The Locomoting State Selectively Amplifies Activity of Sensitizing Neurons in Primary Visual Cortex

Curation statements for this article:
  • Curated by eLife

    eLife logo

    eLife Assessment

    This paper examines how locomotion modulates the activity of different sub-types of cortical neurons in mouse primary visual cortex. The authors convincingly demonstrate that locomotion more strongly increases responses in sensitizing than in depressing excitatory cells. Computational modeling then accounts for these data through locomotor state-dependent modulation of inhibitory synapses. The findings represent an important contribution to our understanding of how internal state and neural dynamics can interact to modulate sensory processing.

This article has been Reviewed by the following groups

Read the full article

Abstract

Sensory processing in the cortex reflects both adaptation to external stimuli and changes in internal state. To investigate how these processes interact in layer 2/3 of mouse V1 we combined calcium imaging, optogenetics and circuit modelling. We find that locomotion preferentially increases the responses of pyramidal cells (PCs) that sensitize during visual stimulation compared to those that depress. A model explains this differential modulation through: (i) variations in the strength of PV and SST connectivity to individual PCs, (ii) broad locomotion-dependent weakening of PC and PV synapses, and (iii) reduced SST inhibition targeting sensitizing PCs. Differences in PV:SST input ratios across sensitizing and depressing PCs can be reproduced by a random- connectivity model based on measured interneuron densities, connection probabilities and synaptic strengths. Thus, stochastic variation in local inhibitory connectivity can explain much of the functional heterogeneity across PCs, while state-dependent modulation of inhibitory synapses reorganizes this balance during locomotion to bias cortical population activity towards sensitizing dynamics. The apparently paradoxical combination of increased PC response but decreased synaptic strength is consistent with a state-dependent gating mechanism that boosts signals leaving V1 while simultaneously preventing disruption of the local excitatory-inhibitory balance required for stable computation.

Article activity feed

  1. eLife Assessment

    This paper examines how locomotion modulates the activity of different sub-types of cortical neurons in mouse primary visual cortex. The authors convincingly demonstrate that locomotion more strongly increases responses in sensitizing than in depressing excitatory cells. Computational modeling then accounts for these data through locomotor state-dependent modulation of inhibitory synapses. The findings represent an important contribution to our understanding of how internal state and neural dynamics can interact to modulate sensory processing.

  2. Reviewer #1 (Public review):

    In this manuscript, Hinojosa and colleagues analysed the changes in V1 visual responses induced by locomotion in head-fixed mice using two-photon calcium imaging. The authors observe that locomotion strongly increases the visual responses of V1 excitatory neurons that exhibit sensitizing responses to visual stimuli. Also, there is an increased response in VIP interneurons, and to a lesser extent, PV interneurons and SST interneurons (non-significant). The authors used a model fitted with data presented in the manuscript, as well as previous knowledge on cortical connectivity among different neuron types. The model suggests that the major component of the increased responses during locomotion is an increase in excitatory drive from external inputs (feedforward, feedback and modulatory), most importantly onto VIP interneurons and excitatory neurons. However, the excitatory drive of local excitatory neurons onto other surrounding excitatory and inhibitory cells is reduced.

    The manuscript is well presented and represents a valuable analysis of how locomotion modulates the activity of different subtypes of cortical neurons. However, major issues should be addressed to strengthen the results.

    Comments on revised version.

    The authors have addressed all our comments satisfactorily. The revisions have strengthened and clarified the manuscript.

  3. Reviewer #2 (Public review):

    This is an interesting paper with important results. The authors, working in V1, have previously, in a 2022 paper, defined sensitizing and depressing excitatory (E) cells as those whose response increases or decreases, respectively, across the 10 seconds of showing a drifting grating stimulus. They showed that sensitizing E cells are dominantly inhibited by SST inhibitory cells, which are dominantly depressing, and that depressing E cells are dominantly inhibited by PV inhibitory cells, which are very largely sensitizing. It's been well established that locomotion greatly increases E-cell firing rates in V1 compared to rest, but much remains to be worked out as to the mechanism. Here, they find that locomotion increases the responses of the sensitizing E cells much more than depressing cells. They develop a model of changes in synaptic weights between rest and locomotion to account for the changes. One reason that sensitizers are increased more by locomotion than depressors is that PV cells, which more strongly inhibit depressors, have increased firing for locomotion, whereas SST cells, which more strongly inhibit sensitizers, don't change their firing rates with locomotion. However, in the mode,l a complex array of postulated changes in connection strengths is also involved.

