Conceived and designed the experiments: LB DB. Performed the experiments: LB P-CC JB. Analyzed the data: LB P-CC JB GP. Contributed reagents/materials/analysis tools: LB GP. Wrote the paper: LB P-CC JB GP DB.
The authors have declared that no competing interests exist.
Resting state networks (RSNs) have been studied extensively with functional MRI in humans in health and disease to reflect brain function in the un-stimulated state as well as reveal how the brain is altered with disease. Rodent models of disease have been used comprehensively to understand the biology of the disease as well as in the development of new therapies. RSN reported studies in rodents, however, are few, and most studies are performed with anesthetized rodents that might alter networks and differ from their non-anesthetized state. Acquiring RSN data in the awake rodent avoids the issues of anesthesia effects on brain function. Using high field fMRI we determined RSNs in awake rats using an independent component analysis (ICA) approach, however, ICA analysis can produce a large number of components, some with biological relevance (networks). We further have applied a novel method to determine networks that are robust and reproducible among all the components found with ICA. This analysis indicates that 7 networks are robust and reproducible in the rat and their putative role is discussed.
Resting state networks (RSNs) in humans is becoming a tool to investigate brain basal states in health and disease
Until recently, RSNs have not been studied in rodents. Consequently, healthy state basal networks are not well understood. Most rodent studies have been carried out on anesthetized animals. Furthermore, these studies use a seed-based approach in which specific questions are tested by using a particular brain region as a seed and determining what else in the brain displays similar temporal pattern. Kannurpatti et al.
To the best of our knowledge, there is only one other study that explores RSNs in awake rodents
An alternative approach in determining RSNs is the application of independent component analysis (ICA -
Optimally, determination of the basic RSNs should be carried out on awake rodents to eliminate the confounding effects of anesthesia. Additionally, a model-free analysis approach such as ICA should be used to define networks. However, as mentioned above, the selection process of significant components in an ICA analysis is rather arbitrary.
We have recently extended a technique (RAICAR -
In this article, we have acquired baseline fMRI data in conscious rats and utilized ICA to determine robust reproducible RSNs. Our results indicate that 7 components meet the criteria and their putative role in brain function is discussed.
All methods and experiments were approved by the Massachusetts General Hospital Subcommittee on Research Animal Care (IACUC-Protocol #2008N00192). Animals were caged 2 per cage and on a 12/12 hours light/dark cycle. Food and water were available ad libitum. Animals were inspected daily for welfare or signs of disease. No animal displayed excessive signs of stress (vocalizing trashing around) during training or imaging session.
Imaging methods have been described elsewhere
Analysis was performed using tools from the FMRIB software library (
Resting state analysis was performed using spatial independent component analysis (ICA) using the tool MELODIC from FSL. There are 2 sources of variability when doing ICA. First, ICA results are produced by optimization of a non-convex function, which means that the results are not unique. Indeed, running ICA several times on the same dataset gives different results. Second, there is inter-subject variability in the ICA results. Given these sources of variability, we would like to find ICA components that are good descriptors of the resting state as well being reproducible across runs and across subjects.
Given our initial population of 13 rats (2 out of 15 were excluded due to excessive motion), we randomly picked 5 rats without replacement out of 13 to create a group of 5 rats. We repeated this process to create 50 different groups of 5 rats. Each group of 5 rats was subjected to a temporal con-catenation based group ICA (concat-ICA). The number of independent components in concat-ICA was fixed at 20 (similar to human RSN studies
For reproducibility of RSN analysis we adopted the following approach. The 50 group ICA maps from individual 5 subject concat-ICAs were subjected to a ranking and reproducibility analysis (RAICAR,
In the original RAICAR algorithm, there is no well-defined cutoff on reproducibility. Hence, we modified the RAICAR algorithm to have an automatically determined reproducibility cutoff
For group averaging and thresholding, the reproducible components (e.g., p-value<0.05) across the 50 groups ICA runs were automatically preserved as described above. These components were then analyzed voxel-wise using a group mean general linear model (GLM). A mixture model (MM) was fitted to the resulting t-statistic maps using algorithms developed in MATLAB
Each component was analyzed to determine brain structures, or parts thereof, that were covered by the network. In addition to the component average statistical value for each brain structure, the percentage of the volume of the structure in the network was calculated. These two metrics were used to classify brain structures within a component in order of statistical significance and relative fraction of the structure involved; components were assigned a specific biological function based on the most salient brain structures in the network.
Brain structures were identified on MRI slices that correspond to a standard Atlas
Two out of 15 rats were excluded from further analysis due to excessive head motion. None were removed because of signs of stress such as vocalization.
The motion parameters for 2 typical rats (Left column) in the study and the 2 rejected rats (Right column) are displayed. Green Line, translation/rotation X-axis, Blue Y-axis, and Red Z-axis.
