Multimodal neural correlates of childhood psychopathology

Article activity feed

1. 

   Version published to 10.1101/2023.03.02.530821v2 on bioRxiv

   Aug 17, 2024
2. 

   eLife

   Aug 15, 2024

   Author response:

   Reviewer #1 - Public Review

   This report describes work aiming to delineate multi-modal MRI correlates of psychopathology from a large cohort of children of 9-11 years from the ABCD cohort. While uni-modal characterisations have been made, the authors rightly argue that multi-modal approaches in imaging are vital to comprehensively and robustly capture modes of large-scale brain variation that may be associated with pathology. The primary analysis integrates structural and resting-state functional data, while post-hoc analyses on subsamples incorporate task and diffusion data. Five latent components (LCs) are identified, with the first three, corresponding to p-factor, internal/externalising, and neurodevelopmental Michelini Factors, described in detail. In addition, associations of these components with primary and …

   Author response:

   Reviewer #1 - Public Review

   This report describes work aiming to delineate multi-modal MRI correlates of psychopathology from a large cohort of children of 9-11 years from the ABCD cohort. While uni-modal characterisations have been made, the authors rightly argue that multi-modal approaches in imaging are vital to comprehensively and robustly capture modes of large-scale brain variation that may be associated with pathology. The primary analysis integrates structural and resting-state functional data, while post-hoc analyses on subsamples incorporate task and diffusion data. Five latent components (LCs) are identified, with the first three, corresponding to p-factor, internal/externalising, and neurodevelopmental Michelini Factors, described in detail. In addition, associations of these components with primary and secondary RSFC functional gradients were identified, and LCs were validated in a replication sample via assessment of correlations of loadings.

   1.1) This work is clearly novel and a comprehensive study of associations within this dataset. Multi-modal analyses are challenging to perform, but this work is methodologically rigorous, with careful implementation of discovery and replication assessments, and primary and exploratory analyses. The ABCD dataset is large, and behavioural and MRI protocols seem appropriate and extensive enough for this study. The study lays out comprehensive associations between MRI brain measures and behaviour that appear to recapitulate the established hierarchical structure of psychopathology.

   We thank Reviewer 1 for appreciating our methods and findings, and we address their suggestions below:

   1.2) The work does have weaknesses, some of them acknowledged. There is limited focus on the strength of observed associations. While the latent component loadings seem reliably reproducible in the behavourial domain, this is considerably less the case in the imaging modalities. A considerable proportion of statistical results focuses on spatial associations in loadings between modalities - it seems likely that these reflect intrinsic correlations between modalities, rather than associations specific to any latent component.

   We appreciate the Reviewer’s comment, and minimized the reporting of correlations between the loadings from the different modalities in the revised Results (specifically subsections on LC1, LC2, and LC3). We now refer to Table S4 in each subsection for this information: “Spatial correlations between modality-specific loadings are reported in Supplementary file 1c.”

   For completeness, we report the intrinsic correlations between the different modalities in Supplementary file 1c (P.19):

   “Lastly, although the current work aimed to reduce intrinsic correlations between variables within a given modality through running a PCA before the PLS approach, intrinsic correlations between measures and modalities may potentially be a remaining factor influencing the PLS solution. We, thus, provided an additional overview of the intrinsic correlations between the different neuroimaging data modalities in the supporting results (Supplementary file 1c).”

   1.3) Assessment of associations with functional gradients is similarly a little hard to interpret. Thus, it is hard to judge the implications for our understanding of the neurophysiological basis of psychopathology and the ability of MRI to provide clinical tools for, say, stratification.

   We now provide additional context, including a rising body of theoretical and empirical work, that outlines the value of functional gradients and cortical hierarchies in the understanding of brain development and psychopathology. Please see P.26.

