2nd Generation
Project Sacha Abou Rachid:
In my doctoral studies, I will study how cortical areas integrate, process and encode sensory information from multiple senses. Since I will focus on the integration of visual and tactile sensory inputs I will also focus on the according primary visual (V1) and somatosensory (S1) cortical areas as well as the higher association areas like the rostrolateral cortical area (RL) in between V1 and S1. Using state-of-the-art two-photon Ca2+ imaging I can study the mechanisms of multisensory integration even in active behaving mice on the level of neuronal populations and also individual neurons and their dendrites. Moreover, I will shed the light on the multisensory impairment in autism spectrum disorder (ASD), which is one criterion in DSM-5 (Diagnostic and Statistical Manual of Mental Disorders, 5th Edition) for the diagnosis and could lead to future research focusing on bridging the gap between the neural mechanisms and the behavioral manifestations.
Inputs from multiple senses can be integrated in the brain and lead to improved performance. Furthermore, multisensory integration has been extensively studied already in the superior colliculus (SC), but which cortical areas are involved in multisensory integration has still to be discovered.
Therefore, I first aim to investigate pyramidal neurons in several cortical areas in a multisensory evidence accumulation task. In a second step, I want to explore the dynamics of dendritic activity during this multisensory experience and how this is changed during learning the task. Finally, I would like to study the neuronal mechanisms of multisensory integration impairment in autism spectrum disorder.
To tackle these questions, a combination of two photon Ca2+ imaging together with behavioral testing will be used. Genetically modified mice expressing a fluorescent calcium sensing protein will be used to image the areas of interest including V1, RL, S1, anterior cingulate cortex (ACC) and anterior lateral motor area (ALM) for further investigation with two photon imaging to reveal the neuronal mechanisms underlying this process. Activity of neurons will be recorded during training in a 2-alternative forced choice task for head-fixed mice involving an initiation period, stimulus presentation, delay (response selection) and motor response presented by water reward based on sensory stimulus. Separating these time periods allows also imaging neuronal responses selective for the sensory stimulus or decision of the animal.
The work of previous RTG students in the lab has revealed that mice are able to learn this task within a few weeks of training. The stable expression of the fluorescent calcium indicators further allows studying the same neurons over the entire learning period from the naïve animal to becoming an expert. I will further compare how these neurons change their activity over time in adaptation to the performance of the mice in the task. Learning improves efficiency by rearranging synapses. By using two-photon imaging of dendritic activity of visual and tactile inputs in the visual cortex and identified higher cortical areas of interest of behaving mice, I aim to uncover the mechanism by which these different sensory information are encoded and integrated at synaptic and dendritic level during a decision making task. Secondly, I aim to investigate how this integration is changed by synaptic plasticity mechanisms during learning the multisensory evidence accumulation task.
Successful integration of various simultaneously perceived perceptual signals is crucial for social behavior. In ASD patients, multisensory integration is affected while it remains unclear how this impairment occurs and what is the neuronal basis behind it. To tackle these questions, I want to explore how these neuronal mechanisms of multisensory integration are impaired in autism spectrum disorder (ASD). By using the same multisensory discrimination task on an established autistic mouse model, the SHANK 3 KO mouse, I will investigate whether multisensory integration is altered in this mouse model. In addition, my task design allows me to test whether the impairment was due to an overload of sensory information (discrimination of multiple unisensory stimuli) or due to a deficient integration of multisensory cues. Finally, I will look into and explore all these questions and concerns at the dendritic and spine level considering our mouse model that strongly alters these structures.
