Dystonia Project (February 2020 - Present)
Dystonia is a neurological disorder involving the basal ganglia and other brain regions that results in painful, involuntary muscle movements and affects over 250,000 people in the United States, making it the third most common movement disorder (Jinnah & Hess, 2017). M
utant mice, known as DRD mice, which exhibit dystonic movements due to a lack of dopamine are used for calcium-imaging samples of population level neurons in the striatum, a region of the basal ganglia (Rose, 2015). However, a current complication is that due to concerns of phototoxicity, the calcium imaging can only be recorded for minutes at a time(Icha, 2017). This concern, combined with the infrequent, dystonic muscle contractions, results in poor and capricious data collection.
I’m currently building a deep-learning computational model that is able to predict when these dystonic movements will occur. To create this model, the positions of the mice are estimated using DeepLabCut, a 3-D marker-less pose estimating tool that uses a deep-learning classifier to mark body segments at each frame of a video feed(Mathis, 2019). I then feed these position vectors to a trained LSTM (Long Short Term Memory Neural Network) so that it can make real-time predictions of dystonic movements. As the model predicts an oncoming episode of muscle flexion, an electronic signal is sent to the calcium-imaging implant that can then be turned on for a safe amount of time, greatly increasing data capturing abilities. This additional data could potentially lead to further discoveries regarding the nature of dystonia and form the infrastructure of a future cure.
LFADS Analysis (August 2020 - Present)
LFADS (Latent Factor Analysis via Dynamical Systems) utilizes a sequential auto-encoder to recreate spiking data in single-trial neural populations. Previously, due to noise and single-trial variability, single-trial data analysis was infeasible. As demonstrated here, LFADS is able to accurately predict observed behavioral variables.
In the Pandarinath Lab, I assist with further LFADS decomposition, creating additional functions to corroborate its usage with RDS (Reza’s Data Structure), a structure to hold neural data. I use these tools to further identify the underlying dynamical systems that govern the activity of neural populations. By identifying the latent factors that determine neural spiking, we can create intricate brain machine interfaces that can normalize the lives of paraplegics and amputees.