Mapping circuit dynamics during function and dysfunction

Article activity feed

1. 

   eLife

   Jul 18, 2022

   Author Response

   Reviewer #1 (Public Review):

   The authors sought to establish a standardized quantitative approach to categorize the activity patterns in a central pattern generator (specifically, the well-studied pyloric circuit in C. borealis). While it is easy to describe these patterns under "normal" conditions, this circuit displays a wide range of irregular behaviors under experimental perturbations. Characterizing and cataloguing these irregular behaviors is of interest to understand how the network avoids these dysfunctional patterns under "normal" circumstances.

   The authors draw upon established machine learning tools to approach this problem. To do so, they must define a set of features that describe circuit activity at a moment in time. They use the distribution of inter-spike-intervals ISIs and spike phases of the LP and …

   Author Response

   Reviewer #1 (Public Review):

   The authors sought to establish a standardized quantitative approach to categorize the activity patterns in a central pattern generator (specifically, the well-studied pyloric circuit in C. borealis). While it is easy to describe these patterns under "normal" conditions, this circuit displays a wide range of irregular behaviors under experimental perturbations. Characterizing and cataloguing these irregular behaviors is of interest to understand how the network avoids these dysfunctional patterns under "normal" circumstances.

   The authors draw upon established machine learning tools to approach this problem. To do so, they must define a set of features that describe circuit activity at a moment in time. They use the distribution of inter-spike-intervals ISIs and spike phases of the LP and PD neuron as these features. As the authors mention in their Discussion section, these features are highly specialized and adapted to this particular circuit. This limits the applicability of their approach to other circuits with neurons that are unidentifiable or very large in number (the number of spike phase statistics grows quadratically with the number of neurons).

   We agree with the reviewer that the size of the feature vectors as described grows quadratically with the number of neurons. The feature sets we describe are most suited for “identified” neurons – neurons whose identity and connectivity are known and can be reliably recorded from multiple animals. The method described here is best suited for systems with small numbers of identified neurons. For other systems, other feature vectors may be chosen, as we have suggested in the Discussion: Applicability to other systems.

   The main results of the paper provide evidence that ISIs and spike phase statistics provide a reasonable descriptive starting point for understanding the diversity of pyloric circuit patterns. The authors rely heavily on t-distributed stochastic neighbor embedding (tSNE), a well-known nonlinear dimensionality reduction method, to visualize activity patterns in a low-dimensional, 2D space. While effective, the outputs of tSNE have to be interpreted with great care (Wattenberg, et al., "How to Use t-SNE Effectively", Distill, 2016. http://doi.org/10.23915/distill.00002). I think the conclusions of this paper would be strengthened if additional machine learning models were applied to the ISI and spike phase features, and if those additional models validated the qualitative results shown by tSNE. For example, tSNE itself is not a clustering method, so applying clustering methods directly to the high-dimensional data features would be a useful validation of the apparent low-dimensional clusters shown in the figures.

   We thank the reviewer for these suggestions, and agree with the reviewer that t-SNE is not a clustering method, and directly clustering on t-SNE embeddings is rife with complexities. Instead we have used t-SNE to generate a visualization that allows domain experts to quickly label and cluster large quantities of data. This makes a previously intractable task feasible, and offers some basic guarantees on quality (e.g., no one data point can have two labels, because labels derive from position of data points in two dimensional space). In addition:

   • We used uMAP, another dimensionality reduction algorithm, to perform the embedding step, and colored points by the original t-SNE embedding. (Figure 3—figure supplement 3). Large sections of the map are still strikingly colored in single colors, suggesting that the manual clustering did not depend on the details of the t-SNE algorithm, but is rather informed by the statistics of the data.

   • We validated our method using synthetic data. We generated synthetic spike trains from different “classes” and embedded the resultant feature vectors using t-SNE. Data from different classes are not intermingled, and form tight “clusters” (Figure 2 -- figure supplement 4).

   • Finally, we attempted to use hierarchical clustering to cluster the raw feature vectors, and were not able to find a reasonable portioning of the linkage tree that separated qualitatively different spike patterns (Figure at the top of this document). We speculate that this is because feature vectors may contain outliers that bias clustering algorithms that attempt to preserve global distance to lump the majority of the data into a single cluster, in order to differentiate outliers from the bulk of the data.