    Comments on revised version.

    To my previous review, I just want to add that the authors have taken the reviews very seriously, responded to them thoroughly and well and with a number of new studies that greatly strengthen the work.

  4. Author response:

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

    Public Reviews:

    Reviewer #1 (Public review):

    (1) Speed and mismatch between locomotion and visual stimulation.

    The authors do not clearly describe the definition of locomotion versus the resting state. The speed should, by itself, have an impact on neuronal responses, especially at the onset of locomotion. Several published studies show that the mismatch between a visual stimulus and the speed of the animal induces specific responses in V1, both in excitatory and subtypes of inhibitory neurons. The authors should address these points upfront in the manuscript, since it is likely a major variable explaining their results.

    This suggestion has revealed interesting new features in our data.

    We have now clarified in the methods that a trial was classified as “locomotion” when an animal ran at a minimum of 3 cm/s for at least 80% of the 10 s stimulus presentation, and was classified as “rest” when running under 3 cm/s during the same fraction of time. Trials with transitions from rest to locomotion and vice versa were rare and were excluded to ensure a stable behavioural state throughout the trial. We have included an example trace in Supplementary Fig. 1 to show how this categorization was done.

    We found a preferred speed at which locomotion most strongly modulates visual processing in layer 2/3, with the strongest effects in sensitizing PCs. To assess how visual responses varied with running speed, we classified neurons based on their average speed during the imaging session. We found that response amplitude does not simply increase monotonically with speed but rather exhibits a maximum at a preferred speed of 10-20 cm/s followed by a second weaker peak at 45-55 cm/s (Fig. 2A). These speed ranges overlap with those typically associated with walking (9 ± 5 cm/s) and trotting (51 ± 9 cm/s) in mice (Bellardita & Kiehn, 2015). Sensitizing PCs consistently underwent a larger increase in visual responses than depressors, with the largest difference occurring at speeds of ~10 cm/s (Fig. 2B,C).

    Locomotion has also been shown to modulate inhibitory interneurons (Dipoppa et al., 2018), but the impact on their visually-evoked responses is less well characterised. We found that visual responses of VIP and PV interneurons were strongly amplified at lower locomotion speeds (Fig. 2D-G), while SSTs were not significantly affected (Fig.2 H,I). Overall, these results reveal a preferred speed at which locomotion most strongly modulates visual processing in layer 2/3, with the strongest effects in sensitizing PCs.

    We did not find a correlation between stimulus onset and response increase. We quantified the onsets preceding each trial based on (Vinck et al., 2015). Although lower speed trials contained more locomotion onsets, as illustrated by the example traces, these events were relatively infrequent and occurred at different time points relative to stimulus presentation, making it difficult to assess their contribution to adaptation.

    The effects of visuomotor mismatch could not be directly assessed based on our experimental paradigm. While previous studies distinguished visual flow from running speed by introducing violations of their relationship in head-fixed mice navigating a virtual-reality corridor (Keller et al., 2012; Saleem et al., 2013), our mice were presented with a fixed stimulus irrespective of their locomotion speed. Consequently, different degrees of mismatch between visual input and locomotion were present continuously rather than being introduced as unexpected events during a trial.

    Action taken

    - At line ~173, we added the speed analysis as a new figure (Fig. 2).

    - At line ~889, we added a new Discussion section interpreting these results in the context of previous literature.

    - At Line ~1030 in the Methods section, we clarified the definition of locomotion and rest periods.

    (2) Use of deconvolution with MLSpike.

    Some results (Figure 2) exclusively depend on the deconvolution of calcium signals into spikes (since the initial peak is not seen in calcium transients). The authors should validate this result either with electrophysiological recordings or with the use of another deconvolution method (e.g. CASCADE), emphasising the limitations of this approach and the limitations of the time resolution of calcium imaging.

    We have now validated our results using an alternative spike inference method CASCADE (Rupprecht et al., 2021), and find that it generates a similar initial peak. This is now shown in Supplementary Fig. 3C, A similar initial increase in amplitude followed by fast depression is a consistent observation in electrophysiological recordings in V1 (Chance et al., 1998; Jin & Glickfeld, 2020; Varela et al., 1997). We have maintained MLSpike as our main spike inference method because CASCADE relies on a mapping learned from external ground truth datasets, whereas MLSpike estimates the calcium-spikes relationship from the dataset being analysed. If adaptation alters the spike-to-calcium transformation, there may exist discrepancies in inferred adaptation dynamics even when both methods recover the initial responses.

