计比较中，我们的任务和无任务区域回路评分的组合优于全脑连接组、默认模式网络无任务连接性和单独的无任务区域回路评分，但不优于以角回为中心的网络连接性；然而，最后一种方法未能提供群集之间可推广的症状差异。替代特征集也在群集、症状和行为表现之间产生了几个可重复的关联，这与之前的文献一致。这表明，尽管我们的方法可能具有优势，但并不否定其他特征选择过程在抑郁生物分型中的潜力。未来的生物分型研究应考虑比较基于任务和无任务数据的每种性能。

Some strengths of our sample are that it represents the entire spectrum of depression and anxiety severity, is almost completely unmedicated (95%) and is recruited from a variety of settings. The sample also features common comorbidities that are often exclusion criteria. However, by including such a diverse population, we potentially reduce our ability to detect additional biotypes that might be more specific to certain clinical settings. It is also possible that some biotypes reflect contributions from comorbidities, which warrants replication in larger transdiagnostic samples. Another possibility is that biotypes are at least partially driven by differences in demographics between datasets. It would not be surprising, for example, if certain age groups belonged more to biotypes characterized by specific brain and clinical dysfunctions, because psychiatric symptoms, treatment response and brain biology all vary with age. We used identical imaging measures to evaluate biotypes across multiple treatments. However, some treatment groups within a biotype were small and could be unduly influenced by comorbidities or treatment design factors; therefore, it is important that the generalizability of our findings be tested by future large treatment studies. We also acknowledge that our imaging measures use a specific set of fMRI tasks that are not widely available. Future replications of our approach will be facilitated by the fact that our tasks are relatively short and easy to implement, as demonstrated by their adoption for large clinical trials such as iSPOT-D, ENGAGE and a recent trial using TMS in treatment-resistant depression²⁹. Future studies could also evaluate whether similar clusters can be derived from different tasks that tap into similar domains and compare the results with ours. Our large sample allowed us to evaluate the generalizability of symptom and behavioral differences in split-half and leave-study-out validations. However, the number of participants of clinical trials was too small to perform such analyses for treatment response (n < 10 for 90% of comparisons; Supplementary Table 8). Future studies should apply our approach to clinical trial data to verify these findings, which should be interpreted prudently until they can be validated in new samples. Finally, the symptom differences between biotypes that we detected were mostly small, with effect sizes ranging from 0.08 to 0.90. The small size of these differences might be a reason why most comparisons did not reach statistical significance when splitting the dataset in two random halves or by study and analyzing each split independently. Small effect sizes in the association between imaging and symptom variables are common³⁰, highlighting the need for consistent measures across studies and for finer-grained clinical measures. In the present study, we show the utility of combining four studies using standardized measures. We recommend interpreting the clinical results that did not survive our validation analyses with caution, but the present study is nevertheless a foundation to further test these results.
我们样本的一些优势在于它代表了抑郁和焦虑严重程度的整个谱系，几乎完全没有用药（95%），并且来自多种环境。样本还具有常见的共病，这些共病通常是排除标准。然而，通过纳入如此多样化的人群，我们可能降低了检测可能更特定于某些临床环境的额外生物类型的能力。还有可能某些生物类型反映了共病的贡献，这需要在更大范围的跨诊断样本中进行重复研究。另一种可能性是生物类型至少部分受到数据集之间人口统计学差异的驱动。例如，如果某些年龄组更倾向于属于特定大脑和临床功能障碍特征的生物类型，这并不令人惊讶，因为精神症状、治疗反应和大脑生物学都随着年龄而变化。我们使用相同的成像测量来评估多个治疗中的生物类型。 然而，某些生物型内的治疗组规模较小，可能会受到合并症或治疗设计因素的不当影响；因此，重要的是通过未来的大型治疗研究来检验我们发现的普遍性。我们也承认，我们的成像测量使用了一组特定的 fMRI 任务，这些任务并不广泛可用。我们的方法的未来复制将得益于我们的任务相对较短且易于实施，正如它们在大型临床试验如 iSPOT-D、ENGAGE 和最近使用 TMS 治疗耐药性抑郁症的试验中被采用所示。未来的研究还可以评估是否可以从不同的任务中推导出类似的聚类，这些任务涉及相似的领域，并将结果与我们的结果进行比较。我们的大样本使我们能够评估在分半和留出研究验证中症状和行为差异的普遍性。然而，临床试验的参与者人数太少，无法进行治疗反应的此类分析（90%的比较中 n < 10；补充表 8）。 未来的研究应将我们的方法应用于临床试验数据，以验证这些发现，这些发现应谨慎解读，直到在新样本中得到验证。最后，我们检测到的生物型之间的症状差异大多较小，效应大小范围从 0.08 到 0.90。这些差异的小规模可能是导致在将数据集分成两个随机部分或按研究进行拆分并独立分析时，大多数比较未达到统计显著性的原因。影像学与症状变量之间的关联效应大小较小是常见的，强调了在研究中需要一致的测量和更细致的临床测量。在本研究中，我们展示了结合四项使用标准化测量的研究的实用性。我们建议对未通过我们验证分析的临床结果进行谨慎解读，但本研究仍然是进一步测试这些结果的基础。

