Modeling and Simulation of Neocortical Micro- and Mesocircuitry. Part I: Anatomy

Reviewer #1 (Public Review):

Summary:

In this study, the authors describe the construction of an extremely large-scale anatomical model of juvenile rat somatosensory cortex (excluding the barrel region), which extends earlier iterations of these models by expanding across multiple interconnected cortical areas. The models are constructed in such a way as to maintain biological detail from a granular scale - for example, individual cell morphologies are maintained, and synaptic connectivity is founded on anatomical contacts. The authors use this model to investigate a variety of properties, from cell-type specific targeting (where the model results are compared to findings from recent large-scale electron microscopy studies) to network metrics. The model is also intended to serve as a platform and resource for the community by being a foundation for simulations of neuronal circuit activity and for additional anatomical studies that rely on the detailed knowledge of cellular identity and connectivity.

Strengths:

As the authors point out, the combination of scale and granularity of their model is what makes this study valuable and unique. The comparisons with recent electron microscopy findings are some of the most compelling results presented in the study, showing that certain connectivity patterns can arise directly from the anatomical configuration, while other discrepancies highlight where more selective targeting rules (perhaps based on molecular cues) are likely employed. They also describe intriguing effects of cortical thickness and curvature on circuit connectivity and characterize the magnitude of those effects on different cortical layers.

The detailed construction of the model is drawn on a wide range of data sources (cellular and synaptic density measures, neuronal morphologies, cellular composition measures, brain geometry, etc.) that are integrated together; other data sources are used for comparison and validation. This consolidation and comparison also represent a valuable contribution to the overall understanding of the modeled system.

Weaknesses:

The scale of the model, which is a primary strength, also can carry some drawbacks. In order to integrate all the diverse data sources together, many specific decisions must be made about, for example, translating findings from different species or regions to the modeled system, or deciding which aspects of the system can be assumed to be the same and which should vary. All these decisions will have effects on the predicted results from the model, which could limit the types of conclusions that can be made (both by the others and by others in the community who may wish to use the model for their own work).

As an example, while it is interesting that broad brain geometry has effects on network structure (Figure 7), it is not clear how those effects are actually manifested. I am not sure if some of the effects could be due to the way the model is constructed - perhaps there may be limited sets of morphologies that fit into columns of particular thicknesses, and those morphologies may have certain idiosyncrasies that could produce different statistics of connectivities where they are heavily used. That may be true to biology, but it may also be somewhat artifactual if, for example, the only neurons in the library that fit into that particular part of the cortex differ from the typical neurons that are actually found in that region (but may not have been part of the morphological sampling). I also wonder how much the assumption that the layers have the same relative thicknesses everywhere in the cortex affects these findings, since layer thicknesses do in fact vary across the cortex.

In addition, the complexity of the model means that some complicated analyses and decisions are only presented in this manuscript with perhaps a single panel and not much textual explanation. I find, for example, that the panels of Figure S2 seem to abstract or simplify many details to the point where I am not clear about what they are actually illustrating - how does Figure S2D represent the results of "the process illustrated in B"? Why are there abrupt changes in connectivity at region borders (shown as discontinuous colors), when dendrites and axons span those borders and so would imply interconnectivity across the borders? What do the histograms in E1 and E2 portray, and how are they related to each other?

Overall, the model presented in this study represents an enormous amount of work and stands as a unique resource for the community, but also is made somewhat unwieldy for the community to employ due to the weight of its manifold specific construction decisions, size, and complexity.

Reviewer #2 (Public Review):

Summary:

The authors build a colossal anatomical model of juvenile rat non-barrel primary somatosensory cortex, including inputs from the thalamus. This enhances past models by incorporating information on the shape of the cortex and estimated densities of various types of excitatory and inhibitory neurons across layers. This is intended to enable an analysis of the micro- and mesoscopic organisation of cortical connectivity and to be a base anatomical model for large-scale simulations of physiology.

Strengths:

• The authors incorporate many diverse data sources on morphology and connectivity.

• This paper takes on the challenging task of linking micro- and mesoscale connectivity.

• By building in the shape of the cortex, the authors were able to link cortical geometry to connectivity. In particular, they make an unexpected prediction that cortical conicality affects the modularity of local connectivity, which should be testable.

