Event Abstract Back to Event Dimensionality reduction of massively recorded activity reveals sequential structure and state-dependency in dissociated neurons Yuichiro Yada1*, Ryohei Kanzaki1 and Hirokazu Takahashi1 1 The University of Tokyo, Department of Mechano-Informatics,, Japan Motivation Structured spontaneous activity may be a substrate of neural computation. Neuronal networks exhibit spatiotemporally similar patterns in spontaneous activities and sensory evoked responses [1], which suggest that the spontaneous activity has some benefits for sensory processing. Such spontaneous activity is a hallmark of a neuronal network, distinguishing the living things from artificial systems. Nevertheless, properties of neural spontaneous activity acquired in a self-organizing manner are still not comprehensively elucidated, and thus are of great interests. To capture neural properties, dimensionality reduction, which characterizes hidden structure in high-dimensional data, has been attracting attention in neuroscience along with advancement in multichannel recording. Remarkably, recent development of CMOS-based high-density microelectrode array (HDMEA), established high-resolution recording of neuronal activity from thousands of sites. Dimensionality reduction of neuronal activity massively recorded with CMOS HDMEA would characterize hidden structures in spontaneous activity. We herein attempted to test the idea on dissociated cortical neurons. Material and Methods All the animal experiments were approved by local ethical committee in the University of Tokyo. Cortical neurons derived from E18 Wistar rat embryos were plated on a CMOS HDEMAs (3brain, Switzerland) (Figs. 1A and B), which has 4,096 electrodes in 2.67 x 2.67 mm2 area [2]. Neurons were incubated in 37.0 oC, 5.0% CO2 atmosphere for about 3 weeks in vitro. Spontaneous activities of dissociated neurons were recorded for 10 minutes simultaneously from 4,096 electrodes with 7 kHz sampling rate. Spikes were detected at each electrode and binned with 10 ms time bin. The obtained spike train was factorized with non-negative matrix factorization (NMF) algorithm [3]; then, neuronal activity was represented with activities of 10 sub-populations. Kullback-Leibler divergence was adopted as a cost function for NMF in this study. Results Dissociated neurons cultured for 3 weeks in vitro were spontaneously active in a synchronized manner (Fig. 1C). With dimensionality reduction, such synchronized activity, i.e. synchronized bursts, were represented as sequential activation of sub-populations (Figs. 1D and E). Notably, repeating sequences were obviously observed in dimensionality-reduced activity. Then, synchronized bursts were clustered with a hierarchical clustering method to identify reproducible sequential patterns of sub-populations. Interestingly, such a repertoire of sub-population patterns was characterized with two properties: partial similarity and state-dependency. Multiple patterns shared a partially similar sequence of sub-population (Fig.1 F), and a particular pattern appeared successively in a period (Fig. 1E). Discussion In this research, we combined a dimensionality reduction technique, i.e. NMF, with massive multi-channel recording from CMOS-based HDMEA. Then, we demonstrated that the method obviously elucidated the structured activity in dissociated neurons. Our results indicate that the method may contribute to find not only repeatable patterns, but also structured fluctuation in neuronal networks: state-dependent spatiotemporal patterns that shared partially similar patterns. State-dependency of neural activity and similarity of neuronal activity across states were also reported in vivo [1]. Our results suggest that dissociated neurons can acquire such in-vivo like properties in a self-organizing manner. Conclusion We demonstrated that dimensionality reduction of high-dimensional data recorded from CMOS HDMEA visualized substantially clear structure of neuronal sub-population. In addition, our methods elucidated not only steady structure, but also structured fluctuation, i.e., state-dependent activity, are preserved in dissociated neurons.