In conclusion, we leveraged personalized regional dysfunction scores grounded in a theoretical taxonomy of brain dysfunction in mood and anxiety disorders to identify six biotypes in a large transdiagnostic sample of unmedicated individuals with depression and anxiety. These biotypes differed significantly in symptom profiles, performance on behavioral testing and responses to multiple treatments. Our results validate a new theory-driven method for depression biotyping as well as a promising approach to advancing precision clinical care in psychiatry.
总之，我们利用基于情绪和焦虑障碍的脑功能障碍理论分类法的个性化区域功能障碍评分，在一大样本中识别出六种生物类型，这些样本来自未服药的抑郁和焦虑个体。这些生物类型在症状特征、行为测试表现和对多种治疗的反应上存在显著差异。我们的结果验证了一种基于理论的新抑郁生物类型方法，以及一种推动精神病学精准临床护理的有前景的方法。

Samples 样本

Data were obtained from four studies: International Study to Predict Optimized Treatment in Depression (iSPOT-D¹⁸, https://clinicaltrials.gov/ct2/show/NCT00693849), Research on Anxiety and Depression study (RAD³⁸), Human Connectome Project for Disordered Emotional States (HCP-DES³⁹) and Engaging self-regulation targets to understand the mechanisms of behavior change and improve mood and weight outcome (ENGAGE⁴⁰, https://clinicaltrials.gov/ct2/show/NCT02246413). Clinical participants from these studies (n = 801) represented the full spectrum of severity of depression and anxiety disorders (see Table 1 and Supplementary Table 1 for details). Healthy controls (iSPOT-D, n = 67; HCP-DES, n = 70) were used as a reference group for building regional circuit scores from the imaging data (see below). Of the 801 clinical participants, 250 completed randomized controlled trials of either antidepressant pharmacotherapy for major depressive disorder (n = 164)¹⁸ or behavioral intervention for clinically substantial depressive symptoms and obesity (n = 86)⁴⁰ (see Supplementary Table 2 for more details).
数据来自四项研究：国际抑郁优化治疗预测研究（iSPOT-D18，https://clinicaltrials.gov/ct2/show/NCT00693849），焦虑与抑郁研究（RAD38），人类连接组项目用于情绪障碍状态（HCP-DES39）以及参与自我调节目标以理解行为变化机制并改善情绪和体重结果（ENGAGE40，https://clinicaltrials.gov/ct2/show/NCT02246413）。这些研究的临床参与者（n = 801）代表了抑郁和焦虑障碍严重程度的全谱（详见表 1 和补充表 1）。健康对照组（iSPOT-D，n = 67；HCP-DES，n = 70）被用作从影像数据构建区域回路评分的参考组（见下文）。在 801 名临床参与者中，250 人完成了针对重度抑郁症的抗抑郁药物随机对照试验（n = 164）18 或针对临床显著抑郁症状和肥胖的行为干预（n = 86）40（详见补充表 2）。

All participants provided written informed consent. Procedures were approved by the Stanford University Institutional Review Board (IRB, protocol nos. 27937 and 41837) or the Western Sydney Area Health Service Human Research Ethics Committee.
所有参与者提供了书面知情同意。程序已获得斯坦福大学机构审查委员会（IRB，协议编号 27937 和 41837）或西悉尼地区健康服务人类研究伦理委员会的批准。

MRI acquisition and preprocessing
MRI 获取与预处理

Details of MRI sequences, fMRI tasks, MRI data quantification and quality control are given in Supplementary Methods.
MRI 序列、fMRI 任务、MRI 数据量化和质量控制的详细信息见补充方法。

Acquisition 获取

Participants underwent the Stanford Et Cere Image Processing System protocol, which probes six brain circuits: default mode circuit, salience circuit, attention circuit, negative affect circuit, positive affect circuit and cognitive control circuit^17,20. The Facial Expressions of Emotion Tasks probed the positive and negative affect circuits and a Go–NoGo task probed the cognitive control circuit. We derived measures of task-free function of the default mode, attention and salience circuits from the task data^41,42. Task-free measures were independent of those obtained from the task conditions (Supplementary Fig. 14).
参与者接受了斯坦福 Et Cere 图像处理系统协议，该协议探测六个脑回路：默认模式回路、显著性回路、注意力回路、负性情感回路、正性情感回路和认知控制回路 17,20。情感面部表情任务探测了正性和负性情感回路，而 Go–NoGo 任务探测了认知控制回路。我们从任务数据中得出了默认模式、注意力和显著性回路的无任务功能测量 41,42。无任务测量与从任务条件获得的测量是独立的（补充图 14）。