Project Simone Albani:
My doctoral thesis research aims to use computational tools to study, at the molecular level, the interactions between drugs and Nav1.7, a voltage-gated sodium channel that is mainly expressed in the membrane of nociceptive and olfactory neurons, with particular focus on the mutated versions of the channel. Loss-of-function mutations in the gene encoding for Nav1.7 (scn9a) can lead to an insensitivity to pain and loss of smell (anosmia), while gain-of-function mutations will cause the appearance of apparently-unmotivated pain sensations and inflammation in the body (the most affected aera depends on the mutation). Structural information about the channel is fundamental to discover drugs that target a specific variant of Nav1.7. Some of these structural properties of voltage-gated sodium channels (e.g., the presence of different binding sites for local anaesthetics, the presence of lateral access pathways) were theorized during the second half of last century, thanks to electrophysiology and mutagenesis experiments. Nowadays, these techniques remain some of the most powerful tools to study in vitro how the ion currents that flow through the channel are affected by drugs and mutations and to formulate hypothesis on the structural functioning of the channel. Research on voltage-gated sodium channels (Navs) has deeply changed in the course of the last 12 years, when a series of new CryoEM structure of bacteria, first, and human, later, Navs has been published. This opened up the possibility to perform docking studies and molecular dynamics simulations of these channels starting from direct observations of the system at a resolution of circa 0.3 nm. In the first part of my research, I used coarse-grained molecular dynamics (CGMD) simulations to rationalize how the insecticide permethrin could differently affect the human Nav 1.7 and its homologue in dogs and cats, despite their high sequence similarity. Namely, permethrin is known to be toxic for cats, but not for humans or dogs, the toxicity being related to the higher persistent currents of the feline channel (preliminary experiments from the Lampert lab). Permethrin was therefore here used as a ‘probe-molecule’ to understand how small sequences’ variations can impact the channel structure and drugs’ response. Initially we established the best protocol setup for molecular simulations, including the choice of the model membrane in which the channels need to be embedded. Our benchmark simulations showed that the interaction between the channel and a complex membrane (i.e., composed of many different lipids rather than just phosphatidyl choline) is necessary to maintain the protein’s stability. Next, we modelled a water-box containing 10 M concentration of permethrin in the presence of cat and human Nav 1.7 embedded in a neuronal membrane. Preliminary data suggests that permethrin’s binding site could differ between the human and feline channel. A map of the most relevant residues for permethrin binding was generated and it will be used to validate these results using computer simulations at higher resolution followed by in vitro mutagenesis experiments. Understanding how genetic variants impact on the structure of Nav channels and their response to drug open up novel directions of personalized medicine: In a multiscale framework, a Nav1.7 variant discovered in a clinical environment could be translated into structural information and used to select the best treatment for the patient.
Project Elisabeta Balla:
The brain is an active perceiver. It is tasked with computing incoming multisensory stimuli in a fast, efficient and sustainable way. An excellent candidate for how this computation might be realized is the predictive coding strategy. Here, higher cortical areas probe the environment through sending refined predictions about incoming sensory input to lower levels of hierarchy. Then, the predicted and encountered signals are compared and only the outcome is propagated through the feedforward path. This is conserving the energetic cost of directly representing all the incoming sensory input per se.
The main aim of my thesis is to shed light on the implementation of such a mechanism in the visual cortex of mice through simultaneously measuring the feedback and feed forward responses to a prediction task. The thesis is structured in a few crucial steps that start by the investigation of cortical responses to a visual Oddball Paradigm in a Wide Field setup in order to narrow down a cortical area of interest where a prediction response arises. Following the Wide Field results, I am targeting the identified cortical area for 2Photon imaging of neuronal responses to a passive visual prediction task (Oddball Paradigm). The final step shall be the visualization of responses from top-down and bottom up projecting neurons in a passive and active (DMS) prediction task and establishing a preparation for trichromatic imaging in the 2Photon setup for which the mice shall be trained appropriately. The final task is intended to encompass combinations of vision and other sensory modalities, which allows for the addressing of multisensory integration of the respective sensory signals at a fairly advanced cognitive level and potentially a shed light on the implementation of predictive coding algorithms at a more behaviourally complex level.
Project Peter Bouss:
In the scope of my doctoral studies, I am focusing on the identification of latent spaces that represent information in the neuronal system. This plays a crucial role in understanding how coordinated activity bridges scales from the level of single cells to the scale of populations of neurons that operate collectively as a network. Sensory processing in the brain requires the information to be represented within the neuronal activity. The temporal and the spatial scales of this representation are in particular important to understand how representation of stimuli from different modalities may be combined into a single coherent percept. For example, the system must be able to integrate stimuli that arrive with a different delay or that are stretched across different temporal extents. Characterizing the neuronal representation in this regard is a fundamental prerequisite to developing hypotheses on cross-modal integration.
During my thesis I am analyzing electrophysiological data for repeating patterns and latent variables. I examine two different perspectives that allow for different temporal scales: the generation of surrogate data and the training of invertible neural networks.
Surrogate data – The first approach deals with finding spatio-temporal patterns of millisecond precision in recordings of multiple spike trains, in our case from the macaque motor cortex. Evaluating the significance of such patterns is a challenging endeavor due to the number of neurons, the variation of neural firing rates, and the number of delay combinations. For this purpose, I investigated the use of surrogate techniques within the framework of SPADE (spike pattern detection and evaluation), a method developed at the INM-6 in the last decade. Surrogates allow for statistical bootstrap testing of the null hypothesis. As I recognized drawbacks of uniform dithering, the classically chosen surrogate technique used in SPADE, we tested it against several alternative surrogates from the literature, as well as new ones. We analyzed not only the effects of these methods on stationary spike trains as a benchmark but went further to create artificial data sets with similar properties as experimental data, thereby modeling the time-varying firing rates, neuronal dead-times, and regularities in inter-spike intervals. As no precise spatio-temporal patterns should be found in this artificial data, applying SPADE and using the different surrogate methods allowed assessing the false-positive patterns. Together with an analysis of experimental data of the macaque motor cortex, my colleague Alessandra Stella and I recently submitted a manuscript as shared first authors to eNeuro.
