Application of deep neural networks on analysis of the neural basis of directional heading

Sammanfattning

The head direction (HD) system spans across several brain regions and encodes an animal’s directional heading through signals generated by HD neurons. Several studies have contributed towards the understanding of the individual HD neurons, the animal’s directional heading they prefer to fire in and their stability. Fewer studies have however focused on how these neurons form larger organisational structures (neural ensembles) and how stable these remain over long periods of time. This study applies an unsupervised deep learning algorithm on one-photon calcium imaging data recorded over a month, from the postsubiculum of 10 mice, with session numbers ranging from 7 to 75. The aim of the study is to assess whether the unsupervised machine learning model can use neural activity patterns to discover and organise neurons into directionally tuned ensembles without head direction angles as labels during training. The main pipeline version used for analysis filtered out non-HD neurons with tuning strength as selection criteria, while another version included all neurons to observe whether the model could organise the directional ensemble structures using an unfiltered population of neurons. This was achieved using a variational autoencoder (VAE) that used each recording session as training data to compress the HD neuron activity into two-dimensional latent representations. K-means clustering was applied to the model output to identify and separate clusters of directional ensembles, with post-hoc validation performed using head direction labels. The results of the VAE model showed clear ring-like structures in all 10 mice representing successful representations of the circular HD structure in absence of head direction angles provided as labels during model training. In all 10 mice and 401 sessions directional ensembles were found (tuning strength > 0.4) with the maximum tuning strength of 0.938. Stability analysis of the ensembles showed varying relationships between ensemble tuning strength and HD neuron counts such as increasing ensemble tuning strength with stable HD neuron counts, stable ensemble tuning strength with decreasing HD neuron counts, and decreasing ensemble strength with decreasing HD neuron counts. The pipeline produced reliable results despite factors such as recording quality decline.These results show that an unsupervised VAE model is able to find structures of directional ensembles from HD neuron population activity without access to labels such as head direction angles. However, longitudinal tracking of neuron membership in ensembles showed moderate to low consistency (mean Jaccard index 0.242 ± 0.046), whether this is a result of recording limitations consistent with one-photon calcium imaging or genuine ensemble reorganisation remains to be confirmed. The low overlap of neurons in matched ensembles limited reliable preferred direction tracking in the ensembles.