Preprocessing 预处理

The MRI data were preprocessed using fMRIprep⁴³. We discarded scans if they contained incidental findings, major artifacts or signal dropouts or had >25% of volumes containing significant frame-wise displacement. An experienced rater (L.T.) also visually checked each scan, leading to the exclusion of 32 participants. Scans removed owing to excessive motion were: Go–NoGo task = 38, Conscious Facial Expressions of Emotion Task = 92, Non-conscious Facial Expressions of Emotion Task = 76 and task free = 51 (see Supplementary Table 17 for the number of scans passing criteria).
MRI 数据使用 fMRIprep 进行了预处理。我们丢弃了包含偶然发现、重大伪影或信号丢失的扫描，或者有超过 25%的体积包含显著的帧位移。一个经验丰富的评估者（L.T.）也对每个扫描进行了视觉检查，导致 32 名参与者被排除。由于过度运动而移除的扫描包括：Go–NoGo 任务 = 38，意识面部情感表达任务 = 92，无意识面部情感表达任务 = 76，以及无任务状态 = 51（有关通过标准的扫描数量，请参见补充表 17）。

Derivation of regional circuit scores
区域回路评分的推导

A summary of how regional circuit scores were obtained is given in the following sections (Fig. 1; see Supplementary Methods for details). We previously demonstrated that this system produces valid and clinically useful individual circuit clinical scores²⁰.
以下部分提供了区域回路评分的获取总结（图 1；详细信息请参见补充方法）。我们之前证明该系统产生有效且临床有用的个体回路临床评分 20。

Extraction of imaging features of interest
提取感兴趣的成像特征

The regions of interest within six circuits of interest were defined from the meta-analytic platform Neurosynth⁴⁴ (see Supplementary Table 18 for search terms and coordinates) and refined by removing regions that did not pass quality control or psychometric criteria. Of the remaining regions, we only retained 29 regions implicated in our theoretical synthesis of dysfunctions in depression and anxiety^20,38. From these regions, we derived 41 features of activation, task-based and task-free connectivity for subsequent analyses²⁰ (see Supplementary Table 18 and Supplementary Tables S5A and S5B in ref. ²⁰ for details on the regions of interest and features). Our focus on regions defined from theory, meta-analyses and anatomy can lead to reliable and reproducible imaging measures. For example, activations within anatomically defined regions of interest have acceptable-to-high within-participant reliability⁴⁵, as does connectivity within established brain networks⁴⁶.
在六个感兴趣的回路中，感兴趣的区域是从元分析平台 Neurosynth 定义的（参见补充表 18 以获取搜索词和坐标），并通过去除未通过质量控制或心理测量标准的区域进行了精炼。在剩余的区域中，我们仅保留了 29 个与我们关于抑郁和焦虑功能障碍的理论综合相关的区域。从这些区域中，我们提取了 41 个激活特征、基于任务和无任务的连接性，以便进行后续分析（参见补充表 18 和参考文献 20 中的补充表 S5A 和 S5B 以获取有关感兴趣区域和特征的详细信息）。我们对从理论、元分析和解剖学定义的区域的关注可以导致可靠和可重复的成像测量。例如，在解剖学定义的感兴趣区域内的激活具有可接受到高的参与者内可靠性，已建立的大脑网络内的连接性也是如此。

All following analyses used RStudio 2022.07.2, R v.4.1.3. Code for these analyses and the regions of interest to derive our imaging features are at https://github.com/leotozzi88/cluster_study_2023.
所有以下分析均使用 RStudio 2022.07.2，R v.4.1.3。用于这些分析的代码以及我们提取成像特征的感兴趣区域可在 https://github.com/leotozzi88/cluster_study_2023 获取。

Imputation of missing values
缺失值的插补

As a result of missing scans and quality control, some regional circuit scores could not be computed for some participants: 7.57% for the default, salience and attention scores, 9.38% for the negative affect sad scores, 9.38% for the negative affect threat conscious scores, 6.72% for the negative affect threat nonconscious scores, 4.05% for the cognitive control scores and 9.38% for the positive affect scores. We imputed these values separately for each scanner by using multiple imputation by chained equations with random forests (R package miceRanger), using one iteration of a predictive mean matching model with the imaging features as the input.
由于缺失扫描和质量控制，一些参与者的区域回路评分无法计算：默认、显著性和注意力评分为 7.57%，负性情感悲伤评分为 9.38%，负性情感威胁意识评分为 9.38%，负性情感威胁非意识评分为 6.72%，认知控制评分为 4.05%，正性情感评分为 9.38%。我们通过使用随机森林的链式方程多重插补（R 包 miceRanger），对每个扫描仪分别插补这些值，使用成像特征作为输入的一次预测均值匹配模型迭代。

Correction for scanner effects
扫描仪效应的校正

We removed the potential confounding effect of between-scanner variability using ComBat^47,48,49, an established method that uses an empirical Bayesian framework to remove batch effects.
我们使用 ComBat47,48,49 去除了扫描仪之间变异的潜在混杂效应，这是一种使用经验贝叶斯框架去除批次效应的成熟方法。

Referencing to a healthy norm
参考健康标准

All imaging features of the clinical participants were expressed in s.d. units relative to the mean and s.d. of healthy controls. These values are henceforth referred to as ‘regional circuit scores’ and represent the amount of dysfunction of each component of each circuit. Subsequent analyses were conducted on the regional circuit scores of the clinical participants only.
所有临床参与者的成像特征均以相对于健康对照组的均值和标准差的标准差单位表示。这些值此后称为“区域回路评分”，代表每个回路中每个组件的功能障碍程度。后续分析仅针对临床参与者的区域回路评分进行。