   The authors do show that the algorithmically defined clusters agree with expert-defined clusters. (Or, at least, they show that one can come up with reasonable post-hoc explanations and interpretations of each cluster). The very large cluster of "regular" patterns -- shown typically in a shade of blue -- actually looks like an archipelago of smaller clusters that the authors have reasoned should be lumped together. Thus, while the approach is still a useful data-driven tool, a non-trivial amount of expert knowledge is baked into the results. A central challenge in this line of research is to understand how sensitive the outcomes are to these modeling choices, and there is unlikely to be a definitive answer.

   We agree with the reviewer entirely.

   Nonetheless, the authors show results which suggest that this analysis framework may be useful for the community of researchers studying central pattern generators. They use their method to qualitatively characterize a variety of network perturbations -- temperature changes, pH changes, decentralization, etc.

   In some cases it is difficult to understand the level of certainty in these qualitative observations. A first look at Figure 5a suggests that three different kinds of perturbations push the circuit activity into different dysfunctional cluster regions. However, the apparent spatial differences between these three groups of perturbations might be due to animal-level differences (i.e. each preparation produces multiple points in the low-D plot, so the number of effective statistical replicates is smaller than it appears at first glance). Similarly, in Figure 9, it is somewhat hard to understand how much the state occupancy plots would change if more animals were collected -- with the exception of proctolin, there are ~25 animals and 12 circuit activity clusters which may not be a favorable ratio. It would be useful if a principled method for computing "error bars" on these occupancy diagrams could be developed. Similar "error bars" on the state transition diagrams (e.g. Fig 6a) would also be useful.

   We agree with the reviewer. Despite this paper containing data from hundreds of animals, the dataset may not be sufficiently large to perform some necessary statistical checks. We agree with the reviewer that a more rigorous error analysis would be useful, but is not trivially done.

   Finally, one nagging concern that I have is that the ISIs and spike phase statistics aren't the ideal features one would use to classify pyloric circuit behaviors. Sub-threshold dynamics are incredibly important for this circuit (e.g. due to electrical coupling of many neurons). A deeper discussion about what is potentially lost by only having access to the spikes would be useful.

   We agree with the reviewer that spike times aren’t the ideal feature to use to describe circuit dynamics. This is especially true in the STG, where synapses are graded, and coupling between cells can persist without spiking. However, the data required simply do not exist, as it requires intracellular recordings, which are substantially harder to perform (and maintain over challenging perturbations) than extracellular recordings.

   Finally, the signal to the muscles – arguably the physiologically and functionally relevant signal – is the spike signal, suggesting that spike patterns from the pyloric circuit are a useful feature to measure. Nevertheless, this is an important point, and we thank the reviewer for raising it, and we have included it in the section titled Discussion: Technical considerations.

   Overall, I think this work provides a useful starting point for large-scale quantitative analysis of CPG circuit behaviors, but there are many additional hurdles to be overcome.

   Reviewer #2 (Public Review):

   This manuscript uses the t-SNE dimensionality reduction technique to capture the rich dynamics of the pyloric circuit of the crab.

   Strengths:

   • The integration of a rich data-set of spiking data from the pyloric circuit

   • Use of nonlinear dimension reduction (t-SNE) to visualise that data

   • Use of clusters from that t-SNE visualisation to create subsets of data that are amenable to consistent analyses (such as using the "regular" cluster as a basis for surveying the types of dynamics possible in baseline conditions)

   • Innovative use of the cluster types to describe transitions between dynamics within the baseline state and within perturbed states (whether by changes to exogenous variables, cutting nerves, or applying neuromodulators)

   • Some interesting main results: o Baseline variability in the spiking patterns of the pyloric circuit is greater within than between animals

   o Transitions to silent states often (always?) pass through the same intermediate state of the LP neuron skipping spikes

   Weaknesses:

   • t-SNE is not, in isolation, a clustering algorithm, yet here it is treated as such. How the clusters were identified is unclear: the manuscript mentions manual curation of randomly sampled points, implying that the clusters were extrapolations from these. This would seem to rather defeat the point of using unsupervised techniques to obtain an unbiased survey of the spiking dynamics, and raises the issue of how robust the clusters are

   We have used t-SNE to visualize the circuit dynamics in a two-dimensional map. We have exploited t-SNE’s ability to preserve local structure to generate an embedding where a domain expert can efficiently manually identify and label stereotyped clusters of activity. As the author points out, this is a manual step, and we have emphasized this in the manuscript. The strength of our approach is to combine the power of a nonlinear dimensionality reduction technique such as t-SNE with human curation to make a task that was previously impossible (identifying and labelling very large datasets of neural activity) feasible.

