Resting-state connectivity biomarkers define neurophysiological
subtypes of depression
Andrew T Drysdale
1,2,3
, Logan Grosenick
4,5
, Jonathan Downar
6
, Katharine Dunlop
6
,
Farrokh Mansouri
6
, Yue Meng
1
, Robert N Fetcho
1
, Benjamin Zebley
7
, Desmond J Oathes
8
,
Amit Etkin
9,10
, Alan F Schatzberg
9
, Keith Sudheimer
9
, Jennifer Keller
9
, Helen S Mayberg
11
,
Faith M Gunning
2,12
, George S Alexopoulos
2,12
, Michael D Fox
13
, Alvaro Pascual-Leone
13
,
Henning U Voss
14
, BJ Casey
15
, Marc J Dubin
1,2
, and Conor Liston
1,2,3
1
Feil Family Brain and Mind Research Institute, Weill Cornell Medical College, New York, New
York, USA
2
Department of Psychiatry, Weill Cornell Medical College, New York, New York, USA
3
Sackler Institute for Developmental Psychobiology, Weill Cornell Medical College, New York,
New York, USA
4
Department of Bioengineering and Center for Mind, Brain and Computation, Stanford University,
Stanford, California, USA
5
Department of Statistics, Columbia University Medical Center, New York, New York, USA
6
Department of Psychiatry, Toronto Western Hospital, Toronto, Canada
7
Department of Psychiatry, Columbia University Medical Center, New York, New York, USA
8
Center for Neuromodulation in Depression and Stress and Department of Psychiatry, University
of Pennsylvania Perelman School of Medicine, Philadelphia, Pennsylvania, USA
9
Department of Psychiatry and Behavioral Science, Stanford University, Stanford, California, USA
10
Veteran Affairs Palo Alto Health Care System, Stanford University, Stanford, California, USA
11
Department of Psychiatry, Emory University School of Medicine, Atlanta, Georgia, USA
12
Institute of Geriatric Psychiatry, Weill Cornell Medical College, New York, New York, USA
Reprints and permissions information is available online at http://www.nature.com/reprints/index.html.
Correspondence should be addressed to C.L. (col2004@med.cornell.edu).
Note: Any Supplementary Information and Source Data files are available in the online version of the paper.
AUTHOR CONTRIBUTIONS
J.D., K.D., F.M., D.J.O., A.E., A.F.S., K.S., J.K., H.S.M., F.M.G., G.S.A., M.D.F., A.P.-L., H.U.V., B.J.C., M.J.D. and C.L. collected
the data. L.G. consulted on all statistical analyses. C.L. designed the protocol for analyzing data pooled across multiple sites and
identifying clusters. A.T.D., R.F. and C.L. designed and implemented the preprocessing pipeline and methods for validating clusters
and optimizing classifiers, and C.L. developed and implemented the method for clustering and classification in a low-dimensional
connectivity-feature space by using canonical correlation analysis (Figs. 1–3). J.D., K.D. and F.M. collected the TMS data. C.L.
analyzed the TMS response data and other clinical data (Figs. 2 and 4) and tested the subtype classifiers on subjects with other
diagnoses (Fig. 5). A.T.D., Y.M. and C.L. implemented the permutation testing. A.T.D., B.Z. and C.L. created the figures and wrote
the manuscript. All authors discussed the results and conclusions and edited the manuscript.
COMPETING FINANCIAL INTERESTS
The authors declare competing financial interests: details are available in the online version of the paper.
HHS Public Access
Author manuscript
Nat Med
. Author manuscript; available in PMC 2017 October 02.