Symptom measures 症状测量

We used self-reported questionnaires to operationalize: ruminative worry (Penn State Worry Questionnaire—Abbreviated total⁵⁰); ruminative brooding (Ruminative Response Scale total⁵¹); anxious arousal (Mood and Anxiety Questionnaire general distress subscale⁵²); negative bias (Depression Anxiety and Stress Scale (DASS) depression subscale); threat dysregulation (DASS anxiety subscale); anhedonia (Snaith–Hamilton Pleasure Scale total⁵³); cognitive dyscontrol (Barratt Impulsiveness Scale attentional impulsiveness subscale⁵⁴); tension (DASS stress subscale); insomnia (Quick Inventory of Depressive Symptomatology—Self-Report Revised (QIDS-SR) sum of items 1–3 (ref. ⁵⁵)); and suicidality (QIDS-SR item 12). In iSPOT-D, we used the Hamilton Depression Rating Scale (HDRS) total score as a measure of depression severity⁵⁶ and, in ENGAGE, we used the Symptom Checklist 20 Depression Scale (SCL-20)⁵⁷. See Supplementary Table 19 for the participants in each sample available for each measure.
我们使用自我报告问卷来操作化：反刍性担忧（宾州州立大学担忧问卷—缩写总分 50）；反刍性沉思（反刍反应量表总分 51）；焦虑唤起（情绪与焦虑问卷一般困扰子量表 52）；负面偏见（抑郁、焦虑和压力量表（DASS）抑郁子量表）；威胁失调（DASS 焦虑子量表）；快感缺失（斯奈斯-汉密尔顿愉悦量表总分 53）；认知失控（巴拉特冲动性量表注意力冲动性子量表 54）；紧张（DASS 压力子量表）；失眠（抑郁症状快速评估量表—自我报告修订版（QIDS-SR）第 1-3 项总和（参考文献 55））；以及自杀意念（QIDS-SR 第 12 项）。在 iSPOT-D 中，我们使用汉密尔顿抑郁评分量表（HDRS）总分作为抑郁严重程度的测量 56，而在 ENGAGE 中，我们使用症状检查表 20 抑郁量表（SCL-20）57。有关每个测量中每个样本参与者的信息，请参见补充表 19。

Clinical diagnoses 临床诊断

DSM-IV-TR (RAD), DSM-5 (HCP-DES) or DSM-IV (iSPOT-D) criteria for major depressive disorder, anxiety disorder, post-traumatic stress disorder or obsessive–compulsive disorder were ascertained by a psychiatrist, general practitioner or researcher using the structured MINI³⁴. In ENGAGE, patients were considered eligible if they scored ≥10 on the PHQ-9, a threshold with 88% specificity for major depressive disorder³⁵, and had a qualifying body mass index (BMI). Comorbidities were ascertained from electronic health records.
DSM-IV-TR（RAD）、DSM-5（HCP-DES）或 DSM-IV（iSPOT-D）标准用于重度抑郁症、焦虑症、创伤后应激障碍或强迫症的诊断由精神科医生、全科医生或研究人员使用结构化的 MINI34 进行确认。在 ENGAGE 研究中，如果患者在 PHQ-9 上得分≥10，且具有合格的身体质量指数（BMI），则被视为符合资格，该阈值对重度抑郁症的特异性为 88%35。共病情况通过电子健康记录进行确认。

Behavioral performance measures
行为表现测量

Cognitive performance was assessed using WebNeuro^37,58,59. We focused on the tests for which our regional circuit scores have been shown to predict performance²⁰: sustained attention (omission errors, commission errors and reaction times in a continuous performance test); executive function (errors and completion time of a maze test); cognitive control (commission errors and reaction times in a Go–NoGo test); explicit emotion identification (reaction time to identify happy, sad, fearful and angry faces); and implicit priming bias by emotion (difference in reaction time in a face identification task when primed implicitly by happy, sad, fearful and angry faces compared with neutral faces). For analyses, we used the test performance referenced to an age-matched norm generated by WebNeuro (z-scores). See Supplementary Table 19 for the number of participants in each sample available for each measure.
认知表现使用 WebNeuro37,58,59 进行评估。我们专注于那些我们的区域回路评分已被证明可以预测表现的测试 20：持续注意力（遗漏错误、误报错误和持续表现测试中的反应时间）；执行功能（迷宫测试中的错误和完成时间）；认知控制（Go–NoGo 测试中的误报错误和反应时间）；显性情绪识别（识别快乐、悲伤、恐惧和愤怒面孔的反应时间）；以及情绪的隐性启动偏差（在面孔识别任务中，当隐性启动快乐、悲伤、恐惧和愤怒面孔时与中性面孔相比的反应时间差异）。在分析中，我们使用了参考 WebNeuro 生成的与年龄匹配的标准（z 分数）的测试表现。有关每个测量可用的每个样本参与者数量，请参见补充表 19。