    We have also added a note in the Methods section (line ~1053) about the limitations of the method due to both the limited time-resolution of calcium imaging and the unvalidated assumptions we used to transform calcium signals into spikes.

    Action taken

    - Added Supplementary Fig. 3C.

    - At line ~216, “Despite the limited time-resolution of calcium imaging, the inferred spiking response revealed the initial phase of fast depressing adaptation that is a consistent feature of electrophysiological recordings (Chance et al., 1998; Fairhall et al., 2001; Varela et al., 1997) (red arrows In Fig. 3A-B). The initial peak was also recovered using the CASCADE (Rupprecht et al., 2021) spike inference method (Supplementary Fig. 3C).”

    - At line ~1053: “The initial component should be interpreted cautiously because of the limited time-resolution of calcium imaging and because spike inference relies on assumptions that were not directly validated in our experiments. Nevertheless, the initial peak was also recovered using a different spike inference method: CASCADE. Calcium traces were analysed using CASCADE (Ref. Rupprecht et al., 2021) with the pretrained Global_EXC_6Hz_smoothing200ms model. CASCADE automatically matched traces to the available noise levels and combined predictions across five network ensembles. The resulting inferred spike probabilities were compared with the MLSpike results.”

    (3) The manuscript is centred around a specific increase in visual responses in sensitizing neurons during locomotion, both in the fraction of responsive neurons and response magnitudes.

    It is hard to tell whether this difference is due to a greater scaling effect of locomotion, a difference in responses during the resting state, or both. The manuscript should further explore and discuss the differences in responses between sensitizing and depressing neurons, both during the resting state and locomotion. Adding metrics and direct comparisons of the magnitudes of fast responses, slow responses, and time integrals between sensitizing and depressing neurons in resting and locomotion states would help to clarify this. Same for fractions of responsive neurons of each type in each condition. E.g., the slow phase is harder to judge from the plots, but the DeltaF/F integral shown in Figure 1G seems to suggest the difference in response magnitude between sensitizing and depressing neurons is largest in locomotion state, rather than resting state. How do these integrals look for inferred firing rates shown in Figure 2?

    The difference between sensitizing and depressing PCs is due to a greater scaling effect of locomotion, not a change in the rest state. We have now clarified this point by adding the metrics suggested. Fig. 1G uses integrated responses and shows that in the resting state there is no significant difference in the visual responses of PCs categorized as sensitizing, intermediate and depressing. Fig. 1H then shows that the increase in the integrated response during locomotion is negatively correlated with the Adaptive Index.

    We have also divided the signal into initial amplitude and average amplitude to represent the fast and slow components respectively. While there are only small variations in initial amplitude that are not correlated with gain increase (Fig. 1I, Supplementary Fig. 2B), average amplitude is just a scaled version of the time integral giving exactly the same results (Supplementary Fig. 2D).

    The fraction of responsive cells of different adaptive types does not significantly vary from rest to locomotion beyond a difference in AI spread which is only the result of an additive change in gain (Fig. 1C, Methods line ~1025). For this reason, we defined the populations in tertiles, so that each population represents the same relative number of cells. We have now made this clearer in the text.

    We now also document metrics for inferred spikes. The new Fig. 3D-F compares the initial peak amplitudes, showing that locomotion causes a larger relative increase in peak response in sensitizing PCs compared to depressors. We have also added in Supplementary Fig. 3B time integrals which again show that sensitizers undergo a larger increase in gain.

    Action taken

    - Adding individual comparisons to Fig. 1G. and Supplementary Fig. 2D.

    - Adding individual comparisons of initial peak amplitude in inferred spikes to Fig. 3D, and time integral to Supplementary Fig. 3B.

    - Adding how the increase in response varies with AI in Fig. 1H and initial amplitude in Fig. 1I and Supplementary Fig. 2B.

    - At line ~142, “These differences were due to sensitizers displaying a larger response during the locomotion state, rather than differences in responses during the resting state (Fig. 1 D-I)”

    - At line ~130, “To investigate how these response changes affected PCs undergoing different forms of adaptation, we split the population into three equal-sized groups based on the distribution of the adaptive index (AI): sensitizing (more negative AI), intermediate, and depressing (more positive AI; Fig. 1C).”

