Modeling and Simulation of Neocortical Micro- and Mesocircuitry. Part II: Physiology and Experimentation

Reviewer #1 (Public review):

This paper presents a model of the whole somatosensory non-barrel cortex of the rat, with 4.2 million morphologically and electrically detailed neurons, with many aspects of the model constrained by a variety of data. The paper focuses on simulation experiments, testing a range of observations. These experiments are aimed at understanding how the multiscale organization of the cortical network shapes neural activity.

Strengths:

(1) The model is very large and detailed. With 4.2 million neurons and 13.2 billion synapses, as well as the level of biophysical realism employed, it is a highly comprehensive computational representation of the cortical network.

(2) Large scope of work - the authors cover a variety of properties of the network structure and activity in this paper, from dendritic and synaptic physiology to multi-area neural activity.

(3) Direct comparisons with experiments, shown throughout the paper, are laudable.

(4) The authors make a number of observations, like describing how high-dimensional connectivity motifs shape patterns of neural activity, which can be useful for thinking about the relations between the structure and the function of the cortical network.

(5) Sharing the simulation tools and a "large subvolume of the model" is appreciated.

Weaknesses:

(1) A substantial part of this paper - the first few figures - focuses on single-cell and single-synapse properties, with high similarity to what was shown in Markram et al., 2015. Details may differ, but overall it is quite similar.

(2) Although the paper is about the model of the whole non-barrel somatosensory cortex, out of all figures, only one deals with simulations of the whole non-barrel somatosensory cortex. Most figures focus on simulations that involve one or a few "microcolumns". Again, it is rather similar to what was done by Markram et al., 2015 and constitutes relatively incremental progress.

(3) With a model like this, one has an opportunity to investigate computations and interactions across an extensive cortical network in an in vivo-like context. However, the simulations presented are not addressing realistic specific situations corresponding to animals performing a task or perceiving a relevant somatosensory stimulus. This makes the insights into the roles of cell types or connectivity architecture less interesting, as they are presented for relatively abstract situations. It is hard to see their relationship to important questions that the community would be excited about - theoretical concepts like predictive coding, biophysical mechanisms like dendritic nonlinearities, or circuit properties like feedforward, lateral, and feedback processing across interacting cortical areas. In other words, what do we learn from this work conceptually, especially, about the whole non-barrel somatosensory cortex?

(4) Most comparisons with in vivo-like activity are done using experimental data for whisker deflection (plus some from the visual stimulation in V1). But this model is for the non-barrel somatosensory cortex, so exactly the part of the cortex that has less to do with whiskers (or vision). Is it not possible to find any in vivo neural activity data from the non-barrel cortex?

(5) The authors almost do not show raw spike rasters or firing rates. I am sure most readers would want to decide for themselves whether the model makes sense, and for that, the first thing to do is to look at raster plots and distributions of firing rates. Instead, the authors show comparisons with in vivo data using highly processed, normalized metrics.

(6) While the authors claim that their model with one set of parameters reproduces many experimentally established metrics, that is not entirely what one finds. Instead, they provide different levels of overall stimulation to their model (adjusting the target "P_FR" parameter, with values from 0 to 1, and other parameters), and that influences results. If I get this right (the figures could really be improved with better organization and labeling), simulations with P_FR closer to 1 provide more realistic firing rate levels for a few different cases, however, P_FR of 0.3 and possibly above tends to cause highly synchronized activity - what the authors call bursting, but which also could be called epileptic-like activity in the network.

(7) The authors mention that the model is available online, but the "Resource availability" section does not describe that in substantial detail. As they mention in the Abstract, it is only a subvolume that is available. That might be fine, but more detail in appropriate parts of the paper would be useful.

Reviewer #2 (Public review):

Summary:

This paper is a companion to Reminann et al. (2022), presenting a large-scale, data-driven, biophysically detailed model of the non-barrel primary somatosensory cortex (nbS1). To achieve this unprecedented scale of a bottom-up model, approximately 140 times larger than the previous model (Markram et al., 2015), they developed new methods to account for inputs from missing brain areas, among other improvements. Isbister et al. focus on detailing these methodological advancements and describing the model's ability to reproduce in vivo-like spontaneous, stimulus-evoked, and optogenetically modified activity.

