Sessions

Social behaviours are crucial for the survival of many vertebrate species by maintaining proper communication between conspecifics. The coordination and fine-tuning of these behaviours require complex neural circuits to form the correct responses to the given social stimuli. Understanding the relationship between brain functions and behaviours requires a detailed analysis of the behaviour itself. This analysis is conventionally achieved by manual scoring, which can be slow and biased.

The rapid development of deep-learning-based computational methods has led to numerous improvements in automated tracking software, such as DeepLabCut (DLC), an open-source pose estimation software that uses machine learning algorithms for tracking. This software can track multiple animals simultaneously, which is useful for studying social behaviours. However, DLC only provides a matrix containing the coordinates of each labelled and tracked body point, so another tool is needed to identify specific behavioural elements. Thus, our laboratory has developed Emerenka, a web-based software that extracts behavioural elements from the output matrix of multi-animal DLC.

Selecting and training the appropriate artificial neural network for our purposes, and the synchronisation of DLC and Emerenka remain unsolved issues. Therefore, my work focuses on properly training the DLC and then finding a way to optimise the determination of behavioural elements. The Cntnap2 KO mouse strain, a well-known model of autism spectrum disorder, was used to test the results. Two groups of Cntnap2 KO mice were tested: one group was treated with a serotonin receptor agonist, while the other received control injections. During the test, both groups were allowed to interact freely with their conspecific partners in an open field arena for 10 minutes under video surveillance. After training the DLC on 200 frames from ten chosen 10-minute-long videos, all the recorded videos were analysed by DLC. Based on the output files, the body points and distances used to ascertain behavioural elements were modified in Emerenka until the software could correctly identify the different social interactions. All analysed videos were evaluated manually, enabling the different methods to be compared.

As a result, no significant difference was found between the manual scoring and the DLC combined Emerenka analysis. This was achieved by finding a proper way to train DLC and by adjusting the parameters in Emerenka multiple times. Furthermore, unlike previous automated behavioural analysis methods, which distinguished only a few elements, we managed to differentiate several kinds of interaction with a methodology easy to follow. The developed technique provides a reliable, efficient, AI-based automated analysis of social behaviour in mice that is unbiased, faster, and less labour-intensive than manual approaches. Thus, it can contribute to a more efficient workflow in laboratories studying rodent social behaviour.