Assemblies, synapse clustering and network topology interact with plasticity to explain structure-function relationships of the cortical connectome

Reviewer #1 (Public review):

This paper investigates the dynamics of excitatory synaptic weights under a calcium-based plasticity rule, in long (up to 10 minutes) simulations of a 211,000-neuron biophysically detailed model of a rat cortical network.

Strengths

(1) A very detailed network model, with a large number of neurons, connections, synapses, etc., and with a huge number of biological considerations implemented in the model.

(2) A carefully developed calcium-based plasticity rule, which operates with biologically relevant variables like calcium concentration and NMDA conductances.

(3) The study itself is detailed and thorough, covering many aspects of the cellular and network anatomy and properties and investigating their relationships to plasticity.

(4) The model remains stable over long periods of simulations, with the plasticity rule maintaining reasonable synaptic weights and not pushing the network to extremes.

(5) The variety of insights the authors derive in terms of relationships between the cellular and network properties and dynamics of the synaptic weights are potentially interesting for the field.

(6) Sharing the model and the associated methods and tools is a big plus.

Weaknesses

(1) Conceptually, there seems to be a missed opportunity here in that it is not clear what the network learns to do. The authors present 10 different input patterns, the network does some plasticity, which is then analyzed, but we do not know whether the learning resulted in anything functionally significant. Did the network learn to discriminate the patterns much better than at the beginning, to capture or anticipate the timing of pattern presentation, detect similarities between patterns, etc.? This is important to understand if one wants to assess the significance of synaptic changes due to plasticity. For example, if the network did not learn much new functionally, relative to its initial state, then the observed plasticity could be considered minor and possibly insufficient. In that case, were the network to learn something substantial, one would potentially observe much more extensive plasticity, and the results of the whole study could change, possibly including the stability of the network. While this could be a whole separate study, this issue is of central importance, and it is hard to judge the value of the results when we do not know what the network learned to do, if anything.

(2) In this study, plasticity occurs only at E-to-E connections but not at others. However, it is well known that inhibitory connections in the cortex exhibit at the very least a substantial short-term plasticity. One would expect that not including these phenomena would have substantial consequences on the results.

(3) Lines 134-135: "We calibrated layer-wise spontaneous firing rates and evoked activity to brief VPM inputs matching in vivo data from Reyes-Puerta et al. (2015)."

(4) Can the authors show these results? It is an important comparison, and so it would be great to see firing rates (ideally, their distributions) for all the cell types and layers vs. experimental data, for the evoked and spontaneous conditions.

(5) That being said, the Reyes-Puerta et al. paper reports firing rates for the barrel cortex, doesn't it? Whereas here, the authors are simulating a non-barrel cortex. Is such a comparison appropriate?

(6) Comparison with STDP on pages 5-7 and Figure 2: if I got this right, the authors applied STDP to already generated spikes, that is, did not run a simulation with STDP. That seems strange. The spikes they use here were generated by the system utilizing their calcium-based plasticity rule. Obviously, the spikes would be different if STDP was utilized instead. The traces of synaptic weights would then also be different. The comparison therefore is not quite appropriate, is it?

(7) Section 2.3 and Figure 5: I am not sure this analysis adds much. The main finding is that plasticity occurs more among cells in assemblies than among all cells. But isn't that expected given what was shown in the previous figures? Specifically, the authors showed that for cells that fire more, plasticity is more prominent. Obviously, cells that fire little or not at all won't belong to any assemblies. Therefore, we expect more plasticity in assemblies.

(8) Section 2.4 and Figure 6: It is not clear that the results truly support the formulation of the section's title ("Synapse clustering contributes to the emergence of cell assemblies, and facilitates plasticity across them") and some of the text in the section. What I can see is that the effect on rho is strong for non-clustered synapses (Figure 6C and Figure S8A). In some cases, it is substantially higher than what is seen for clustered synapses. Furthermore, the wording "synapse clustering contributes to the emergence of cell assemblies" suggests some kind of causal role of clustered synapses in determining which neurons form specific cell assemblies. I do not see how the data presented supports that. Overall, it appears that the story about clustered synapses is quite complicated, with both clustered and non-clustered synapses driving changes in rho across the board.