• The author's analysis of the model led to the interesting prediction that layer 5 neurons connect local modules, which may be testable in the future, and provide a basis to link from detailed anatomy to functional computations.

• The visualisation of the anatomy in various forms is excellent.

• A subnetwork of the model is openly shared (but see question below).

Weaknesses:

• Why was non-barrel S1 of the juvenile rat cortex selected as the target for this huge modelling effort? This is not explained.

• There is no effort to determine how specific or generalisable the findings here are to other parts of the cortex.

• Although there is a link to physiological modelling in another paper, there is no clear pathway to go from this type of model to understand how the specific function of the modelled areas may emerge here (and not in other cortical areas).

• In a few places the manuscript could be improved by being more specific in the language, for example:
- "our anatomy-based approach has been shown to be powerful", I would prefer instead to read about specific contributions of past papers to the field, and how this builds on them.
- similarly: "ensuring that the total number of synapses in a region-to-region pathway matches biology." Biology here is a loose term and implies too much confidence in the matching to some ground truth. Please instead describe the source of the data, including the type of experiment.

• Some of the decisions seem a little ad-hoc, and the means to assess those decisions are not always available to the reader e.g.
- pg. 10. "Based on these results, we decided that the local connectome sufficed to model connectivity within a region.". What is the basis for this decision? Can it be formalised?
- "In the remaining layers the results of the objective classification were used to validate the class assignments of individual pyramidal cells. We found the objective classification to match the expert classification closely (i.e., for 80-90% of the morphologies). Consequently, we considered the expert classification to be sufficiently accurate to build the model." The description of the validation is a little informal. How many experts were there? What are their initials? Was inter-rater or intra-rater reliability assessed? What are these numbers? The match with Kanari's classification accuracy should be reported exactly. There are clearly experts among the author list, but we are all fallible without good controls in place, and they should be more explicit about those controls here, in my opinion.
- "Morphology selection was then performed as previously (Markram et al., 2015), that is, a morphology was selected randomly from the top 10% scorers for a given position." A lot of the decisions seem a little ad-hoc, without justification other than this group had previously done the same thing. For example, why 10% here? Shouldn't this be based on selecting from all of the reasonable morphologies?

• I would like to know if one of the key results relating to modularity and cortical geometry can be further explored. In particular, there seem to be sharp changes in the data at the end of the modelled cortical regions, which need to be explored or explained further.

• The shape of the juvenile cortex - a key novelty of this work - was based on merely a scalar reduction of the adult cortex. This is very surprising, and surely an oversimplification. Huge efforts have gone into modelling the complex nonlinear development of the cortex, by teams including the developing Human Connectome Project. For such a fundamental aspect of this work, why isn't it possible to reconstruct the shape of this relatively small part of the juvenile rat cortex?

• The same relative laminar depths are used for all subregions. This will have a large impact on the model. However, relative laminar depths can change drastically across the cortex (see e.g. many papers by Palomero-Gallagher, Zilles, and colleagues). The authors should incorporate the real laminar depths, or, failing that, show evidence to show that the laminar depth differences across the subregions included in the model are negligible.

• The authors perform an affine mapping between mouse and rat cortex. This is again surprising. In human imaging, affine mappings are insufficient to map between two individual brains of the same species and nonlinear transformations are instead used. That an affine transformation should be considered sufficient to map between two different species is then very surprising. For some models, this may be fine, but there is a supposed emphasis here on biological precision in terms of anatomical location.

• One of the most interesting conclusions, that the connectivity pattern observed is in part due to cooperative synapse formation, is based on analyses that are unfortunately not shown.

• Open code:
- Why is only a subvolume available to the community?
- Live nature of the model. This is such a colossal model, and effort, that I worry that it may be quite difficult to update in light of new data. For example, how much person and computer time would it take to update the model to account for different layer sizes across subregions? Or to more precisely account for the shape of the juvenile rat cortex?

Reviewer #3 (Public Review):

This manuscript reports a detailed model of the rat non-barrel somatosensory cortex, consisting of 4.2 million morphologically and biophysically detailed neuron models, arranged in space and connected according to highly sophisticated rules informed by diverse experimental data. Due to its breadth and sophistication, the model will undoubtedly be of interest to the community, and the reporting of anatomical details of modeling in this paper is important for understanding all the assumptions and procedures involved in constructing the model. While a useful contribution to this field, the model and the manuscript could be improved by employing data more directly and comparing simple features of the model's connectivity - in particular, connection probabilities - with relevant experimental data.

