Title:	Scalable High-Dimensional Computational Phenotyping of Mouse Behaviour Across Social, Defensive and Metabolic Paradigms
DNr:	Berzelius-2026-203
Project Type:	LiU Berzelius
Principal Investigator:	Stefanos Stagkourakis <stefanos.stagkourakis@ki.se>
Affiliation:	Karolinska Institutet
Duration:	2026-06-26 – 2027-01-01
Classification:	30105
Homepage:	https://www.stagkourakislab.org/
Keywords:

Abstract

This project will establish a scalable AI-driven computational neuroethology pipeline for high-dimensional phenotyping of mouse behaviour across social, defensive, aggression-related, and metabolism-related paradigms. Modern behavioural neuroscience increasingly generates large video datasets from freely moving animals, often recorded across multiple cameras, experimental days, behavioural contexts, and treatment conditions. However, manual behavioural scoring remains slow, subjective, low-throughput, and restricted to predefined categories. This creates a major bottleneck for discovering how neural activity, hormonal state, metabolic condition, pharmacological intervention, and prior experience shape behavioural organisation. We will combine markerless pose estimation, unsupervised behavioural segmentation, latent-state modelling, and supervised classification to transform high-resolution behavioural videos into quantitative representations of movement structure, behavioural motifs, transition dynamics, and individual differences. The pipeline will be applied across several ongoing projects in the laboratory, including social interaction, resident-intruder aggression, fear and defensive behaviour, and metabolism-related behavioural states such as food deprivation, refeeding, high-fat diet exposure, and chronic glucocorticoid manipulation. Aggression in the resident–intruder paradigm will serve as the initial benchmark dataset because it provides a complex multi-animal behavioural context with existing annotations, large-scale videography, clear phenotype structure, and well-defined computational bottlenecks. The main scientific goal is to develop a generalisable framework that can identify behavioural states without relying exclusively on human-defined labels, quantify how animals diverge into distinct behavioural phenotypes, and test whether experience or intervention shifts animals between behavioural states. This is particularly important for studying maladaptive behavioural states, where clinically or biologically relevant phenotypes may emerge not from a single behaviour, but from changes in behavioural sequences, transitions, persistence, escalation, and context dependence. By extracting behavioural “syllables”, latent trajectories, transition probabilities, and predictive feature sets, the project will allow us to compare behavioural organisation across paradigms and identify shared or distinct computational signatures of internal state. Access to GPU-accelerated computing is essential because the pipeline requires repeated training and benchmarking of pose-estimation models, inference across hundreds of high-resolution videos, optimisation of unsupervised sequence models, and downstream classifier development across multiple experimental conditions. These analyses exceed the practical capacity of local workstations and require high-memory GPU resources for efficient and reproducible execution. The expected impact is both methodological and biological. Methodologically, the project will provide a reusable, scalable, and reproducible workflow for large-scale behavioural phenotyping. Biologically, it will enable discovery of behavioural signatures that are difficult to detect by manual observation alone, thereby improving our understanding of how internal state and neural circuit function generate adaptive and maladaptive behaviour. In the longer term, this framework will support integrated analyses linking behaviour to neural recordings, endocrine measurements, pharmacological manipulations, and metabolic physiology.