How the layer-dependent ratio of excitatory to inhibitory cells shapes cortical coding in balanced networks

Reviewer #1 (Public review):

Summary:

The authors seek to understand the role of different ratios of excitatory to inhibitory (EI) neurons, which in experimental studies of the cerebral cortex have been shown to range from 4 to 9. They do this through a simulation study of sparsely connected networks of excitatory and inhibitory neurons.

Their main finding is that the participation ratio and decoding accuracy increase as the E/I ratio decreases. This suggests higher computational complexity.

This is the start of an interesting computational study. However, there is no analysis to explain the numerical results, although there is a long literature of reduced models for randomly connected neural networks which could potentially be applied here. (For example, it seems that the authors could derive a mean field expression for the expected firing rate and variance - hence CV - which could be used to target points in parameter space (vs. repeated simulation in Figures 1,2).) The paper would be stronger and more impactful if this was attempted.

Strengths:

Some issues I appreciated are:

(1) The use of a publicly available simulator (Brian), which helps reproducibility. I would also request that the authors supply submission or configuration scripts (if applicable, I don't know Brian).

(2) A thorough exploration of the parameter space of interest (shown in Figure 2).

(3) A good motivation for the underlying question: other things being equal, how does the E/I ratio impact computational capacity?

Weaknesses:

(1) Lack of mathematical analysis of the network model

Major issues I recommend that the authors address (not sure whether these are "weaknesses"):

(1) In "Coding capacity in different layers of visual cortex" the authors measure PR values from layers 2/3 and 4 in VISp and find that layer 2/3 has a higher PR than layer 4.

But in Dahmen et al. 2020 (https://doi.org/10.1101/2020.11.02.365072 ), the opposite was found (see Figure 2d of Dahmen et al.): layer 2 had a lower PR than layer 4. Can the authors explain how that difference might arise? i.e. were they analyzing the same data sets? If so why the different results? Could it have to do with the way the authors subsample for the E/I ratio?

From the Methods of that paper: "Visual stimuli were generated using scripts based on PsychoPy and followed one of two stimulus sequences ("brain observatory 1.1" and
"functional connectivity"). We focused on spontaneous neural activity registered while the animal was not performing any task. In each session, the spontaneous activity condition lasted 30 minutes while the animal was in front of a screen of mean grey luminance. We, therefore, analyzed 26 of the original 58 sessions corresponding to the "functional connectivity" subdataset as they included such a period of spontaneous activity. " This suggests to me they may have analyzed recordings with the other stimulus sequence; however, the hypothesis that E/I ratio should modulate dimensionality would not seem to "care" about which stimulus sequence was used.

(2) In Discussion (pg. 20, line 383): "They showed that brain regions closer to sensory input, like the thalamus, have higher dimensionality than those further away, such as
the visual cortex. " How is this consistent with the hypothesis that "higher dimensionality might be linked to more complex cognitive functions"?

(3) What is the probability of connection between different populations? e.g. the probability of there being a synaptic connection between any two E cells? I could not find a statement about this. It should be included in the Methods.

(4) pg. 27, line 540: "Synchronicity within the network" For each cell pair, the authors use the maximum cross-correlation over time lag. I don't think I have seen this before. Can the authors explain why they use this measurement, vs (a) integrated cross-correlation or (b) cross-correlation at some time scale? Also, it seems like this fails to account for neuron pairs for which there is a strong inhibitory correlation.

(5) "When stimulated, a time-varying input, μext(t), is applied to 2,000 randomly selected excitatory neurons. " I would guess that computing PR would depend on the overlap of the 500 neurons analyzed and this population. Do the authors check or control for that?

5b) Related: to clarify, are the 500 neurons chosen from the analysis equally likely to be E or I neurons?

Reviewer #2 (Public review):

Summary:

Alizadeh et al. investigate how varying cellular E/I (excitatory/inhibitory) composition impacts coding across cortical layers. They build on findings from a recent study (Huang et al., 2022) that demonstrated a decrease in the fraction of inhibitory neurons from L2/3 to L4. Using a network of excitatory and inhibitory leaky integrate-and-fire neurons, they systematically assess how these anatomical features influence the dimensionality of network activity and coding capacity. Their key finding is that increasing the proportion of inhibitory neurons enhances the dimensionality of activity and improves the encoding of time-varying stimuli.

Strengths:

The authors use a clear methodology and well-established model of network activity that allows them to relate network parameters to the coding properties. They systematically evaluate the impact of the key features of the inhibitory population. Thus, in addition to changing the fraction of inhibitory cells, they control for the inhibitory firing threshold of inhibitory neurons and connection strength between inhibitory and excitatory cells. Furthermore, they show their modeling results are aligned with the analysis of the spiking activity in L2/3 vs. L4 from the Allen Institute data.

Weaknesses:

One general shortcoming of this approach is that it focuses on a small preselected number of network features. For example, it is unclear to what extent the results would be affected by other aspects of the organization of cortical columns, such as subclasses of inhibitory cells (SOM, VIP, PV), specific differences in synapses, realistic population sizes, or even connectivity between layers. Similarly, the models of L2/3 and L4 are constrained based on a limited set of observations, and it has not been demonstrated whether the same findings hold true for V1 recordings analyzed by the authors.

The modeling relies on anatomical data from the barrel cortex, but the decoding comparison is based on V1 data. This raises questions about how anatomical differences between regions may influence the conclusions.

The coding capacity appears inversely correlated with the firing rate, which in this study is largely influenced by the properties of the inhibitory population. It would be important to confirm that the observed changes in coding capacity and participation ratio are not solely driven by firing rate changes.