Published in final edited form as:
Nat Med
. 2017 January ; 23(1): 28–38. doi:10.1038/nm.4246.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
13
Berenson-Allen Center for Noninvasive Brain Stimulation and Harvard Medical School, Boston,
Massachusetts, USA
14
Department of Radiology, Weill Cornell Medical College, New York, New York, USA
15
Department of Psychology, Yale University, New Haven, Connecticut, USA
Abstract
Biomarkers have transformed modern medicine but remain largely elusive in psychiatry, partly
because there is a weak correspondence between diagnostic labels and their neurobiological
substrates. Like other neuropsychiatric disorders, depression is not a unitary disease, but rather a
heterogeneous syndrome that encompasses varied, co-occurring symptoms and divergent
responses to treatment. By using functional magnetic resonance imaging (fMRI) in a large
multisite sample (
n
= 1,188), we show here that patients with depression can be subdivided into
four neurophysiological subtypes (‘biotypes’) defined by distinct patterns of dysfunctional
connectivity in limbic and frontostriatal networks. Clustering patients on this basis enabled the
development of diagnostic classifiers (biomarkers) with high (82–93%) sensitivity and specificity
for depression subtypes in multisite validation (
n
= 711) and out-of-sample replication (
n
= 477)
data sets. These biotypes cannot be differentiated solely on the basis of clinical features, but they
are associated with differing clinical-symptom profiles. They also predict responsiveness to
transcranial magnetic stimulation therapy (
n
= 154). Our results define novel subtypes of
depression that transcend current diagnostic boundaries and may be useful for identifying the
individuals who are most likely to benefit from targeted neurostimulation therapies.
Depression is a heterogeneous clinical syndrome that is diagnosed when a patient reports at
least five of nine symptoms. This allows for several hundred unique combinations of
changes in mood, appetite, sleep, energy, cognition and motor activity. Such remarkable
heterogeneity reflects the consensus view that there are multiple forms of depression, but
their neurobiological basis remains poorly understood
1,2
. So far, most efforts to characterize
depression subtypes and develop diagnostic biomarkers have begun by identifying clusters
of symptoms that tend to co-occur, and by then testing for neurophysiological correlates.
These pioneering studies have defined atypical, melancholic, seasonal and agitated subtypes
of depression associated with characteristic changes in neuroendocrine activity, circadian
rhythms and other potential biomarkers
3–5
. Still, the association between clinical subtypes
and their biological substrates is inconsistent and variable at the individual level, and unlike
diagnostic biomarkers in other areas of medicine, they have not yet proven useful for
differentiating individual patients from healthy controls or for reliably predicting treatment
response at the individual level.
An alternative to subtyping patients on the basis of co-occurring clinical symptoms is to
identify neurophysiological subtypes, or biotypes, by clustering subjects according to shared
signatures of brain dysfunction
6
. This type of approach has already begun to yield insights
into how differing biological mechanisms may give rise to overlapping, heterogeneous
clinical presentations of psychotic disorders
6,7
. Neuroimaging biomarkers of abnormal brain
function have proven utility in the assessment of pain
8
and have also shown promise for
depression, for both the prediction of treatment response
9–13
and treatment selection
14
.
Drysdale et al. Page 2
Nat Med
. Author manuscript; available in PMC 2017 October 02.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Resting-state fMRI (rsfMRI) is an especially useful modality because it can be used easily in
diverse patient populations to quantify functional network connectivity in terms of
correlated, spontaneous MR signal fluctuations. Depression is associated with dysfunction
and abnormal functional connectivity in frontostriatal and limbic brain networks
15–20
, in
accordance with morphological and synaptic changes in chronic stress models in
rodents
21–24
. These studies raise the intriguing possibility that fMRI measures of
connectivity could be leveraged to identify novel subtypes of depression with stronger
neurobiological correlates that predict treatment responsiveness.
To this end, we developed a method for defining depression subtypes by clustering subjects
according to distinct, whole-brain patterns of abnormal functional connectivity in resting-
state networks, unbiased by assumptions about the involvement of particular brain regions,
and tested it in a large, multisite data set. Our analyses revealed four biotypes that were
defined by homogeneous patterns of dysfunctional connectivity in frontostriatal and limbic
networks, and that could be diagnosed with high sensitivity and specificity in individual
subjects. Importantly, these biotypes were also prognostically informative, predicting which
patients responded to repetitive transcranial magnetic stimulation (TMS), a targeted
neurostimulation therapy.
