Compensation of Hyperexcitability with Simulation-Based Inference

Joint Public Review:

Summary:

This manuscript couples a 32-parameter model with simulation-based inference (SBI) to identify parameter changes that can compensate for three canonical hyperexcitability perturbations (interneuron loss, recurrent-excitatory sprouting, and intrinsic depolarisation). The study demonstrates a careful implementation of SBI and offers a practical ranking of "compensatory levers" that could, in principle, guide therapeutic strategies for epilepsy and related network disorders.

Strengths:

(1) By analysing three mechanistically distinct hyper-excitable regimes within the same modelling and inference framework, the work reveals how different perturbations require different compensatory interventions.

(2) The authors adopt posterior estimation to systematically rank the efficiency of different mechanisms in balancing hyperexcitability.

(3) Code and data are available.

Comments on revised version:

I appreciate the authors' extensive efforts in revising the manuscript and responding to the previous review. The revised version is substantially improved in clarity, organization, and presentation. In particular, the addition of schematic figures, the reorganization of the Methods section, the improved explanation of the model, and the inclusion of replication analyses all strengthen the manuscript.

The manuscript remains entirely computational, and therefore its conclusions should be interpreted as predictions generated by a specific model rather than validated biological mechanisms. I believe the work has the potential to make a useful methodological contribution. However, several concerns remain regarding validation, interpretation of inferred posteriors, organization of the manuscript, and presentation.

Major comments:

(1) The manuscript states that simulation-based calibration showed the amortized posterior estimator was unreliable (85-88), but these results are not shown. The manuscript explicitly states that simulation-based calibration demonstrated substantial failures of the amortized posterior estimator, yet the corresponding analyses are not presented. Since these results motivate the transition to sequential NPE and are central to assessing inference reliability, they should be reported quantitatively, either in the main text or supplementary material.

(2) The authors present two independently trained estimators and show strong agreement between them. This is a useful robustness analysis. However, the rebuttal occasionally presents this as addressing concerns regarding cross-validation and generalization. The new analysis does not constitute cross-validation in the usual sense and does not directly assess generalization to held-out targets or posterior accuracy.
I recommend that the authors explicitly describe Figure 4 as a reproducibility analysis and avoid presenting it as a substitute for validation.

(3) Posterior correlations are useful for generating hypotheses about compensatory mechanisms, but they should not be interpreted as direct evidence of compensation. The compensatory interpretation should instead be supported by the perturbation analyses (e.g., Figure 6), which provide mechanistic validation.

The manuscript consistently treats posterior correlations and conditional posterior shifts as direct evidence of compensatory mechanisms. These are consistent with compensatory mechanisms, but they do not by themselves establish that the corresponding biological parameters causally compensate for the perturbation. I recommend clarifying this distinction and emphasizing that the conditional posterior analyses generate hypotheses regarding compensation, which are then partially supported by the perturbation experiments shown later in the manuscript.

The language throughout the manuscript should therefore be softened.

(4) The manuscript repeatedly suggests that the inferred conditional distributions may be useful for identifying precise interventions or guiding personalized treatments (examples include lines 24-29, lines 217-223, lines 242-246, lines 277-282, lines 283-286). These claims go beyond what is directly demonstrated.

The study does not evaluate treatment outcomes, patient-specific inference, intervention efficacy, or clinical decision-making. Rather, it demonstrates differences in inferred parameter distributions within a computational model. While these results are valuable and may generate clinically relevant hypotheses, they do not yet establish predictive utility for treatment selection or precision medicine. I therefore recommend substantially softening these translational claims and emphasizing that the current findings generate hypotheses that could be tested experimentally in future work.

(5) The revised manuscript still mixes presentation of findings with interpretation.

For example, lines 217-226 largely continue to describe findings from Figure 6 and would fit better in the Results section. The Discussion would be strengthened by focusing more exclusively on biological implications, limitations, and future directions.