Strengths:

The model generated a series of predictions that are currently impossible in vivo, as summarized in Table S1. Additionally, the tools used in this study are made available online, fostering community-based exploration. Together with the companion paper, this study makes significant contributions by detailing the model's constraints, validations, and potential caveats, which are likely to serve as a basis for advancing further research in this area.

Weaknesses:

That said, I have several suggestions to improve clarity and strengthen the validation of the model's in vivo relevance.

Major:

(1) For the stimulus-response simulations, the authors should also reference, analyze, and compare data from O'Connor et al. (2010; https://pubmed.ncbi.nlm.nih.gov/20869600/) and Yu et al .(2016; https://pubmed.ncbi.nlm.nih.gov/27749825/) in addition to Yu et al. 2019, which is the only data source the authors consider for an awake response. The authors mentioned bias in spike rate measurements, but O'Connor et al. used cell-attached recordings, which do not suffer from activity-based selection bias (in addition, they also performed Ca2+ imaging of L2/3). This was done in the exact same task as Yu et al., 2019, and they recorded from over 100 neurons across layers. Combining this data with Yu et al., 2019 would provide a comprehensive view of activity across layers and inhibitory cell types. Additionally, Yu et al. (2016) recorded VPM neurons in the same task, alongside whole-cell recordings in L4, showing that L4 PV neurons filter movement-related signals encoded in thalamocortical inputs during active touch. This dataset is more suitable for extracting VPM activity, as it was collected under the same behavior and from the same species (Unlike Diamond et al., 1992, which used anesthetized rats). Furthermore, this filtering is an interesting computation performed by the network the authors modeled. The validation would be significantly strengthened and more biologically interesting if the authors could also reproduce the filtering properties, membrane potential dynamics, and variability in the encoding of touch across neurons, not just the latency (which is likely largely determined by the distance and number of synapses).

(2) The authors mention that in the model, the response of the main activated downstream area was confined to L6. Is this consistent with in vivo observations? Additionally, is there any in vivo characterization of the distance dependence of spiking correlation to validate Figure 8I?

(3) Across the figures, activity is averaged across neurons within layers and E or I cell types, with a limited description of single-cell type and single-cell responses. Were there any predictions regarding the responses of particular cell types that significantly differ from others in the same layer? Such predictions could be valuable for future investigations and could showcase the advantages of a data-driven, biophysically detailed model.

(4) 2.4: Are there caveats to assuming the OU process as a model for missing inputs? Inputs to the cortex are usually correlated and low-dimensional (i.e., communication subspace between cortical regions), but the OU process assumes independent conductance injection. Can (weakly) correlated inputs give rise to different activity regimes in the model? Can you add a discussion on this?

(5) 2.6: The network structure is well characterized in the companion paper, where the authors report that correlations in higher dimensions were driven by a small number of neurons with high participation ratios. It would be interesting to identify which cell types exhibit high node participation in high-dimensional simplices and examine the spiking activity of cells within these motifs. This could generate testable predictions and inform theoretical cell-type-specific point neuron models for excitatory/inhibitory balanced networks and cortical processing.

Minor:

(1) Since the previous model was published in 2015, the neuroscience field has seen significant advancements in single-cell and single-nucleus sequencing, leading to the clustering of transcriptomic cell types in the entire mouse brain. For instance, the Allen Institute has identified ~10 distinct glutamatergic cell types in layer 5, which exceeds the number incorporated into the current model. Could you discuss 1) the relationship between the modeled me-types and these transcriptomic cell types, and 2) how future models will evolve to integrate this new information? If there are gaps in knowledge in order to incorporate some transcriptome cell types into your model, it would be helpful to highlight them so that efforts can be directed toward addressing these areas.

(2) For the optogenetic manipulation, it would be interesting if the model could reproduce the paradoxical effects (for example, Mahrach et al. reported paradoxical effects caused by PV manipulation in S1; https://pubmed.ncbi.nlm.nih.gov/31951197/). This seems a more relevant and non-trivial network phenomenon than the V1 manipulation the authors attempted to replicate.