A multiscale model of striatum microcircuit dynamics

Reviewer #1 (Public Review):

Summary:

The authors aimed to develop a mean-field model that captures the key aspects of activity in the striatal microcircuit of the basal ganglia. They start from a spiking network of individual neuron models tuned to fit striatal data. They show that an existing mean-field framework matches the output firing rates generated by the spiking network both in static conditions and when the network is subject to perfectly periodic drive. They introduce a very simplified representation of dopaminergic cortico-striatal plasticity and show that simulated dopamine exposure makes model firing rates go up or down, in a way that matches the design of the model. Finally, they aim to test the performance of the model in a reinforcement learning scenario, with two very simplified channels corresponding to the selection between two actions. Overall, I do not find that this work will be useful for the field or provide novel insights.

Strengths:

The mean-field model dynamics match well with the spiking network dynamics in all scenarios shown. The authors also introduce a dopamine-dependent synaptic plasticity rule in the context of their reinforcement learning task, which can nicely capture the appropriate potentiation or depression of corticostriatal synapses when dopamine levels change.

Weaknesses:

From the title onwards, the authors refer to a "multiscale" model. They do not, in fact, work with a multiscale model; rather, they fit a spiking model to baseline data and then fit a mean-field model to the spiking model. The idea is then to use the mean-field model for subsequent simulations.

The mean-field modeling framework that is used was already introduced previously by the authors, so that is not a novel aspect of this work in itself. The model includes an adaptation variable for each population in the network. Mean-field models with adaptation already exist, and there is no discussion of why this new framework would be preferable to those. Moreover, as presented, the mean-field model is not a closed system. It includes a variable w (in equation 7) that is never defined.

Overall, the paper shows that a mean-field model behaves similarly to a spiking model in several scenarios. A much stronger result would be to show that the mean-field model captures the activity of neurons recorded experimentally. The spiking model is supposedly fit to data from recordings in some sort of baseline conditions initially, but the quality of this fit is not adequately demonstrated; the authors just show a cursory comparison of data from a single dSPN neuron with the activity of a single model dSPN, for one set of parameters.

The authors purport to test their model via its response to "the main brain rhythms observed experimentally". In reality, this test consists of driving the model with periodic input signals. This is far too simplistic to achieve the authors' goals in this part of the work.

The work also presents model responses to simple simulations of dopamine currents, treated as negative or positive inputs to different model striatal populations. These are implemented as changes in glutamate conductance and possibly in an additional depolarizing/hyperpolarizing current, so the results that are shown are guaranteed to occur by the direct design of the simulation experiment; nothing new is learned from this. The consideration of dopamine also points out that the model is apparently designed and fit in a way that does not explicitly include dopamine, even though the fitting is done to control (i.e., with-dopamine) data, so it's not clear how this modeling framework should be adapted for dopamine-depleted scenarios.

For the reinforcement learning scenario, the model network considered is extremely simplified. Moreover, the behavior generated is unrealistic, with action two selected several times in succession independent of reward outcomes and then an instant change to a pattern of perfectly alternating selection of action 1 and action 2.

Finally, various aspects of the paper are sloppily written. The Discussion section is especially disappointing, because it is almost entirely a summary of the results of the paper, without an actual discussion of their deeper implications, connections to the existing literature, predictions that emerge, caveats or limitations of the current work, and natural directions for future study, as one would expect from a usual discussion section.

Reviewer #2 (Public Review):

Summary:

The present article by Tesler et al proposes a 3-population model of the striatum input-output function including the direct pathway (D1) striatal projection neurons (dSPNs), the indirect pathway (D2) striatal projection neurons (iSPNs), and the fast-spiking striatal interneurons. The authors derive a mean-field version of the model where the firing rate of each population follows the transfer function obtained from a spiking (AdEx) neuron model for each cell population. They report the response of the mean-field circuit to oscillatory inputs from the cortex, the effect of dopamine on dSPNs and iSPNs, and how a simple reinforcement learning rule at cortico-striatal synapses would adapt the model's output in the face of 2 distinct inputs.

Strengths:

The model is simple and easy to understand.