    (4) There is something counterintuitive about how the changes in inhibition onto sensitizing and depressing neurons during locomotion explain the reported activity changes.

    Sensitizers receive reduced SST input and increased PV input during locomotion. If SSTs depress and PVs sensitize (and this is the main reason why sensitizers, which receive dominant input from SSTs sensitize, and vice-versa), how is it possible that this switch does not alter the sensitizing or depressing nature of these neurons' responses in locomotion? Are these changes insufficient to flip the dominant SST-PV drive? Figure 6D-E seems to show there is a flip, at least for sensitizers. How do authors explain this? Do authors think this is related to the narrowing of the adaptive index distribution shown in Figure 1C?

    The additional dimensionality reduction analyses has simplified the model and the PV: SST ratio remains approximately unchanged in sensitizing PCs (Rest = 1.1; Loco = 1.1) and decreases in depressing PCs, without reversing their relative inhibitory balance (Rest = 2.7; Loco = 1.8).

    The narrowing of the adaptive index distribution in Fig. 1C is the result of an additive gain increase from rest to locomotion. We calculated the adaptive index as AI = (R1 − R2)/(R1 + R2) where R1 is the average response over the period 0.5 - 2.5 s after stimulus onset and R2 is the average at 8-10 s. Adding the same constant to R1 and R2 will make all AIs closer to 0 narrowing the adaptive index distribution as observed in figure 1C. We have now clarified this in the Methods section.

    Action taken

    - At line ~1027, “As a result, adding the same constant to R1 and R2 will make all AIs closer to 0 which is what we observe from rest to locomotion.”

    (5) Presentation of the experimental data and the model.

    The manuscript introduces the results of interneuron recordings during the description of the model. Similarly, the results of optogenetic manipulations are presented inside the model's description. It would be clearer to present all experimental data first and introduce the model later, fitting it to all experimental evidence previously presented.

    We understand that a clear separation between experimental and modelling results is often preferred in papers that combine these approaches but in our case modelling and experimental data are highly interdependent and we believe that an overlapping presentation make it easier for the reader to appreciate the links. One example is Fig. 3G-L that shows experimental results validating a key feature of the model: the use of average response dynamics for each population of interneuron. Similarly, the results in Fig. 4 validate the use of the VIP response dynamics as the template for the slow modulatory input to layer 2/3. Then the results of optogenetic experiments in Fig. 5 are used to narrow down fits to the model. We have retained the integrated structure but revised the presentation throughout to clarify which findings are experimental, which motivate modelling assumptions and which are derived from the model.

    Reviewer #2 (Public review):

    (13) Lines 573-576 assert that given low numbers of PV and SST cells and low probabilities of their connection to PCs, simple random connectivity can give rise to the distribution of PV:SST input ratios on PCs inferred from the model - this is a hypothesis, and it would be simple to test it in simulations.

    This prompt has triggered a major addition to the paper. We find that variations in PV: SST input ratio predicted by the model are largely reproduced by simulations in which PV and SST connections are assigned randomly according to experimentally measured cell densities, connection probabilities and synaptic strengths. We now include this finding in the Abstract and Introduction and evidence it in a new section of the Results, “Random local connectivity as a driver of functional heterogeneity”, and the accompanying Fig. 8. This random connectivity rule provides a simple and parsimonious explanation for variations in PV and SST inputs, although in the Discussion we point out that we cannot rule out the possibility that more specific local connectivity rules also contribute to variations in the PV: SST input ratios that cause variations in adaptation across the PC population.

    In the model, they postulate that synapses within the 6-cell-type network - sensitizing, intermediate, and depressing E cells, and PV, SST, and VIP I cells - and from three sources of external input to each of the six types all change between rest and locomotion (except that connections between the E cells don't depend on their types). There are a lot of degrees of freedom, and this makes interpretation of the results difficult. I would have liked to have seen more efforts to constrain the degrees of freedom. For example, there seems to be very little difference between the three E cell types in any of the three types of external input received. Why not constrain them all to get the same external input and see if it significantly affects model fit? Or what if synapses from the three types of external input are left unchanged, and only change their strengths between rest and locomotion? How well could this do? During optimization, why not constrain the changes between rest and locomotion, for example, by putting an L1 penalty on the changes or the relative changes, trying to force them to be sparse, and see whether there are roughly equally good fits? And then, if the main changes are in a small set of synapses, can the authors isolate changes to that small set and do roughly equally well? What about looking at the principal components of the weight changes across models, to isolate patterns of change that are most important?