   “Initially demonstrated at the level of intrinsic functional connectivity (Margulies et al., 2016), follow up work confirmed a similar cortical patterning using microarchitectural in-vivo MRI indices related to cortical myelination (Burt et al., 2018; Huntenburg et al., 2017; Paquola et al., 2019), post-mortem cytoarchitecture (Goulas et al., 2018; Paquola et al., 2020, 2019), or post-mortem microarray gene expression (Burt et al., 2018). Spatiotemporal patterns in the formation and maturation of large-scale networks have been found to follow a similar sensory-to-association axis; moreover, there is the emerging view that this framework may offer key insights into brain plasticity and susceptibility to psychopathology (Sydnor et al., 2021). In particular, the increased vulnerability of transmodal association cortices in late childhood and early adolescence has been suggested to relate to prolonged maturation and potential for plastic reconfigurations of these systems (Paquola et al., 2019; Park et al., 2022b). Between mid-childhood and early adolescence, heteromodal association systems such as the default network become progressively more integrated among distant regions, while being more differentiated from spatially adjacent systems, paralleling the development of cognitive control, as well as increasingly abstract and logical thinking. [...] This suggests that neurodevelopmental difficulties might be related to alterations in various processes underpinned by sensory and association regions, as well as the macroscale balance and hierarchy of these systems, in line with previous findings in several neurodevelopmental conditions, including autism, schizophrenia, as well as epilepsy, showing a decreased differentiation between the two anchors of this gradient (Hong et al., 2019). In future work, it will be important to evaluate these tools for diagnostics and population stratification. In particular, the compact and low dimensional perspective of gradients may provide beneficial in terms of biomarker reliability as well as phenotypic prediction, as previously demonstrated using typically developing cohorts (Hong et al. 2020) On the other hand, it will be of interest to explore in how far alterations in connectivity along sensory-to-transmodal hierarchies provide sufficient graduality to differentiate between specific psychopathologies, or whether they, as the current work suggests, mainly reflect risk for general psychopathology and atypical development.”

   1.4) The observation of a recapitulation of psychopathology hierarchy may be somewhat undermined by the relatively modest strength of the components in the imaging domain.

   We thank the Reviewer for this comment, and now expressed this limitation in the revised Discussion, P.23.

   “The p factor, internalizing, externalizing, and neurodevelopmental dimensions were each associated with distinct morphological and intrinsic functional connectivity signatures, although these relationships varied in strength.”

   1.5) The task fMRI was assessed with a fairly basic functional connectivity approach, not using task timings to more specifically extract network responses.

   In the revised Discussion on P.24, we acknowledge that more in-depth analyses of task-based fMRI may have offered additional insights into state-dependent changes in functional architecture.

   “While the current work derived main imaging signatures from resting-state fMRI as well as grey matter morphometry, we could nevertheless demonstrate associations to white matter architecture (derived from diffusion MRI tractography) and recover similar dimensions when using task-based fMRI connectivity. Despite subtle variations in the strength of observed associations, the latter finding provided additional support that the different behavioral dimensions of psychopathology more generally relate to alterations in functional connectivity. Given that task-based fMRI data offers numerous avenues for analytical exploration, our findings may motivate follow-up work assessing associations to network- and gradient-based response strength and timing with respect to external stimuli across different functional states.”

   1.6) Overall, the authors achieve their aim to provide a detailed multimodal characterisation of MRI correlations of psychopathology. Code and data are available and well organised and should provide a valuable resource for researchers wanting to understand MRI-based neural correlates of psycho-pathology-related behavioural traits in this important age group. It is largely a descriptive study, with comparisons to previous uni-modal work, but without particularly strong testing of neuroscience hypotheses.

   We thank the Reviewer for recognizing the detail and rigor of data-driven study and extensive code and data documentation.

   Reviewer #2 - Public Review

   In "Multi-modal Neural Correlates of Childhood Psychopathology" Krebets et al. integrate multi-modal neuroimaging data using machine learning to delineate dissociable links to diverse dimensions of psychopathology in the ABCD sample. This paper had numerous strengths including a superb use of a large resource dataset, appropriate analyses, beautiful visualizations, clear writing, and highly interpretable results from a data-driven analysis. Overall, I think it would certainly be of interest to a general readership. That being said, I do have several comments for the authors to consider.

   We thank Dr Satterthwaite for the positive evaluation and helpful comments.

   2.1) Out-of-sample testing: while the permutation testing procedure for the PLS is entirely appropriate, without out-of-sample testing the reported effect sizes are likely inflated.

   As discussed in the editorial summary of essential revisions, we agree that out-of-sample prediction indeed provides stronger estimates of generalizability. We assess this by applying the PCA coefficients derived from the discovery cohort imaging data to the replication cohort imaging data. The resulting PCA scores and behavioral data were then z-scored using the mean and standard deviation of the replication cohort. The SVD weights derived from the discovery cohort were applied to the normalized replication cohort data to derive imaging and behavioral composite scores, which were used to recover the contribution of each imaging and behavioral variable to the LCs (i.e., loadings). Out-of-sample replicability of imaging (mean r=0.681, S.D.=0.131) and behavioral (mean r=0.948, S.D.=0.022) loadings was generally high across LCs 1-5. This analysis is reported in the revised manuscript (P.18).

