Dynamic organization of visual cortical networks revealed by machine learning applied to massive spiking datasets
Curation statements for this article:-
Curated by eLife
eLife assessment
This study presents a useful method for using multi-electrode spike recordings to track the time-varying functional connectivity between neurons. However, the evidence is incomplete: a demonstration of the utility of the method relative to conventional approaches is needed. If such a demonstration is made, this could be a tool for gaining insight into circuit structure.
This article has been Reviewed by the following groups
Listed in
- Evaluated articles (eLife)
- Neuroscience (eLife)
Abstract
Complex cognitive functions in a mammalian brain are distributed across many anatomically and functionally distinct areas and rely on highly dynamic routing of neural activity across the network. While modern electrophysiology methods enable recording of spiking activity from increasingly large neuronal populations at a cellular level, development of probabilistic methods to extract these dynamic inter-area interactions is lagging. Here, we introduce an unsupervised machine learning model that infers dynamic connectivity across the recorded neuronal population from a synchrony of their spiking activity. As opposed to traditional population decoding models that reveal dynamics of the whole population, the model produces cellular-level cell-type specific dynamic functional interactions that are otherwise omitted from analysis. The model is evaluated on ground truth synthetic data and compared to alternative methods to ensure quality and quantification of model predictions. Our strategy incorporates two sequential stages – extraction of static connectivity structure of the network followed by inference of temporal changes of the connection strength. This two-stage architecture enables detailed statistical criteria to be developed to evaluate confidence of the model predictions in comparison with traditional descriptive statistical methods. We applied the model to analyze large-scale in-vivo recordings of spiking activity across mammalian visual cortices. The model enables the discovery of cellular-level dynamic connectivity patterns in local and long-range circuits across the whole visual cortex with temporally varying strength of feedforward and feedback drives during sensory stimulation. Our approach provides a conceptual link between slow brain-wide network dynamics studied with neuroimaging and fast cellular-level dynamics enabled by modern electrophysiology that may help to uncover often overlooked dimensions of the brain code.
Article activity feed
-
-
-
eLife assessment
This study presents a useful method for using multi-electrode spike recordings to track the time-varying functional connectivity between neurons. However, the evidence is incomplete: a demonstration of the utility of the method relative to conventional approaches is needed. If such a demonstration is made, this could be a tool for gaining insight into circuit structure.
-
Reviewer #1 (Public Review):
Summary:
This work proposes a new method, DyNetCP, for inferring dynamic functional connectivity between neurons from spike data. DyNetCP is based on a neural network model with a two-stage model architecture of static and dynamic functional connectivity.
This work evaluates the accuracy of the synaptic connectivity inference and shows that DyNetCP can infer the excitatory synaptic connectivity more accurately than a state-of-the-art model (GLMCC) by analyzing the simulated spike trains. Furthermore, it is shown that the inference results obtained by DyNetCP from large-scale in-vivo recordings are similar to the results obtained by the existing methods (jitter-corrected CCG and JPSTH). Finally, this work investigates the dynamic connectivity in the primary visual area VISp and in the visual areas using …
Reviewer #1 (Public Review):
Summary:
This work proposes a new method, DyNetCP, for inferring dynamic functional connectivity between neurons from spike data. DyNetCP is based on a neural network model with a two-stage model architecture of static and dynamic functional connectivity.
This work evaluates the accuracy of the synaptic connectivity inference and shows that DyNetCP can infer the excitatory synaptic connectivity more accurately than a state-of-the-art model (GLMCC) by analyzing the simulated spike trains. Furthermore, it is shown that the inference results obtained by DyNetCP from large-scale in-vivo recordings are similar to the results obtained by the existing methods (jitter-corrected CCG and JPSTH). Finally, this work investigates the dynamic connectivity in the primary visual area VISp and in the visual areas using DyNetCP.
Strengths:
The strength of the paper is that it proposes a method to extract the dynamics of functional connectivity from spike trains of multiple neurons. The method is potentially useful for analyzing parallel spike trains in general, as there are only a few methods (e.g. Aertsen et al., J. Neurophysiol., 1989, Shimazaki et al., PLoS Comput Biol 2012) that infer the dynamic connectivity from spikes. Furthermore, the approach of DyNetCP is different from the existing methods: while the proposed method is based on the neural network, the previous methods are based on either the descriptive statistics (JSPH) or the Ising model.
