Skip to main content

Biological neural networks: from structure to function

Periodic Reporting for period 1 - NeuArc2Fun (Biological neural networks: from structure to function)

Reporting period: 2016-03-01 to 2018-02-28

The NeuArc2Fun project proposal was to set at the connection between theoretical and experimental neuroscience. The goal was the investigation of information processing in neural networks with feedback. Technically, it aimed to adapt and improve analysis tools for newly collected data by nowadays recording and imaging techniques, which have much improved recently. The whole approach relies on the estimation and interpretation of interactions between large (mesoscopic) populations of neurons from their activity that is simultaneously recorded. This includes electrophysiology (e.g. Utah electrode array), calcium imaging or fMRI (considering whole-brain activity). In particular, the focus of NeuArc2Fun was on “large” networks (i.e. from 20 to hundreds of nodes), which is suitable for state-of-the-art measurements. It produced novel analysis tools for recurrently connected neural networks, bridging the structural and functional levels (towards information processing).

The key for a useful formalism is finding the adequate balance between the mathematical tractability and biological realism of the model. To address this trade-off problem, NeuArc2Fun focuses on the mesoscopic level, i.e. scales at which many interacting neural populations can be simultaneously recorded by current state-of-the-art experimental techniques, such as electrode arrays. The advantage of this model-based approach is the ability to make predictions about the role of each component of the model –in particular, its heterogeneous connectivity– in shaping neural activity. A particular focus was on bridging several disciplines in a common comprehensive formalism: dynamic system, graph theory, statistics and information theory.

The primary application of the framework was targeted at electrophysiological data recorded in monkeys from the laboratory of Prof. Thiele in Newcastle University. Modeling such data is particularly challenging because they exhibit a large variability over repeated trials in the same condition, which hinders the extraction of consistent condition-specific information.
The project has achieved most of its objectives and milestones for the period, with relatively minor deviations. More precisely, the training objectives have been fully completed and the proposed scientific project has been fully realized for the theoretical part, but has been adapted for the application to experimental data. The NeuArc2Fun project has produced excellent publications: 4 articles as first author (including 1 as single author), 2 from direct supervision (me as last author) and 3 from further collaboration and co-supervision. 4 publications are in high-impact journals (Neuroimage and Human Brain Mapping), bringing visibility on my research direction. All publications are either in gold open access or in green open access (using the prepint websites arxiv and biorxiv, as well as the UPF university OpenAIRE system). The computer code of the theoretical papers has been written in the open-source language Python and made publicly available on a dedicated website (

On the theoretical side, the proposal planned 2 articles in collaboration with Drs Tauste Campo and Zamora-Lopez from the CNS group (1 article with each), linking dynamic system (my speciality) with statistics and with graph theory respectively. The first one has been published in Network Neuroscience in 2017 ( and the second one has just been accepted by Physical Review E (March 2018; preprint on arxiv: A third paper was published in Biological Cybernetics ( formally linking network dynamics with information theory.

On the application side, it turned out that electrophysiological data are more difficult to analyze with the developed network model than fMRI data. This explains why 1 publication (same article in Network Neuroscience mentioned above) focuses on electrophysiological data from the lab of Prof. Alex Thiele (Newcastle University), whereas the other application papers concern fMRI data. Using other types data in the network model was in fact mentioned in the proposed risk management. Nonetheless, the application to electrophysiological data is still ongoing work and will lead to another submitted paper about the connectivity between cortical layers, tentatively in 2018.

The 2-year MSCA fellowship also gave me the opportunity to (co)supervise 5 PhD and master students. Among those, 4 led to publications or submitted manuscripts: Katharina
Glomb (2 published in Neuroimage, and j.neuroimage.2017.12.074) Niels Reuter (me as last author, published in Human Brain Mapping, Vicente Pallares (me as last author, minor revisions in Neuroimage, available on biorxiv and Murat Demirtas (submitted, available on biorxiv I also gave internal seminars within the CNS group at UPF in the context of the Journal Club (October 2016 and November 2017) to review recent advances in the literature and Group Meeting (September 2016, June and September 2017) where I presented my own work.

The diffusion of gained knowledge and visibility of the work was supported by attending 9 conferences, including 2 oral presentations in plenary sessions for Neural Coding 2016 and
Coupling and Causality in Complex Systems 2017. There was also the organization a workshop on connectivity analysis (extraction of “fingerprint” of brain activity;
events/workshop_CNS2017.html) in CNS 2017 (Antwerp, Belgium) and the co-organization of a workshop in CNS 2016 (Jeju, South Korea; MAMC_workshop_CNS2016/_node.html including an oral presentation). I also visited laboratories to develop new collaborations (for new types of data, beyond fMRI and electrophysiology), during 4 laboratory visits (including 1 to finish a paper with Niels Reuter in Maastric
Until recently, data analyses focused on one of the 3 following points: dynamics, statistics (including machine learning) or graph analysis. A goal of NeuArc2Fun was to develop a multidisciplinary approach that links these fields and tools. Together with my collaborators, I combined our knowledge to develop a comprehensive formalism for network-oriented analysis of neural data. In practice, the first step is the fitting of a network model to multivariate signals (recorded simultaneously) to capture their dynamics. This estimates interactions between the nodes in the network, which are neuronal populations near electrodes in multi-electrode arrays or brain regions for fMRI measurements. These interactions define the communication scheme between the nodes in the network, which can then be analyzed and quantitatively compared across task or behavioral conditions. This approach provides a novel flexibility in the analysis from the link level (single interaction between 2 nodes) to the network level (functional communities of nodes). Adapting analysis tools from different fields has created a common ground and language to quantitatively relate concepts in neural information processing (e.g. causality, segregation versus integration).
General overview of the network-orented analysis framework.