   “Generalizability of reported findings was also assessed by directly applying PCA coefficients and latent components weights from the PLS analysis performed in the discovery cohort to the replication sample data. Out-of-sample prediction was overall high across LCs1-5 for both imaging (mean r=0.681, S.D.=0.131) and behavioral (mean r=0.948, S.D.=0.022) loadings.”

   2.2) Site/family structure: it was unclear how site/family structure were handled as covariates.

   Only unrelated participants were included in discovery and replication samples (see P.6). The site variable was regressed out of the imaging and behavioral data prior to the PLS analysis using the residuals from a multiple linear model which also included age, age2, sex, and ethnicity. This is now clarified on P.29:

   “Prior to the PLS analysis, effects of age, age2, sex, site, and ethnicity were regressed out from the behavioral and imaging data using a multiple linear regression to ensure that the LCs would not be driven by possible confounders (Kebets et al., 2021, 2019; Xia et al., 2018). The imaging and behavioral residuals of this procedure were input to the PLS analysis.”

   2.3) Anatomical features: I was a bit surprised to see volume, surface area, and thickness all evaluated - and that there were several comments on the correspondence between the SA and volume in the results section. Given that cortical volume is simply a product of SA and CT (and mainly driven by SA), this result may be pre-required.

   As suggested, we reduced the reporting of correlations between the loadings from the different modalities in the revised Results (specifically subsections on LC1, LC2, and LC3). Instead, we now refer to Table S4 in each subsection for this information: “Spatial correlations between modality-specific loadings are reported in Supplementary file 1c.”

   We also reran the PLS analysis while only including thickness and surface area as our structural metrics, to account for potential redundancy of these measures with volume. This analysis and associated findings are reported on P.36 and P.19:

   “As cortical volume is a result of both thickness and surface area, we repeated our main PLS analysis while excluding cortical volume from our imaging metrics and report the consistency of these findings with our main model.”

   “Third, to account for redundancy within structural imaging metrics included in our main PLS model (i.e., cortical volume is a result of both thickness and surface area), we also repeated our main analysis while excluding cortical volume from our imaging metrics. Findings were very similar to those in our main analysis, with an average absolute correlation of 0.898±0.114 across imaging composite scores of LCs 1-5.”

   2.4) Ethnicity: the rationale for regressing ethnicity from the data was unclear and may conflict with current best practices.

   We thank the Reviewer for this comment. In light of recent discussions on including this covariate in large datasets such as ABCD (e.g., Saragosa-Harris et al., 2022), we elaborate on our rationale for including this variable in our model in the revised manuscript on P.30:

   “Of note, the inclusion of ethnicity as a covariate in imaging studies has been recently called into question. In the present study, we included this variable in our main model as a proxy for social inequalities relating to race and ethnicity alongside biological factors (age, sex) with documented effects on brain organization and neurodevelopmental symptomatology queried in the CBCL.”

   We also assess the replicability of our analyses when removing race and ethnicity covariates prior to computing the PLS analysis and correlating imaging and behavioral composite scores across both models. We report resulting correlations in the revised manuscript (P.37, 19, and 27):

   “We also assessed the replicability of our findings when removing race and ethnicity covariates prior to computing the PLS analysis and correlating imaging and behavioral composite scores across both models.”

   “Moreover, repeating the PLS analysis while excluding this variable as a model covariate yielded overall similar imaging and behavioral composites scores across LCs to our original analysis. Across LCs 1-5, the average absolute correlations reached r=0.636±0.248 for imaging composite scores, and r=0.715±0.269 for behavioral composite scores. Removing these covariates seemed to exert stronger effects on LC3 and LC4 for both imaging and behavior, as lower correlations across models were specifically observed for these components.”

   “Although we could consider some socio-demographic variables and proxies of social inequalities relating to race and ethnicity as covariates in our main model, the relationship of these social factors to structural and functional brain phenotypes remains to be established with more targeted analyses.”

   2.5) Data quality: the authors did an admirable job in controlling for data quality in the analyses of functional connectivity data. However, it is unclear if a comparable measure of data quality was used for the T1/dMRI analyses. This likely will result in inflated effect sizes in some cases; it has the potential to reduce sensitivity to real effects.

