We have fulfilled the main objectives by addressing these major aims: (1) to understand how different inputs contribute to neuronal output firing of subicular neurons and (2) monitoring free behaviour at unprecedented detail to investigate how behavioural modules contribute to dendritic integration on the single cell level and whether the prediction of membrane potential is feasible exclusively from behavioural observation. To reach these goals we have performed in-vivo patch-clamp recordings of subicular neurons in navigating mice and two photon Ca2+ imaging of subicular neurons during a spatial navigation task, while we inactivated specific input pathways from EC and CA1. Our analyses revealed two predominant synaptic input patterns that were tuned to the running speed and location of the mice. We found using acute brain slices and Channelrhodopsin assisted circuit mapping, that different input pathways target different dendritic subregions. Following DREADD based synaptic inactivation (by CNO perfusion through the cranial window) in vivo with simultaneous behavioural assessment. We also found that different input pathways have differential effects on neuronal tuning and membrane potential changes during spatial navigation. The behavioural assessment was possible via a development of an unsupervised machine learning algorithm: Given its complex and highly dynamic nature, the quantification of animal behaviour is still an open challenge in neuroscientific research. The recent development of markerless pose estimation methods has revolutionized the field. It allowed for the first time to track the animal pose from video recordings without any interference with the ongoing activity of the subject. We developed Variational Animal Motion Embedding (VAME), an unsupervised probabilistic deep learning framework that uncovers behavioural structures in markerless pose estimation data and segments it into clearly separated discrete motifs. We demonstrated the efficacy of our framework by performing unsupervised behavioural quantification of mouse lines during free behaviour. Furthermore, we showed that the behavioural representation learned by VAME was highly sensitive to the frequency as well as the phase of the behavioural signals, and therefore formed an ideal toolkit for segmentation of rodent movement, such as exploratory behaviour or specific types of locomotion. We showed that the distribution of phenotype-specific motifs can be used to identify subtle differences between mice, that was otherwise difficult to find to human observers. Moreover, we showed how a hierarchical representation of the motif’s usage allowed to group motifs into bigger categories called “communities” allowing to explore the relationship between motifs on a larger scale. Finally, we showed that VAME obtained the highest scores when comparing the obtained motif segmentations of current state-of-the-art behaviour quantification approaches to ground truth labels created by human experts, confirming that this method improved the state-of-the-art behavioural quantification systems currently used (Luxem et al., BioRXiv, Luxem et al.). We used this VAME paradigm for quantification of treadmill-based and free behaviour. It allowed for detection of distinct subpopulations of subicular in proximal and distal subiculum during treadmill-based spatial navigation.