   To address the question of how robust the manually identified clusters are, we have:

   1. used another dimensionality reduction technique, uMAP, to generate an embedding and colored points by the original t-SNE map (Figure 3 – figure supplement 3). To rough approximation, the coloring reveals that a similar clustering exists in this uMAP embedding.

   2. We generated synthetic spike trains from pre-determined spike pattern classes and used the feature vector extraction and t-SNE embedding procedure as described in the paper. We found that this generated a map (Figure 2—figure supplement 4) where classes of spike patterns were well separated in the t-SNE space.

   • the main purpose and contribution of the paper is unclear, as the results are descriptive, and mostly state that dynamics in some vary between different states of the circuit; while the collated dataset is a wonderful resource, and the map is no doubt useful for the lab to place in context what they are looking at, it is not clear what we learn about the pyloric circuit, or more widely about the dynamical repertoire of neural circuits

   • in some places the contribution is noted as being the pipeline of analysis: unfortunately as the pipeline used here seems to rely in manual curation, it is of limited general use; moreover, there are already a number of previous works that use unsupervised machine-learning pipelines to characterise the complexity of spiking activity across a large data-set of neurons, using the same general approach here (quantify properties of spiking as a vector; map/cluster using dimension reduction), including Baden et al (2016, Nature), Bruno et al (2015, Neuron), Frady et al (2016, Neural Computation).

   • Some key limitations are not considered:

   o the omission of the PY neuron activity means that the map as given is incomplete: potentially there are many more states, and hence transitions, within or beyond those already found that correspond to changes in PY neuron activity

   We agree with the reviewer that the omission of the PY neurons’ activity means that the map is incomplete. There are likely many more states, and hence many more transitions, than the ones we have identified. In addition, we note that there are other pyloric neurons whose activity is also missing (AB, IC, LPG, VD). However, measuring just LP and PD allows us to monitor the activity of the most important functional antagonists in the system (because they are effectively in a half-center oscillator because PD is electrically coupled to AB). In general, the more neurons one measures, the richer the description of the circuit dynamics will be. Collecting datasets at this scale (~500 animals) from all pyloric neurons is challenging, and we have revised the manuscript to make this important point (see Discussion: Technical considerations).

   o The use of long, non-overlapping time segments (20s) - this means, for example, that the transitions are slow and discrete, whereas in reality they may be abrupt, or continuous.

   We agree with the reviewer. There are tradeoffs in choosing a bin size in analyzing time series – choosing longer bins can increase the number of “states” and choosing shorter bins can increase the number of transitions. We chose 20s bins because it is long enough to include several cycles of the pyloric rhythm, even when decentralized, yet was short enough to resolve slow changes in spiking. We have included a statement clarifying this (see Discussion: Technical considerations).

   o tSNE cannot capture hierarchical structure, nor has a null model to demonstrate that the underlying data contains some clustering structure. So, for example, distances measured on the map may not be strictly meaningful if the data is hierarchical.

   We agree with the reviewer. t-SNE can manifest clusters when none exist (Section 4 of https://distill.pub/2016/misread-tsne/) and can obscure or merge true clusters. We have restricted analyses that rely on distances measured in the map to cases where there are qualitative differences in behavior (e.g., with decentralization, Fig 7) or have compared distances within subsets of data where a single parameter is changed (e.g., pH or temperature, Fig 5). The only conclusion we draw from these distance measures is that data are more (or less) spread out in the map, which we use as a proxy for variability. We have included a statement discussion limitations of using t-SNE (Discussion: Comparison with other methods).

   • the Discussion does not include enough insight and contextualisation of the results.

   We have completely rewritten the discussion to address this.

   Reviewer #3 (Public Review):

   Gorur-Shandilya et al. apply an unsupervised dimensionality reduction (t-SNE) to characterize neural spiking dynamics in the pyloric circuit in the stomatogastric ganglion of the crab. The application of unsupervised methods to characterize qualitatively distinct regimes of spiking neural circuits is very interesting and novel, and the manuscript provides a comprehensive demonstration of its utility by analyzing dynamical variability in function and dysfunction in an important rhythm-generating circuit. The system is highly tractable with small numbers of neurons, and the study here provides an important new characterization of the system that can be used to further understand the mapping between gene expression, circuit activity, and functional regimes. The explicit note about the importance of visualization and manual labeling was also nice, since this is often brushed under the rug in other studies.