(9) Section 2.5 and Figure 7: Can we be certain that it is the edge participation that is a particularly good predictor of synaptic changes and/or strength, as opposed to something simpler? For example, could it be the overall number of synapses, excitatory synapses, or something along these lines, that the source and/or target neurons receive, that determine the rho dynamics? And then, I do not understand the claim that edge participation allows one to "delineate potentiation from depression". The only related data I can find is in Figure 7A3, about which the authors write "this effect was stronger for potentiation than depression". But I don't see what they mean. For both depression and facilitation, the changes observed are in the range of ~12% of probability values. And even if the effect is stronger, does it mean one can "delineate" potentiation from depression better? What does it mean, to "delineate"? If it is some kind of decoding based on the edge participation, then the authors did not show that.

(10) "test novel predictions in the MICrONS (2021) dataset, which while pushing the boundaries of big data neuroscience, was so far only analyzed with single cells in focus instead of the network as a whole (Ding et al., 2023; Wang et al., 2023)." That is incorrect. For example, the whole work of Ding et al. analyzes connectivity and its relation to the neuron's functional properties at the network level.

Reviewer #2 (Public review):

Summary:

This paper aims to understand the effects of plasticity in shaping the dynamics and structure of cortical circuits, as well as how that depends on aspects such as network structure and dendritic processing.

Strengths:

The level of biological detail included is impressive, and the numerical simulations appear to be well executed. Additionally, they have done a commendable job in open-sourcing the model.

Weaknesses:

The main result of this work is that activity in their network model remains stable without the need for a homeostatic mechanism. However, as the authors acknowledge, this has been demonstrated in previous studies (e.g., Higgins et al. 2014). In those studies, stability was attributed to calcium-based rules combined with calcium concentrations at in vivo levels and background neuronal activity. Since the authors use the same calcium-based rule, it is unclear what new result, if any, is being presented. If the authors are suggesting that the mechanism in their simulations differs, that should be stated clearly, and evidence supporting that claim should be provided.

The other findings discussed in the paper are related to a characterization of the dependency of plastic changes on network structure. While this analysis is potentially interesting, it has the following limitations.

First, I believe the authors should include an analysis of the generality and specificity of their results. All the findings seem to be derived from a single run of the simulation. How do the results vary with different network initializations, simulation times, or parameter choices?

Second, the presentation of the results is difficult to follow. The characterization comes across as a long list of experiments, making it hard to identify a central message or distinguish key findings from minor details. The authors provide little intuition about why certain outcomes arise, and the complexity of the simulation makes it challenging - if not impossible - to determine which model elements are essential for specific results and which mechanisms drive emergent properties. Additionally, the text often lacks crucial details. For instance, the description of k-edge participation should be expanded, and an explanation of what this method quantifies should be included. Overall, I believe the authors should focus on a smaller set of significant results and provide a more in-depth discussion.

The comparison of the model with the MICrONS dataset could be improved. In Figure 7B, the authors should show how the same quantification looks in a network model without plasticity. In Figure 8B, the data aligns with the model before plasticity, so it's unclear how this serves as a verification of the theoretical predictions.

Reviewer #3 (Public review):

Summary:

Ecker et al. utilized a biologically realistic, large-scale cortical model of the rat's non-barrel somatosensory cortex, incorporating a calcium-dependent plasticity rule to examine how various factors influence synaptic plasticity under in vivo-like conditions. Their analysis characterized the resulting plastic changes and revealed that key factors, including the co-firing of stimulus-evoked neuronal ensembles, the spatial organization of synaptic clusters, and the overall network topology, play an important role in affecting the extent of synaptic plasticity.

Strengths:

The detailed, large-scale model employed in this study enables the evaluation of diverse factors across various levels that influence the extent of plastic changes. Specifically, it facilitates the assessment of synaptic organization at the subcellular level, network topology at the macroscopic level, and the co-activation of neuronal ensembles at the activity level. Moreover, modeling plasticity under in vivo-like conditions enhances the model's relevance to experiments.

Weaknesses:

(1) The authors claimed that, under in vivo-like conditions and in the presence of plasticity, firing rates and weight distributions remain stable without additional homeostatic mechanisms during a 10-minute stimulation period. However, the weights do not reach the steady state immediately after the 10-minute stimulation. Therefore, extended simulations are necessary to substantiate the claim.

(2) Another major limitation of the paper lies in its lack of mechanistic insights into the observed phenomena (particularly on aspects that are typically impossible to assess in traditional simplified models, like layer-specific and layer-to-layer pathways-specific plasticity changes), as well as the absence of discussions on the potential computational implications of the corresponding observed plastic changes.