   We agree that data quality was not accounted for in our analysis of T1w- and diffusion-derived metrics. We now accounted for T1w image quality by adding manual quality control ratings to the regressors applied to all structural imaging metrics prior to performing the PLS analysis, and reported the consistency of this new model with original findings. See P.36, P.19:

   “We also considered manual quality control ratings as a measure of T1w scan quality. This metric was included as a covariate in a multiple linear regression model accounting for potential confounds in the structural imaging data, in addition to age, age2, sex, site, ethnicity, ICV, and total surface area. Downstream PLS results were then benchmarked against those obtained from our main model.”

   “Considering scan quality in T1w-derived metrics (from manual quality control ratings) yielded similar results to our main analysis, with an average correlation of 0.986±0.014 across imaging composite scores.”

   As for diffusion imaging, we also regressed out effects of head motion in addition to age, age2, sex, site, and ethnicity from FA and MD measures and reported the consistency with our original results (P.36, P.19):

   “We tested another model which additionally included head motion parameters as regressors in our analyses of FA and MD measures, and assessed the consistency of findings from both models.”

   “Additionally considering head motion parameters from diffusion imaging metrics in our model yielded consistent results to those in our main analyses (mean r=0.891, S.D.=0.103; r=0.733-0.998).”

   Reviewer #3 - Public Review

   In this study, the authors utilized the Adolescent Brain Cognitive Development dataset to investigate the relationship between structural and functional brain network patterns and dimensions of psychopathology. They identified multiple components, including a general psychopathology (p) factor that exhibited a strong association with multimodal imaging features. The connectivity signatures associated with the p factor and neurodevelopmental dimensions aligned with the sensory-to-transmodal axis of cortical organization, which is linked to complex cognition and psychopathology risk. The findings were consistent across two separate subsamples and remained robust when accounting for variations in analytical parameters, thus contributing to a better understanding of the biological mechanisms underlying psychopathology dimensions and offering potential brain-based vulnerability markers.

   3.1) An intriguing aspect of this study is the integration of multiple neuroimaging modalities, combining structural and functional measures, to comprehensively assess the covariance with various symptom combinations. This approach provides a multidimensional understanding of the risk patterns associated with mental illness development.

   We thank the Reviewer for acknowledging the multimodal approach, and for the constructive suggestions.

   3.2) The paper delves deeper into established behavioral latent variables such as the p factor, internalizing, externalizing, and neurodevelopmental dimensions, revealing their distinct associations with morphological and intrinsic functional connectivity signatures. This sheds light on the neurobiological underpinnings of these dimensions.

   We are happy to hear the Reviewer appreciates the gain in understanding neural underpinnings of dimensions of psychopathology resulting from the current work.

   3.3) The robustness of the findings is a notable strength, as they were validated in a separate replication sample and remained consistent even when accounting for different parameter variations in the analysis methodology. This reinforces the generalizability and reliability of the results.

   We appreciate that the Reviewer found our robustness and generalizability assessment convincing.

   3.4) Based on their findings, the authors suggest that the observed variations in resting-state functional connectivity may indicate shared neurobiological substrates specific to certain symptoms. However, it should be noted that differences in resting-state connectivity between groups can stem from various factors, as highlighted in the existing literature. For instance, discrepancies in the interpretation of instructions during the resting state scan can influence the results. Hence, while their findings may indicate biological distinctions, they could also reflect differences in behavior.

   For the ABCD dataset, resting-state fMRI scans were based on eyes open and passive viewing of a crosshair, and are thus homogenized. We acknowledge, however, that there may still be state-to-state fluctuations contributing to the findings, and this is now discussed in the revised Discussion, on P.28. Note, however, that prior literature has generally also suggested rather modest impacts of cognitive and daily variation on resting-state functional networks, compared to much more dominating inter-individual and inter-group factors.

   “Finally, while prior research has shown that resting-state fMRI networks may be affected by differences in instructions and study paradigm (e.g., with respect to eyes open vs closed) (Agcaoglu et al., 2019), the resting-state fMRI paradigm is homogenized in the ABCD study to be passive viewing of a centrally presented fixation cross. It is nevertheless possible that there were slight variations in compliance and instructions that contributed to differences in associated functional architecture. Notably, however, there is a mounting literature based on high-definition fMRI acquisitions suggesting that functional networks are mainly dominated by common organizational principles and stable individual features, with substantially more modest contributions from task-state variability (Gratton et al. 2018). These findings, thus, suggest that resting-state fMRI markers can serve as powerful phenotypes of psychiatric conditions, and potential biomarkers (Abraham et al., 2017; Gratton et al., 2020; Parkes et al., 2020).”