Treatment 治疗

In iSPOT-D, participants were randomized to one of three treatments: escitalopram (selective serotonin reuptake inhibitor (SSRI)), sertraline (SSRI) or venlafaxine XR (selective serotonin–norepinephrine reuptake inhibitor (SNRI))¹⁸. In ENGAGE, participants were randomized to either a behavioral intervention combining problem-solving, behavioral activation and weight loss (Integrated Coaching for Better Mood and Weight, I-CARE) or usual care (U-CARE)^19,40. No treatment was administered in HCP-DES and RAD, so these studies were not considered in the treatment analyses.
在 iSPOT-D 中，参与者被随机分配到三种治疗之一：艾司西酞普兰（选择性 5-羟色胺再摄取抑制剂（SSRI））、舍曲林（SSRI）或文拉法辛 XR（选择性 5-羟色胺-去甲肾上腺素再摄取抑制剂（SNRI））18。在 ENGAGE 中，参与者被随机分配到结合问题解决、行为激活和减重的行为干预（改善情绪和体重的综合辅导，I-CARE）或常规护理（U-CARE）19,40。在 HCP-DES 和 RAD 中没有进行治疗，因此这些研究未被纳入治疗分析。

Identification of depression biotypes
抑郁症生物型的识别

To identify biotypes within our clinical participants, we used hierarchical clustering of their 41 regional circuit scores. We selected the optimal number of clusters using six convergent sources of evidence: the elbow method; two procedures proposed by Dinga et al.¹⁴ to evaluate biotypes of depression (simulation-based significance testing of the silhouette index and stability using crossvalidation); permutation-based significance testing of the silhouette index; split-half reliability of the cluster profiles; and the match of the solution to a theoretical framework¹⁷ (Fig. 2).
为了识别我们临床参与者中的生物类型，我们对他们的 41 个区域回路评分进行了层次聚类。我们使用六个收敛证据选择了最佳聚类数量：肘部法；Dinga 等人提出的两种程序 14 来评估抑郁症的生物类型（基于模拟的轮廓指数显著性测试和使用交叉验证的稳定性）；基于置换的轮廓指数显著性测试；聚类轮廓的分半信度；以及解决方案与理论框架的匹配 17（图 2）。

Hierarchical clustering 层次聚类

For each pair of clinical participants, we first computed the correlation coefficient of their 41 imaging-derived regional circuit scores (Fig. 1). Then, we computed the dissimilarity between each pair of clinical participants as 1 − this correlation (see ref. ⁶⁰ for a similar approach). We used the between-individual dissimilarity matrix as input to hierarchical clustering using the average as agglomeration method.
对于每对临床参与者，我们首先计算了他们 41 个影像衍生的区域回路评分的相关系数（图 1）。然后，我们将每对临床参与者之间的差异计算为 1 减去该相关性（参见参考文献 60 以获取类似方法）。我们使用个体之间的差异矩阵作为输入，采用平均法作为聚类方法进行层次聚类。

Elbow method 肘部法则

The first source of evidence that we used to choose the optimal number of clusters was the elbow method, based on a plot showing the within-cluster sum of distances between participants for solutions between 2 and 15 clusters (Fig. 2a).
我们选择最佳聚类数的第一个证据来源是肘部法则，基于一个图表，该图表显示了参与者之间的聚类内距离总和，聚类数在 2 到 15 之间（图 2a）。

Simulation-based significance testing of silhouette
基于模拟的轮廓显著性测试

We tested the probability of our observed average silhouette index occurring under the null hypothesis of no clusters (that is, of the data coming from a multinormal distribution)¹⁴. For clusters between 2 and 15, we conducted 10,000 simulation runs, in which we drew 801 participants from a multinormal distribution that had the same mean and covariance for each regional circuit score as our data. These simulated participants were then used as input in hierarchical clustering, as described above, and the average silhouette index across participants was calculated. Thus, we obtained null distributions for these average silhouette indices. Finally, we calculated the proportion of average silhouette indices generated under the null that were greater than the one we obtained from our data (P value). We considered statistically significant solutions with numbers of clusters for which P < 0.05 (Fig. 2b).
我们测试了在没有聚类的零假设下（即数据来自多元正态分布）观察到的平均轮廓指数发生的概率 14。对于 2 到 15 个聚类，我们进行了 10,000 次模拟运行，在每次运行中，我们从一个多元正态分布中抽取 801 名参与者，该分布的每个区域回路评分的均值和协方差与我们的数据相同。这些模拟参与者随后被用作层次聚类的输入，如上所述，并计算了参与者的平均轮廓指数。因此，我们获得了这些平均轮廓指数的零分布。最后，我们计算了在零假设下生成的平均轮廓指数中大于我们从数据中获得的那个的比例（P 值）。我们认为 P < 0.05 的聚类数量的统计显著性解决方案是显著的（图 2b）。

Permutation-based significance testing of silhouette
基于置换的轮廓显著性检验

For each number of clusters between 2 and 15, we shuffled each brain circuit score across subjects 10,000×, then repeated the hierarchical clustering as described above and calculated the average silhouette index. Thus, we obtained null distributions for these average silhouette indices. Finally, we calculated the proportion of average silhouette indices generated under the null that were greater than the one we obtained from our data (P value). We considered statistically significant solutions with numbers of clusters for which P < 0.05 (Fig. 2c).
对于 2 到 15 个聚类的每个数量，我们在受试者之间随机打乱每个脑回路评分 10,000 次，然后重复上述层次聚类并计算平均轮廓指数。因此，我们获得了这些平均轮廓指数的零分布。最后，我们计算了在零假设下生成的平均轮廓指数中大于我们从数据中获得的那个的比例（P 值）。我们认为 P < 0.05 的聚类数量是统计显著的解（图 2c）。

