Computerized Behavioural Analysis in Mice. At Memory Dynamics in Neuroinformatics

Abstract

Autism is a cluster of behavioural abnormalities that manifest as impaired social behaviour, perseverance behaviours and altered memory processes. The study of memory processes in mice is used as a model for normal and pathological cognitive functioning. The euchromatic methyltransferase 1 heterozygous knockout (ehmt1+/ ) mouse is a model for Kleefstra syndrome, a condition characterized by autism and intellectual disability. Studies on memory processes in ehmt1+/ have mainly focused on episodic memory, with mixed results on whether this is improved or equal to healthy subjects. In this study, ehmt1+/ mice were subjected to the Object-Space task, a novel paradigm to distinguish episodic from semantic-like memory. In this task mice are exposed to multiple trials involving objects placed dynamically but with overlapping regularity (overlapping condition), objects placed all in the same location over trials (stable), or objects randomly placed each trial. Over trials, mice may acquire the spatial patterns in the first two conditions but not last. However, one major challenge is extracting comprehensive behavioural information from video data of mice performing such a task. This thesis describes a computerized method to categorize various behaviours (i.e. Object Exploration, Wall Exploration, and Corner Sitting) from video data of mice performing the Object-Space task. The method involves a model that uses techniques such as kinetic action recognition, transfer learning, and pose estimation to categorize behaviours in both a supervised and an unsupervised manner. The former implements optic flow over multiple frames in order to learn what constitutes a behavioural module of Object Exploration. The latter implements recent developments in deep learning for pose estimation to define both Wall Exploration and Corner Sitting behaviours as a geometrical configuration of limbs. Visual inspection of these models combined show it to be highly accurate in time in terms of sensitivity and specificity of action classification. Moreover, this behavioural categorization model was used to describe an array of behaviours (e.g. object exploration time, object discrimination index) of mice performing trials in all conditions of the Object-Space task. This array of behaviours could be used to predict genotype (i.e. ehmt1+/ or ehmt1+/+) of a mice based on a single video of a trial in both the overlapping and stable condition, but not in the control condition. One especially interesting finding is that the models to predict genotype used more memory related behaviours (e.g. discrimination index) to predict genotype in the overlapping condition, whereas the models to predict genotype in the stable condition mainly used general behaviours (e.g. total exploration time). Further inspection of these behaviours between genotypes show that ehmt1+/ mice may display increased memory expression behaviours over healthy controls. This indicates that memory processes in this Kleefstra mice model might be improved, and not characterized by intellectual disability as previously thought.