ReachingBot: an automated and scalable benchtop device for highly parallel Single Pellet Reach-and-Grasp training and assessment in mice

Abstract

The single pellet reaching and grasp (SPRG) task is a behavioural assay widely used to study motor learning, control and recovery after nervous system injury in animals. The manual training and assessment of the SPRG is labour intensive and time consuming and has led to the development of multiple devices which automate the SPRG task. Current state-of-the-art desktop methods either still require attendance, manual classification of trial outcome, or expensive locally-installed hardware such as graphical processing units (GPUs). Here, using robotics, computer vision, and machine learning analysis of videos, we describe a novel cost-effective benchtop device that can be left unattended, presents pellets to mice automatically, video records each trial, and, using two supervised learning algorithms, classifies the outcome of each trial automatically with an accuracy of greater than 94% without the use of GPUs. Finally, the device is simple in design with few components meaning manufacturing at scale is straightforward and, with few moving parts, reliable and robust. Our devices can also be operated using our cross-platform Graphical User Interface (GUI), meaning no knowledge of programming is required by its users. We show that these devices can train 30 mice with them collectively performing ~83,000 trials over 3 months, saving users an estimated 8 and half hours of labour per day. Over five weeks, most mice undertook more trials per session and retrieved more pellets successfully. 21 out of 30 mice retrieved at least 40% of pellets successfully in at least one session during the training period. Devices measured motor deficits induced in mice by a focal ischaemic stroke; some mice showed large persistent deficits whilst others showed only transient deficits. This highlights the heterogeneity in reaching outcomes following stroke. We conjecture that reach-and-grasp is represented in motor cortex bilaterally but with greater asymmetry in some mice than in others. We predict that bilateral lesions of motor cortex would cause long-lasting deficits in reach-and-grasp in mice. We propose a strategy for preclinical evaluation of novel therapeutics that improve reach-and-grasp by pre-screening a large cohort of mice automatically and excluding those that fail to achieve pre-specific success rates, which generates a cohort of mice trained with consistent performance levels, suitable for randomization to treatment arms in a preclinical study. Well-powered sample sizes are easily achievable. Highly parallel automated training and assessment should accelerate the development of new therapies for movement disorders.