Final Report Summary - DYNVIB (Dynamic effective connectivity of the Virtual Brain)
The project DynViB “Dynamic Effective Connectivity of the Virtual Brain” focused on understanding through advanced analyses of empirical and simulated neural activity data how dynamic communication and interactions between different brain regions, as the ones required for performing behavior and cognition can stem out of multi-scale brain anatomical circuits which can be considered as “hard-wired” on fast time-scales. A typical example of a brain function requiring restructuring of inter-areal interactions is constituted by selective attention, in which information conveyed by the activity of populations encoding aspects of e.g. a visual scene which are directly relevant to the current task should be preferentially routed toward higher-order brain regions, and in which the selection of which should be the prioritary sender populations is achieved via top-down signals which implement an “attentional spotlight”. Thus, understanding some of the circuit mechanisms which may underly the flexibility of attention, a higher cognitive function which everyone uses in his daily life, is one of the driving motivations of DynViB. The objectives of the project were twofold:
• On one side, deepening the characterization of structured patterns of Dynamic Functional Connectivity in empirical data of spontaneous activity, e.g. from whole-brain imaging of the resting state, too often investigated only by averaging interaction networks analyses over extremely long time-windows, thus blurring away precious dynamical informations.
• On the other side, connecting this observed temporal evolution of functional connectivity networks to the actual emergent complex dynamics of the underlying multi-scale anatomical circuits, resorting to simulations of a whole Virtual Brain model as a chief heuristic tool.
The main tenet underlying DynViB is indeed that the “Chronnectome” —i.e. the characteristic repertoire of time-dependent functional states of a neural system (Calhoun et al., 2014)— is the visible manifestation of the “Dynome” —i.e. the collection of possible collective (oscillatory?) dynamical regimes that a given neural circuit can generate (Kopell et al., 2014)— that the underlying structural “Connectome” —i.e. the description of synaptic connections — is engendering. A fundamental role in this scientific endeavour is played by connectome-based models, i.e. computational models in which a realistic anatomy based on actual experimental reconstructions is used to constrained simulations of noise-driven neural activity, mimicking the spontaneous activation of neural systems at different scales. A open-source neuroinformatics platform developed by the Institute for Systems Neuroscience at Aix-Marseille University (the DynViB host unit), known as The Virtual Brain (Sanz-Leon et al., 2013) greatly simplifies the task of systematically exploring the features of emergent dynamics in different parameter ranges. Such Virtual Brain models are available already today, waiting for the more detailed models that the EU Human Brain Project is promising to deliver.
DynViB has achieved several important results, a selection of which is here summarized:
• We have succeeded in generating the first ever whole-brain mean-field model, based on a DTI-reconstructed cortical connectome, of the switching chronnectome of the resting state (Hansen et al., 2015). By tuning our Virtual Brain model in a novel critical regime which makes the emergent dynome as rich as possible, given the many simplifications in our model, the noise-driven sampling of the internal repertoire of states naturally lead to a richly structured non-stationary functional connectivity, as observed in empirical human resting state fMRI data. Our model also reproduces with a non-trivial accuracy, during this resting state sampling, several empirically observed functional networks, such as e.g. the intrinsic Dorsal Attention Network, which thus appears to be the manifestation of the transient visit of a specific connectome-constrained emergent dynamical supspace (see Figure 1A in attachment).
• We have considerably advanced the methodology for analyzing dynamic functional connectivity (FCD) and functional-to-structural connectivity relations in neural systems at different scales and based on different types of imaging signals and neural recordings. We have introduced a new recurrency matrix representation, known as FCD matrix (Hansen et al., 2015) as well as additional biomarkers of FCD, based on random walk and hypergraph representations, which allow quantifying the statistical features of dynamical inter-areal interactions, beyond their mere observation. We have illustrated recently the power of these metrics starting to apply them for the study of how FCD is reorganized in human aging (see Figure 1B in attachment), discovering their tremendous and partially unanticipated biomarking potential (Battaglia et al., 2015a). Considering more microscopic networks, we have crowd-sourced the development of machine-learning algorithms for the reverse engineering of structural connectivity based on the imaging of state-dependent neural activity, organizing a IEEE-sponsored algorithmic challenge, attracting over 150 world-wide participants and aiming at —and highly successful in— out-performing the previous state-of-the-art established by our own Transfer Entropy algorithms (see Battaglia et al., 2015b).