The top panel displays components sorted according to reproducibility level. The red line marks the 90% cutoff as determined in the top panel. With a 90% cutoff, 7 components are above the threshold; with a 95% cutoff 6 components survive. The bottom panel depicts the normalized reproducibility and the adjusted cut-off for 90%.
Components (C1–C7) are ordered according to their reproducibility degree. Component 1 has significant cerebellar structures; Component 2 includes medial and lateral cortical structures resembling the human default mode network; Component 3 includes a basal-ganglia-hypothalamus network; Component 4 encompasses basal-ganglia-thalamus-hippocampus circuitry; Component 5 represents an autonomic pathway; Component 6 represents the sensory network; and Component 7 groups interoceptive structures to form a network. All components have been thresholded according to a mixture model approach-see
This network involved structures mostly in the cerebellum (cerebellar lobules 2,3,4,5) and the brainstem (periaqueductal grey (PAG) raphe nuclei), (
The predominant regions involved are in cortical and subcortical regions with essentially only the PAG showing some brainstem involvement (
This component is comprised mainly of subcortical structures that include components of the basal ganglia (globus pallidus, extension of the amygdala (exA), septal region, CPu, claustrum), thalamus, and hypothalamus (
Regions involved in this component are shown in
This network is also predominantly subcortical in nature (
This component is essentially only cortical in nature (
This network includes sensory cortices as described in component 6 as well as secondary visual structures and the insular and cingulate cortices (
These components did not achieve statistical significance for robustness and reproducibility at the 90% confidence level. Inspection of them revealed that some were artifacts (9, 13, and 20 involved ventricles); 14 captured some edge of the brain; 15–19 networks seemed to already appear as part of statistically significant components. Only networks #8 and 10 reflected some biological relevant structures (CPu and cingulate/cortical ribbon, respectively). Please see
Here we report 7 resting state networks of a model of awake rats that are robust and reproducible. These networks were chosen statistically based on their probability of reproducibility. The networks could serve as a base to further study changes induced pharmacologically or in disease models.
Performing awake animal imaging eliminates issues related to anesthesia
The approach utilized here allows objective identification of the most reproducible RSN components. Nevertheless, it is possible that some biologically relevant components are only moderately reproducible and, hence, this analysis might miss those components. In principle, they could be recovered by relaxing the p-value threshold on reproducibility. Furthermore, the degree of reproducibility does not necessarily imply higher biological importance.
Resting state networks represent low frequency brain fluctuations that correspond to functionally relevant networks
In our study, seven robust and reproducible components were found in awake rats. The 7 observed networks seem to be relevant and show some similarity to the networks observed in healthy humans, specifically those defined in Smith et al.
Reported here for the first time for rats, probably was not observed in other studies because imaging did not cover the cerebellum. It is highly reproducible in rats, furthermore, this network clearly relates to component 5 of Smith et al. in humans
Zhang et al.
These include networks that encompass basal ganglia (BG, CPu, Globus Pallidus), thalamic nuclei, the amygdala, septal nucleus, and the hippocampus (see Joel and Weiner, 1994
Regions such as the hypothalamus are well characterized in their role in autonomic control, but other forebrain regions are also involved
Combined, these mainly represent sensorimotor components. This is a network commonly seen in part (bilateral S1-sparse
This network is similar to the sensory network described above and observed in several RSN rat studies with the inclusion of insular and cingulate cortices. Liang et al. observed that one of their modules displayed some instability and further analysis resulted in a splitting of the module into 2 that encompassed the cortical ribbon observed here for components 6 and 7
As indicated in the results, most of them have no biological relevance, only two seem to have some biological significance and involve the CPu and the cingulate, both structures have been identified in several other networks described above. Nevertheless, these two components might reflect sub-networks within these structures that could achieve statistical significance with a larger cohort of rats.
Performing studies in awake rats might induce changes in “natural” RSNs, as restraining could influence the animal's stress as well as emphasize certain networks such as environmental monitoring. As described above, stress was minimized by repeated exposure of the animals to restraint and scanner noises. Excessive (chronic) restraining, however, might be inductive of constant, elevated stress
Given the obvious differences in brain morphology across species, comparisons with humans should be taken with caution. The observed networks in rodents seem to parallel those in humans. However, their assignments or interpretation should be evaluated by altering them (chemically or genetically, for example) and relating behavioral changes with RSN changes.
Robust reproducible networks in awake rats were identified. Some are in agreement with those observed in other rat RSN studies. Several of the observed networks are similar to those observed in healthy humans. The elimination of the use of anesthesia might enhance significantly the opportunity to study brain alterations in well-controlled preclinical models. This approach may provide a useful addition to the use of imaging in evaluating drug effects, disease states, or genetic and strain differences.
(TIF)
(TIF)
(TIF)
(DOCX)
(DOCX)
(DOCX)
(DOCX)
(DOCX)
(DOCX)
(DOCX)