Assessment of cluster stability using crossvalidation
使用交叉验证评估聚类稳定性

To evaluate whether the clustering was stable under small perturbations to the data¹⁴, we repeated the clustering procedure 801×, each time with one participant left out. For each run and for each solution between 2 and 15 clusters, we calculated the similarity of the new cluster assignments to the original ones using the ARI (Fig. 2d). We then repeated this procedure while holding out 20% of the sample instead of one participant (Fig. 2e).
为了评估在对数据进行小扰动时聚类是否稳定，我们重复了 801 次聚类过程，每次排除一名参与者。对于每次运行以及 2 到 15 个聚类之间的每个解决方案，我们使用 ARI 计算新聚类分配与原始分配的相似性（图 2d）。然后，我们在保留 20%样本而不是一名参与者的情况下重复了这个过程（图 2e）。

Matching of clusters to a theoretical framework
将簇与理论框架匹配

We identified the primary circuit dysfunction of each cluster by averaging the values of regional circuit scores by circuit and modality (task-based activity, task-based connectivity, task-free connectivity) and identifying the measures that showed a >0.5 s.d. absolute mean difference compared with the healthy norm. We then compared the profile of circuit dysfunction of each cluster with those hypothesized in a theoretical framework of circuit dysfunction in depression and anxiety^16,17.
我们通过按回路和模态（任务相关活动、任务相关连接、无任务连接）平均区域回路评分的值，识别了每个簇的主要回路功能障碍，并确定了与健康标准相比显示>0.5 标准差绝对均值差异的指标。然后，我们将每个簇的回路功能障碍特征与抑郁和焦虑的回路功能障碍理论框架中假设的特征进行了比较 16,17。

Split-half replication of cluster profiles
分半复制聚类轮廓

First, we split our dataset into two random samples of equal size. Then, we ran our clustering procedure on the first half-split. Then, we assigned each participant in the second split to one of the clusters obtained in the first half-split. To do so, we computed the mean circuit scores across all participants belonging to each cluster in the first half-split. Then, we calculated Pearson’s correlation coefficient between each participant’s brain circuit scores and these cluster-averaged scores. Each out-of-sample participant was assigned to the cluster for which this correlation was highest. Finally, we identified the primary circuit dysfunctions of each cluster in each split as described above (>0.5 s.d. absolute mean difference compared with the healthy reference data) and examined whether they replicated the circuit profiles found in the whole sample visually and by computing Pearson’s correlation coefficient of the mean profile dysfunction profile of each cluster between splits (Fig. 2f).
首先，我们将数据集分成两个相等大小的随机样本。然后，我们在第一半部分上运行了聚类程序。接着，我们将第二部分的每个参与者分配到第一半部分获得的一个聚类中。为此，我们计算了第一半部分中每个聚类所有参与者的平均回路评分。然后，我们计算了每个参与者的脑回路评分与这些聚类平均评分之间的皮尔逊相关系数。每个样本外参与者被分配到相关性最高的聚类。最后，我们识别了每个聚类在每个分割中的主要回路功能障碍，如上所述（与健康参考数据相比，绝对均值差异>0.5 标准差），并检查它们是否在视觉上和通过计算每个聚类在分割之间的均值功能障碍轮廓的皮尔逊相关系数上复制了在整个样本中发现的回路特征（图 2f）。

Clinical characterization of biotypes
生物型的临床特征描述

We characterized our final clustering solution by using external clinical measures independent of cluster inputs: symptoms, clinical diagnoses, performance on behavioral tests and treatment response. Importantly, we also replicated our findings in split-half and leave-study-out analyses (Fig. 2g–l).
我们通过使用与聚类输入无关的外部临床指标来表征我们的最终聚类解决方案：症状、临床诊断、行为测试表现和治疗反应。重要的是，我们还在分半和留出研究分析中复制了我们的发现（图 2g–l）。

Comparison of symptoms between biotypes
生物型之间症状的比较

For each symptom, we compared the median severity of participants in each biotype to the median severity of participants who were not in the biotype using Wilcoxon’s tests. As insomnia and suicidality were assessed using only three and one QIDS-SR items, respectively, we used a χ² test to compare the fraction of participants in the biotype who endorsed any of the items (total value >0) compared with participants who were not in the biotype. For Wilcoxon’s tests, we calculated the effect size r as the z statistic divided by the square root of the sample size and we considered significant tests for which P < 0.05 (Fig. 2h,j).
对于每个症状，我们使用 Wilcoxon 检验比较了每种生物型参与者的中位严重程度与不在该生物型的参与者的中位严重程度。由于失眠和自杀意念仅使用了三个和一个 QIDS-SR 项目进行评估，因此我们使用χ2 检验比较了生物型参与者中支持任何项目（总值>0）的参与者比例与不在该生物型的参与者的比例。对于 Wilcoxon 检验，我们计算了效应大小 r，作为 z 统计量除以样本大小的平方根，并且我们认为 P < 0.05 的检验是显著的（图 2h,j）。

