The beta2-adrenergic receptor (B2AR) belongs to the family of G protein-coupled receptors, one of the major drug targets. G protein-coupled receptors are integral membrane proteins that convert external signals into intracellular responses. Two already known drugs employed in the treatment of several respiratory diseases are salmeterol and salbutamol. They show a high affinity to B2AR, however, their binding pathways have not yet been fully characterized. Along this project we will shed light on the binding process by means of coarse-grained molecular dynamics simulations using the Martini 3.0 force field. This methodology enables us to study the binding pathway of both drugs in an unbiased way.

In all eukaryotes, the Unfolded Protein Response (UPR) is a molecular program that maintains the protein folding homeostasis in the endoplasmic reticulum (ER). The UPR plays a crucial role in health and disease. Stress sensors proteins on the ER membrane activate the UPR. The evolutionary most conserved sensor is the protein IRE1, which activates the UPR by forming dimers and larger assemblies. In particular, IRE1's luminal domain (LD) interacts with unfolded proteins and these interactions promote oligomerization by an unresolved mechanism. The direct binding mode of peptides is still not understood.

My work aims to elucidate the structure and assembly mechanism of large supramolecular assemblies of human IRE1 and probe its binding to unfolded proteins. These events are crucial for IRE1's functions but are not yet understood.

We employed a multiscale approach, performing atomistic and coarse-grained (CG) molecular dynamic (MD) simulations.

For investigating the formation of clusters of dimers of IRE1, we used the coarse-grained Martini 3 force field. We obtained encouraging results: hIRE cLD dimers can form clusters where contacts are mediated by disordered regions.

Peptide binding experiments, in atomistic and CG, were successful and led us to propose a new model for the direct binding of peptides and unfolded proteins.

Further analysis will be needed to extrapolate relevant dimer-dimer conformations from our simulations and to assess the effect of peptides on the dynamics of hIRE1 cLD dimer.

We outline a new model in which generalised uncertainty relations are obtained without modified commutation relations. While existing models introduce modified phase space volumes for the canonical degrees of freedom, we introduce new degrees of freedom for the background geometry. The background is treated as a genuinely quantum object, with an associated state vector, and the model naturally gives rise to the extended generalised uncertainty principle (EGUP). Importantly, this approach solves (or rather, evades) well known problems associated with modified commutators, including violation of the equivalence principle, the ‘soccer ball’ problem for multi-particle states, and the velocity dependence of the minimum length. However, it implies two radical conclusions. The first is that space must be quantised on a different scale to matter and the second is that the fundamental quanta of geometry are fermions. We explain how, in the context of the model, this gives rise to an effective dark energy density, without contradicting established results including the no go theorems for multiple quantisation constants, which still hold for species of material particles, and the spin-2 nature of gravitons.

Modern supervised machine learning (ML) techniques have demonstrated great utility for the classification of electroencephalography (EEG) signals. However, the large amounts of annotated training data typically required are not available in many medical settings. In this talk, we will discuss a few possibilities applying the unsupervised learning approaches for EEG representation learning in the context of epilepsy. Specifically, we apply a variational autoencoder (VAE) network to learn a compressed representation of the input EEG signals. This could be potentially used in the tasks of early diagnosis of epilepsy, treatment effect detection, or disease progression understanding. We also show large intersubject variability in the data that imposes great challenges.

The b-value from the Gutenberg-Richter relation captures the ratio between small and large earthquakes triggering for a region. Several researchers have observed fluctuation in b-values prior to an earthquake, especially during the aftershock sequence analysis. By leveraging the ever-increasing earthquake database and the development in the field of deep learning computations, this study focuses on analyzing the spatial-temporal b-value series for Japan.

We follow a two step approach, in the first step a spatial-temporal b-value series is calculated from the earthquake catalog, which takes the shape of a series of [32 × 32] pixel images of spatial b-value distributions. On this we train an autoencoder, which compresses and then decompresses the input to learn the normal behavior and relationships within the data. In the second step, we then take the pixel by pixel reconstruction error as input for a Dilated ResNet-like classifier, which predicts an earthquake probability. We expect that this study will improve the understanding of earthquake occurrences which could further be beneficial for Early Warning, rapid response and mitigation plans, especially for sustainable human habitats in earthquake-prone regions.

The human tyrosine-protein kinase Met is a trans-membrane receptor placed on the plasma membrane. Also called c-Met, it is fundamental for both cell migration and replication in epithelial tissues of our bodies. At the same time, the c-Met receptor plays a crucial role in tumoral development by enhancing metastasis growth via its known overexpression on many kinds of tumoral cells. To gain insight into the cell-cell signalling process, we put our focus on the activation mechanism of the ectodomain of the receptor by the non-native ligand InlB. We exploited atomistic MD simulations to study the influence of the ligand on the receptor by simulating its upper ectodomain both bound to the ligand and in isolation. We also simulated these models in the case of full glycosylation. The analysis of the obtained trajectories revealed the signalling-competent conformation assumed by the upper domains of the ectodomain while identifying the angle among them as the discriminant between signalling- and non-signalling-competent conformations. Furthermore, we observed alternating bridging interactions among the glycans that reduce the conformational plasticity of the isolated ectodomain. For the data analysis, we combined traditional strategies with a machine learning approach called Diffusion Maps. In parallel, to challenge our conclusions, we modelled and simulated two more models. The collected data successfully support our findings, in light of which we plan to study the rearrangements of the whole ectodomain and how this interacts with its surroundings.