A similar issue appears later in the discussion comparing posterior correlations and conditional distributions. Much of this section effectively reinterprets Figures 2 and 3 rather than discussing broader implications.

(6) The discussion around lines 271-282 overstates what can be concluded from the inferred posteriors.
The statement that correlations "discover broadly applicable mechanisms" whereas conditionals "identify specific mechanisms" is stronger than the presented evidence supports. Likewise, the conclusion that conditional distributions are more useful for precision treatments is speculative and not directly demonstrated.

I recommend reformulating these statements as interpretations or hypotheses rather than conclusions.

(7) Around line 84, the manuscript introduces q(theta|x) without clearly defining θ, x, or q. Readers unfamiliar with SBI may struggle to follow the notation. All quantities should be defined when first introduced.

(8) The manuscript equates larger KS distances between conditional posteriors with greater compensatory potential. While KS distance provides a useful measure of posterior redistribution, it is not obvious that it should be interpreted as a measure of biological efficacy.

(9) The manuscript would benefit from a discussion of parameter identifiability. The inference problem maps 32 model parameters to 7 summary statistics, implying substantial non-identifiability. While complete identifiability analysis is likely beyond the scope of the current work, this limitation should be discussed explicitly.

All in all, the revised manuscript is significantly improved and addresses several concerns raised in the previous review. However, important issues remain as discussed above.

Author response:

The following is the authors’ response to the original reviews.

Joint Public Review:

Summary:

This manuscript couples a 32-parameter model with simulation-based inference (SBI) to identify parameter changes that can compensate for three canonical hyperexcitability perturbations (interneuron loss, recurrent-excitatory sprouting, and intrinsic depolarisation). The study demonstrates a careful implementation of SBI and offers a practical ranking of "compensatory levers" that could, in principle, guide therapeutic strategies for epilepsy and related network disorders.

Strengths:

(1) By analysing three mechanistically distinct hyper-excitable regimes within the same modelling and inference framework, the work reveals how different perturbations require different compensatory interventions.

(2) The authors adopt posterior estimation to systematically rank the efficiency of different mechanisms in balancing hyperexcitability.

(3) Code and data are available.

We thank the reviewers for their positive comments on our manuscript.

Weaknesses:

(1) A highly dense presentation of the simulated models and undefined symbols makes it hard for readers outside the modelling community to follow the biological message. An illustration of the models, accompanied by some explanations and references to the main equations and parameters discussed in this paper, would make the first section much more straightforward.

Thank you for this feedback. To clarify our methods, we have added Figure 7, which illustrates the dynamics of the point neurons and their synapses. We have also added explanations and definitions of variables right where they appear. These variables were previously defined only in a table on a different page.

We also moved the methods section to the back of the paper, as is common in many modern manuscripts. We hope that relegating method details to the end makes the manuscript more accessible.

(2) This methodology appears to be a brute-force approach, requiring millions of simulations to tune 32 parameters in a network of 500-700 cells. It isn't scalable. Moreover, the authors did not use cross-validation, which, with a relatively low increase in computational cost, would provide a quantitative measure as to how well it generalizes; this combination raises doubts about both scalability and reliability.

Scalability is indeed a key challenge of SBI methods. Amortized neural posterior estimation (NPE) is a brute-force approach in that it samples solely from the prior distribution, which is extremely wide. Many of these samples are therefore not very informative for the biologically plausible dynamics we are interested in, which is a downside of amortized NPE. However, amortized NPE is extremely scalable because once the estimator is trained, it can estimate the parameter distribution of any given output dynamic. We tried to build an amortized NPE for our simulator, but simulation-based calibration (a method to validate posterior estimates using additional simulations) showed that the estimators were unreliable.