   3.5) The authors conducted several analyses to investigate the relationship between imaging loadings associated with latent components and the principal functional gradient. They found several associations between principal gradient scores and both within- and between-network resting-state functional connectivity (RSFC) loadings. Assessing the analysis presented here proves challenging due to the nature of relating loadings, which are partly based on the RSFC, to gradients derived from RSFC. Consequently, a certain level of correlation between these two variables would be expected, making it difficult to determine the significance of the authors' findings. It would be more intriguing if a direct correlation between the composite scores reflecting behavior and the gradients were to yield statistically significant results.

   We thank the Reviewer for the comment, and agree that investigating gradient-behavior relationships could offer additional insights into the neural basis of psychiatric symptomatology. However, the current analysis pipeline precludes this direct comparison which is performed on a region-by-region basis across the span of the cortical gradient. Indeed, the behavioral loadings are provided for each CBCL item, and not cortical regions.

   The Reviewer also evokes concerns of potential circularity in our analysis, as we compared imaging loadings, which are partially based on RSFC, and gradient values generated from the same RSFC data. In response to this comment, we cross-validated our findings using an RSFC gradient derived from an independent dataset (HCP), showing highly consistent findings to those presented in the manuscript. This correlation is now reported in the Results section P.15.

   “A similar pattern of findings was observed when cross-validating between- and within-network RSFC loadings to a RSFC gradient derived from an independent dataset (HCP), with strongest correlations seen for between-network RSFC loadings for LC1 and LC3 (LC1: r=0.50, pspin<0.001; LC3: r=0.37, pspin<0.001).”

   We furthermore note similar correlations between imaging loadings and T1w/T2w ratio in the same participants, a proxy of intracortical microstructure and hierarchy (Glasser et al., 2011). These findings are now detailed in the revised Results, P.15-16:

   “Of note, we obtain similar correlations when using T1w/T2w ratio in the same participants, a proxy of intracortical microstructure and hierarchy (Glasser et al., 2011). Specifically, we observed the strongest association between this microstructural marker of the cortical hierarchy and between-network RSFC loadings related to LC1 (r=-0.43, pspin<0.001).”

   3.6) Lastly, regarding the interpretation of the first identified latent component, I have some reservations. Upon examining the loadings, it appears that LC1 primarily reflects impulse control issues rather than representing a comprehensive p-factor. Furthermore, it is worth noting that within the field, there is an ongoing debate concerning the interpretation and utilization of the p-factor. An insightful publication on this topic is "The p factor is the sum of its parts, for now" (Fried et al, 2021), which explains that the p-factor emerges as a result of a positive manifold, but it does not necessarily provide insights into the underlying mechanisms that generated the data.

   We thank the Reviewer for this comment, and added greater nuance into the discussion of the association to the p factor. We furthermore discuss some of the ongoing debate about the use of the p factor, and cite the recommended publication on P.27.

   “Other factors have also been suggested to impact the development of psychopathology, such as executive functioning deficits, earlier pubertal timing, negative life events (Brieant et al., 2021), maternal depression, or psychological factors (e.g., low effortful control, high neuroticism, negative affectivity). Inclusion of such data could also help to add further mechanistic insights into the rather synoptic proxy measure of the p factor itself (Fried et al., 2021), and to potentially assess shared and unique effects of the p factor vis-à-vis highly correlated measures of impulse control.”

   Read the original source
3. 

   eLife

   Aug 10, 2024

   eLife assessment

   This important study explores and delineates multivariate mappings between brain structure and functional measures with latent dimensions of psychopathology. This work provides solid evidence for the existence of such mappings and charts the relationship between different neurobiological measures and distinct dimensions of psychopathology. This work will be of broad interest within the neuroscience field.