Weaknesses:

Feedforward inhibition from FSI and interconnections between dSPNs and iSPNs does not seem to have any significant impact on the input-output response of dSPNs and iSPNs to cortical inputs. Therefore, all of the results shown can be derived relatively easily from the basic knowledge we have about mean-field neuronal models and their responses to external inputs: all populations have an output that linearly follows the input. Concerning the reinforcement learning paradigm, showing that 2 distinct inputs can be associated with opposite outputs based on a tri-partite synaptic learning rule does not appear new either. As it is, it's unclear to me how this model contributes to new knowledge concerning striatal neuronal activity. Moreover, the assumptions made concerning the effect of dopamine and the synaptic plasticity rules appear rather simplistic and relatively outdated.

Many of the goals set in the introduction do not appear met:

"understanding and modelling the complex dynamics and functions of the striatum constitutes a very relevant and challenging task".
I'm not sure if the authors aim to understand and model the complex dynamics of the striatum here: there are no complex dynamics that are revealed or explained in the model, as the dSPNs and iSPNs mainly appear to have a linear relationship to their inputs (with added noise) in 3 for example. I did not find any non-trivial dynamics highlighted in the presentation of the results either.

"modelling and studying the functions of the striatum and its associated neuronal dynamics requires to investigate these cellular/microcircuits mechanisms, and how the small-scale mechanisms affect large-scale behavior"
I also did not find a statement about the effect of cellular/microcircuit mechanisms on behavior or large-scale activity in the results or discussion. The effects of micro-circuits are rather transparent as dSPNs and iSPNs do not seem to differ from feedforward responses to cortical inputs.

"existing mean-fields are based on generic models (sometimes inspired by cortical circuits) [7, 8], which do not consider the rich and specific cellular and synaptic variability observed along brain regions."
The authors argue here that specific input-output relationships of striatal neurons may contribute to the circuit dynamics. However, the input-output they derive from a spiking neuron model (AdEx) in Figure 2, are very typical IF curves used in most mean-field models. Apart from a slight saturation effect at large rates (which is incorporated in many mean-field models and may not even be relevant here given the max firing of these cells), the I-F curve looks exactly like what is expected from the most basic rate model neuron with a rectifying transfer function in the presence of synaptic noise. What cellular or synaptic properties would the authors like to highlight here? Linking to molecular and cellular parameters, as advertised in the intro, seems much beyond the current achievements of the present model.

"This approach permits an efficient transition between scales and, furthermore, it allows to explore the effects of cellular parameters at the network level, as we will show for the case of dopaminergic effects in the striatum."
If the authors mean the excitation of D1 SPNS and the inhibition of D2 SPNs by dopamine, this statement seems slightly oversold. It's very well known that dopaminergic effects cannot simply be resumed by a change in excitability as it acts on non-linear currents and complex synaptic parameters. They model it as follows: "To model these effects of dopamine in dSPN cells we will assume the increase of excitability due to D1 activation in dPSNs can be described as an increase in the glutamatergic conductance (Qe in our model) together with the action of a depolarizing current" Which basically means an additional excitatory input and a depolarizing current. The expected effect on the firing rate of these 2 effects is rather simple and does not require circuit modelling I believe.

This effect of dopamine is referred to in the discussion as: "This analysis allowed us to show how modifications at the cellular level can be incorporated within the mean-field model which can in turn predict and capture the emergent changes at the network level generated by them, and in addition has provided further validation to our model."
Again, I don't see any emergent property or model validation here. Maybe the authors can be a bit more precise about what emergent property they refer to.

"In addition it illustrates how changes at the cellular level can lead to emerging effects at the network level, which can be captured by the mean-field model"
I did not find any description of 'emerging effects at the network level" in the results or discussion. Maybe the authors could elaborate on what they mean here.

"shows the capabilities of the model to reproduce specific brain functions"
The capacity of a network to associate stim A to a positive output and stim B to a negative one through reward-driven synaptic plasticity is rather well described and is a bit far from 'specific brain functions'. Concerning the discussion, it highlights how the model 'could be useful' rather than highlighting any strength of the model or relation to existing work. In particular, the (large) literature on circuit modelling in the striatum and BG circuits is not cited at all beyond self-citations, except in one book chapter (Houk et al, 1995) and one paper (Bogacz, 2020).

"The RL model proposed can very easily be improved to capture more biologically complex scenarios"
Why did the authors not implement such an 'easy' improvement?