Participants
COORDINATE-MDD is an international consortium consisting of raw individual MRI data with deep phenotypic characterization in MDD (Fu et al., 2022). Ethical approvals were acquired by institutional review boards for each study site. The subset of MDD participants included in the present study satisfied the following inclusion criteria: 1) DSM-based diagnosis of MDD; 2) in current depressive episode of at least moderate severity, defined as a 17-item Hamilton Rating Scale for Depression score equal to or greater than 14; 3) medication-free at the time of scanning; and exclusion criteria: 1) current comorbid psychiatric, medical or neurological disorders; 2) treatment resistant depression, defined as not achieving clinical response to two or more trials of antidepressant medications.
The present study consists of a total of 685 MDD participants from 10 studies: Canadian Biomarker Integration Network in Depression (CAN-BIND) (N=92, (MacQueen et al., 2019)), Establishing Moderators and Biosignatures of Antidepressant Response in Clinical Care (EMBARC) (N=257, (Madhukar H. Trivedi et al., 2016)), Huaxi MR Research Center SCU (HMRRC) (N=111, (J. Zhang et al., 2011)), King’s College London (KCL) (N=20, (Wise et al., 2018)), Manchester Remedi (N=40, (Arnone, McIntosh, Ebmeier, Munafò, & Anderson, 2012; Dutta et al., 2019)), Laureate Institute for Brain Research (LIBR) (N =554, (Misaki, Suzuki, Savitz, Drevets, & Bodurka, 2016; Victor et al., 2018)), Oxford (N=39, (Vai et al., 2016)), Predictors of Remission in Depression to Individual and Combined Treatments (PReDICT) (N=63, (Dunlop et al., 2012)), Stanford SNAP (N=8, (Sacchet, Livermore, Iglesias, Glover, & Gotlib, 2015)), STratifying Resilience and Depression Longitudinally (STRADL) (N=1, (Habota et al., 2019)); and a total of 699 healthy control (HC) participants from 10 studies: CAN-BIND (N=23), EMBARC (N=39), KCL (N=20), LIBR (N=141), Manchester Blame (N=46), Manchester Remedi (N=30), Oxford (N=31), HMRRC SCU (N=139), Stanford SNAP (N=50), STRADL (N=180). The pooled age range was 18-65 years for MDD and 16-72 for healthy control participants. MDD diagnosis was based on DSM-IV or DSM-IV-TR diagnostic criteria. The number of MDD participants who were treatment-naïve is 128. Information about ethnicity (self-reported) can be found in Table 1. Missing information is either because data was not collected or was not shared. Image protocols, scanner acquisition parameters and study characteristics can be found in Table 1 and Supplementary Materials. Each study was approved by the local ethics committee and all participants gave written consent to participate and share de-identified data according to each institution’s local legislative and/or ethical policies. Ethical approval numbers: Manchester (Stockport Research Ethics Committee 07/H1012/76), SNAP (IRB approval 12104), EMBARC (STU 092010-151), Oxford (REC reference 11/SC/0224), LIBR (WCG IRB 1136261 & 1136947), STRADL (NHS Tayside committee 14/SS/0039), PReDICT (Emory IRB # 00024975), KCL (Bromley NHS REC 13/LO/0904) and SCU (IRB 2020(54)).
Longitudinal treatment outcomes were available in a subset of 5 prospective clinical treatment trials: CAN-BIND (N=81), EMBARC (N=207), Oxford (N=35), Manchester (N=36) and PReDICT (N=63). The treatments were a selective serotonin reuptake inhibitor (SSRI) antidepressant medication: citalopram (Manchester), escitalopram (CAN-BIND, Oxford, PReDICT) or sertraline (EMBARC); a serotonin noradrenergic reuptake inhibitor antidepressant (SNRI) medication, duloxetine (PReDICT); placebo (EMBARC); or cognitive behavioral therapy (CBT) (PReDICT). Treatment duration was 6 weeks (Oxford), 8 weeks (CAN-BIND, EMBARC, Manchester) or 12 weeks (PReDICT). Depressive symptom severity was assessed by clinician-rated scales: 17-item Hamilton Rating Scale (HAMD) (EMBARC, Oxford, PReDICT) (Hamilton, 1967) and Montgomery-Åsberg Depressive Ratings Scale (MADRS) (CAN-BIND, Manchester) (Montgomery & Asberg, 1979). MADRS ratings were converted into HAMD rating using conversion tables (Leucht, Fennema, Engel, Kaspers-Janssen, & Szegedi, 2018). Symptom ratings were acquired at baseline and following treatment for all studies (Table 1). Trial registration numbers are: CAN-BIND (NCT01655706), EMBARC (NCT01407094), PReDICT (NCT00360399). Oxford and Manchester do not have clinical trial registration because it was not a national or funder requirement at the time.
