Deducing ensemble dynamics and information flow from the whole-brain imaging data

Abstract

Recent development of large-scale activity imaging of neuronal ensembles provides opportunities for understanding how activity patterns are generated in the brain and how information is transmitted between neurons or neuronal ensembles. However, methodologies for extracting the component properties that generate overall dynamics are still limited. In this study, the results of time-lapse 3D imaging (4D imaging) of head neurons of the nematode C. elegans were analyzed by hitherto unemployed methodologies. By combining time-delay embedding with independent component analysis, the whole-brain activities were decomposed to a small number of component dynamics. Results from multiple samples, where different subsets of neurons were observed, were further combined by matrix factorization, revealing common dynamics from neuronal activities that are apparently divergent across sampled animals. By this analysis, we could identify components that show common relationships across different samples and those that show relationships distinct between individual samples. We also constructed a network model building on time-lagged prediction models of synaptic communications. This was achieved by dimension reduction of 4D imaging data using the general framework gKDR (gradient kernel dimension reduction). The model is able to decompose basal dynamics of the network. We further extended the model by incorporating probabilistic distribution, resulting in models that we call gKDR-GMM and gKDR-GP. The models capture the overall relationships of neural activities and reproduce the stochastic but coordinated dynamics in the neural network simulation. By virtual manipulation of individual neurons and synaptic contacts in this model, information flow could be estimated from whole-brain imaging results.