• We have explored the role of self-organized oscillatory dynamics in the dynamic routing of information in brain multi-areal circuits. In simplified formal models, in which individual areas are described as simple noise-driven phase oscillators, we have derived fully analytical expressions for the amount of information shared and transferred between any two regions or communities of regions in an arbitrarily complex modular connectome, showing that Information Routing Patterns depend on the collective oscillatory configuration (Kirst, Timme and Battaglia, 2015). In a smaller scale, but more realistic spiking model of a few interacting brain areas at the edge of synchrony we have proved the robustness of the state-dependency of information transfer when oscillations are not periodic but are stochastically bursting, transient and sparsely synchronized as observed in vivo (Palmigiano et al. 2014).
• Toward the Virtual Brain multi-scale modelling of large-scale coherence networks and their involvement in attention, we have systematically studied a “Virtual Region”, in which a meso-scale connectome of the multi-layer canonical circuit of cortex is shown to characteristically constrain and regulate the local generations of multi-frequency oscillations and the way in which they coordinate between different cortical regions, making possible frequency-selective communications in bottom-up and top-down directions (Helmer et al., 2015). We are currently investigating an extension to whole brain simulations, which embed both meso- and macro-scale connectomes in a same model architecture. Such computational effort is casting light on the connectome-constrained dynomic determinants of the fluctuations of functional vs structural hierarchy of cortical areas identified in attentional experiments (Bastos et al., 2015).
In short, DynViB has lead to substantial advancements towards the quantitative description and implementation of multi-scale computational models of the Functional Connectivity Dynamics of neural circuits including the whole Human brain. Besides their intrinsic interest for fundamental science (how can cognitive flexibility be implemented? Which circuit mechanisms underly its emergence, and also its impairments? What is the role of dynamic complexity into all this?), we believe that DynViB achievements will prove extremely useful: from a data analysis methodology perspective, to design qualitatively novel and theory-inspired biomarkers of brain state alterations in healthy aging but also e.g. for early diagnosis of Alzheimer’s Disease; from a computational perspective, introducing improved design criteria for whole-brain mean-field top-down model, including future personalized ones, with potential groundbreaking role in 3P medicine applications.
• On one side, deepening the characterization of structured patterns of Dynamic Functional Connectivity in empirical data of spontaneous activity, e.g. from whole-brain imaging of the resting state, too often investigated only by averaging interaction networks analyses over extremely long time-windows, thus blurring away precious dynamical informations.
• On the other side, connecting this observed temporal evolution of functional connectivity networks to the actual emergent complex dynamics of the underlying multi-scale anatomical circuits, resorting to simulations of a whole Virtual Brain model as a chief heuristic tool.
The main tenet underlying DynViB is indeed that the “Chronnectome” —i.e. the characteristic repertoire of time-dependent functional states of a neural system (Calhoun et al., 2014)— is the visible manifestation of the “Dynome” —i.e. the collection of possible collective (oscillatory?) dynamical regimes that a given neural circuit can generate (Kopell et al., 2014)— that the underlying structural “Connectome” —i.e. the description of synaptic connections — is engendering. A fundamental role in this scientific endeavour is played by connectome-based models, i.e. computational models in which a realistic anatomy based on actual experimental reconstructions is used to constrained simulations of noise-driven neural activity, mimicking the spontaneous activation of neural systems at different scales. A open-source neuroinformatics platform developed by the Institute for Systems Neuroscience at Aix-Marseille University (the DynViB host unit), known as The Virtual Brain (Sanz-Leon et al., 2013) greatly simplifies the task of systematically exploring the features of emergent dynamics in different parameter ranges. Such Virtual Brain models are available already today, waiting for the more detailed models that the EU Human Brain Project is promising to deliver.