RESULTS
Frontostriatal and limbic connectivity define four depression biotypes
We began by designing and implementing a preprocessing procedure (Online Methods) to
control for motion-, scanner- and age-related effects in a multisite data set that comprised
rsfMRI scans for 711 subjects (the ‘training data set’,
n
= 333 patients with depression;
n
=
378 healthy controls). No subjects had comorbid substance-abuse disorders, and patients and
controls were matched for age and sex. Data that support our approach to controlling for
motion-related Blood-oxygen-level dependent (BOLD) signal effects, a particularly
important source of rsfMRI artifact
25–27
, are presented in Supplementary Figure 1. After co-
registering the functional volumes to a common (Montreal Neurological Institute (MNI))
space, we applied an extensively validated parcellation system
28
to delineate 258 functional
network nodes that spanned the whole brain and had stable signals across all sites and scans
in this data set (Fig. 1a). Next, we extracted BOLD signal residual time series and calculated
correlation matrices between each node, which provided an unbiased estimate of the whole-
brain architecture of functional connectivity in each subject (Fig. 1b).
Each correlation matrix comprised 33,154 unique connectivity features, which thus
necessitated a protocol for selecting a subset of relevant, nonredundant connectivity features
for use in clustering. We reasoned that biologically meaningful depression subtypes would
be best characterized by a subset of connectivity features that were significantly correlated
with depressive symptoms. Therefore, to select connectivity features for use in clustering,
we used canonical correlation analysis (Online Methods) to define a low-dimensional
representation of connectivity features that were associated with weighted combinations of
clinical symptoms, as quantified by the 17-item Hamilton Depression Rating Scale
(HAMD), a commonly used, clinician-rated assessment. To ensure that cluster discovery
was not confounded by site-related differences in subject recruitment criteria or by other
Drysdale et al. Page 3
Nat Med
. Author manuscript; available in PMC 2017 October 02.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
unidentified variables, the cluster-discovery analysis was restricted to a subset of patients
(the ‘cluster-discovery subset’,
n
= 220 of the 333 patients with depression) from two sites
with identical inclusion and exclusion criteria and statistically equivalent depression-
symptom scores (see Supplementary Tables 1–3 for details). This analysis identified linear
combinations of connectivity features (analogous to principal components) that predicted
two distinct sets of depressive symptoms (Fig. 1c,d). The first connectivity component
(canonical variate) defined a combination of predominantly frontostriatal and orbitofrontal
connectivity features that were correlated with anhedonia and psychomotor retardation (Fig.
1c, Supplementary Fig. 2 and Supplementary Table 4). The second component defined a
distinct set of predominantly limbic connectivity features involving the amygdala, ventral
hippocampus, ventral striatum, subgenual cingulate and lateral prefrontal control areas, and
that was correlated with anxiety and insomnia (Fig. 1d). Thus, this empirical, data-driven
approach to feature selection and dimensionality reduction identified two sets of functional
connectivity features that were correlated with distinct clinical-symptom combinations.
We then tested whether abnormalities in these connectivity feature sets tended to cluster in
patient subgroups. Multiple statistical learning approaches are available for discovering
notable structure in large data sets (‘unsupervised learning’). Here we chose to use
hierarchical clustering—a standard approach that has been used extensively in the biological
sciences
29,30
—to discover clusters of patients, by assigning them to nested subgroups with
similar patterns of connectivity (Online Methods). This analysis revealed four patient
clusters defined by distinct and relatively homogeneous patterns of connectivity along these
two dimensions (Fig. 1e,f) and comprising 23.6%, 22.7%, 20.0% and 33.6% of the 220
patients with depression, respectively. This four-cluster solution was optimal for defining
relatively homogeneous subgroups that were maximally dissimilar from each other
(maximizing the ratio of between-cluster to within-cluster variance), while ensuring
individual cluster sample sizes that provided sufficient statistical power to detect biologically
meaningful differences (Supplementary Fig. 3). Therefore, we focused our subsequent
analyses on characterizing and validating these four putative subtypes of depression.
