Computational protocol: Genetic and Environmental Contributions to Functional Connectivity Architecture of the Human Brain

Similar protocols

Protocol publication

[…] Image processing steps were performed using the Connectome Computation System (CCS, ;, which integrates the functionalities of AFNI (), FSL (), and FreeSurfer () with shell and MATLAB scripts. The structural images were first de-noised using a spatially adaptive nonlocal means filter (; ) and further processed to reconstruct the brain cortical surface (; , ; , ). The rfMRI functional images were preprocessed with the following steps: 1) deleting the first 5 EPI volumes; 2) de-spiking time series; 3) correcting slice timing; 4) aligning each volume to the mean EPI image; 5) normalizing the 4D global mean intensity (10 000); 6) band-pass (0.01–0.1 Hz) filtering the time series; 7) estimating a rigid transformation from individual functional images to the individual anatomical image using the GM-WM boundary-based registration (BBR) algorithm (), and 8) a nonlinear transformation between individual anatomical space to the MNI152 template (FNIRT in FSL, 1 mm isotropic). [...] The analytic pipelines are illustrated in Figure . First, generalized ranking and averaging independent component analysis by reproducibility (gRAICAR) (, ) was applied to the template dataset (N = 105), with all participants having a mean FD < 0.2 mm. This provided a set of unbiased templates that reflected common ICNs that would be detectable in the main dataset. In this way, the registration accuracy of the ICNs between the templates and the main dataset was maximized, since the 2 datasets were acquired using the same imaging parameters such as field strength and resolution. Specifically, after preprocessing and quality control procedures described above, spatial independent component analysis was applied to the rfMRI dataset using the MELODIC module of FSL software (). The numbers of components were automatically estimated by MELODIC runs. The resultant spatial components from individual participants were pooled, and gRAICAR matched the components from different participants and allocated them into clusters based on their spatial similarity measured using mutual information. For each cluster, a participant could contribute at most one component. A total intersubject spatial similarity among the member component maps (each representing a subject) was calculated to indicate the cross-subject consistency of the cluster. The member component maps in the cluster were weighted averaged to generate a cluster-wise component map. The weights in the averaging were determined by the degree centrality of the individual component maps in the cluster. The degree centrality significance was determined using permutation tests in which components from individual participants were randomly matched, yielding probability values that reflect contributions of individual participants to the clusterwise components. The effectiveness of gRAICAR has been demonstrated in multiple studies (; ; ). Ranking the clusterwise component maps by their cross-subject consistency, we identified 7 ICNs that represented distributed areas in gray matter and were significantly contributed to by at least 60% of the template subjects. Figure 1. [...] As demonstrated in Figure , the fluctuation level of individual ICNs was quantified using variance of the corresponding representative time courses. Temporal variance has been demonstrated to be closely relevant to the functioning of ICNs (; ; ). For each of the ICNs, we first calculated the cross-twin correlation of the variance metrics for MZ and DZ twin pairs separately. The resultant cross-twin correlation coefficients were compared to provide an indication of the relative contribution of genetic and environmental factors to the ICN fluctuations (; ).Further, we used an ACE model to quantify the genetic and environmental contributions to the ICN fluctuations. In a standard ACE model, the variance of a trait can be decomposed into additive genetic (A), shared or common (C) environment, and nonshared (unique) environmental (E) effects, the latter of which includes measurement error (). This model takes advantage of the differences in genetic relatedness between MZ twins who share 100% of their genes, and DZ twins who share, on average, 50% of their genes (; ). Based on these assumptions, one can derive the contributions of A and C effects using SEM, as implemented in OpenMx (), with covariates of age and sex modeled as regressions or deviation effects on the mean. In SEM, A, C, and E are represented by latent variables, and their relative variances (denoted as a2, c2, and e2, respectively) reflect the proportion of their contributions to the cross-subject covariance of the observed variables (i.e., variance metrics of ICNs).Since the variance metric of ICNs may not follow a normal distribution, we used a nonparametric approach to examine the significance of A and C effects (). The first step is model identification, which is used to examine the significance of the A and C contributions in the model. Besides the ACE model, the path coefficients for 2 submodels, AE and CE, were estimated. The differences in likelihoods between the full model and the 2 submodels were calculated. A significant likelihood difference between ACE and CE models indicated that the A factor is critical in the full model, while a significant likelihood difference between ACE and AE model suggested a significant contribution of C factor.Null distributions of the likelihood difference were generated by permuting the MZ/DZ labels and the composition of the twins with constraints on exchangeability. Specifically, if a twin from twin pair A was randomly paired with a twin from twin pair B, then the other twins from twin pairs A and B were paired. In addition, the sex and age of the 2 twins forming a twin pair were constrained to be the same, so that age and sex could still be included as covariates in the ACE model. This permutation was repeated 5000 times, and the ACE model, as well as AE and CE submodels, was estimated. The likelihood differences between the full model and submodels were collected to form 2 null distributions, one for ACE-AE and the other for ACE-CE.The significance of A and C factors for the original dataset was then evaluated within the null distributions. Under the rule of parsimony in model identification, if the C factor was significant but A was not, we used the CE model to estimate the C effect; if the A factor was significant but C was not, we used the AE model to evaluate the A effect. If neither A nor C was significant in these tests, we used the ACE model. For each of the 7 ICNs, we separately conducted the above procedure for model identification. The threshold for significance was P < 0.005 (single-tail test). Once the model was selected, the relative variance of A, C, and E factors (a2, c2, and e2) was evaluated from the identified model. […]

Pipeline specifications

Software tools AFNI, FreeSurfer, RAICAR, OpenMx
Applications Miscellaneous, Functional magnetic resonance imaging
Organisms Homo sapiens