Sequential NPE is not a brute-force approach because it samples from posterior estimates, which are narrower than the prior. Because the amortized NPE failed, we use sequential NPE to create the two estimators for the baseline and the hyperexcitable condition described in the paper. While millions of prior samples are used to generate the initial posterior estimate, which is then sequentially refined, the sequential refinement requires only 80,000 additional simulations. This requires a significant amount of computational resources, which is why we consider the results worth reporting, but we make the simulator, the simulation results, and the trained estimators available, so other researchers can use or train their own estimators without running millions of simulations. We hope our rewrites make the advantages and disadvantages of the approach clearer.

Regarding reliability and cross-validation, we agree that our initial submission has fallen short. We presented results from only one density estimator per condition, which we considered sufficient given the large number of samples. In the revised version, we present the key results from two additional density estimators trained on partially new training data (Figure 4).

(3) Several parameters remain so broadly distributed after fitting that the model cannot say with confidence which specific changes matter. Therefore, presenting them as "compensatory levers" is somewhat questionable.

It is indeed difficult to determine which changes matter because of the simulator’s complexity. Especially the marginal correlation coefficient (Figure 2 C) are small and the pairwise histograms are broad (Figure 2A), because all other parameters are unconstrained. But the conditional correlation coefficients are larger (Figure 2D) and narrower (Figure 2B). We have added the histograms in Figure 2B in the revised version to highlight the difference. We cannot provide a definitive threshold for correlation coefficients to discriminate between important and unimportant mechanisms. Therefore, compensatory mechanisms discovered with SBI should be validated mechanistically, as we do in Figure 5.

(4) Every conclusion is drawn from simulated data; without testing the predictions on recordings, we have no evidence that the proposed interventions would work in real neural tissue. Because today we cannot diagnose which of the three modelled pathological regimes is actually present in vivo, the paper's recommendations cannot yet be used to guide therapy.

This is indeed an unfortunate drawback of our current work. We are working to apply this approach to constrain microcircuit simulators with data from epilepsy patients. But that work is currently ongoing and will not fit into the present manuscript.

Recommendations for the authors:

Beyond the issues I wrote above, which are methodological, I would like to raise my concern about the way this manuscript is written:

We highly appreciate this editorial feedback on clarity and style. Such feedback is rare and we have worked to address each point to improve the manuscript.

(1) Paragraphs - several paragraphs start with: "To quantify/identify/find specific compensatory mechanisms of hyperexcitability with simulation-based inference". It is a good idea to orient the reader with the specific goal of each section, but it is not helpful to repeat the overall message of the paper in every paragraph. Several paragraphs open with "However," or "Additionally,". Please restructure sentences so that connectors appear after a clear topic sentence.

We have done major rewrites to improve the readability of our manuscript. We start paragraphs with more specific context sentences, rather than the broad research goal, and also made paragraphs much shorter with clearer main messages.

(2) Section 1: The way you present NPE, it would seem like it's specific to neuroscience (and it's not). The paragraph starting at line 26 is not clear. Please revise it. Line 29 - missing a "." before the next sentence begins. Avoid phrases like "for the longest time".

We now stress that NPE, like SBI, is used across scientific domains.

(3) Section 2 was tough to read. Please present each equation in its own numbered display, followed immediately by a plain-language explanation of every symbol and parameter. Provide an illustrative diagram: a small schematic of the AdEx neuron, synaptic connections, and the three perturbations. Even a simple block figure will orient nonexperts. Keep critical methodological decisions (priors, summary statistics, simulation length) in the main text, but move voluminous tables of parameter bounds, learning rates, and hardware specs to the supplement. Remove mentions of which Python functions you used. Readers care about algorithmic choices, not function names. Please reserve specific code references for the GitHub README.

We have added schematic panels at the beginning of Figures 3 & 4 and added Figure 7, which illustrates the neuron and synapse models of the simulator. We also made major rewrites to the methods section to remove programmatic implementation details and define variables where they appear.

In general, I think it would be a good idea to have an editor to polish syntax, verb tense consistency, and punctuation. A thorough language edit will improve the readability and impact of this manuscript.

We have attempted to improve the points raised by the reviewer. In particular, we have carefully rewritten verb tense and punctuation throughout the revised manuscript.