Introduction: Understanding how brain regions interact to perform a specific task is very challenging. EEG and MEG are two noninvasive imaging modalities that allow the measurement of brain activation with high temporal resolution. Several works in EEG/MEG source reconstruction show that estimating brain activation can be improved by considering spatio-temporal constraints but only few of them use structural informa on to do so. We present a source estimation algorithm that uses brain structural connectivity, obtained from diffusion MRI (dMRI), to constrain the EEG/MEG source reconstruction. Contrarily to most source reconstruction methods which reconstruct activation for each time instant, the proposed method estimates an initial reconstruction for the first time instants and a multivariate auto-regressive model that explains the data in further time instants. This auto-regressive model can be thought as an estimation of the effective connectivity between brain regions. Methods: We use diffusion MRI (dMRI) in conjunction with EEG/MEG measurements to reconstruct brain activity. dMRI is processed using probabilistic fiber tracking from FSL. These fiber tracks are used in two ways: 1) to parcellate the cortex into functional regions (defined as having an homogeneous connectivity profile) and 2) to create a network of connexions between these regions. The parameters were tuned to obtain cortical regions of around 100 mm2, which is about the minimum size of detectable activations with M/EEG. We further assume a constant activation (a scalar value) per cortical region at a given time instant. This allows to greatly reduce the dimension of the source space from the ~10k nodes of the cortical mesh to ~600 extracted cortical regions. This spatial source model is completed by a temporal multivariate auto-regressive (MAR) model where a region activation at time t is obtained as linear combination of the activations at the p previous time instants t=1, ..., t-p of brain regions to which it is connected. The unknowns in this model are the region activations J(t) for the first t=1..p time instants and the coefficients A of the linear combination. Given some M/EEG measurements for a time window of size T, the goal is to estimate these unknowns to fit these data. To do so, we introduce a slightly modified MxNE criterion U(J) (Gramfort etal, 2012) that promotes spatial sparsity along with temporal continuity of the activations. A second criterion V(A) measures how well the estimated activations J(t), t=1..T are obeying the MAR model. The optimisa on procedure iterates between two steps, which alternatively improve the estimates of J(t), t=1..T to decrease U(J) and the linear coefficients A, given J(t), t=1..T. The process stops when a maximum number of iterations is reached or when there is no significant changes between two iterations. Results: The method was evaluated using the real dataset described in Wakeman et al (2015). MEG/EEG data were simultaneously recorded during a face recogni on task where a subject is shown famous, unknown or scrambled faces. The dataset also contains dMRI andT1 images. Four MAR models were tested:p∈ {1,2,3,4}. We show that the acquired data can be explained by our model with few regions as soon as p>1 (Fig. 1). Reconstructions using EEG and MEG show a clear negative peak at the FG around 200ms for p>1 which matches what can be found in the literature (Fig. 2). The non-null coefficient in the final A can be considered as the effective connectivity used during the task. Conclusions: We have presented a way of reconstructing brain activation and effective connectivity between the brain regions using an extension of the MxNE solver. Sources are constrained to follow a MAR model of order p. We have shown that such a model can fit real M/EEG measurements with relatively few activated regions as soon as p>1 and that the recovered activated regions are coherent with the task used to acquire the dataset.