Synaptic and dendritic architecture of different types of hippocampal somatostatin interneurons

GABAergic inhibitory neurons fundamentally shape the activity and plasticity of cortical circuits. A major subset of these neurons contains somatostatin (SOM); these cells play crucial roles in neuroplasticity, learning, and memory in many brain areas including the hippocampus, and are implicated in several neuropsychiatric diseases and neurodegenerative disorders. Two main types of SOM-containing cells in area CA1 of the hippocampus are oriens-lacunosum-moleculare (OLM) cells and hippocampo-septal (HS) cells. These cell types show many similarities in their soma-dendritic architecture, but they have different axonal targets, display different activity patterns in vivo, and are thought to have distinct network functions. However, a complete understanding of the functional roles of these interneurons requires a precise description of their intrinsic computational properties and their synaptic interactions. In the current study we generated, analyzed, and make available several key data sets that enable a quantitative comparison of various anatomical and physiological properties of OLM and HS cells in mouse. The data set includes detailed scanning electron microscopy (SEM)-based 3D reconstructions of OLM and HS cells along with their excitatory and inhibitory synaptic inputs. Combining this core data set with other anatomical data, patch-clamp electrophysiology, and compartmental modeling, we examined the precise morphological structure, inputs, outputs, and basic physiological properties of these cells. Our results highlight key differences between OLM and HS cells, particularly regarding the density and distribution of their synaptic inputs and mitochondria. For example, we estimated that an OLM cell receives about 8,400, whereas an HS cell about 15,600 synaptic inputs, about 16% of which are GABAergic. Our data and models provide insight into the possible basis of the different functionality of OLM and HS cell types and supply essential information for more detailed functional models of these neurons and the hippocampal network.

The manuscript is very well written and illustrated.The set of data provided in this study is very important for our understanding how neurons of similar / comparable morphology may differentially influence signal propagation.The study shows that details of input size and distribution matter.
We are grateful to Reviewer 1 for their constructive comments and hope that the additional computational work included in this revision further demonstrates the potential of this large dataset.
However, my enthusiasm is diminished by the lacking experimental evidences on single input conductances, which could have been measured by e.g.extracellular minimal stimulation along the somato-dendritic axis (dendritic recordings would have been ideal, however, difficult to perform).Moreover, it remains unclear how the different input properties and their distribution may contribute to spatial and temporal signal integration and the control of action potential generation in the two cell types.In other words, what is the functional difference of both cell types, how do the different input distributions and sizes manifest in potential physiological differences in signal integration?
We agree with the Reviewer that these are important questions, and we answer them in detail below.
What is the real single input conductance in both cell types?One option would be to stimulate with minimal intensity synaptic inputs (glutamatergic / GABAergic) at distinct distances between soma and dendrite and measure IPSC / EPSC sizes.The idea behind the proposal is to get a bit more realistic estimates on the conductances used in the single cell models.
We agree with the Reviewer that knowing the correct values of synaptic conductances would allow more realistic modeling of how synaptic inputs affect the output of these neurons.As far as we know, the precise values of synaptic conductances that characterize the various input pathways targeting these cells have not been measured.Precise measurement of synaptic conductances would require dendritic patch-clamp recordings in combination with local activation of synapses at specific locations (we note that injection of a high-osmolarity sucrose solution rather than electrical stimulation has been used in most such experiments, probably because it allows better control of the location and strength of the stimulation).Somatic recordings alone would not be sufficient to measure the true synaptic currents and conductances due to space clamp issues and the (unknown) attenuation of synaptic currents between the synapse and the soma.On the other hand, the necessary dendritic recordings (especially from thin distal dendrites) would be technically extremely challenging, for which only a few laboratories have the necessary expertise available.Therefore, we decided to use approximate values for the synaptic conductances in our study, which led to somatic postsynaptic currents and potentials that are in agreement with experimentally measured responses (see Methods: Modeling: Model testing section of the manuscript).Nevertheless, the models (which we will also share publicly) can be easily updated and re-run if more precise data on these synaptic conductances become available.

