A convolutional neural network for estimating synaptic connectivity from spike trains

Abstract

The recent increase in reliable, simultaneous high channel count extracellular recordings is exciting for physiologists and theoreticians, because it offers the possibility of reconstructing the underlying neuronal circuits. We recently presented a method of inferring this circuit connectivity from neuronal spike trains by applying the generalized linear model to crosscorrelograms, GLMCC. Although the GLMCC algorithm can do a good job of circuit reconstruction, the parameters need to be carefully tuned for each individual dataset. Here we present another algorithm using a convolutional neural network for estimating synaptic connectivity from spike trains, CoNNECT. After adaptation to very large amounts of simulated data, this algorithm robustly captures the specific feature of monosynaptic impact in a noisy cross-correlogram. There are no user-adjustable parameters. With this new algorithm, we have constructed diagrams of neuronal circuits recorded in several cortical areas of monkeys. * shinomoto.shigeru.6e@kyoto-u.ac.jpRecently, we developed an estimation method that works well in balancing the conflicting demands of reducing FPs and reducing FNs [11]. The estimation method we call GLMCC (GLM cross-correlation) nonetheless has a shortcoming: the estimation results are sensitive to the model parameters, and therefore the parameters need to be tuned for the spiking data. Here, we develop another algorithm: Convolutional Neural Network for Estimating synaptic Connectivity from spike Trains (CoNNECT). The two premises are that a convolutional neural network is good at capturing the features important for distinguishing among images (in this case, cross-correlograms) and that the cross-correlogram image will contain sufficient information from which to infer the presence of monosynaptic connectivity. This new algorithm is easy to use, and it works robustly with data arising from different cortical regions in non-human primates. The convolutional neural network algorithm has tens of thousands of internal parameters [12][13][14][15]. The parameters are adjusted using hundreds of thousands of pairs of spike trains generated with a large-scale simulation of the circuitry of realistic model neurons. To reproduce large fluctuations in real spike trains, we added external fluctuations to the model neurons in the simulation.CoNNECT provides reasonable inference. It does not, however, give a rationale for why the result was derived, whereas our previous algorithm GLMCC does because it fits an interaction kernel to the cross-correlogram. These methods, therefore, have different strengths and weaknesses

A General LSTM-based Deep Learning Method for Estimating Neuronal Models and Inferring Neural Circuitry

“…Researchers have spared great efforts in designing parameter optimization approaches for decades(Prinz et al, 2003; Schneider et al, 2004; Keren et al, 2005; Gerken et al, 2006; Druckmann et al, 2007; Gurkiewicz & Korngreen, 2007; Menon et al, 2009; Hendrickson et al, 2011; Rossant et al, 2011; Santana et al, 2011; Bahl et al, 2012; Svensson et al, 2012; Markram et al, 2015; Eliasmith et al, 2016; Rumbell et al, 2016; Masoli et al, 2017; Neymotin et al, 2017; Gouwens et al, 2018; Kobayashi et al, 2019; Ladenbauer et al, 2019; Billeh et al, 2020; Endo et al, 2020; Gonçalves et al, 2020).…”

“…By performing likelihood maximization, the connectivity weights of the neural circuit can be estimated. (Kobayashi et al, 2019; Endo et al, 2020) estimate neural connectivity through the cross-correlogram(CC) from the spike trains of two neurons. (Kobayashi et al, 2019) fits the CC through a generalized linear model(GLM), and (Endo et al, 2020) estimates neural connectivity from CC by training a CNN.…”

“…(Kobayashi et al, 2019; Endo et al, 2020) estimate neural connectivity through the cross-correlogram(CC) from the spike trains of two neurons. (Kobayashi et al, 2019) fits the CC through a generalized linear model(GLM), and (Endo et al, 2020) estimates neural connectivity from CC by training a CNN. While the above three approaches can only estimate a rough synaptic weight and the existence of the connection, which they cannot estimate the connectivity strength of multiple synaptic receptors (NMDAR and AMPAR for example).…”

“…Huys et al, 2006; Q. J. Huys & Paninski, 2009; Pozzorini et al, 2015; Ladenbauer et al, 2019) or neural connectivity estimation(Kobayashi et al, 2019; Ladenbauer et al, 2019; Endo et al, 2020) alone. Sequential neural posterior estimation (SNPE)-based algorithms(Gonçalves et al, 2020) and some search-based algorithms are task-independent, while they require massive and careful feature engineering on raw neural responses, which is time-consuming and non-trivial.…”