Weaknesses:
Although the paper proposes a new method, DyNetCP, for inferring the dynamic functional connectivity, its strengths are neither clear nor directly demonstrated in this paper. That is, insufficient analyses are performed to support the usefulness of DyNetCP.
First, this paper attempts to show the superiority of DyNetCP by comparing the performance of synaptic connectivity inference with GLMCC (Figure 2). However, the improvement in the synaptic connectivity inference does not seem to be convincing. While this paper compares the performance of DyNetCP with a state-of-the-art method (GLMCC), there are several problems with the comparison. For example:
(1) This paper focused only on excitatory connections (i.e., ignoring inhibitory neurons).
(2) This paper does not compare with existing neural network-based methods (e.g., CoNNECT: Endo et al. Sci. Rep. 2021; Deep learning: Donner et al. bioRxiv, 2024).
(3) Only a population of neurons generated from the Hodgkin-Huxley model was evaluated.
Thus, the results in this paper are not sufficient to conclude the superiority of DyNetCP in the estimation of synaptic connections. In addition, this paper compares the proposed method with the standard statistical methods Jitter-corrected CCG (Figure 3) and JPSTH (Figure 4). Unfortunately, these results do not show the superiority of the proposed method. It only shows that the results obtained by the proposed method are consistent with those obtained by the existing methods (CCG or JPSTH). This paper also compares the proposed method with standard statistical methods, such as jitter-corrected CCG (Figure 3) and JPSTH (Figure 4). It only shows that the results obtained by the proposed method are consistent with those obtained by the existing methods (CCG or JPSTH), which does not show the superiority of the proposed method.
In summary, although DyNetCP has the potential to infer synaptic connections more accurately than existing methods, the paper does not provide sufficient analysis to make this claim. It is also unclear whether the proposed method is superior to the existing methods for estimating functional connectivity, such as jitter-corrected CCG and JPSTH. Thus, the strength of DyNetCP is unclear.
-
Reviewer #2 (Public Review):
Summary:
Here the authors describe a model for tracking time-varying coupling between neurons from multi-electrode spike recordings. Their approach extends a GLM with static coupling between neurons to include dynamic weights, learned by a long-short-term-memory (LSTM) model. Each connection has a corresponding LSTM embedding and is read out by a multi-layer perceptron to predict the time-varying weight.
Strengths:
This is an interesting approach to an open problem in neural data analysis. I think, in general, the method would be interesting to computational neuroscientists.
Weaknesses:
It is somewhat difficult to interpret what the model is doing. I think it would be worthwhile to add some additional results that make it more clear what types of patterns are being described and how.
Major Issues:
Simulation …
Reviewer #2 (Public Review):
Summary:
Here the authors describe a model for tracking time-varying coupling between neurons from multi-electrode spike recordings. Their approach extends a GLM with static coupling between neurons to include dynamic weights, learned by a long-short-term-memory (LSTM) model. Each connection has a corresponding LSTM embedding and is read out by a multi-layer perceptron to predict the time-varying weight.
Strengths:
This is an interesting approach to an open problem in neural data analysis. I think, in general, the method would be interesting to computational neuroscientists.
Weaknesses:
It is somewhat difficult to interpret what the model is doing. I think it would be worthwhile to add some additional results that make it more clear what types of patterns are being described and how.
Major Issues:
Simulation for dynamic connectivity. It certainly seems doable to simulate a recurrent spiking network whose weights change over time, and I think this would be a worthwhile validation for this DyNetCP model. In particular, I think it would be valuable to understand how much the model overfits, and how accurately it can track known changes in coupling strength. If the only goal is "smoothing" time-varying CCGs, there are much easier statistical methods to do this (c.f. McKenzie et al. Neuron, 2021. Ren, Wei, Ghanbari, Stevenson. J Neurosci, 2022), and simulations could be useful to illustrate what the model adds beyond smoothing.
Stimulus vs noise correlations. For studying correlations between neurons in sensory systems that are strongly driven by stimuli, it's common to use shuffling over trials to distinguish between stimulus correlations and "noise" correlations or putative synaptic connections. This would be a valuable comparison for Figure 5 to show if these are dynamic stimulus correlations or noise correlations. I would also suggest just plotting the CCGs calculated with a moving window to better illustrate how (and if) the dynamic weights differ from the data.
-
-
-