2.
The here provided single cell models would be ideally suited to examine in more detail differences in the temporal and spatial synaptic integration of excitatory inputs and the influence of inhibition on the dendritic integrative processes in dependence of dendritic location.The impact of inhibition on the very distal dendrites of OLM cells should be stronger than at more proximal sides.Based on Figure 8, the E/I ratio seems to switch between proximal and distal dendrites of OLM cells but appears to be constant in HL neurons.For precise investigations of the role of inhibition in the integration of EPSPs it is of course important to have good estimates on the reversal potential of IPSCs.
We share the Reviewer's interest in all of these functional issues, and hope that, eventually, the data produced in this study will contribute to answering these important questions.The primary goal of the present Resources article is to describe our multi-modal dataset that we hope will be utilized in various modeling studies and inspire further experiments.We believe that a comprehensive modeling study of spatio-temporal synaptic integration and action potential generation in the two cell types would deserve at least one full paper.Nevertheless, we already showed in the previous version of the paper how our data can be used to construct morphologically detailed models and to test some of the consequences of the experimentally measured synaptic distributions.We have now added several new modeling results that demonstrate additional uses of our experimental data.Most importantly, we now present fully active models of OLM and HS cells, and characterize their spike output in response to current injections and random as well as spatially organized and temporally patterned synaptic inputs (see Results sections: Modeling spiking responses in OLM and HS neurons: model construction; Modeling the responses of active neurons to synaptic inputs; Supplementary Figures 8, 9, 10, and 11).We have also examined one of the functional consequences of the location-dependence of the E/I balance in our passive models in the simulated high-conductance state, showing that the local equilibrium potential in the dendrites has a distinct dependence on the distance from the soma in the two cell types (Supplementary Figure 7).

3.
As stated by the authors in the discussion, more information on the impact of the synaptic input distribution on the recruitment of HL / OLM cells would be perfect.May be the authors see a possibility to make some assumptions on the distribution of voltage gated conductances similar to published data on OLM cells; although the reviewer fully understands the difficulty on making such assumptions and the difficulty to do the related experiments.
Following the suggestion of the Reviewer, we have now constructed and simulated active models of both cell types (Results section: Modeling spiking responses in OLM and HS neurons: model construction).More specifically, we added a variety of voltage-gated conductances to our models based on earlier modeling studies by the group of Frances Skinner, and tuned their maximal conductance values by fitting the responses to a series of depolarizing and hyperpolarizing current steps using features extracted from our physiological recordings.Next, we validated these active models using additional features (Supplementary Figures 8 and  9).Then we examined the action potential output of the models in response to synaptic inputs, varying the number and location of the inputs, and also looked at how the spike output is modulated by time-varying input rates (Results Section: Modeling the responses of active neurons to synaptic inputs; Supplementary Figures 10 and 11).

Reviewer: 2
The paper by Takács et al., titled "Synaptic and dendritic architecture of two types of hippocampal somatostatin interneurons", describes extensive profiling of two cell types present among the somatostatin (SST) positive inhibitory neurons in the hippocampus, termed OLM and HS interneurons.The authors characterize the abundance of these two neuron types among the SST interneurons.They quantify the morphologies of the two cell types, including dendritic and axonal features and synaptic targeting, for both the incoming and outgoing synapses.Further, intrinsic electrophysiology of these cells is assessed in slice recordings.Finally, the authors fit morphologically detailed models (but, as they describe them, biophysically simplified, since active conductances are excluded) for these two cell types and investigate responses of the cell models to synaptic inputs.
Overall, the paper has many appealing characteristics.It summarizes a large body of carefully obtained data.The data are multimodal, i.e., come from a variety of experimental techniques to shed light onto multiple properties of the cells.And, modeling is used to integrate the data and obtain further insights through simulations.The paper is well written and clear.