Comparison of behavioral performance between biotypes
生物型之间行为表现的比较

For each of our behavioral performance measures, we compared the median performance of participants in each biotype with the median performance of participants who were not in the biotype using Wilcoxon’s tests. We calculated the effect size r as the z statistic divided by the square root of the sample size and we considered significant tests for which P < 0.05 (Fig. 2i,k).
对于我们每个行为表现指标，我们使用 Wilcoxon 检验比较了每种生物类型参与者的中位表现与不在该生物类型中的参与者的中位表现。我们计算了效应大小 r，作为 z 统计量除以样本大小的平方根，并且我们认为 P < 0.05 的检验具有显著性（图 2i,k）。

Comparison of treatment response between biotypes
生物型之间治疗反应的比较

To obtain a comparable measure of symptom severity across our clinical trial datasets, we first scaled the measures of total HDRS scores (collected in iSPOT-D) and SCL-20 scores (collected in ENGAGE) between 0 and 1 based on the minimum and maximum values of each scale. We defined response as a decrease of at least 50% of symptom severity from baseline to follow-up and remission as follow-up HDRS ≤ 7 or SCL-20 ≤ 0.5. Then, for each treatment modality and each biotype, the severity of symptoms after treatment of participants in the biotype was compared with the median symptom severity of clinical participants not in the biotype using Wilcoxon’s tests. For these tests, we excluded biotypes in which only five or fewer participants received a treatment. We calculated the effect size r as the z statistic divided by the square root of the sample size and considered significant tests for which P < 0.05. (Fig. 2l).
为了获得我们临床试验数据集中症状严重程度的可比测量，我们首先根据每个量表的最小值和最大值将总 HDRS 评分（在 iSPOT-D 中收集）和 SCL-20 评分（在 ENGAGE 中收集）缩放到 0 到 1 之间。我们将反应定义为从基线到随访症状严重程度至少降低 50%，将缓解定义为随访 HDRS ≤ 7 或 SCL-20 ≤ 0.5。然后，对于每种治疗方式和每种生物类型，比较生物类型参与者治疗后症状的严重程度与不在该生物类型中的临床参与者的中位症状严重程度，使用 Wilcoxon 检验。对于这些检验，我们排除了只有五名或更少参与者接受治疗的生物类型。我们计算效应大小 r 为 z 统计量除以样本大小的平方根，并将 P < 0.05 的检验视为显著。（图 2l）。

Split-half replication of clinical associations
临床关联的分半复制

We replicated the significant comparisons of behavior and symptoms between biotypes found in the complete sample by splitting the sample into two random halves, repeating the clustering procedure on the first half and then assigning participants in the second half to the clusters obtained in the first half, as described above. We then conducted Wilcoxon’s tests as described above in each split and considered a result replicable if it was significant both in the original sample and in each of the split-half samples (for the second split, we conducted a confirmatory one-sided test).
我们通过将样本分成两个随机部分，重复对第一部分进行聚类程序，然后将第二部分的参与者分配到第一部分获得的聚类中，从而复制了在完整样本中发现的生物类型之间行为和症状的显著比较。然后，我们在每个分割中进行了上述的 Wilcoxon 检验，并认为如果结果在原始样本和每个分割样本中均显著，则该结果是可复制的（对于第二个分割，我们进行了确认性单侧检验）。

Leave-study-out replication of clinical associations
留出研究的临床关联复制

For each of the four studies included in our dataset, we replicated the significant comparisons of behavior and symptoms between biotypes by splitting the sample into two subsets: one containing the participants who were not from that study and one containing participants from that study. Then, we repeated the clustering procedure on the first split and assigned participants in the second subset to the clusters obtained in the first split, as described above. We then conducted Wilcoxon’s tests as described above and considered a result replicable if it was significant in each of the two splits when holding out at least one study (for the second split, we conducted a confirmatory one-sided test).
对于我们数据集中包含的四项研究，我们通过将样本分成两个子集来复制生物类型之间行为和症状的显著比较：一个子集包含不来自该研究的参与者，另一个子集包含来自该研究的参与者。然后，我们在第一个分割上重复聚类程序，并将第二个子集中的参与者分配到第一个分割中获得的聚类，如上所述。随后，我们进行了如上所述的 Wilcoxon 检验，并认为如果在保留至少一项研究的情况下，在两个分割中均显著，则结果是可复制的（对于第二个分割，我们进行了确认性单侧检验）。

Comparison of diagnoses between biotypes
生物类型之间的诊断比较

To evaluate whether the clusters reflected traditional diagnostic categories, we used χ² tests to compare the proportion of clinical participants in each biotype who met criteria for major depressive disorder, generalized anxiety disorder, obsessive–compulsive disorder, post-traumatic stress disorder, panic disorder or social phobia.
为了评估这些簇是否反映了传统的诊断类别，我们使用χ2 检验比较每种生物类型中符合重度抑郁症、广泛性焦虑症、强迫症、创伤后应激障碍、惊恐障碍或社交恐惧症标准的临床参与者的比例。