The manuscript is well-written overall but contains a substantial number of confusing or unclear statements, and some important information is not provided.

Below, major concerns are listed, followed by more specific but still important issues.

MAJOR ISSUES

(1) Cortical connectivity.

Section 2.3, "Local, mid-range and extrinsic connectivity modeled separately", and Figure 4: I am confused about what is done here and why. The authors have target data for connectivity (Figure 4B1). But then they use an apposition-based algorithm that results in connectivity that is quite different from the data (Figure 4B2, C). They then use a correction based on the data (Figure 4E) to arrive at a more realistic connectivity. Why not set the connectivity based on the data right away then? That would seem like a more straightforward approach.

The same comment applies to Section 2.4., "Specificity of axonal targeting": the distributions of synapses on different types of target cell compartments were not well captured by the original model based on axon-dendrite overlap and pruning, so the authors introduced further pruning to match data specificity. While details of this process and what worked and what didn't may be interesting to some, overall it is not surprising, as it has been well known that cell types exhibit connectivity that is much more specific than "Peters rule" or its simple variations. The question is, since one has the data, why not use the data in the first place to set up the connectivity, instead of using the convoluted process of employing axon-dendrite overlap followed by multiple corrections?

Most importantly, what is missing from the whole paper is the characterization of connection probabilities, at least for the local circuit within one area. Such connection probabilities can be obtained from the data that the authors already use here, such as the MICRONS dataset. Another good source of such data is Campagnola et al., Science, 2022. Both datasets are for mouse V1, but they provide a comprehensive characterization across all cortical layers, thus offering a good benchmark for comparison of the model with the data. It would be important for the authors to show how connection probabilities realized in their model for different cell types compared to these data.

(2) Section 2.5, "Structure of thalamic inputs" and Figure 6.

The text in section 2.5 should provide more details on what was done - namely, that the thalamic axons were generated based on the axon density profiles and then synapses were established based on their overall with cortical dendrites. Figure S10 where the target axon densities from data and the model axon densities are compared is not even mentioned here. Now, Figure S10 only shows that the axon densities were generated in a way that matches the data reasonably well. However, how can we know that it results in connectivity that agrees with data? Are there data sources that can be used for that purpose? For example, the authors show that in their model "the peaks of the mean number of thalamic inputs per neuron occur at lower depths than the peaks of the synaptic density". Is this prediction of the model consistent with any available data?

Most importantly, the authors should show how the different cell types in their model are targeted by the thalamic inputs in each layer. Experimental studies have been done suggesting specificity in targeting of interneuron types by thalamic axons, such as PV cells being targeted strongly whereas SST and VIP cells being targeted less.

(3) "We have therefore made not only the model but also most of our tool chain openly available to the public (Figure 1; step 7)."
In fact it is not the whole model that is made publicly available, but only about 5% of it (211,000 out of 4,200,000 neurons). Also, why is "most" of the tool chain made openly available, and not the whole tool chain?

OTHER ISSUES

"At each soma location, a reconstruction of the corresponding m-type was chosen based on the size and shape of its dendritic and axonal trees (Figure S6). Additionally, it was rotated to according to the orientation towards the cortical surface at that point."

After this procedure, were cells additionally rotated around the white matter-pia axis? If yes, then how much and randomly or not? If not, then why not? Such rotations would seem important because otherwise additional order potentially not present in the real cortex is introduced in the model affecting connectivity and possibly also in vivo physiology (such as the dynamics of the extracellular electric field).

The term "new in vivo reconstructions" for the 58 neurons used in this paper in addition to "in vitro reconstructions" is a misnomer. It is not straightforward to see where the procedure is described, but then one finds that the part of Methods that describes experimental manipulations is mostly about that (so, a clearer pointer to that part of Methods could be useful). However, the description in Methods makes it clear that it is only labeling that is done in vivo; the microscopy and reconstruction are done subsequently in vitro. I would recommend changing the terminology here, as it is confusing. Also, can the authors show reconstructions of these neurons in the supplementary figures? Is the reconstruction shown in Figure 4A representative?

