Letting behaviour speak: how long-term, continuous home-cage observation and deep learning capture stress dynamics

Stress-related disorders such as anxiety and depression affect millions and arise from complex interactions between stress exposure and genetic predispositions. Conventional behavioural assays of stress often capture only brief snapshots of behaviour, missing the continuous, dynamic nature of the stress response. This limits the ability to detect subtle, temporally evolving phenotypes relevant to stress resilience and vulnerability.

Here we present an ecologically relevant deep phenotyping pipeline that leverages recent advances in open‑source software and computational methods to monitor home‑cage activity continuously for over 24 hours following acute restraint stress. Long-term recordings of freely moving mice are combined with neural network-based animal pose estimation (e.g., DeepLabCut) to capture fine-grained dynamics. This approach enables continuous comparison of stress-induced behaviour across the full response, while increasing accessibility to improved translational animal model analyses [1]. Animals were restrained within their home cage, and a fresh restrainer was left in the cage to preserve contextual memory of the stress experience and allow more nuanced phenotyping of resilience. Importantly, this deep phenotyping supports a translational framework, aiming to mirror individualized approaches in psychiatry [2].

Home-cage activity and restrainer interaction were quantified across time and normalized using z-scoring. Animals were grouped into control-like, stress-susceptible, and stress-resilient clusters based on behavioural features. Hourly statistical comparisons between groups were conducted using the Python “statannotations” library with the Brunner–Munzel (generalized Wilcoxon) test. Supervised machine-learning behavioural classifiers implemented in pretrained DeepOF models were used to quantify specific behaviours. Comparisons between control and restraint-stressed animals were performed using the Mann–Whitney U test with Benjamini–Hochberg correction for multiple testing.

Alterations in home-cage activity and restrainer-interaction persisted long after a single stress exposure. Notably, changes in restrainer-interaction of previously stressed mice uncovered distinct behavioural signatures associated with stress resilience.Distance travelled and time on top of the restrainer differed between clustered groups of control-like animals, susceptible animals, and resilient animals. Susceptible animals consistently showed lower z-scored values across multiple timepoints, especially during the activity phase (~6-18 hours post 10 A.M. stress, p<0.05 at discrete timepoints). Stronger effects persisted after the stress in susceptible animals compared to control-like when stressed at a time closer to the activity phase (spanning ~2-16 hours post 4 P.M. stress, p<0.05-0.001). In contrast, resilient animals remained consistently similar to controls across time. Machine-learning classifiers shows that stressed animals spent significantly less time exploring the cage, climbing the cage walls, and sniffing over the full 12 hour activity phase (p<0.01), supporting the sensitivity of combined tracking approaches to uncover the full potential of long-term recordings.

Together, this deep-learning based home cage monitoring analysis framework provides a scalable and open-source approach for capturing nuanced behavioural phenotypes, aligning with individualized medicine approaches emerging as necessary in the psychiatric field. By improving both resolution and translation, this method advances the study of stress resilience and vulnerability in psychiatric disorders.

References

1. Jurek, B., Schlegel, P., Zglinicki, B., Aman, L., Kovarova, V., Narayan, S., Florea, R., Ziuzia, P., Badura, A., Schmidt, M. V., & Slezak, M. (2025). Long-term deep phenotyping of behavioral traits in mice using homecage monitoring. Neuroscience & Biobehavioral Reviews, 180, 106453.
2. Bordes, J., Miranda, L., Müller‐Myhsok, B., & Schmidt, M. V. (2023). Advancing social behavioral neuroscience by integrating ethology and comparative psychology methods through machine learning. Neuroscience & Biobehavioral Reviews, 151, 105243–105243.