Navigating a complex environment is assumed to require stable cortical representations of environmental stimuli. Previous experimental studies, however, show substantial ongoing remodeling at the level of synaptic connections, even under stable conditions. It remains unclear, how these changes affect sensory representations on the level of neuronal populations during basal conditions and how learning influences these dynamics.

Using chronic neuronal population activity data recorded in awake mouse auditory cortex, we find that sound representations are subject to a significant ongoing remodeling across the timespan of days under basal conditions. Fear conditioning introduces a bias into these ongoing dynamics, resulting in a differential generalization both on the level of neuronal populations and on the behavioral level. This means that sounds that are perceived similar to the conditioned stimulus (CS+) show an increased co-mapping to the same response mode the CS+ is mapped to. This differential generalization is also observed in animal behavior, where sounds similar to the CS+ result in the same freezing behavior as the CS+, whereas dissimilar sounds do not. We conclude that learning-induced plasticity leading to a representational linkage between the conditioned stimulus and non-conditioned stimuli weaves into ongoing dynamics of the brain rather than acting on an otherwise static substrate.

The problem of stereo matching, i.e. finding the correct correspondences between features in left and right image for stereo-vision has been studied for many years and from different perspectives. Computations in the neocortex critically depend (in distinction to deep learning structures) on feedback connections as well as lateral connections between cells at the same hierarchical level. While the first disparity selectivity in V1 is reasonably well explained by the classical disparity energy model, it does not address the question of finding the corresponding image points in both images by eliminating 'false matches'. Inside an object's boundaries this can be done by assuming that the disparity should vary smoothly over small distances. As has been proposed previously, we assume such a Gestalt-law to be implemented via lateral connections that form an associative filter network in layer 2/3 of neocortex. We develop an algorithm that is inspired by the columnar structure of neocortex and that is able to perform stereo-matching for textured surfaces. Over and above previous work we aim to extract information about local disparity gradients with the help of lateral sets of connections in the form of net-fragments. This problem of finding a coherent neural representation within ambiguous and noisy input is omnipresent in the neocortex and independent from the modality of stereo-vision. The algorithm we have developed should be applicable to different modalities as well as able to integrate inputs from different submodalities.

Extra-embryonic membranes are believed to be a significant factor for the evolutionary success of insects. However, the fruit fly Drosophila melanogaster, unlike the majority of insects, develops only a single, dorsally located, vestigial and hence functionally limited membrane, the amnioserosa. In consequence, the red flour beetle Tribolium castaneum has become a frequently used model to investigate the morphogenetic principles and function of extra-embryonic structures. In this species, formation of the amnion and serosa, two distinct extra-embryonic tissues, has already been associated with several genes, for example Zerknüllt 1. However, the functional spectrum of this gene has not yet been fully characterized. Currently, data are only available from parental RNAi-based knock-down experiments (van der Zee et al. 2005, Current Biology; Panfilio et al. 2013, Biology Open; Jain et al. 2020, Nature Communications). However, Zerknüllt 1 has never been investigated at the gene and morphogenetic level, and wild-type as well as knock-down development have never been compared with a knock-out phenotype. We have combined CRISPR/Cas9 genome editing (Gilles et al. 2015, Development) with our AGOC vector concept (Strobl et al. 2018, eLife) to create a stable Zerknüllt 1 knock-out Tribolium line via insertional mutagenesis. In detail, we inserted 3×P3-based eye marker cassettes via homology-directed repair into the Zerknüllt 1 coding sequence to impede proper expression. Preliminary imaging data of DAPI-stained homozygous knock-out embryos suggest a similar phenotype as found in previous knock-down experiments, i.e., the embryos appear to be lacking a serosa and are covered only by a dorsal amnion. Next, we will generate double transgenic lines that also carry fluorescent protein expression cassette suitable for live imaging experiments to dynamically characterize the knock-out development cascade during gastrulation.

Developmental Artificial Intelligence seeks to create agents with cognitive architectures and open-ended learning abilities similar to those of human infants. In our research, we explore one fundamental skill acquired by children in their early developmental stages: sensorimotor coordination. Previous work has shown that active visual control can be learned merely from the motivation of efficiently encoding incoming visual information. Here, we aim to extend these results to multiple sensory modalities, beginning with the integration of vision and proprioception during the learning of eye-hand coordination. To this end we use MIMo, an infant model embodied in a physics simulator. MIMo's brain consists of a hierarchical architecture with lower-level unimodal controllers and higher-level multimodal ones. MIMo is trained using intrinsically motivated reinforcement learning: the controllers at each hierarchical level have the objective of efficiently encoding the information at their respective levels of abstraction. Initial results reveal that vision alone is insufficient to achieve eye-hand coordination, thus highlighting the importance multimodal integration. Ongoing work focuses on establishing a satisfactory integration mechanism.

Presentation of the workshops at the FIGSS Retreat in Haus Bergkranz.