Network Analysis and Multivariate Pattern Recognition Methods to Study Drug Effects on Functional Connectivity in Mice Based on Resting State fMRI

Abstract

Resting state fMRI aims at establishing the functional relationship between two spatially distinct brain regions based on temporal correlation of the respective signals. The method has been applied to study functional connectivity (FC) across species and was found of special value for studies in mice fMRI studies due to its experimental simplicity, which allowed eliminating confounds typically observed in stimulus-evoked fMRI studies. Yet, analysis of fMRI data in animals face a number of problems. First, fMRI studies in mice typically require the use of anaesthetics, which are known to alter responses to stimuli or functional networks at rest. Proper interpretation of fMRI data collected in animals under anaesthesia animals requires investigation of the effects of these drugs on brain processing per se. Second, since fMRI analysis tools have been typically developed for processing of human fMRI data, translation to animals including mice may be challenging due to anatomical and physiological differences across species. It therefore appears appropriate to evaluate several state of the art analysis tools for human fMRI for their potential use in mice fMRI. In the first study, we have used Dual Regression analysis Network Modelling to test its feasibility in mouse resting-state fMRI analysis to investigate effects of two commonly used anaesthetics, isoflurane and medetomidine, on rs-fMRI derived functional networks, and in particular to study to what extent anaesthesia affected the interaction within and between brain networks. The analysis revealed both similarities and specific differences in network patterns of the two groups. Under isoflurane anaesthesia, intra- and interhemispheric cortical interactions have been predominantly observed, with only minor interactions involving subcortical structures. In particular, cortico—thalamic connectivity appeared significantly attenuated in line with previous observations. In contrast, medetomidine anesthetized mice displayed significant subcortical functional connectivity including interactions between cortical and thalamic ICA components. Combining the two anaesthetic drugs at low dose resulted in network interactions that by large constituted the superposition of the interactions observed for each of the two agents alone. In conclusion, the study revealed that with suitable adaptations the DR based network modelling can be used for analyzing mouse fMRI data and the results are comparable to those obtained with classical seed based analysis. In a second study, we investigated whether the method was sensitive enough to detect changes in FC in mouse brain in response to varying the dose dependent effects of isoflurane using resting state fMRI. Stationary FC analysis was complemented by analysis of dynamic functional connectivity (dFC), i.e. looking for short-term changes in the interaction of brain functional networks. Stationary network analysis using FSL Nets revealed that increasing isoflurane dose led to a reduction of functional connectivity between the bilateral homotopic cortical regions as well as between cortical and thalamic areas. In addition, dFC analysis revealed a dominance of functional states (dFS) exhibiting pronounced modular structure in mice anaesthetized with a low dose of isoflurane, while at high isoflurane levels dFS showing widespread unstructured correlation displayed highest weights. This indicates that spatial segregation across brain functional networks is lost upon increasing dose of the anaesthetic drug. In conclusion, by combining the results of stationary and dynamic FC analysis of mouse resting-state fMRI data we found that increasing isoflurane levels led to loss of modular network organization, which includes the loss of strong bilateral interactions between homotopic brain areas. In a third study, we evaluated to what extent machine learning methods could be applied for unsupervised classification of subjects according to their resting-state fMRI derived FC pattern. Features extracted from stationary as well as dynamic functional connectivity analysis derived from mice exposed to the anaesthetic isoflurane at different doses were subjected to machine learning algorithms for both support vector machines (SVM) and deep belief networks (DBN). The results show that we were able to successfully classify, i.e. assignments to group above chance level, between anaesthetic doses using features extracted from stationary and dynamic functional connectivity analysis. Not surprisingly, the classification accuracy increased when comparing extreme groups, e.g. lowest and highest dose of isoflurane. The features extracted from dFC analysis were found to be more discriminative with regard to the different anaesthetic doses than those derived from stationary FC. This illustrates the potential of using dFC features. A major limitation regarding the use of machine learning in the context of our study was small sample size (N=12 per group), which led to an accuracy of less than 70% for most comparisons. In conclusion, classification based on machine learning tools yielded results that were clearly above chance levels though classification accuracy was likely compromised by the small size of the training data sets. Future studies are needed to assess the value of machine learning in mouse fMRI. Finally, we applied the Random Forest (RF) classification to detect differences in the pharmacological MRI (phMRI) response of rats to treatment with an analgesic drug (buprenorphine) at two doses as compared to control (saline). RF analysis was able to identify drug effects based on differential phMRI responses in the hippocampus, amygdala, nucleus accumbens, superior colliculus, and the lateral and posterior thalamus for drug versus saline. These structures have high levels of mu opioid receptors associated to the drug response. In addition, these regions are involved in aversive signalling, which is inhibited by mu opioids. In conclusion, the results demonstrate that buprenorphine mediated phMRI responses comprise characteristic features that allow a supervised differentiation from placebo treated rats as well as the proper allocation to the respective drug dose group using the RF method. In this thesis, network analysis methods originally developed for analysis of human fMRI data have been applied for processing of mouse fMRI data. Networks identified were found biologically meaningful as were the within and between network interactions. In particular, dFC analysis appears an attractive approach yielding insight into network interactions that appear indicative of specific condition. Classification attempts revealed that machine learning approaches work in principle, though accuracy is largely compromised by the typical small size of data sets available in animal fMRI. In this context, the establishment of open-access databases, where researches can deposit their original data, may become attractive.