   Read the original source
4. 

   eLife

   Aug 10, 2024

   Reviewer #1 (Public Review):

   This report describes work aiming to delineate multi-modal MRI correlates of psychopathology from a large cohort of children of 9-11 years from the ABCD cohort. While uni-modal characterisations have been made, the authors rightly argue that multi-modal approaches in imaging are vital to comprehensively and robustly capture modes of large-scale brain variation that may be associated with pathology. The primary analysis integrates structural and resting-state functional data, while post-hoc analyses on subsamples incorporate task and diffusion data. Five latent components (LCs) are identified, with the first three, corresponding to p-factor, internal/externalising, and neurodevelopmental Michelini Factors, described in detail. In addition, associations of these components with primary and secondary RSFC …

   Reviewer #1 (Public Review):

   This report describes work aiming to delineate multi-modal MRI correlates of psychopathology from a large cohort of children of 9-11 years from the ABCD cohort. While uni-modal characterisations have been made, the authors rightly argue that multi-modal approaches in imaging are vital to comprehensively and robustly capture modes of large-scale brain variation that may be associated with pathology. The primary analysis integrates structural and resting-state functional data, while post-hoc analyses on subsamples incorporate task and diffusion data. Five latent components (LCs) are identified, with the first three, corresponding to p-factor, internal/externalising, and neurodevelopmental Michelini Factors, described in detail. In addition, associations of these components with primary and secondary RSFC functional gradients were identified, and LCs were validated in a replication sample via assessment of correlations of loadings.

   This work is clearly novel and a comprehensive study of associations within this dataset. Multi-modal analyses are challenging to perform, but this work is methodologically rigorous, with careful implementation of discovery and replication assessments, and primary and exploratory analyses. The ABCD dataset is large, and behavioural and MRI protocols seem appropriate and extensive enough for this study. The study lays out comprehensive associations between MRI brain measures and behaviour that appear to recapitulate the established hierarchical structure of psychopathology.

   The work does have weaknesses, some of them acknowledged. There is limited focus on the strength of observed associations. While the latent component loadings seem reliably reproducible in the behavourial domain, this is considerably less the case in the imaging modalities. A considerable proportion of statistical results focuses on spatial associations in loadings between modalities - it seems likely that these reflect intrinsic correlations between modalities, rather than associations specific to any latent component. Assessment of associations with functional gradients is similarly a little hard to interpret. Thus, it is hard to judge the implications for our understanding of the neurophysiological basis of psychopathology and the ability of MRI to provide clinical tools for, say, stratification. The observation of a recapitulation of psychopathology hierarchy may be somewhat undermined by the relatively modest strength of the components in the imaging domain. The task fMRI was assessed with a fairly basic functional connectivity approach, not using task timings to more specifically extract network responses.

   Overall, the authors achieve their aim to provide a detailed multimodal characterisation of MRI correlations of psychopathology. Code and data are available and well organised and should provide a valuable resource for researchers wanting to understand MRI-based neural correlates of psycho-pathology-related behavioural traits in this important age group. It is largely a descriptive study, with comparisons to previous uni-modal work, but without particularly strong testing of neuroscience hypotheses.

   Read the original source
5. 

   eLife

   Aug 10, 2024

   Reviewer #2 (Public Review):

   In "Multi-modal Neural Correlates of Childhood Psychopathology" Krebets et al. integrate multi-modal neuroimaging data using machine learning to delineate dissociable links to diverse dimensions of psychopathology in the ABCD sample. This paper had numerous strengths including a superb use of a large resource dataset, appropriate analyses, beautiful visualizations, clear writing, and highly interpretable results from a data-driven analysis. Overall, I think it would certainly be of interest to a general readership.

   That being said, I do have several comments for the authors to consider.

   - Out-of-sample testing: while the permutation testing procedure for the PLS is entirely appropriate, without out-of-sample testing the reported effect sizes are likely inflated.

   - Site/family structure: it was unclear how …

   Reviewer #2 (Public Review):

   In "Multi-modal Neural Correlates of Childhood Psychopathology" Krebets et al. integrate multi-modal neuroimaging data using machine learning to delineate dissociable links to diverse dimensions of psychopathology in the ABCD sample. This paper had numerous strengths including a superb use of a large resource dataset, appropriate analyses, beautiful visualizations, clear writing, and highly interpretable results from a data-driven analysis. Overall, I think it would certainly be of interest to a general readership.

   That being said, I do have several comments for the authors to consider.

   - Out-of-sample testing: while the permutation testing procedure for the PLS is entirely appropriate, without out-of-sample testing the reported effect sizes are likely inflated.

   - Site/family structure: it was unclear how site/family structure were handled as covariates.

