smart-Kage: Autonomous behavioural phenotyping in a rodent home-cage environment

Abstract

Comprehensive understanding of the brain involves detailed examinations of animal behaviour together with underlying neural computations from which it emerges. Whereas great progress has been made in technology used to record neural activity, studies of behaviour have thus far lacked an equally methodical and concerted development that could support/enable insights into high-resolution neural data. Automated home-cage monitoring present a valuable paradigm that could fill this gap. However, current commercial systems still face significant challenges: they are limited in the range of cognitive performances they are designed to test, often involve unnatural training routines affecting both animal welfare as well as result reproducibility and cannot support long-term studies, relevant for our understanding of progressive processes, such as neurodegenerative diseases and ageing. To address these challenges we have developed the smart-Kage, a novel home-cage system for autonomous phenotyping of rodent cognition and behaviour. Its robust and portable design allows easy adoption across different laboratories, while its fully automated operation limits the involvement of experimenters, both key steps towards a standardized solution of the reproducibility crisis in behavioural neuroscience. The system incorporates automated versions of gold-standard T-maze alternation, novel object recognition, and object-in-place recognition cognitive tests, while also monitoring locomotion, drinking and quiescence patterns, all within the ethologically-relevant, environmentally-enriched setting of the mouse home cage. The critical aspect of this system - and the focus of this thesis - is the machine learning pipeline, integrating convolutional neural networks, random forests and unsupervised clustering into a single software framework, capable of processing large amounts of data collected during longitudinal studies. We demonstrated the ability of our system for non-invasive diagnosis of underlying brain damage by successfully discriminating between mice with hippocampal, medial-entorhinal and sham lesions along with predicting the genotype of mice serving as models of Alzheimer’s disease with high accuracy. This technology could potentially accelerate basic and applied research in neurodegenerative diseases as well as enable large-scale behavioral screening for genes and neural circuits underlying spatial memory and other cognitive processes.