Comparison of covariates of no interest between biotypes
不同生物类型之间无关协变量的比较

To verify that biotypes were not driven by scanner effects, we used χ² tests to evaluate whether the proportion of participants in each cluster was different across scanners. Similarly, we used χ² tests to examine the effects of gender and dataset and a one-way analysis of variance (ANOVA) to test whether different biotypes had different age distributions.
为了验证生物类型是否受到扫描仪效应的影响，我们使用χ2 检验评估每个簇中参与者的比例在不同扫描仪之间是否存在差异。同样，我们使用χ2 检验来检查性别和数据集的影响，并使用单因素方差分析（ANOVA）测试不同生物类型是否具有不同的年龄分布。

Comparison of brain circuit scores to other biotyping inputs
与其他生物类型输入的脑回路评分比较

We selected three alternative feature sets, each used in a recent paper identifying biotypes of depression using resting state fMRI (to our knowledge, no prior publication has used task fMRI): whole-brain functional connectivity from the Power atlas⁴; functional connectivity in the default mode network⁵; and a functional connectivity of the angular gyrus⁷. We evaluated these features using the same criteria that we used for our own: (1) solution outperforms null hypothesis of no clusters (simulated data); (2) solution outperforms null hypothesis of no clusters (permuted data); (3) ARI (leave-one-out mean); (4) ARI (leave-20%-out mean); (5) generalizable cluster profiles across random split-half; (6) generalizable symptom differences across random split-half; (7) generalizable behavior differences across random split-half; (8) generalizable symptom differences across leave-study-out; (9) generalizable behavior differences across leave-study-out; and (10) biotypes differ in treatment response. For each of the alternative sets of features, we evaluated the number of clusters reported in the original paper and six clusters (the number that we chose in our analysis). We also conducted two statistical tests comparing clustering performance using our features with other features. First, a resampling test: we sampled 80% of participants, used each set of features to cluster their data and computed the corresponding average silhouette index over 10,000 iterations. For each set of alternative features, we considered as P_resample the fraction of samplings in which the silhouette index was higher than the one obtained with our features. Then a permutation test: after clustering each of the imaging feature sets, we randomly permuted the cluster assignments 10,000× and computed a silhouette score for each. This provided us with null distributions of the silhouette index for each feature set. We then calculated the difference between the null distribution of the silhouette index obtained using our features and each of the null distributions obtained from alternative features. We considered as P_permute the proportion of permutations in which the difference between the two null distributions was greater than that between the silhouette indices of the real solutions. We considered our features to provide a better clustering when P_permute < 0.05 and P_resample < 0.05.
我们选择了三组替代特征集，每组特征集均用于最近一篇利用静息态 fMRI 识别抑郁症生物类型的论文（据我们所知，之前没有任何出版物使用任务 fMRI）：来自 Power atlas 的全脑功能连接 4；默认模式网络中的功能连接 5；以及角回的功能连接 7。我们使用与我们自己相同的标准评估这些特征：（1）解决方案优于无聚类的零假设（模拟数据）；（2）解决方案优于无聚类的零假设（置换数据）；（3）ARI（留一法均值）；（4）ARI（留 20%法均值）；（5）随机分半的可推广聚类特征；（6）随机分半的可推广症状差异；（7）随机分半的可推广行为差异；（8）留研究法的可推广症状差异；（9）留研究法的可推广行为差异；（10）生物类型在治疗反应上存在差异。对于每组替代特征集，我们评估了原始论文中报告的聚类数量和我们分析中选择的六个聚类（我们选择的数量）。 我们还进行了两项统计测试，比较使用我们的特征与其他特征的聚类性能。首先是重抽样测试：我们抽取了 80%的参与者，使用每组特征对他们的数据进行聚类，并计算了 10,000 次迭代的平均轮廓指数。对于每组替代特征，我们将轮廓指数高于使用我们特征获得的轮廓指数的抽样比例视为预抽样。然后是置换测试：在对每组成像特征集进行聚类后，我们随机置换聚类分配 10,000 次，并为每次计算轮廓得分。这为我们提供了每组特征集的轮廓指数的零分布。然后，我们计算了使用我们的特征获得的轮廓指数的零分布与每组替代特征获得的零分布之间的差异。我们将置换中两个零分布之间的差异大于真实解的轮廓指数之间的差异的比例视为 Ppermute。当 Ppermute < 0 时，我们认为我们的特征提供了更好的聚类。05 和预采样< 0.05。

Finally, we compared our original results to results obtained using only our task-free brain circuit scores, choosing as the number of clusters six (the number we chose in our analysis using all features) and two (the number of clusters with task-free dysfunction identified in our analyses).
最后，我们将原始结果与仅使用无任务脑回路评分获得的结果进行了比较，选择的聚类数为六（我们在使用所有特征的分析中选择的数量）和二（在我们的分析中识别出的无任务功能障碍的聚类数）。

Reporting summary 报告摘要

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.
有关研究设计的更多信息，请参阅与本文相关的《自然》期刊报告摘要。