Aligning Transregional Neural Dynamics with Transformer-based Variational Autoencoders*
Shenghui Wu, Xiang Zhang, Yifan Huang, Yiwen Wang · 2024
Spike prediction models can reveal how different brain regions communicate via neural spiking activity. Meanwhile, low-dimensional latent variables have been widely used to describe the evolvement of neural activities in single-region analysis. However, correlations of these latent dynamics for transregional neural activity have rarely been studied. Here, we propose a unified architecture to analyze and exploit the correlation of latent dynamics between two cortical areas for spike prediction from upstream to downstream areas. The method is validated on neural population activity from the medial prefrontal cortex (mPFC) and the primary motor cortex (M1) of a Sprague Dawley rat during the two-lever discrimination task. We separately train two Transformer-based variational autoencoders (tVAEs) for mPFC and M1 neurons to extract the latent variables from their self-history spike trains. Then, we align the latent variables from the two brain regions with a regression model. By cascading the tVAE encoder for mPFC neurons and the tVAE decoder for M1 neurons through aligned latent variables, we achieve the prediction from mPFC spike train history to M1 future neural activity. The results show that the tVAEs can extract latent dynamics from the mPFC and M1 spike ensembles that resemble behavioral trajectories. We also demonstrate that the mPFC and M1 neural activity have shared latent dynamics and can be linearly aligned for spike prediction. Consequently, our method can be applied to study the evolving relationship of transregional latent dynamics and contribute to the design of future neural prostheses.