Image preprocessing
Each participant’s quality-controlled (QC) T1-weighted MRI image was preprocessed with a containerized processing pipeline. Preprocessing steps consisted of correction for magnetic field intensity inhomogeneity correction followed by multi-atlas skull-stripping (Tustison et al., 2010). Images were segmented using a state-of-the-art multi-atlas, label fusion method (MUSE) to derive 259 pre-defined anatomical regions of interest (ROI) of the segmented tissue maps (Doshi et al., 2016). Voxel-wise regional volumetric maps (RAVENS) were generated for each tissue volume (Davatzikos, Genc, Xu, & Resnick, 2001) by spatially aligning the skull-stripped images to a template in MNI-space using a registration method (Ou, Sotiras, Paragios, & Davatzikos, 2011) and harmonizing for site, age, and sex effects (Pomponio et al., 2020).
Application of HYDRA to identify neuroanatomical dimensions
HYDRA is a semi-supervised machine learning clustering method to distinguish patients from controls by combining multiple linear classifiers, whereby each hyperplane separates a dimension of patients from the control group resulting in a “1-to-k” mapping (Varol et al., 2017). Therefore, HYDRA clusters disease effects by comparing brain patterns to those of healthy controls rather than by comparing patients with one another. The Adjusted Rand Index (ARI) is used to identify the optimal number of dimensions (k) from a range between 2 and 5. Since HYDRA is a multivariate method, we applied it to the raw MUSE ROIs. To evaluate the robustness of the optimal k clusters scheme, we performed additional analyses. Firstly, we used split-sample analyses to evaluate the robustness of the optimal k dimension solution, to assess whether the dimensions in each half exhibit similar neuroanatomical patterns, given that the two halves have similar cohort characteristics in terms of age, sex, and site. Secondly, we conducted leave-site-out cross-validation to examine if the dimensions were being driven by any one particular site. Lastly, a permutation test was performed to test the statistical significance with the optimal k dimension scheme.
Voxel-wise RAVENS regional tissue volumes
Voxel-wise RAVENS gray and white matter maps (Davatzikos et al., 2001) were used to identify the brain regions that differentiate each HYDRA dimension from the healthy control group. Statistical parametric maps estimating deviations from healthy controls for each dimension were calculated using regionally linear multivariate discriminative statistical mapping (MIDAS) (Varol, Sotiras, & Davatzikos, 2018) with age and sex as covariates and filtering out non-significant voxels (pFDR<0.05).
HYDRA and the voxel-wise analyses were run in MATLAB 2018A.
Statistics
Demographic and clinical variables
Group comparisons for demographic (age, sex, years of education) and clinical variables (age of onset, years of illness, duration of current episode in weeks) were examined across the HYDRA dimensions using Mann-Whitney U tests for continuous variables (e.g. age) and Chi-Square tests for categorical variables (e.g. sex).
Evaluation of HYDRA dimensions and their treatment response to antidepressant and placebo
The subset consisted of 4 cohorts of MDD participants from the prospective, longitudinal clinical treatment trials which had included healthy control participants from the same sites: CAN-BIND (N=81), EMBARC (N=207), Oxford (N=35), Manchester (N=36). Treatment was SSRI antidepressant: citalopram (Manchester), escitalopram (CAN-BIND, Oxford) or sertraline (EMBARC); or placebo (EMBARC). Treatment duration was 6 weeks (Oxford) or 8 weeks (CAN-BIND, EMBARC and Manchester).
Of the 5 cohorts with longitudinal treatment outcomes, PReDICT (N=63) had only included MDD participants. As robustness of the optimal dimensional clustering involves comparison of the patterns between patients and healthy controls, we could not be certain about the results for the 5 cohorts, therefore we present the results for 4 cohorts here, and the results including PReDICT are presented in the Supplementary Materials (Supplementary Figure 3 and 4).
To examine interactions between HYDRA dimension and treatment group, we used a linear regression model with the percentage change in the clinician-rated depressive symptom scale (continuous) as the outcome variable and HYDRA dimension (categorical, 2 groups) and treatment group (categorical, 2 groups: SSRI and placebo) as the independent variables whilst controlling for age, sex and site. Percentage change in score was calculated as follows: (pre-treatment baseline score – post-treatment score)/pre-treatment score x 100. The effect size (Cohen’s f2= 0.06) of the interaction term has a F statistic of 3.607 based on our analysis using a linear regression model. With a sample size of 359, assuming that we adjust for 6 additional covariates in the model and the same effect size, we have over 99% power to detect a significant interaction term between treatment and HYDRA dimension under 5% Type-I error. P=0.05 (two-sided) was selected as the threshold for significance.
The linear regression models were conducted using the statsmodels 0.13.1 Python module (Seabold et al., 2010). Power analyses, Mann-Whitney U tests and Chi-Square tests were conducted in R version 4.2.2.
In a machine learning analysis, we trained a support vector machine to classify patients between the identified HYDRA dimensions and performed an additional linear regression using the calculated hyperplane distance in place of the dimension label.