We are grateful to Reviewer 2 for their constructive comments and hope that the additional computational work included in this revision further demonstrates the potential of this large dataset.
There are, however, some cons as well.To me, the major issue is that modeling did not use active conductances, therefore resulting in models that cannot generate action potentials and cannot be used for any studies beyond those using weak subthreshold stimuli.For a relatively high-profile publication, this is not sufficient, especially because the paper did not build models for that many cells (only 4 each for OLM and HS cells), and neither did it profile electrophysiology in a particularly large number of cells.I would therefore suggest that, to make their paper sufficiently strong, the authors should add models with active conductances in the soma (and, perhaps, the axon initial segments).I understand that adding active conductances in dendrites would be way too computationally expensive, so I am not suggesting that.But, developing models with active conductances in the soma would be a necessary addition, which would improve the paper drastically.
Following the suggestion of the Reviewer, we have now constructed and simulated active models of both cell types (Results section: Modeling spiking responses in OLM and HS neurons: model construction).We decided to include voltage-gated conductances in all compartments of the model neurons.More specifically, we added a variety of voltage-gated conductances to our models based on earlier modeling studies of OLM cells by the group of Frances Skinner, and (in the absence of specific data) we assumed that the same set of conductances was present in HS cells as well.We then tuned the maximal conductance values of the voltagegated channels by fitting the responses of each model neuron to a series of depolarizing and hyperpolarizing current steps using features extracted from our physiological recordings of the two cell types.
With such an addition, the authors can simulate spiking responses to current injections into the soma as well as to synaptic inputs, for example, at theta frequencies.These simulation results will show whether OLM and HS cells exhibit the same or distinct spiking responses under such circumstances, which will be an important contribution of the modeling.As it stands now, the differences observed with modeling subthreshold phenomena are all rather minor and are likely not highly relevant for physiological situations where cells produce many action potentials.Obviously, we don't know whether the results will be different between the OLM and HS, but it is important to explore this.That will elevate the results of the paper from less relevant to in vivo physiology to highly relevant.
We first examined the responses of the active models to somatic current injections and compared them with the experimental data using a large number of electrophysiological features, including many that had not been used during the parameter optimization process (Supplementary Figures 8 and 9).Next, we investigated the action potential output of the models in response to synaptic input, varying the number and location of the inputs (Results section: Modeling the responses of active neurons to synaptic inputs; Supplementary Figure 10).Finally, we also looked at how the spike output of the models is modulated by scattered synaptic input whose rate varies at the theta frequency (Results section: Modeling the responses of active neurons to synaptic inputs; Supplementary Figure 11).Interestingly, our simulations showed substantial variability in spiking responses within each cell type, but no major difference was found between the two cell types.

Other comments.
Page 5, bottom, where synapses onto the OLM and HS cells are characterized: It would be interesting to know how many (what fraction) of the excitatory input synapses are targeting spines vs. shaft and soma, and the same for inhibitory input synapses.
We are grateful for this suggestion, as it has resulted in an interesting set of additional data, which is now presented in the Results section under the title: "Distribution of glutamatergic and GABAegic synapses on dendritic shafts and spines".In addition, it has also generated further calculations in the Results section under the title: "Modeling synaptic distributions in OLM and HS neurons".
Page 9, Proportions of OLM and HS cells within the somatostatin-positive interneurons of CA1 str.oriens: It would be good to provide numbers for each mouse, since it's only 3 in each case.One wonders whether the proportions are well conserved across animals.Now we provided the proportions for each mouse in the text under the subheading "Proportions of OLM and HS cells within the somatostatin-positive interneurons of CA1 str.oriens" and we also added these data to the figure legends of Supplementary Fig.The reason why we do not normally show and compare individual voltage traces from experiments and models is that, due to the way we fit our models, such comparisons are not completely straightforward and even potentially misleading.First, we actually use extracted features rather than raw voltage traces to compare model outputs to experimental data during the optimization, so we think that showing this comparison in the figures is a better representation of the quality of the results than plotting the raw traces.Second, the target of the optimization is not a single measurement from one cell, but multiple repeated measurements from a group of cells, and we use the statistics of the extracted features (their means and standard deviations in particular) to define the error function for the optimization.Third, we perform multiple runs of the optimization with different random seeds to obtain a population of model neurons in each case.We found it more informative to represent this variability in the experimental data and the behavior of the models using the feature statistics.Nevertheless, we agree with the Reviewer that a visual representation of the traces also carries useful information, so we have added examples of simulated traces (from our new active models) to Figure 7.
Page 13, last paragraph: The part about mitochondria is interesting, but more information should be shown if the authors want to include it in the paper.There should be a reference to the data on the volume occupied by mitochondria in the Tables (which are mentioned earlier, but readers would forget by now).Now we added references in the text to the data tables: proportion of mitochondria in Table 1 and  Supplementary Table 2, 3. And, where the authors say that they did not see a change in the behavior of model neurons, they should show these results in a figure or table .Plotting the results of simulations that included mitochondrial volume produced figures that looked virtually identical on paper to the results of the original simulations, so we decided not to include them in the paper.
Page 19, lines 12-25: A good starting point can be to use the same ionic conductances for the OLM cells and HS cells.As the recordings presented here show, the electrophysiological responses of these two cell types are not drastically different, suggesting that same conductances might work fine.
We thank the Reviewer for this suggestion, which we followed in the modeling work.We have corrected it.Table 4. "Physiological and model parameters and statistics": These do not seem to be model parameters, or maybe I don't understand something.
2. Page 11, modeling subthreshold responses: The authors only show summary statistics for modeling subthreshold responses to negative current injections in these cells.It would be useful to show in the main figures examples of the actual simulated traces, overlapped with the experimental traces.

Fig
Fig.1L: "Complete Neurolucida reconstruction of OLM and HS dendritic trees and (Fig.5, 6)."It looks like some text is missing after the "and".