    We thank the reviewer for these suggestions; the additional dimensionality reduction analyses simplified the model and strengthened the conclusions drawn from it. We applied dimensionality reduction after the initial global sampling step by identifying weights that changed substantially from rest to locomotion. Because the model generated large ensembles of good solutions, conventional significance testing would be strongly influenced by the number of solutions and would not provide a meaningful measure of the magnitude of the differences. We therefore quantified each change using effect sizes, more specifically the normalized Mann–Whitney U statistic, Unorm = U/(n1n2), equivalent to the Vargha–Delaney A effect size (Vargha and Delaney, 2000) (See Methods). Based on this criterion, we fixed nine weights showing no consistent change, and achieved similar model performance ( Values: Rest original = 2.9±0.0; Rest reduced = 2.9±0.0; Loco original = 9.2±0.01; Loco reduced = 8.1±0.03;) (Supplementary Table 3).

    We then examined whether any additional weights could be fixed. For this purpose, we performed principal component analysis on the original unrestricted solutions and examined the three weights with borderline effect size (w5, w6 and w12). PC1 explained approximately 30% of the variance across solutions. Most weights fixed according to their effect sizes contributed little to PC1, whereas the weights that were kept variable generally made larger contributions (Supplementary Table 2). Among the borderline weights, w12 made only a small contribution to PC1 and was therefore fixed, while w5 and w6 made substantial contributions and remained free to vary. The final reduced model therefore contained nine free weights, while maintaining a fit comparable to that of the unrestricted model. We have added the selection criteria and the results of these analyses to the Results and Methods sections.

    We followed a similar approach to determine which inputs meaningfully distinguished depressing and sensitizing PCs. In this case, the effect-size and PCA analyses did not identify a clear subset of parameters, likely because this comparison involved fewer variable inputs. We therefore tested reduced models with different combinations of fixed and variable weights and selected the simplest model that maintained a good quality fit. This result supports local SST and PV inhibition as the principal modelled distinction between depressing and sensitizing PCs. However, because external inputs were not measured experimentally, the reduced model should not be interpreted as evidence that these inputs are necessarily identical. We have clarified both the procedure and this limitation in the revised text.

    In relation to the L1 penalty, we tried an approach similar to that described (Dipoppa et al., 2018) by using the locomotion weights as initial conditions for the rest traces and introducing penalties at later optimization stages. However, the local optimization algorithms failed to reach distant regions of parameter space containing the good fitting solutions for the rest condition. Consequently, the outcome depended strongly on initialization and did not provide a reliable assessment of which changes could be constrained through an L1 penalty. We therefore fitted rest and locomotion using independent global searches and subsequently reduced the model using the effect-size, PC1, and model-refitting procedure described above. We believe this approach provides a robust and interpretable reduction because it has both allowed us to describe circuit changes during internal-state transitions (the present paper) and, more recently, it has made a series of predictions about different learning states that have been confirmed by optogenetic tests (Hinojosa et al., 2026).

    Action taken

    - Adding Supplementary Table 2 showing the effect sizes and PC1 values.

    - At line ~441, “To simplify the model, we then examined which parameters differed consistently between behavioural states using effect sizes and principal component analysis. This allowed 10 parameters to be constrained to the same values at rest and during locomotion without substantially reducing the quality of either fit (Supplementary Table 2; see Methods).”

    - At line ~608, “The simplest model that maintained a good-quality fit kept excitatory drive from PCs and all external inputs fixed and equal across the three PC types, restricting their differences to their local inhibitory connections (Fig. 7D,E)”.

    - At line ~885, “This reduced model provided the simplest account of the data; however, because external inputs were not measured experimentally, their equality across PC subpopulations should not be interpreted as evidence that they are biologically identical.”

