Authors: Hao Zhu, Xiaoyu Tong, Nancy B. Carlisle, Hua Xie, Corey J. Keller, Desmond J. Oathes, Charles B. Nemeroff, Gregory A. Fonzo, Yu Zhang
Categories: Article
Source: bioRxiv
Authors: Hao Zhu, Xiaoyu Tong, Nancy B. Carlisle, Hua Xie, Corey J. Keller, Desmond J. Oathes, Charles B. Nemeroff, Gregory A. Fonzo, Yu Zhang
Major depressive disorder (MDD) is a prevalent psychiatric disorder characterized by substantial clinical and neurobiological heterogeneity. Conventional studies that solely focus on clinical symptoms or neuroimaging metrics often fail to capture the intricate relationship between these modalities, limiting their ability to disentangle the complexity in MDD. Moreover, patient neuroimaging data typically contains normal sources of variance shared with healthy controls, which can obscure disorder-specific variance and complicate the delineation of disease heterogeneity.
We employed contrastive principal component analysis to extract disorder-specific variations in fMRI-based resting-state functional connectivity (RSFC) by contrasting MDD patients (N=233) with age-matched healthy controls (N=285). We then applied sparse canonical correlation analysis to identify latent dimensions in the disorder variations by linking the extracted contrastive connectivity features to clinical symptoms in MDD patients.
Two significant and generalizable dimensions linking distinct brain circuits and clinical profiles were discovered. The first dimension, associated with an apparent “internalizing-externalizing” symptom dimension, was characterized by self-connections within the visual network and also associated with choice reaction times of cognitive tasks. The second dimension, associated with personality facets such as extraversion and conscientiousness typically inversely associated with depression symptoms, is primarily driven by self-connections within the dorsal attention network. This “depression-protective personality” dimension is also associated with multiple cognitive task performances related to psychomotor slowing and cognitive control.
Our contrastive RSFC-based dimensional approach offers a new avenue to dissect clinical heterogeneity underlying MDD. By identifying two stable, neurophysiology-informed symptom dimensions in MDD patients, our findings may enhance disease mechanism insights and facilitate precision phenotyping, thus advancing the development of targeted therapeutics for precision mental health.
Major depressive disorder (MDD) is a highly prevalent mental disorder with a lifetime prevalence of over 20% in women and 11% in men, impacting a substantial number of individuals globally. MDD is diagnosed based on the Diagnostic and Statistical Manual of Mental Disorders 5^th^ edition^1^. Unfortunately, such symptom-based diagnoses include patients with a bewildering mixture of symptoms and have resulted in substantial clinical and neurobiological heterogeneity, obscuring the underlying mechanisms of the cognitive and behavioral dysfunctions in MDD patients. This oversight may contribute to suboptimal treatment efficacy^2,3^ and hinder the development of more effective therapeutics, necessitating the dissection of MDD heterogeneity.
Prior studies have attempted to examine the clinical heterogeneity within MDD utilizing symptom-based dimensional approaches or subtyping analyses^4–6^. For instance, efforts have been made to identify symptom dimensions shared among patients and their associations with personality^7^, treatment outcomes^8^, or comorbid mental health conditions^9^. While current clinical measures are useful for inexpensively characterizing a patient symptom profile, they fall short in capturing underlying neurophysiological variations, limiting the discovery of neurobiological basis underlying the clinical heterogeneity. A recent systematic review^6^ reported that there is wide diversity of identified symptom dimensions and subtypes among previous studies, indicating the failure of addressing disease heterogeneity and the necessity of a more reliable and objective delineation of MDD pathological dimensions.
Neuroimaging techniques, such as functional magnetic resonance imaging (fMRI), have demonstrated promise in probing neurobiology of various psychiatric disorders. However, many previous neuroimaging studies have followed a case-control design that focuses on the group difference in aberrant brain circuits between MDD patients and healthy individuals, either during the resting state^10–12^ or in response to specific tasks^13^. Such an approach can only extract population-level information and has a limited ability to characterize the heterogeneity in MDD^14^. Recent research efforts have shifted to neuroimaging-based dimensional analyses to examine variations in brain structure and function beyond the conventional group-level comparison^15,16^. In particular, leveraging functional connectivity that measures neural coupling between brain regions, recent studies have revealed novel subtypes among mood and anxiety disorders^17,18^, which are not identifiable via a conventional case-control approach or clinically-defined MDD categories. To further enhance the clinical relevance of subtype or dimension discovery, increasing research efforts have delved deeper into not only dimensional patterns of brain circuits but also their behavioral or cognitive profiles^17,19–21^. Typically employing machine learning techniques such as canonical correlation analysis (CCA) or partial least squares, these studies have identified latent dimensions by jointly examining functional connectivity and symptom/behavioral measures in a data-driven manner^19–21^. These approaches revealed patterns of association between neurophysiological characteristics and specific symptom combinations, offering insights into the underlying mechanisms behind these disorders.
Neuroimaging data from patients contains both disorder-specific variation and disorder-irrelevant variation shared with healthy controls^22^. Given the high variability observed across individual-level brain metrics, one challenge in characterizing heterogeneity among clinical populations lies in the fact that it is usually muddied by overall normative variability. This has substantially decreased the signal-to-noise ratio and leads to suboptimal identification of meaningful patterns. In addressing this challenge, contrastive learning emerges as a powerful tool for training models to selectively retain group-specific features while filtering out confounding information inherent in the broader population. This process enables a focused analysis by isolating and emphasizing the distinctive characteristics pertinent to the target group. Encouragingly, some recent works have integrated the contrastive learning technique into the exploration of brain biomarkers in psychiatric disorders such as autism^22,23^, demonstrating its unique advantages in uncovering the intricate nuances of brain morphology^22^ and functional connectivity^23^ associated with psychiatric conditions. Contrastive learning holds great potential in capturing the inadequately explored pathology-relevant variations for defining symptom dimensions more precisely within MDD patients. Hence, we aim to shed light on the underlying heterogeneity specific to MDD through the identification of linked dimensions between clinical symptoms and MDD-specific neurophysiology components obtained by contrastive learning.
In this study, we applied a contrastive learning-based brain-symptom analytical framework^23^ to identify latent dimensions that link resting-state functional connectivity (RSFC) and clinical symptoms in MDD patients. This framework comprises two main steps. First, we conducted the contrastive principal component analysis (cPCA) on the fMRI-based RSFC data of MDD patients (N=233), with age-matched healthy controls (N=285) as background data, i.e. normative sources of variance to be removed from consideration in subtyping analyses. This process allowed us to extract contrastive RSFC features that disentangled disorder-specific variations from those shared with the healthy population. We then applied sparse CCA between these contrastive RSFC features and representative symptoms to reveal informative latent dimensions that link MDD-specific brain functional variations and clinical symptoms. Our study successfully identified two robust latent dimensions showing distinct neural circuit patterns and clinical profiles, primarily involving internalizing-externalizing symptoms and depression-protective personality. Rigorous stability analysis and cross-validation further confirmed their robustness and generalizability. We also examined associations of the identified dimensions with performance on various cognitive tasks, offering valuable insights into MDD-related cognitive dysfunction. Collectively, our contrastive RSFC-based dimensional approach provides a new avenue to examine neurophysiology-informed symptom dimensions for an improved understanding of mechanisms underlying MDD symptom heterogeneity. This may lead to a more objective classification of psychiatric conditions, thereby advancing targeted therapeutics for precision mental health.
The patient population used in our study include MDD patients from the Establishing Moderators and Biosignatures of Antidepressant Response for Clinical Care for Depression (EMBARC) dataset^24^. EMBARC is a large, randomized placebo-controlled clinical trial for examining biomarkers for MDD and antidepressant treatment response. Written informed consent was obtained from each participant under the Institutional Review Board (IRB) approved protocols at each of the four study sites, including University of Texas Southwestern Medical Center, Columbia University/Stony Brook, Massachusetts General Hospital, University of Michigan, University of Pittsburgh, and McLean Hospital. Subjects were required to meet the SCID criterion for an MDD episode and have a Quick Inventory of Depressive Symptomatology score^25^ of ≥14 at both screening and randomization visits. This dataset includes 296 MDD patients who were randomly prescribed sertraline or placebo for eight weeks, with the primary outcome of treatment response measured by the 17-item Hamilton Depression Rating Scale (HAMD17)^26^. For our brain-symptom linked dimension analysis, we included 25 baseline clinical scales and symptoms of depression (the 17-item HAMD; Quick Inventory of Depressive Symptomatology); Childhood Trauma Questionnaire subscales (emotional abuse, emotional neglect, physical abuse, physical neglect, sexual abuse)^27^; suicide-related symptoms (Concise Associated Symptoms Tracking^28^; Concise Health Risk Tracking-Self Report: propensity and risk scores^29^); Mood and Anxiety Symptoms Questionnaire subscales (anxious arousal, anhedonic depression, general distress)^30^; Mood Disorder Questionnaire^31^; personalities (NEO-Five Factor Inventory subscales, including neuroticism, extraversion, openness, agreeableness, and conscientiousness^32^; Standardised Assessment of Personality – Abbreviated Scale); Self-Administered Comorbidity Questionnaire^33^; Snaith-Hamilton Pleasure Score^34^; and Social Adjustment Scale overall mean score. Due to incomplete clinical measures, 63 subjects were excluded. The remaining 233 subjects (aged 18–65) were used for our analyses. Demographics of these subjects are listed in Supplementary Table S1.
We compiled and harmonized 285 healthy controls (aged 18–77.5) as a background group for cPCA. Demographic information of these subjects is summarized in Supplementary Table S2. These subjects were selected by matching the age of the healthy population with the patient population (Kolmogorov–Smirnov test p = 0.052, Figure S1). Among them, 27 are from the EMBARC dataset and others are from the following three datasets. Specifically, we used 83 healthy controls from the University of California Los Angeles Consortium for Neuropsychiatric Phenomics (UCLA-CNP)^35^, which was approved by the IRB at University of California, Los Angeles and the Los Angeles County Department of Mental Health. 72 subjects were selected from the Amsterdam Open MRI Collection - Population Imaging of Psychology (AOMIC-PIOP1)^36^ dataset, which was approved by the faculty’s ethical committee at the University of Amsterdam. The remaining 103 subjects were from the Leipzig Study for Mind-Body-Emotion Interactions (LEMON)^37^ dataset, which was approved by the ethics committee at the medical faculty of the University of Leipzig.
MRI data were acquired using 3T MRI systems at four different sites. At each site, resting-state fMRI data were scanned via T2* weighted images using a single-shot gradient echo-planar pulse sequence lasting for six minutes, with parameter repetition time 2000 ms, echo time 28 ms, flip angle 90°, matrix size 64 × 64, voxel size 3.2 × 3.2 × 3.1 mm^3^, and 39 axial slices. Each subject underwent one or two runs of fMRI scans within a single day.
MRI data were acquired on two 3T Siemens Trio scanners at UCLA. Resting-state MRI data were scanned using a T2*-weighted echoplanar imaging sequence lasting for 304 seconds, with the following repetition time 2000 ms, echo time 30 ms, flip angle 90°, matrix size 64 × 64, voxel size 3 × 3 × 4 mm^3^, and 34 slices.
MRI data were acquired on a Philips 3T scanner. During the resting state scans, participants were instructed to keep their gaze fixated on a fixation cross in the middle of the screen with a gray background and to let their thoughts run freely. Resting-state fMRI scanning lasted six minutes (i.e., 480 volumes with a repetition time 750 ms).
MRI data were scanned using a 3T Siemens Verio scanner. Resting-state fMRI data were scanned using a T2*-weighted gradient echo planar imaging sequence lasting for 15 minutes and 30 seconds. The sequence parameters were specified as repetition time 1400 ms, echo time 30 ms, flip angle 69°, matrix size 88 × 88, voxel size 2.3 × 2.3 × 2.3 mm^3^, and 64 slices.
All the resting-state fMRI data were preprocessed using the fMRIPrep pipeline^38^ and aggregated into 100 regions-of-interest (ROIs) level time series according to the Schaefer parcellation^39^. RSFC features were then calculated as Pearson’s correlation coefficient in fMRI time series between every pair of ROIs. For MDD patients with multiple fMRI runs, we took the average of the RSFC across runs to generate the RSFC feature data used for the subsequent analysis. Given that the fMRI data from healthy controls were derived from different studies, we applied the well-established ComBat harmonization technique^40^ to their RSFC data to mitigate site effects. During the removal of site effects, age and gender were designated as biological covariates to be preserved, with EMBARC as the reference batch.
In our study, we employed cPCA^41^ on the RSFC data of MDD patients as the target data and the RSFC data of healthy controls as the background data, aiming to pinpoint the disorder-specific foreground components by contrasting the healthy population. cPCA operates under the assumption that the target data comprises both domain-specific (foreground) information and domain-unrelated (background) variance. Utilizing the covariance matrix Cp derived from the target data (MDD patients in our case) and the covariance matrix Cbg derived from the background data (healthy controls), cPCA identifies the linear components most closely associated with the foreground data by subtracting the background covariance matrix from the target covariance matrix to obtain the foreground covariance Cfg=Cp-αCbg, where α represents a hyperparameter quantifying the degree of contrast. This hyperparameter was determined based on the cross-validation results through grid search. The top 200 contrastive principal components, which explain more than 70% of the data variance, were used in subsequent analyses. This selection was made because the dimensionality of informative components is constrained by the sample size of patients.
We first applied PCA to the clinical variables and retained 80% of the total variance to reduce dimensionality. We then applied sparse canonical correlation analysis (sCCA)^42^ to identify latent dimensions linking contrastive connectivity features with clinical symptoms*.* sCCA optimizes the correlation between two data matrices, thus yielding symptom dimensions maximally associated with neuroimaging features. Specifically, we utilized sCCA to analyze the disorder-specific components extracted from cPCA and corresponding clinical measures, aiming to identify neurophysiology-informed symptom dimensions. Sparsity constraint is imposed on both connectivity features and clinical measures to improve interpretability and alleviate overfitting. The overall minimization objective is formulated as -Cov(Au,Bv)+λ1|u|1+λ2|v|1, where A and B are data matrices of connectivity features and clinical measures respectively, u,v are the FC and symptom dimension loadings, and λ1, λ2 are sparsity hyperparameters on the loadings. We performed grid search to find the hyperparameters with best performance.
We performed 10 repetitions of 10-fold cross-validation to evaluate the generalizability and stability of identified latent dimensions. In each fold, 90% of the patients and all healthy subjects were compiled as the training set of cPCA. As the order of identified CCA dimensions might vary across folds due to training data variance, we employed the dimensions identified from the whole dataset as reference dimensions to align the dimensions acquired from different folds. Pearson’s correlation coefficient was then computed to assess the generalizability of the identified dimensions on the validation set. Intra-class correlation coefficient was employed to confirm the stability of dimension loadings across cross-validation folds.
Furthermore, to evaluate the significance of the identified dimensions, we conducted permutation tests 1000 times. Specifically, we randomly permuted the clinical symptoms for each patient to others and followed the same procedures to generate the Pearson correlation R values on cross validation test set as described above. Test R values from dimensions with the same ranking across different folds were aggregated to generate the null distribution.
Network-level connectivity provides a broader view of which neural systems and their interactions contribute most to the identified latent dimensions. To measure this, we first retained only the top 10% loadings with the highest absolute values, setting the rest to zero to exclude the less important and stable connections. Then, we averaged all absolute loadings within and between each pair of networks to obtain the network-level connectivity importance.
We investigated the correlation between the FC scores of identified dimensions and the cognitive task performance of MDD patients. This analysis may provide insights into how different cognitive abilities, such as psychomotor slowing^43^, cognitive control^44^, working memory^45^, reward learning^46^, and resolution and adjustment behavior in response to emotional conflict^47^, relate to these brain-symptom latent dimensions. These cognitive and emotional measurements are selected because they are potential predictors for MDD treatment response^24^, thus having the probability to be correlated with the identified MDD-specific dimensions. The specific task item we included can be found in Table S3.
We applied the contrastive learning-based brain-symptom association identification framework to the 233 MDD patients. Two dimensions with generalizable association between neurophysiology and symptom profile have been identified through 10 rounds of 10-fold cross-validation and permutation test. The composition and essence of the first identified dimension was interpreted, based upon loadings, to indicate an axis of internalizing-externalizing symptoms (Rtrain = 0.551, Rcv = 0.287, ppermutation < 0.001, Figures 2, S4). A higher internalizing-externalizing symptoms dimension score indicates a greater tendency towards internalizing behavior such as lack of pleasure or energy, and a lower tendency towards externalizing behavior such mania and anger attack. The second identified dimension incorporates the concept of depression-protective personality (Rtrain = 0.437, Rcv = 0.207, ppermutation = 0.001, Figure 2, S4). Individuals with higher scores on this personality dimension score exhibited higher scores on personality facets typically negatively associated with depression symptoms, such as extraversion and conscientiousness, while a low dimension score indicates higher scores on personality facets typically associated with greater depression symptoms, such as neuroticism^48^. Both dimensions significantly outperformed the performance yielded from its non-contrastive counterpart integrating standard PCA and sCCA (paired t-test: internalizing-externalizing symptoms p=5.2×10^−4^; depression-protective personality p = 0.008), demonstrating the unique advantage of contrastive learning for improving the identification of brain-symptom associations. Afterward, we calculated the intra-class correlation coefficient (ICC) of dimension composition in each cross-validation fold to assess the stability of the dimensions. These two dimensions both showed high stability. For internalizing-externalizing symptoms, the ICC value is 0.87 (95% CI: [0.87, 0.88]) for FC loadings, 0.92 (95% CI: [0.88, 0.96]) for symptom loadings. For depression-protective personality, the ICC value is 0.94 (95% CI: [0.94, 0.94]) for FC loadings, 0.94 (95% CI: [0.90, 0.97]) for symptom loadings (Figure S4).
Next, we investigated the most important ROI-level and network-level connectivity within each dimension. (Figure 3). In the internalizing-externalizing symptoms dimension, the correlation between the left and right cuneus and other regions, particularly the left fusiform gyrus, the left inferior temporal gyrus, and the right fusiform gyrus, contributes the most. In terms of anti-correlation, the connection between the left and right posterior cingulate cortex and other regions, mainly the left superior occipital gyrus and the right middle occipital gyrus, contributes the most. On the network level, the dimension is characterized by the self-connection of visual network, the connection between limbic network and sensorimotor network, and the connection between limbic network and dorsal attention network. In depression-protective personality dimension, the correlation between the right middle cingulate cortex and the left superior parietal cortex contributes the most. In terms of anti-correlation, the key connections are between the left fusiform gyrus, the left inferior temporal gyrus, and the left superior parietal cortex, as well as with the left inferior frontal gyrus and the left middle frontal gyrus. On network level, this dimension is highly related to self-connections within the dorsal attention network.
Internalizing-externalizing symptoms score showed a significant correlation with reaction time of choice reaction time task (r=0.20, pfdr =0.0084) and reaction time difference in a flanker task (r=0.18, pfdr =0.024). This indicates that a higher score on this dimension is associated with longer reaction time, suggesting lower performance. Reaction time difference in the flanker task measures selective attention and inhibitory function; longer reaction times indicate a greater interference effect from incongruent stimuli. Depression-protective personality dimension score had significant correlations with multiple cognitive tasks, including the accuracy of A-not-B task (r=−0.24, pfdr=0.0016), reaction time of choice reaction time task (r=0.31, pfdr =2.8×10^−5^), and both the accuracy difference (r=−0.15, pfdr =0.046) and reaction time difference (r=0.20, pfdr =0.0084) in the flanker task. This suggests that this dimension has a wider range of correlations with different cognitive tasks measuring working memory, interference adjustments, and cognitive control. No significant correlations were observed between these two dimensions and the word fluency task, emotion conflict task, or probabilistic reward task (Figure 4).
We also examined the capabilities of these dimensions in predicting antidepressant treatment response using chi-square test. However, no significant performance was observed for either dimension in differentiating remission versus non-remission based on whether the FC dimension scores were above or below the median (internalizing-externalizing symptoms p=0.27 for sertraline arm, p=1.00 for placebo arm; depression-protective personality p=0.40 for sertraline arm, p=0.49 for placebo arm) (Supplementary Table S4).
In this study, we implemented a data-driven framework that integrated contrastive machine learning with multivariate correlation analysis to uncover MDD-specific brain functional variations and their associations with clinical symptoms. We identified two robust and generalizable latent dimensions linking neurophysiological and clinical profiles, offering novel and objective biomarkers for dissecting heterogeneity in MDD.
While neuroimaging data conveys important information about circuit dysfunction in MDD patients, it also contains undesired variance shared with healthy individuals. Conventional dimensional methods obtain and identify behavior or symptom-related biomarkers from clinical populations, which may result in suboptimal findings without removing the variance shared by healthy populations. Alternatively, contrastive learning-based frameworks can better extract disorder-specific neurophysiology components through data distillation. In our results, the dimensions derived from contrastive learning show a reduced correlation with age compared to those without contrastive learning (Figures S2, S3). This suggests that contrastive learning successfully removed some disorder-unrelated components, such as age. Other than the undesired variance, MDD itself also exhibits substantial neurobiological and clinical heterogeneity. Previous studies based on case-control analysis or dimensional approaches considering only clinical symptoms or neuroimaging may therefore inadequately address the heterogeneity^6,49^. Our framework links both modalities, addresses the underlying heterogeneity, and provides interpretability.
For the two neurophysiology-informed symptom dimensions we have identified, internalizing-externalizing symptoms dimension is associated with key brain regions such as the cuneus and fusiform gyrus, consistent with findings from studies on neurodevelopmental trajectories of internalizing-externalizing symptoms^50^. Our results suggest that these associations likely persist into adulthood. Our network-level analysis on this dimension aligns with the literature, which identifies the visual network and dorsal attention network as strong predictors of internalizing-externalizing traits^51^. Our analysis of the behavioral task is also consistent with findings in the literature that associate cuneus activation with facilitating vigilance^52^.
The depression-protective personality dimension is marked by personalities such as extraversion, conscientiousness and neuroticism. Studies have identified a strong association between key brain regions, such as the fusiform gyrus, and personality disorders^53,54^, which is consistent with our findings. A recent study also identified the significant correlations between dorsal attention network and personalities including neuroticism and agreeableness^55^, and borderline personality disorder^56^. Additionally, dorsal attention network has also shown great importance in multiple cognitive control tasks^57,58^. All these results align with our findings to the depression-protective personality dimension.
Despite our contributions, this study has several limitations. First, while cPCA is effective for extracting disorder-specific components, it requires fine-tuning of hyperparameters to control the degree of contrast with the background data (i.e., healthy population). Future research could focus on developing automatic frameworks for extracting contrastive connectivity features. Additionally, our study harmonized data from different sources to enlarge the sample size, enhancing the robustness of the identified latent dimensions, followed by verifying their generalizability through rigorous cross-validation. However, future work should further involve independent datasets with comparable clinical measurements to replicate our findings. Lastly, given the recognized importance of subcortical regions and cerebellum in depression^59,60^, extending the analysis to these regions may provide a more comprehensive understanding of brain dysfunctions associated with MDD psychopathology.
In summary, this study employed a combination of contrastive learning and sparse canonical component analysis on a depression patient dataset, unveiling two generalizable disorder-specific dimensions linking neurophysiology and symptom profiles. Our findings hold potential for advancing the understanding of disease mechanisms, facilitating precise diagnosis, formulating individualized treatment plans, and fostering the development of innovative treatment methods.