   Major concern:

   While the specific analysis pipeline clearly identifies qualitatively distinct regimes of spike patterns in the LP/PD neurons, it is not clear how much of this is due to t-SNE itself vs the initial pre-processing and feature definition (ISI and spike phase percentiles). Analyses that would help clarify this would be to check whether the same clusters emerge after (1) applying ordinary PCA to the feature vectors and plotting the projections of the data along the first two PCs, or (2) defining input features as the concatenated binned spike rates over time of the LP & PD neurons (which would also yield a fixed-length vector per 20 s trial), and then passing these inputs to PCA or tSNE. As the significance of this work is largely motivated by using unsupervised vs ad hoc descriptors of circuit dynamics, it will be important to clarify how much of the results derive from the use of ISI and phase representation percentiles, etc. as input features, vs how much emerge from the dimensionality reduction.

   We agree with the reviewer that is important to clarify how much of our results come from the data itself, and how we parameterize them using ISIs and phases, and how much comes from the choice of t-SNE as a dimensionality reduction algorithm. We have addressed this concern in the following ways:

   1. We used principal components analysis on the feature vectors and measured triadic differences in features such as the period and duty cycle of the PD neuron. We found that triadic differences were lower in the t-SNE embedding than in the first two PCA features, or in shuffled t-SNE embeddings (Figure 2– Figure supplement 2), suggesting that the embedding is creating a useful representation that captures key features of the data.

   2. We have used uMAP to reduce the dimensionality of the feature matrix to two dimensions and found that it too preserved the coarse features of the embedding that we observe with t-SNE. Coloring the uMAP embedding by the t-SNE labels revealed that the overall classification scheme was intact (Fig 3 – figure supplement 3).

   3. We generated a synthetic dataset and applied the unsupervised part of our algorithm to it (conversion to ISIs, phases, etc., then t-SNE). We colored the points in the t-SNE embedding by the category in the synthetic dataset. We found that categories were well separated in the t-SNE plot, and each cluster tended to have a single color. This validates the overall power of our approach and shows that it can recover clustering information in large spike sets (Figure 2—figure supplement 4).

   4. We have run k-means and hierarchical clustering on the feature vectors directly and shown that our method is superior to these naïve clustering algorithms running on the feature vectors. We speculate that this is because these clustering methods attempt to partition the full space using global distances, at the expense of distance along the manifold on which the data is located. Algorithms like t-SNE are biased towards local distances, and discount global distances between points outside a neighborhood, and are this better suited here.

   Read the original source
2. 

   eLife

   Jun 14, 2022

   Evaluation Summary:

   This manuscript presents a method to characterize diverse neural activity patterns arising from a small invertebrate circuit. This is of practical interest to invertebrate neuroscientists. The application of unsupervised methods to characterize qualitatively distinct regimes of spiking neural circuits is very interesting. The challenges and lessons learned in this study are therefore of broader interest to those seeking to quantitatively characterize large sets of neural data across many subjects. The survey could be improved by further validation of the derived clusters.

   (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 agreed to share their name with the authors.)

   Read the original source
3. 

   eLife

   Jun 14, 2022

   Reviewer #1 (Public Review):

   The authors sought to establish a standardized quantitative approach to categorize the activity patterns in a central pattern generator (specifically, the well-studied pyloric circuit in C. borealis). While it is easy to describe these patterns under "normal" conditions, this circuit displays a wide range of irregular behaviors under experimental perturbations. Characterizing and cataloguing these irregular behaviors is of interest to understand how the network avoids these dysfunctional patterns under "normal" circumstances.

   The authors draw upon established machine learning tools to approach this problem. To do so, they must define a set of features that describe circuit activity at a moment in time. They use the distribution of inter-spike-intervals ISIs and spike phases of the LP and PD neuron as these …

   Reviewer #1 (Public Review):

   The authors sought to establish a standardized quantitative approach to categorize the activity patterns in a central pattern generator (specifically, the well-studied pyloric circuit in C. borealis). While it is easy to describe these patterns under "normal" conditions, this circuit displays a wide range of irregular behaviors under experimental perturbations. Characterizing and cataloguing these irregular behaviors is of interest to understand how the network avoids these dysfunctional patterns under "normal" circumstances.

   The authors draw upon established machine learning tools to approach this problem. To do so, they must define a set of features that describe circuit activity at a moment in time. They use the distribution of inter-spike-intervals ISIs and spike phases of the LP and PD neuron as these features. As the authors mention in their Discussion section, these features are highly specialized and adapted to this particular circuit. This limits the applicability of their approach to other circuits with neurons that are unidentifiable or very large in number (the number of spike phase statistics grows quadratically with the number of neurons).