    - At line ~1202, In Methods “To reduce the number of free parameters and attempt to fit a simpler model, we measured the magnitude of the difference between rest and locomotion for each parameter…”

    In terms of comparing to previous works, when optogenetic manipulations of SST and PV are done to test various hypotheses, I would like to see some discussion of what is already known from the authors' 2022 paper and what they are adding or testing that wasn't known or tested from that paper. And Dipoppa et al. (2018) also found weight changes to account for the difference between rest and locomotion. They were looking at a fixed point of responses of neurons across retinotopic space to stimuli of various sizes with only one E-cell type, whereas they are accounting for trajectories across time considering 3 E-cell subtypes but without variation in stimuli or retinotopic position of neurons, so the efforts are somewhat different, but still, it would be good to see a bit more discussion of what is in agreement or in contradiction in the conclusions.

    We have now added a section in the Discussion comparing our results with those of Dipoppa et al. (2018). In Heintz et al. 2022 we found that the proportion of depressors and sensitizers did not change significantly during locomotion, concluding through a qualitative analysis that this may be mediated by VIP->SST and SST->PV pathways. Here we have moved this analysis to a quantitative model to better understand how changes in external inputs to layer 2/3 interact with recurrent inhibition.

    Dipoppa et al. (2018) show that locomotion increases external inputs and reduces recurrent excitation, resulting in a reduced effective SST inhibition. We have come to a similar conclusion, but with more information about the dynamics of external inputs. A key difference, however, is that we find that the increase in VIP response amplitude is mostly mediated by external inputs rather than local PCs, We have included these comparisons in the Discussion.

    Action taken

    - At line ~803,” In a previous, smaller dataset, we qualitatively examined the effects of locomotion across the PC population and found that sensitizing and depressing responses remained balanced, consistent with recruitment of the VIP→SST and SST→PV pathways (Heintz et al., 2022). Our expanded quantitative analysis and modelling now suggest that locomotion acts through changes in both external inputs and recurrent connection weights. Similar circuit changes were inferred by modelling the initial responses of V1 neurons to stimuli of different sizes (Dipoppa et al., 2018). That study also predicted increased external excitation and reduced recurrent excitation, resulting in weaker effective SST inhibition, but attributed the locomotion dependent change in VIP responses to recurrent input from PCs rather than external drive…”

    (1) The main result is that sensitizers increase their responses with locomotion ~2X (for dF/F) or about 3.5X (for spikes) more than depressors. But there are other differences between sensitizers and depressors, for example sensitizers have smaller initial stimulus responses at rest, and depressors have larger. What if cells were divided into tertiles by initial stimulus response at rest? Would the authors see the same differences in the effects of locomotion? If so, can they establish whether the difference is really attached to the adaptation properties rather than to, for example, the initial responses, for example, by comparing the regression of response increase against AI vs the regression of response increase against initial resting response? And there might be other controls to be done for other features in which sensitizers and depressors differ.

    Thanks for highlighting this. In the manuscript, cells were categorized as sensitizers, intermediates, or depressors separately according to their adaptation index (AI) during rest and locomotion. Consequently, the traces shown in the previous Fig. 1D–F were not calculated from completely exclusive neuronal populations. We have now preserved cell identity by classifying neurons according to their AI during locomotion alone. This analysis reproduced the greater locomotion-induced gain increase in sensitizers (Fig. 1G), while revealing no difference between sensitizers and depressors in their initial response amplitude at rest (Fig. 1I).

    We further examined the relationships between the locomotion-induced response increase and either AI or initial response amplitude at rest. Response increase was strongly correlated with AI (Fig. 1H; R=−0.77, P<0.001), whereas its relationship with initial response amplitude differed between sensitizers and depressors (Supplementary Fig. 2B). When cells were binned separately according to their initial resting response, sensitizers showed greater response increases than depressors across the overlapping amplitude range. Response increase was positively correlated with initial amplitude among sensitizers but not depressors (Supplementary Fig. 2B). Thus, although initial amplitude influenced the magnitude of the response increase within sensitizers, it could not explain their greater locomotion-induced amplification relative to depressors, which was more strongly associated with their adaptive index.

    Action taken

    - Adding Fig. 1H.

    - Adding Fig. 1I

    - Adding Supplementary Fig. 2B.

    (2) Lines 103 and following: the authors refer to a "second notable change" which is the narrower distribution of adaptive effects, but I think this is trivial. The adaptive index is AI=(R1-R2)/(R1+R2), where R1 is response 0.5-2.5s after stimulus onset and R2 over 8-10s. But if the change is additive, as suggested by the dF/F figures (and I believe the distributions of AI here are based on dF/F measurements) -- adding the same constant to R1 and R2 will shrink |AI| without changing the sign of AI. So this would seem to just be a signature of a change that is primarily additive rather than multiplicative.