Invertible neural networks – The second perspective seeks to use invertible neural networks (INN), also known as normalizing flows, as an analysis tool for electro- and optophysiological recordings. An invertible neural network is able to learn a generic data distribution, from which it then can draw samples. First approaches toward analysis of EEG data have already been performed in the literature. Before starting with analyses, I did not only implement a neural network containing just the essential building blocks but I am analyzing the properties of INNs for different data distributions.
The motivation is the following: Since neural networks are known to operate as black boxes, we seek to improve our understanding of them, such that they can be used more reliably in a scientific context. Therefore, before training a network on our data, we need to consider which types of data distributions it is able to learn, in what situations it may fail, and how we could mitigate those.
As a first challenge, I am dealing with data with non-trivial topology. In this scenario, networks can easily fill the holes of the data distribution or finish training at a local but not global minimum. To mitigate this risk, I introduce a quantitative measure to detect topological critical points. We are currently investigating under which circumstances such a situation arises in electrophysiological data.
A second challenge is to transform data regions with different volumes. In this regard, we compare the difference between additive and affine coupling layers, where one allows for non-constant volume transformations, i.e. non-constant Jacobi determinants, while the other does not. While the first one is generally used for an interpretable latent space, the latter one typically yields better accuracies. We investigate the reason for this behavior and visualize it in low-dimensional test cases.
With the knowledge of these two analyses, we introduce an extra loss term that promotes a low-dimensional latent space. Concretely, we use the participation ratio established as a reliable measure of dimensionality.
Given these insights, we plan to use the resulting INNs to identify latent non-linear dimensions in datasets with different temporal scales, namely, spike trains recorded with Utah arrays (together with Sonja Grün), EEG, and calcium imaging data (together with Simon Musall and Björn Kampa).
Identifying and characterizing the latent spaces that span neuronal representations of information will then allow us to study the interplay of information from multiple sensory inputs that are processed by a cortical region.
Project Alice Despatin:
My doctoral thesis research will study the cortico-striatal areas and projections involved in multimodal and multisensory decision-making. Multisensory stimuli are part of every environment and constitute the basis of decisions in all animals. The posterior parietal cortex (PPC) and the anterior lateral motor cortex (ALM) both receive inputs from multiple primary sensory areas and have projection to the dorsal striatum (dStr). PPC-dStr projection neurons also appear to be involved in the history bias, a tendency to replicate previous decisions independently of sensory inputs. I will study how decisions are made from the history bias as well as multisensory inputs. To do so I will research how this information is conveyed between PPC, ALM and dStr and which neuronal categories are involved (intratelencephalic (IT) neurons, pyramidal tract (PT) neurons or uncategorized networks).
First, I will focus on the role of PPC projection neurons to ALM and dStr in mice during a multisensory discrimination task with a delay. Combining behavioural analysis with two-photon imaging of either PT or IT neurons will help to understand the role of neuronal categories of PPC in integrating multisensory inputs and history bias in order to make decisions.
Then, I will focus on the inputs of both PPC and primary sensory areas (visual – V1 and sensory – S1) to ALM. ALM seems to receives sensory inputs twice, once via PPC and once directly from V1 and S1. The hypothesis is that ALM could thus play a role in decision-making by comparing direct sensory inputs and inputs from PPC including history bias. ALM could also be involved in working memory and store sensory inputs or decision during the delay phase. I will use Neuropixels recording of ALM neurons in association with optogenetic inhibition to define the importance of PPC and primary sensory areas inputs for decision-making.
Finally, I aim to define the role of dStr and ALM in decision implementation. ALM neurons projecting to dStr are known to be important for decision-making. I want to show the exact spatiotemporal relation between these areas allowing decision execution by using a combination of two-photon imaging in ALM and Neuropixels recording in dStr. I will also study the importance of IT and PT neurons in ALM-dStr projection on decision-making. This will allow a better understanding of the pathway followed by external information – multisensory inputs – and internal information – history of previous decisions – on multiscale: neuronal networks and systems and their impact on behaviour.