   The main results of the paper provide evidence that ISIs and spike phase statistics provide a reasonable descriptive starting point for understanding the diversity of pyloric circuit patterns. The authors rely heavily on t-distributed stochastic neighbor embedding (tSNE), a well-known nonlinear dimensionality reduction method, to visualize activity patterns in a low-dimensional, 2D space. While effective, the outputs of tSNE have to be interpreted with great care (Wattenberg, et al., "How to Use t-SNE Effectively", Distill, 2016. http://doi.org/10.23915/distill.00002). I think the conclusions of this paper would be strengthened if additional machine learning models were applied to the ISI and spike phase features, and if those additional models validated the qualitative results shown by tSNE. For example, tSNE itself is not a clustering method, so applying clustering methods directly to the high-dimensional data features would be a useful validation of the apparent low-dimensional clusters shown in the figures.

   The authors do show that the algorithmically defined clusters agree with expert-defined clusters. (Or, at least, they show that one can come up with reasonable post-hoc explanations and interpretations of each cluster). The very large cluster of "regular" patterns -- shown typically in a shade of blue -- actually looks like an archipelago of smaller clusters that the authors have reasoned should be lumped together. Thus, while the approach is still a useful data-driven tool, a non-trivial amount of expert knowledge is baked into the results. A central challenge in this line of research is to understand how sensitive the outcomes are to these modeling choices, and there is unlikely to be a definitive answer.

   Nonetheless, the authors show results which suggest that this analysis framework may be useful for the community of researchers studying central pattern generators. They use their method to qualitatively characterize a variety of network perturbations -- temperature changes, pH changes, decentralization, etc.

   In some cases it is difficult to understand the level of certainty in these qualitative observations. A first look at Figure 5a suggests that three different kinds of perturbations push the circuit activity into different dysfunctional cluster regions. However, the apparent spatial differences between these three groups of perturbations might be due to animal-level differences (i.e. each preparation produces multiple points in the low-D plot, so the number of effective statistical replicates is smaller than it appears at first glance). Similarly, in Figure 9, it is somewhat hard to understand how much the state occupancy plots would change if more animals were collected -- with the exception of proctolin, there are ~25 animals and 12 circuit activity clusters which may not be a favorable ratio. It would be useful if a principled method for computing "error bars" on these occupancy diagrams could be developed. Similar "error bars" on the state transition diagrams (e.g. Fig 6a) would also be useful.

   Finally, one nagging concern that I have is that the ISIs and spike phase statistics aren't the ideal features one would use to classify pyloric circuit behaviors. Sub-threshold dynamics are incredibly important for this circuit (e.g. due to electrical coupling of many neurons). A deeper discussion about what is potentially lost by only having access to the spikes would be useful.

   Overall, I think this work provides a useful starting point for large-scale quantitative analysis of CPG circuit behaviors, but there are many additional hurdles to be overcome.

   Read the original source
4. 

   eLife

   Jun 14, 2022

   Reviewer #2 (Public Review):

   This manuscript uses the t-SNE dimensionality reduction technique to capture the rich dynamics of the pyloric circuit of the crab.

   Strengths:
   - The integration of a rich data-set of spiking data from the pyloric circuit
   - Use of nonlinear dimension reduction (t-SNE) to visualise that data
   - Use of clusters from that t-SNE visualisation to create subsets of data that are amenable to consistent analyses (such as using the "regular" cluster as a basis for surveying the types of dynamics possible in baseline conditions)
   - Innovative use of the cluster types to describe transitions between dynamics within the baseline state and within perturbed states (whether by changes to exogenous variables, cutting nerves, or applying neuromodulators)

   - Some interesting main results:
   o Baseline variability in the spiking …

   Reviewer #2 (Public Review):

   This manuscript uses the t-SNE dimensionality reduction technique to capture the rich dynamics of the pyloric circuit of the crab.