Biotype-specific clinical profiles predicted by frontostriatal and limbic network dysfunction
To understand the neurobiological basis of these biotypes, we began by testing for
differences in the whole-brain architecture of functional connectivity between patients (
n
=
220) and age-, sex- and site-matched healthy controls (
n
= 378) and for connectivity features
that differed between patient subgroups. We observed a common neuroanatomical core of
pathology underlying all four biotypes and encompassing areas spanning the insula,
orbitofrontal cortex, ventromedial prefrontal cortex and multiple subcortical areas (Fig. 2a,b
and Supplementary Table 5)—all of which have been implicated in depression
previously
15–20
. This led us to ask whether these connectivity features predicted the severity
of ‘core’ symptoms that were present in almost all patients, regardless of biotype. We found
that, of the 17 symptoms quantified by the HAMD, three were present in almost all patients
with depression (>90%): mood (“feelings of sadness, hopelessness, helplessness,” 97.1%),
anhedonia (96.7%) and anergia or fatigue (93.9%). Across subjects, regardless of biotype,
abnormal connectivity in this shared neuroanatomical core (as indexed by the first principal
Drysdale et al. Page 4
Nat Med
. Author manuscript; available in PMC 2017 October 02.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript
component in a principal-component analysis (PCA)) was correlated with severity scores on
these three symptoms (Fig. 2c;
r
= 0.72–0.82).
In addition, we found that, superimposed on this shared pathological core, distinct patterns
of abnormal functional connectivity differentiated the four biotypes (Fig. 2d,e) and were
associated with specific clinical-symptom profiles (Fig. 2f). For example, as compared to
controls, reduced connectivity in frontoamygdala networks, which regulate fear-related
behavior and reappraisal of negative emotional stimuli
31–33
, was most severe in biotypes 1
and 4, which were characterized in part by increased anxiety. By contrast, hyperconnectivity
in thalamic and frontostriatal networks, which support reward processing, adaptive motor
control and action initiation
20,34–37
, were especially pronounced in biotypes 3 and 4 and
were associated with increased anhedonia and psychomotor retardation. And reduced
connectivity in anterior cingulate and orbitofrontal areas supporting motivation and
incentive-salience evaluation
38–40
was most severe in biotypes 1 and 2, which were
characterized partly by increased anergia and fatigue.
Importantly, although the connectivity-based biotypes revealed in our analysis were
associated with differences in clinical symptoms, they did not simply reflect differences in
overall depression severity. Although overall depression severity scores were modestly but
significantly decreased in biotype 2 as compared to the other three groups (by 15–16%),
there were no significant differences in severity between biotypes 1, 3 and 4 (Fig. 2g; see
Supplementary Fig. 4 for convergent findings in independent data acquired from subjects not
included in the cluster-discovery analysis). Furthermore, they did not simply recapitulate
subtypes derived strictly from clinical-symptom measures; whereas clustering according to
functional connectivity features in random patient subsamples yielded stable clustering
outcomes, clustering according to clinical symptoms yielded unstable outcomes with
relatively low longitudinal stability over time (Supplementary Fig. 5).
Functional connectivity biomarkers for diagnosing depression biotypes
By reducing diagnostic heterogeneity, we reasoned that clustering could be leveraged to
develop classifiers for the diagnosis of depression biotypes solely on the basis of fMRI
measures of functional connectivity, which have shown promise in smaller-scale, single-site
studies of depression
41–43
and other neuropsychiatric disorders
44,45
, but that have not
performed as well when tested in multisite data sets
44
. To this end, we developed classifiers
for each depression biotype, testing and optimizing standard, extensively used methods for
brain parcellation, subject clustering, feature selection and classification to identify
empirically the most successful approach to clustering and classification (Fig. 3a and Online
Methods). Throughout, clustering analysis was performed in the same cluster-discovery
sample (
n
= 220), whereas classification of patients versus controls was optimized in the full
training data set (
n
= 333 patients;
n
= 378 controls), and leave-one-out cross-validation and
permutation testing were used to assess performance and significance (Supplementary Fig.
6; for additional analysis confirming the stability of cluster assignments, see Supplementary
Fig. 3d–f). The optimization process yielded progressive improvements in classifier
performance (Fig. 3b). Support-vector machine (SVM) classifiers (using linear kernel
functions) performed best, yielding overall accuracy rates of up to 89.2% for the clusters
Drysdale et al. Page 5
Nat Med
. Author manuscript; available in PMC 2017 October 02.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript