Abstract
High rates of comorbidity, shared risk, and overlapping therapeutic mechanisms have led psychopathology research towards transdiagnostic dimensional investigations of clustered symptoms. One influential framework accounts for these transdiagnostic phenomena through a single general factor, sometimes referred to as the ‘p’ factor, associated with risk for all common forms of mental illness. Here we build on past research identifying unique structural neural correlates of the p factor by conducting a data-driven analysis of connectome wide intrinsic functional connectivity. We demonstrate that higher p factor scores and associated risk for common mental illness maps onto hyper-connectivity between visual association cortex and both frontoparietal and default mode networks. These results provide initial evidence that the broad risk for common forms of mental illness is associated with patterns of inefficient connectome wide intrinsic connectivity supporting executive control and self-referential processes, which are often impaired across categorical disorders.
Emerging research has identified a general factor of psychopathology that accounts for shared risk among internalizing, externalizing, and thought disorders across diverse samples1,2. Moreover, this general psychopathology or ‘p’ factor3 provides a framework for explaining the high rates of comorbidity as well as the shared genetic variance among categorical mental disorders4,5. As such, the p factor represents a potentially useful avenue for better understanding the shared and unique etiology of common mental illness. However, the biological mechanisms through which the p factor confers general risk for psychopathology remain unclear. Identifying such mechanisms is necessary for effectively leveraging the p factor to derive novel targets for clinical intervention and prevention.
Clinical neuroscience has begun to adapt transdiagnostic methodologies to accelerate the search for common neurobiological abnormalities across disorders6. For example, a recent large meta-analysis of six categorical disorders reported a shared pattern of reduced gray matter volume in a distributed network supporting attention and cognitive control7. In addition, we have recently examined the structural neural correlates of the p factor specifically8. In our work, higher p factor scores and thus risk for common mental illness was associated with reduced gray matter volumes in the occipital lobes and neocerebellum. Furthermore, higher p factor scores were associated with reduced fractional anisotropy in pontine pathways linking the neocerebellum with the thalamus and prefrontal cortex. Our observed patterns along with those of the recent meta-analysis suggest that higher p factor scores contribute to broad risk for common forms of mental illness through alterations in networks critical for feed-forward monitoring of information processing and executive control. However, the putative functional consequences of these observed structural associations have not yet been examined.
Resting-state functional connectivity has emerged as a powerful tool in clinical neuroscience because it can be readily administered across patient populations9,10, demonstrates trait-like stability11 as well as moderate heritability12,13, and represents a powerful probe of the intrinsic architecture of neural networks that play a primary role in shaping task-based network activity and associated behaviors14. In addition, altered intrinsic functional connectivity within the default mode network (DMN), and frontoparietal network (FPN), both of which are linked to higher order cognition, have been broadly linked to psychopathology across categorical disorders15–17. Thus, resting-state measures of intrinsic network connectivity represent one avenue for extending the structural associations of the p factor to variability in functional neural dynamics representing mechanisms through which risk may emerge.
Here, we investigate intrinsic functional connectivity correlates of the p factor in a volunteer sample of 614 university students from the Duke Neurogenetics Study. While our previous research in this sample identified discrete structural correlates of the p factor in the occipital lobes, neocerebellum, and pons, we opted for a whole-brain exploratory analysis of intrinsic connectivity to better capture possible functional differences beyond these regions and impose minimal assumptions about the nature of p factor associations in the brain. Thus, we performed a Connectome-Wide Association Study (CWAS)18 of the p factor, a data driven method for identifying resting-state seeds whose whole brain connectivity are associated with the p factor in the absence of a priori assumptions about brain regions or networks.
Results
Demographics
From the 614 participants who completed two resting-state scans, 605 had data that survived quality control procedures. Of these, 336 were women, and the mean age was 20.23±1.19 years old. Scores for the p factor ranged from 76.71to 191.96 with a mean of 99.80, sd of 15.39.
Multi-dimensional matrix regression
Of the 1015 ROIs investigated, multi-dimensional matrix regression (MDMR) analysis revealed four regions with whole-brain connectivity patterns significantly associated with p factor scores: left lingual gyrus, right middle occipital gyrus, and two adjacent parcels of the left middle occipital gyrus (figure 1).
Follow-up intrinsic connectivity analyses
The follow-up connectivity analyses of each seed identified through MDMR revealed the primary network associations for each seed as well as their pattern of whole-brain connectivity associated with p factor scores. These analyses showed striking convergence across MDMR selected ROIs wherein the mean whole-brain pattern of connectivity for each seed showed subtle variation, but largely outlined the canonical resting-state visual processing network19. The connectivity of each ROI with visual and somatosensory regions decreased with increasing p factor scores, while the connectivity between each ROI and transmodal association regions20 increased with increasing p scores (figure 2).
Further analyses were conducted to better characterize the above consistent patterns of p factor associations with the intrinsic connectivity of all seeds by averaging the independent whole- brain connectivity maps. The resulting average z-scores were summarized for each of the 7 Yeo networks21 to quantify their respective contribution to the associations with p factor scores (figure 3). These analyses revealed the DMN and FPN as the major networks for which intrinsic functional connectivity was positively correlated with p factor scores. In contrast, a more modest but notable negative correlation was observed between p factor scores and the intrinsic functional connectivity between the visual association cortex and somatomotor network.
Discussion
Here, we provide a novel extension of prior structural neural correlates of the p factor to the intrinsic architecture of the whole-brain functional connectome. Our unconstrained connectome wide MDMR analysis revealed a circumscribed relationship between p factor scores and the whole-brain intrinsic connectivity of nodes in visual association cortex, which is consistent with our earlier work finding a negative correlation between p factor scores and gray matter volume in the occipital cortex8. Further investigation of the patterns of intrinsic connectivity driving this relationship primarily implicated hyper-connectivity between visual association cortex and heteromodal frontoparietal and default mode networks. While the visual association cortex and the heteromodal default mode and frontoparietal networks represent opposite ends of the sensory processing hierarchy,20 their dynamic interactions have been shown to be an important component of successful goal-directed behavior22–24.
The frontoparietal network in particular has been linked to the core cognitive faculty of executive control,14,25,26 which contributes to mental health and general well-being by shaping successful goal directed behavior27. Fittingly, disrupted FPN activity has been linked to psychopathology across categorical disorders including schizophrenia28, depression,29 and bipolar disorder30. Building off of this body of research, an emerging theory suggests that the relative integrity of the FPN and associated executive control mechanisms are fundamental for the capacity to self-regulate, manage symptoms, and succeed in treatment across disorders.15. Our current findings are consistent with this framework by demonstrating that higher p factor scores regardless of diagnosis are associated with relative hyper-connectivity of the FPN with the visual association cortex.
Regulation and suppression of incoming sensory information is an important component of goal directed behavior and has been linked to functional connectivity between the FPN and visual association cortex31,32. Although speculative, our finding may indicate more effortful or less efficient integration of bottom-up sensory information with attentional demands and executive control processes in those at higher risk for mental illness. This pattern is further consistent with our earlier structural analyses linking higher p factor scores with lower structural integrity of cerebello-thalamo-cerebro-cortical circuits supporting feed-forward executive control and monitoring of goal-directed behaviors8.
In addition to the frontoparietal network, our analyses implicate hyper-connectivity between the visual association cortex and default mode network as a function of higher p factor scores. The default mode network has been generally linked to introspective thought, autobiographical memory, and future oriented thought33. Interestingly, DMN activity is suppressed in attention demanding tasks33,34 and altered DMN activity has been broadly observed across categorical psychiatric disorders16,35. Visual association cortex connectivity with the DMN has been suggested to be important in the suppression of internally generated distracting information31. Against this background, our observed association between higher p factor scores and hyper-connectivity between the DMN and visual association cortex suggests that broad risk for mental illness may be related to more effortful or less efficient regulation of internally generated thoughts and information that could have functional consequences in the context of competing attentional demands between internal generated thought and incoming sensory information.
While providing initial evidence that broad risk for all forms of common mental illness is manifest as alterations in the intrinsic connectivity of functional neural networks, our analyses were exploratory by design and replication is needed. Given prior research implicating the frontoparietal and default mode networks across categorical disorders, we focused our above discussion on the potential relevance of intrinsic connectivity between visual association cortex and these networks in the emergence of broad risk for mental illness. While the intrinsic connectivity of these networks also exhibited an outsized influence on the association with p factor scores, variation between visual association cortex and other resting-state networks contributed as well, albeit more modestly (figure 3). MDMR uses information from all whole-brain connections in selecting seeds, and the inferential significance comes from the aggregate of connections rather than any one in particular. Thus, formally testing the relative contributions of different networks is not typically conducted. While we think future studies of the p factor will benefit from using our observations of intrinsic connectivity between visual association cortex and both DMN and FPN as an a priori starting point, the potential relevance of other networks should not be ignored until the patterns reported herein are replicated.
Additional limitations, which can be addressed in future research, include the relatively limited range of psychopathology, especially severe forms including psychosis, represented in our volunteer sample of young adults. Future research should extend our analyses to more diverse populations including individuals with severe mental illness. Our current analyses were also limited to the intrinsic connectivity of nodes within the cerebrum as our resting-state fMRI acquisition protocol did not afford full coverage of the cerebellum, including the neocerebellar subregion identified in our earlier structural analyses. Thus, we are unable to determine the relationship between p factor scores and the intrinsic functional connectivity of the cerebellum. We anticipate that current state-of-the-art multiband image acquisition protocols will routinely allow for full coverage of the cerebellum and, subsequently, direct analyses of how its intrinsic connectivity scales as a function of p factor scores. The observational nature of our study represents another limitation as we cannot establish causal links between p factor scores and intrinsic connectivity. Longitudinal designs may better address causality and temporal order of these phenomena. Future research employing transcranial magnetic stimulation, closed-loop fMRI, and intervention designs can further map causal relationships.
These limitations notwithstanding, our current work provides initial evidence for unique connectome wide functional signatures of the p factor. Consistent with emerging transdiagnostic and dimensional research into the neural basis of psychopathology7,8,36, our analyses reveal that increased broad risk for all common forms of mental illness is associated with higher intrinsic connectivity between visual association cortex and both frontoparietal and default mode networks. Such hyper-connectivity suggests that increased risk for psychopathology may be manifest as greater effortful or less efficient executive control as well as poor regulation of self-referential information processing. These patterns place alterations of the functional connectome squarely in the middle of converging theories of network dysfunction in psychopathology, further suggesting the p factor as a promising tool in clinical neuroscience.
Online Methods
Participants
Data were available from 614 university students who successfully completed the Duke Neurogenetics Study (DNS). All participants provided informed consent in accordance with the Duke University Medical Center Institutional Review Board guidelines before participation. All participants were in good general health and free of the following study conditions: (1) medical diagnoses of cancer, stroke, head injury with loss of consciousness, untreated migraine headaches, diabetes requiring insulin treatment, chronic kidney or liver disease; (2) use of psychotropic, glucocorticoid or hypolipidemic medication; and (3) conditions affecting cerebral blood flow and metabolism (e.g., hypertension).
Clinical Diagnosis
Current and lifetime DSM-IV Axis I disorder or select Axis II disorders (antisocial personality disorder and borderline personality disorder), was assessed with the electronic Mini International Neuropsychiatric Interview37 and Structured Clinical Interview for the DSM-IV subtests38 respectively. Importantly, neither current nor lifetime diagnosis were exclusion criterion, as the DNS seeks to establish broad variability in multiple behavioral phenotypes related to psychopathology. Allowing for a broad spectrum of symptoms is particularly critical for accurately deriving p factor scores. Nevertheless, no participants, regardless of diagnosis, were taking any psychoactive medication during or at least 14 days prior to their participation. Of the 605 participants with data included in our analyses, 133 individuals had at least one DSM-IV diagnosis, including 76 with alcohol use disorders, 24 with non-alcohol substance use disorders, 33 with major depression disorder, 26 with bipolar disorder, 7 with panic disorder (no agoraphobia), 9 with panic disorder including agoraphobia, 4 with social anxiety disorder, 8 with generalized anxiety disorder, 10 with obsessive compulsive disorder, and 7 with eating disorders.
Derivation of p factor scores
In previous work8, our group replicated the p factor in the DNS using confirmatory factor analysis of self-report and diagnostic interview measures of internalizing, externalizing, and thought disorder symptoms. These p factor scores were extracted using the standard regression method from those analyses, and standardized to a mean of 100 (SD = 15), with higher scores indicating a greater propensity to experience all forms of psychiatric symptoms. The current analyses were conducted in a subsample of 614 subjects of the original 1246 participants for whom there was resting-state fMRI data.
Image acquisition
Each participant was scanned using one of two identical research-dedicated GE MR750 3 T scanners equipped with high-power high-duty-cycle 50-mT/m gradients at 200 T/m/s slew rate, and an eight-channel head coil for parallel imaging at high bandwidth up to 1MHz at the Duke-UNC Brain Imaging and Analysis Center. A semi-automated high-order shimming program was used to ensure global field homogeneity. A series of 34 interleaved axial functional slices aligned with the anterior commissure-posterior commissure plane were acquired for full- brain coverage using an inverse-spiral pulse sequence to reduce susceptibility artifacts (TR/TE/flip angle=2000 ms/30 ms/60; FOV=240mm; 3.75×3.75×4mm voxels; interslice skip=0). Four initial radiofrequency excitations were performed (and discarded) to achieve steady-state equilibrium. For each participant, 2 back-to-back 4-minute 16-second resting state functional MRI scans were acquired. Participants were instructed to remain awake, with their eyes open during each resting state scan. To allow for spatial registration of each participant’s data T1-weighted images were obtained using a 3D Ax FSPGR BRAVO with the following parameters: TR = 8.148 ms; TE = 3.22 ms; 162 axial slices; flip angle, 12°; FOV, 240 mm; matrix =256×256; slice thickness = 1 mm with no gap; and total scan time = 4 min and 13 s.
Image Processing
Anatomical images for each subject were skull-stripped, intensity-normalized, and nonlinearly warped to a study-specific average template in the standard stereotactic space of the Montreal Neurological Institute template using the ANTs SyN registration algorithm39,40. Time series images for each subject were despiked, slice-time-corrected, realigned to the first volume in the time series to correct for head motion using AFNI tools41, coregistered to the anatomical image using FSL’s Boundary Based Registration42, spatially normalized into MNI space using the non- linear ANTs SyN warp from the anatomical image, resampled to 2mm isotropic voxels, and smoothed to minimize noise and residual difference in gyral anatomy with a Gaussian filter set at 6-mm full-width at half-maximum. All transformations were concatenated so that a single interpolation was performed.
Time-series images for each participant were furthered processed to limit the influence of motion and other artifacts. Voxel-wise signal intensities were scaled to yield a time series mean of 100 for each voxel. Motion regressors were created using each subject’s 6 motion correction parameters (3 rotation and 3 translation) and their first derivatives43,44 yielding 12 motion regressors. White matter (WM) and cerebrospinal fluid (CSF) nuisance regressors were created using CompCorr45. Images were bandpass filtered to retain frequencies between .008 and .1 Hz, and volumes exceeding 0.25mm frame-wise displacement or 1.55 standardized DVARS46,47 were censored. Nuisance regression, bandpass filtering and censoring for each time series was performed in a single processing step using AFNI’s 3dTproject. Participants were excluded if they had less than 185 TRs left after censoring (resulting in inadequate degrees of freedom to perform nuisance regressions), resulting in a final sample of 605 subjects.
CWAS
To make the analysis computationally tractable, time-series were extracted from a parcellated atlas instead of using voxelwise data. We used the Lausanne atlas parcellated into 1015 equally sized regions through the program easy_lausanne (github.com/mattcieslak/easy_lausanne). Time-series data for each subject were then processed using CWAS. Described extensively elsewhere18, CWAS consists of 3 processing steps. First, beginning with a single ROI time-series, seed-based connectivity analysis is conducted to generate a whole-brain functional connectivity map for each participant. Second, the average distance (1 minus the Pearson correlation) between each pair of participant’s functional connectivity maps is computed, resulting in a distance matrix encoding the multivariate similarity between each participant’s connectivity map. Finally, multi-dimensional matrix regression (MDMR) is used to generate a pseudo-F statistic quantifying the strength of the association between the phenotype of interest, here p factor scores, and the distance matrix created in the second step. The advantage of MDMR is allowing covariates to be entered into the regression and utilizing non-parametric permutation to generate p-values for each ROI. These three steps are repeated for each of the 1015 ROIs, resulting in a whole-brain map that represents the association between p factor scores and whole-brain connectivity at each ROI. CWAS was performed to identify seed regions with whole- brain patterns of connectivity are related to p factor scores. Participant sex was included as a covariate, and 500,000 permutations were performed to generate p-values. To minimize false positives across the 1,015 ROIs, a false discovery rate48 (FDR) correction was applied.
Seed-based analyses
MDMR identifies a set of ROIs with patterns of whole-brain connectivity associated with p factor scores. However, it is still unclear how the connectivity of these ROIs relates to the scores. Previous research using CWAS18,36,49 has demonstrated the utility of using traditional seed-based connectivity follow-up analyses to better understand the networks and brain regions that drive the associations discovered through MDMR. Similar analyses were performed here for each ROI identified via MDMR. Seed-based connectivity maps were created and correlations were converted to Z statistics via the Fischer R to Z transform. Whole-brain correlations between these connectivity values and p factor scores were calculated, including sex as a covariate. Importantly, these follow-up analyses do not represent independent statistical tests as they were performed posthoc to the family wise error controlled MDMR findings. Accordingly, these followup analyses maps are not thresholded to visualize all information that was relevant to the MDMR step.
Acknowledgements and Disclosures
We thank the Duke Neurogenetics Study participants and the staff of the Laboratory of NeuroGenetics. The Duke Neurogenetics Study received support from Duke University as well as US-National Institutes of Health grants R01DA033369 and R01DA031579. ARK, and ARH received further support from US-National Institutes of Health grant R01AG049789. MLE was supported by the National Science Foundation Graduate Research Fellowship under Grant No. NSF DGE-1644868. The authors declare no competing financial interests.