Brain activity is a highly dynamic process across different temporal and spatial scales. The control of cognition and behaviour relies on computations performed at different location in the brain, and relies on the integration of complex information. Resting-state fMRI has been proposed to resolve the functional connectome of the brain in an effort to understand function in the healthy and disordered brain. However, while the brain is highly dynamic, the vast majority of the resting-state fMRI data analysis relies on the stationary assumption, i.e. a constant interaction between region pairs over the duration of the recording. Here, we investigated the dynamic functional networks of the anesthetised mouse brain. We used sliding window correlation to estimate the dynamics between any region pairs, and dictionary learning to perform dimensionality reduction to obtain a set of 20 reproducible dynamic functional networks.C57BL/6 mice were anesthetized with isoflurane 0.5% and medetomidine 0.05mg/kg bolus and 0.1mg/kg/h infusion as in (Grandjean et al., 2014). Functional imaging took place on a 9.4 Bruker biospec with a linear volume coil for excitation and a 2x2 phased-array receiver cryogenic coil. Multi-echo gradient-echo EPI were acquired with TE=11, 17, 23ms, TR=1500ms, FA=60°, 600 volumes, and an in plane resolution of 0.3mm². ME-EPI were reconstructed and denoised using meica.py script (Kundu et al., 2012) (AFNI, http://afni.nimh.nih.gov/) , and coregistered to the AMBMC mouse template. Independent component analysis was performed to parcellate the brain using Melodic with 30 components (FSL, http://fsl.fmrib.ox.ac.uk/). The components identified were further separated between hemispheres. A static functional network was reconstructed from the signal from 12 selected components. Louvain algorithm for weighted networks was used to identify 3 modules, a sub-cortical network (SuCN), a lateral cortical network (LCN), and a set of regions related to the default mode network (DMN). Sliding window correlation (SWC) with window size =30, and step size =2, was used to estimate the changes in correlations between region pairs over time. The SWC matrix was demeaned, and detrended with third order polynomials. Dictionary learning algorithm with 30 folds each made of 200 iterations was used to estimate simple building blocks of whole-brain connectivity, the atoms. The patterns thus extracted explained 50% of the variance in the SWC matrix.Static functional connectivity of the ICA-parcellated mouse brain indicate strongest pattern of correlations between inter-hemispheric homotopic region pairs (Figure 1: e.g. left to right anterior parietal cortex; aPc). The regions could be organized into 3 distinct modules, the sub-cortical network (SuCN) including the dorsal, ventral, and lateral striatum (dStr, vStr, lStr) and anterior parietal cortex (aPc), the lateral cortex network (LCN) including the medial and posterior parietal cortex (mPc, pPc), and motor cortex (Mc), and finally a default mode network (DMN) which included the prefrontal cortex (PFC), the cingulate/retrosplenial cortex (Cg/Rs), dorsal and ventral hippocampus (dHc, vHc), and thalamus. Dynamic functional networks were estimated using dictionary learning, a set of 20 atoms were estimated and accounted for 50% of the total variance of the SWC (Fig 2). The patterns revealed interactions, in particular between and within the modules, for instance atom #3 indicate interactions between the LCN and the SuCN, while atom #14 indicates interactions within the SuCN. Further, the thalamus is highlighted in atom #17, while there was little evidence of static interactions between the thalamus and the other regions (Fig 1). Finally the atoms presented high values of similarities between the different folds of the dictionary learning process, underlying the reproducibility of the atoms estimated (Fig 3).Dynamic functional connectivity analysis in the mouse brain is highly attractive due to its relative simple architecture as compared to humans, but also in the light of the many transgenic lines available. Here we show that estimation of dynamic functional networks reveals rich patterns of interactions masked in the standard static connectivity analysis. Moreover, the patterns highlighted in the mouse present salient features, such as between and within modules groups of interactions, which may indeed represent crucial elements necessary for the segregation and integration of information, which are necessary for brain computation. The mouse is a model of choice to investigate the nature and mechanisms underlying dynamic functional networks in the healthy brain, as well as in the many murine models of brain disorders.