DynViB has achieved several important results, a selection of which is here summarized:
• We have succeeded in generating the first ever whole-brain mean-field model, based on a DTI-reconstructed cortical connectome, of the switching chronnectome of the resting state (Hansen et al., 2015). By tuning our Virtual Brain model in a novel critical regime which makes the emergent dynome as rich as possible, given the many simplifications in our model, the noise-driven sampling of the internal repertoire of states naturally lead to a richly structured non-stationary functional connectivity, as observed in empirical human resting state fMRI data. Our model also reproduces with a non-trivial accuracy, during this resting state sampling, several empirically observed functional networks, such as e.g. the intrinsic Dorsal Attention Network, which thus appears to be the manifestation of the transient visit of a specific connectome-constrained emergent dynamical supspace (see Figure 1A in attachment).
• We have considerably advanced the methodology for analyzing dynamic functional connectivity (FCD) and functional-to-structural connectivity relations in neural systems at different scales and based on different types of imaging signals and neural recordings. We have introduced a new recurrency matrix representation, known as FCD matrix (Hansen et al., 2015) as well as additional biomarkers of FCD, based on random walk and hypergraph representations, which allow quantifying the statistical features of dynamical inter-areal interactions, beyond their mere observation. We have illustrated recently the power of these metrics starting to apply them for the study of how FCD is reorganized in human aging (see Figure 1B in attachment), discovering their tremendous and partially unanticipated biomarking potential (Battaglia et al., 2015a). Considering more microscopic networks, we have crowd-sourced the development of machine-learning algorithms for the reverse engineering of structural connectivity based on the imaging of state-dependent neural activity, organizing a IEEE-sponsored algorithmic challenge, attracting over 150 world-wide participants and aiming at —and highly successful in— out-performing the previous state-of-the-art established by our own Transfer Entropy algorithms (see Battaglia et al., 2015b).
• We have explored the role of self-organized oscillatory dynamics in the dynamic routing of information in brain multi-areal circuits. In simplified formal models, in which individual areas are described as simple noise-driven phase oscillators, we have derived fully analytical expressions for the amount of information shared and transferred between any two regions or communities of regions in an arbitrarily complex modular connectome, showing that Information Routing Patterns depend on the collective oscillatory configuration (Kirst, Timme and Battaglia, 2015). In a smaller scale, but more realistic spiking model of a few interacting brain areas at the edge of synchrony we have proved the robustness of the state-dependency of information transfer when oscillations are not periodic but are stochastically bursting, transient and sparsely synchronized as observed in vivo (Palmigiano et al. 2014).
• Toward the Virtual Brain multi-scale modelling of large-scale coherence networks and their involvement in attention, we have systematically studied a “Virtual Region”, in which a meso-scale connectome of the multi-layer canonical circuit of cortex is shown to characteristically constrain and regulate the local generations of multi-frequency oscillations and the way in which they coordinate between different cortical regions, making possible frequency-selective communications in bottom-up and top-down directions (Helmer et al., 2015). We are currently investigating an extension to whole brain simulations, which embed both meso- and macro-scale connectomes in a same model architecture. Such computational effort is casting light on the connectome-constrained dynomic determinants of the fluctuations of functional vs structural hierarchy of cortical areas identified in attentional experiments (Bastos et al., 2015).
In short, DynViB has lead to substantial advancements towards the quantitative description and implementation of multi-scale computational models of the Functional Connectivity Dynamics of neural circuits including the whole Human brain. Besides their intrinsic interest for fundamental science (how can cognitive flexibility be implemented? Which circuit mechanisms underly its emergence, and also its impairments? What is the role of dynamic complexity into all this?), we believe that DynViB achievements will prove extremely useful: from a data analysis methodology perspective, to design qualitatively novel and theory-inspired biomarkers of brain state alterations in healthy aging but also e.g. for early diagnosis of Alzheimer’s Disease; from a computational perspective, introducing improved design criteria for whole-brain mean-field top-down model, including future personalized ones, with potential groundbreaking role in 3P medicine applications.