In the Discussion, "This was taken into account during the modeling of the anatomical composition, e.g. by using three-dimensional, layer-specific neuron density profiles that match biological measurements, and by ensuring the biologically correct orientation of model neurons with respect to the orientation towards the cortical surface. As local connectivity was derived from axo-dendritic appositions in the anatomical model, it was strongly affected by these aspects.
However, this approach alone was insufficient at the large spatial scale of the model, as it was limited to connections at distances below 1000μm."

As mentioned above, it is not clear that this approach was sufficient for local connectivity either. It would be great if the authors showed a systematic comparison of local connection probabilities between different cell types in their model with experimental data and commented here in the Discussion about how well the model agrees with the data.

In the Discussion: "The combined connectome therefore captures important correlations at that level, such as slender-tufted layer 5 PCs sending strong non-local cortico-cortical connections, but thick-tufted layer 5 PCs not." (Also the corresponding findings in Results.)

If I understand this statement correctly, it may not agree with biological data. See analysis from MICRONS dataset in Bodor et al., https://www.biorxiv.org/content/10.1101/2023.10.18.562531v1.

Table 2 is confusing. What do pluses and minuses mean? What does it mean that some entries have two pluses? This table is not mentioned anywhere else in the text. If pluses mean some meaningful predictions of the model, then their distribution in the table seems quite liberal and arbitrary. It is not clear to me that the model makes that many predictions, especially for type-specificity and plasticity. Also, why is the hippocampus mentioned in this table? I don't see anything about the hippocampus anywhere else in the paper.

In the Discussion, "Thus, we made the tools to improve our model also openly available (see Data and Code availability section)."
As mentioned before, the authors themselves write that they made "most of our tool chain openly available to the public", but not all of it.

Table S2 has multiple question marks. It is not clear whether the "predictions" listed in that table are truly well-thought-out and/or whether experimental confirmations are real.

Introduction: It would be quite appropriate to cite here Einevoll et al., Neuron, 2019 ("The Scientific Case for Brain Simulations").

Author response:

We thank the reviewers for their insightful comments on our model and manuscript. In this provisional response, we would like to comment on some of the issues raised and how we plan to address them.

First, the reviewers correctly pointed out that only a small part of the full model was openly available. We have now rectified this and the full model is available at: https://dataverse.harvard.edu/dataverse/sscx.

Next, we would like to comment on the perceived lack of clarity of certain descriptions in the manuscript. We note that individual techniques and parts of the model have been developed, justified, and validated in previous publications. This left us with the question of how much of the contents of those papers we should re-describe. Too much, and the manuscript becomes overly long; too little, and the reader cannot gain a sufficient understanding of the model building process. The reviewers' comments made it clear that some aspects of the model should be described in more detail and we plan to address this in a revision. Crucially, one missing item raised by all reviewers was a comparison of local connection probabilities to the literature. This will be provided in the revision. Additionally, the reviewers questioned our decision to use a connectivity algorithm that is not based on direct parameterization of target connection probabilities. While this is a limitation of the algorithm we employed, it also has unique strengths, providing non-random aspects of connectivity that have been proven to be impossible to model with algorithms that enforce given connection probabilities or degree distributions. We plan to explain this better in a revision.

We will also comment on the challenges associated with the interpretation of experimentally measured connection probabilities and employing them for the parameterization of a biophysically detailed model spanning millimeters.

The reviewers also suggested several aspects of the model that could be improved. Whilst we see merit with all of them, we would like to briefly comment on model completeness in general. First, this model - and any model - can probably never be considered complete. Instead, the model has to be continuously refined, which one reviewer phrased as the "live nature" of the model. However, to demonstrate the model's utility and justify the expense of modeling, we also have to use the model in projects that explore specific scientific questions. To undertake and complete such a project, one must select and "freeze" a given version of the model-- otherwise the project will never conclude. Further, we believe that it is advantageous if several projects use the same version of the model. In that case, a reader who is already familiar with the model from one paper may find it easier to understand other papers using the same model. The goal of this manuscript is to describe the version of the model that we used in several ongoing and concluded follow-up projects, including its limitations and opportunities for refinement. As such, we do not plan to add further improvements to the model for this reviewed pre-print. We will, however, continue to refine the model outside of the scope of this publication. Since we believe the development and bottom up models are best done in a community driven manner, we encourage interested parties to participate.

We invite anyone with ideas of how the model could be refined to contact us to discuss how we could integrate these changes into the model together using our tools.