   - Anatomical features: I was a bit surprised to see volume, surface area, and thickness all evaluated - and that there were several comments on the correspondence between the SA and volume in the results section. Given that cortical volume is simply a product of SA and CT (and mainly driven by SA), this result may be pre-required.

   - Ethnicity: the rationale for regressing ethnicity from the data was unclear and may conflict with current best practices.

   - Data quality: the authors did an admirable job in controlling for data quality in the analyses of functional connectivity data. However, it is unclear if a comparable measure of data quality was used for the T1/dMRI analyses. This likely will result in inflated effect sizes in some cases; it has the potential to reduce sensitivity to real effects.

   Read the original source
6. 

   eLife

   Aug 10, 2024

   Reviewer #3 (Public Review):

   In this study, the authors utilized the Adolescent Brain Cognitive Development dataset to investigate the relationship between structural and functional brain network patterns and dimensions of psychopathology. They identified multiple components, including a general psychopathology (p) factor that exhibited a strong association with multimodal imaging features. The connectivity signatures associated with the p factor and neurodevelopmental dimensions aligned with the sensory-to-transmodal axis of cortical organization, which is linked to complex cognition and psychopathology risk. The findings were consistent across two separate subsamples and remained robust when accounting for variations in analytical parameters, thus contributing to a better understanding of the biological mechanisms underlying …

   Reviewer #3 (Public Review):

   In this study, the authors utilized the Adolescent Brain Cognitive Development dataset to investigate the relationship between structural and functional brain network patterns and dimensions of psychopathology. They identified multiple components, including a general psychopathology (p) factor that exhibited a strong association with multimodal imaging features. The connectivity signatures associated with the p factor and neurodevelopmental dimensions aligned with the sensory-to-transmodal axis of cortical organization, which is linked to complex cognition and psychopathology risk. The findings were consistent across two separate subsamples and remained robust when accounting for variations in analytical parameters, thus contributing to a better understanding of the biological mechanisms underlying psychopathology dimensions and offering potential brain-based vulnerability markers.

   Strengths:
   - An intriguing aspect of this study is the integration of multiple neuroimaging modalities, combining structural and functional measures, to comprehensively assess the covariance with various symptom combinations. This approach provides a multidimensional understanding of the risk patterns associated with mental illness development.

   - The paper delves deeper into established behavioral latent variables such as the p factor, internalizing, externalizing, and neurodevelopmental dimensions, revealing their distinct associations with morphological and intrinsic functional connectivity signatures. This sheds light on the neurobiological underpinnings of these dimensions.

   - The robustness of the findings is a notable strength, as they were validated in a separate replication sample and remained consistent even when accounting for different parameter variations in the analysis methodology. This reinforces the generalizability and reliability of the results.

   Weaknesses:

   - Based on their findings, the authors suggest that the observed variations in resting-state functional connectivity may indicate shared neurobiological substrates specific to certain symptoms. However, it should be noted that differences in resting-state connectivity between groups can stem from various factors, as highlighted in the existing literature. For instance, discrepancies in the interpretation of instructions during the resting state scan can influence the results. Hence, while their findings may indicate biological distinctions, they could also reflect differences in behavior.

   - The authors conducted several analyses to investigate the relationship between imaging loadings associated with latent components and the principal functional gradient. They found several associations between principal gradient scores and both within- and between-network resting-state functional connectivity (RSFC) loadings. Assessing the analysis presented here proves challenging due to the nature of relating loadings, which are partly based on the RSFC, to gradients derived from RSFC. Consequently, a certain level of correlation between these two variables would be expected, making it difficult to determine the significance of the authors' findings. It would be more intriguing if a direct correlation between the composite scores reflecting behavior and the gradients were to yield statistically significant results.

   - Lastly, regarding the interpretation of the first identified latent component, I have some reservations. Upon examining the loadings, it appears that LC1 primarily reflects impulse control issues rather than representing a comprehensive p-factor. Furthermore, it is worth noting that within the field, there is an ongoing debate concerning the interpretation and utilization of the p-factor. An insightful publication on this topic is "The p factor is the sum of its parts, for now" (Fried et al, 2021), which explains that the p-factor emerges as a result of a positive manifold, but it does not necessarily provide insights into the underlying mechanisms that generated the data.

   Read the original source
7. 

   Version published to 10.1101/2023.03.02.530821v1 on bioRxiv

   Mar 3, 2023