Introduction A single pyramidal neuron can compute XOR using its dendritic structure [1]. This suggests that neuronal morphology is closely related to computational capability. Since neurons exhibit diverse morphologies [2], individual neurons may possess distinct computational capabilities. To reveal the relationship between neuronal morphology and computational capability, simulation experiments using neuron models with diverse morphologies are effective. However, constructing realistic neuron models that capture the morphologies of neurons is difficult because the essential features that characterize neuronal morphology are poorly understood. This study aims to investigate whether a Variational Auto-Encoder (VAE) is effective for feature extraction.
Methods A toy neuron dataset consisting of single-branch neurons was created to train the VAE. The toy neurons were generated using several morphometric features, including node type (elongation, branch, or terminal), angle, and elongation length. The VAE was trained to map input neuronal morphologies to a low-dimensional latent space and reconstruct them from the space. The space was analyzed using latent traversal [3]. In this method, the VAE was first inputted a toy neuron morphology from the dataset, which was then reconstructed. Next, the value of one variable spanning the space was varied while the others were fixed. We then evaluated which morphometric features the latent variable represented based on changes in the reconstructed morphology.
Results The VAE successfully reconstructed morphologies that resembled the toy neurons. The reconstructed morphologies gradually varied in response to changes in the latent variables. Those changes reflected the morphometric features used to create the toy neurons. When the input toy neuron data were replaced with another neuron, the features represented by each latent variable sometimes differed; however, across all variables in the latent space, all the features were extracted.
Discussion These results suggest that the VAE is a useful approach for extracting morphometric features. The dependence of the extracted features on the input morphology suggests that the VAE may implicitly cluster training data in the latent space and extract cluster-specific morphometric features. Future work is to apply the proposed approach to real neuronal data, such as pyramidal neurons.
References [1] Gidon, A., et al. (2020). Dendritic action potentials and computation in human layer 2/3 cortical neurons. Science. 367:83–87. [2] Peng, H., et al. (2021). Morphological diversity of single neurons in molecularly defined cell types. Nature. 598 (7879):174–181. [3] Burgess, C. P., et al. (2018) Understanding disentangling in beta-VAE. arXiv preprint arXiv:1804.03599.
Acknowledgement This research was supported by AMED under Grant Number JP25wm0625418h0001.