   Strengths:
   - The integration of a rich data-set of spiking data from the pyloric circuit
   - Use of nonlinear dimension reduction (t-SNE) to visualise that data
   - Use of clusters from that t-SNE visualisation to create subsets of data that are amenable to consistent analyses (such as using the "regular" cluster as a basis for surveying the types of dynamics possible in baseline conditions)
   - Innovative use of the cluster types to describe transitions between dynamics within the baseline state and within perturbed states (whether by changes to exogenous variables, cutting nerves, or applying neuromodulators)

   - Some interesting main results:
   o Baseline variability in the spiking patterns of the pyloric circuit is greater within than between animals
   o Transitions to silent states often (always?) pass through the same intermediate state of the LP neuron skipping spikes

   Weaknesses:
   - t-SNE is not, in isolation, a clustering algorithm, yet here it is treated as such. How the clusters were identified is unclear: the manuscript mentions manual curation of randomly sampled points, implying that the clusters were extrapolations from these. This would seem to rather defeat the point of using unsupervised techniques to obtain an unbiased survey of the spiking dynamics, and raises the issue of how robust the clusters are
   - the main purpose and contribution of the paper is unclear, as the results are descriptive, and mostly state that dynamics in some vary between different states of the circuit; while the collated dataset is a wonderful resource, and the map is no doubt useful for the lab to place in context what they are looking at, it is not clear what we learn about the pyloric circuit, or more widely about the dynamical repertoire of neural circuits
   - in some places the contribution is noted as being the pipeline of analysis: unfortunately as the pipeline used here seems to rely in manual curation, it is of limited general use; moreover, there are already a number of previous works that use unsupervised machine-learning pipelines to characterise the complexity of spiking activity across a large data-set of neurons, using the same general approach here (quantify properties of spiking as a vector; map/cluster using dimension reduction), including Baden et al (2016, Nature), Bruno et al (2015, Neuron), Frady et al (2016, Neural Computation).

   - Some key limitations are not considered:
   o the omission of the PY neuron activity means that the map as given is incomplete: potentially there are many more states, and hence transitions, within or beyond those already found that correspond to changes in PY neuron activity
   o The use of long, non-overlapping time segments (20s) - this means, for example, that the transitions are slow and discrete, whereas in reality they may be abrupt, or continuous.
   o tSNE cannot capture hierarchical structure, nor has a null model to demonstrate that the underlying data contains some clustering structure. So, for example, distances measured on the map may not be strictly meaningful if the data is hierarchical.
   - the Discussion does not include enough insight and contextualisation of the results.

   Read the original source
5. 

   eLife

   Jun 14, 2022

   Reviewer #3 (Public Review):

   Gorur-Shandilya et al. apply an unsupervised dimensionality reduction (t-SNE) to characterize neural spiking dynamics in the pyloric circuit in the stomatogastric ganglion of the crab. The application of unsupervised methods to characterize qualitatively distinct regimes of spiking neural circuits is very interesting and novel, and the manuscript provides a comprehensive demonstration of its utility by analyzing dynamical variability in function and dysfunction in an important rhythm-generating circuit. The system is highly tractable with small numbers of neurons, and the study here provides an important new characterization of the system that can be used to further understand the mapping between gene expression, circuit activity, and functional regimes. The explicit note about the importance of …

   Reviewer #3 (Public Review):

   Gorur-Shandilya et al. apply an unsupervised dimensionality reduction (t-SNE) to characterize neural spiking dynamics in the pyloric circuit in the stomatogastric ganglion of the crab. The application of unsupervised methods to characterize qualitatively distinct regimes of spiking neural circuits is very interesting and novel, and the manuscript provides a comprehensive demonstration of its utility by analyzing dynamical variability in function and dysfunction in an important rhythm-generating circuit. The system is highly tractable with small numbers of neurons, and the study here provides an important new characterization of the system that can be used to further understand the mapping between gene expression, circuit activity, and functional regimes. The explicit note about the importance of visualization and manual labeling was also nice, since this is often brushed under the rug in other studies.

   Major concern:

   While the specific analysis pipeline clearly identifies qualitatively distinct regimes of spike patterns in the LP/PD neurons, it is not clear how much of this is due to t-SNE itself vs the initial pre-processing and feature definition (ISI and spike phase percentiles). Analyses that would help clarify this would be to check whether the same clusters emerge after (1) applying ordinary PCA to the feature vectors and plotting the projections of the data along the first two PCs, or (2) defining input features as the concatenated binned spike rates over time of the LP & PD neurons (which would also yield a fixed-length vector per 20 s trial), and then passing these inputs to PCA or t-SNE. As the significance of this work is largely motivated by using unsupervised vs ad hoc descriptors of circuit dynamics, it will be important to clarify how much of the results derive from the use of ISI and phase representation percentiles, etc. as input features, vs how much emerge from the dimensionality reduction.

   Read the original source
6. 

   Version published to 10.7554/elife.76579 on eLife

   Mar 18, 2022
7. 

   Version published to 10.1101/2021.07.06.451370 on bioRxiv

   Jul 7, 2021