    Also, if the authors do decide that they are going to focus on spikes after showing the raw dF/F, then this analysis should be repeated for spikes.

    We thank the reviewer for pointing out that the narrower AI distribution can arise mathematically from an approximately additive increase in responses. We agree that it should therefore not be presented as a separate biological effect of locomotion. We have removed this point and its associated interpretation from the Results and now clarify this property of the AI calculation in the Methods.

    Additionally, we have included the corresponding AI distributions calculated from inferred spikes (Supplementary Fig. 3A). These distributions appear less smooth because spike inference algorithm excluded some of the cells and produces quantized response estimates, but the overall pattern is preserved. We have added this clarification to the manuscript.

    Action taken

    - Adding Fig. 1H.

    - Adding Fig. 1I.

    - At line ~1025, “The adaptive index (AI) was calculated from the GCaMP6f signal as AI = (R1 − R2)/ (R1 + R2) where R1 is the average response over the period 0.5 - 2.5 s after stimulus onset and R2 is the average at 8-10 s. As a result, adding the same constant to R1 and R2 will make all AIs closer to 0 which is what we observe from rest to locomotion.”

    (3) Figure 2, F is supposed to be D minus E, but it doesn't look like it. For example, the initial response under locomotion is very similar in sensitizers and depressors, so the initial difference in F should be small, but it's not; and at rest, depressors initially have larger responses than sensitizers, whereas later depressors have smaller responses than sensitizers, yet the difference at rest is positive at all times. Something seems wrong here.

    We apologize for the confusion. Previous Figure 2F did not represent the difference between sensitizing and depressing PCs from panels D and E. Instead, it showed the time-varying difference between locomotion and rest states of sensitizers (blue, in figure 2D) and depressors (green, in figure 2E). Thus, panel F showed within-population modulation by behavioural state, rather than differences between sensitizing and depressing neurons. After revising the paper, we do not think that this result is central to the main conclusions, so we have transferred the figure to Supplementary Fig. 3E. We have nonetheless amended the figure legend to better explain its nature.

    Action taken

    - At line ~1379, Supplementary Fig. 3E legend. “Time course of the locomotion-induced change in firing rate in sensitizing and depressing PCs.”

    Reviewer #3 (Public review):

    (1) Key concern is the usage of dF/F signals for all analyses, especially when comparing responses.

    (1a) Figure 1G: Comparison of sensitisers and depressors. It is important to consider what the baseline rates are when making these comparisons, especially when comparing the degree of effects between different cell types. For example, if baseline rates for sensitizers were overall higher, it would mean the difference in gain of response would be lower, and could affect the results in the opposite direction of what is claimed. One option to account for this would be to z-score the overall responses, using the same normalization for locomotion and rest. We also suggest plotting differences in sensitisers, intermediates, and depressors as a function of firing rate. Matching for firing rate across each PC categorization and calculating delta AI for each matched firing rate bin.

    (1b) Figure 2A-F: The above is an even more significant issue when it comes to estimating spiking rates. The methods do not state how dF/F is calculated. If these are based on using the pre-stimulus as the reference, the algorithms for spike rate used might not be appropriate if this were used. Using pre-stimulus referencing could result in the estimate going into the wrong range in the calculation of the spike rate.

    (1c) In both cases above, it could be a problem if baseline firing rates are different between cell types, or states (locomotion/stationary). The latter is established to have effects on many cell types measured, and so needs to be accounted for very carefully.

    The increase in gain does not depend on differences in noise levels across PC populations. We agree with the reviewer that noise levels must be considered to calculate the rate of gain change. We have now calculated the z-score on the original traces, before rest and locomotion trials were segregated, to account for a potential difference in baseline activity. Supplementary Fig. 2E and F reveal that baseline activity is similar in both locomotion and rest conditions.

    Initial response amplitude at rest does not differ between sensitizing and depressing PCs. In the previous manuscript we reported a difference in initial response during rest that was the result of dividing the rest and locomotion traces in sensitizers, intermediates and depressors based on AI calculated separately in each state. We have now preserved the cell identity in these classifications by sorting according to AI during locomotion alone: initial response amplitude during rest does not differ (Fig. 1I). We also binned sensitizing and depressing PCs separately according to their response amplitude during rest. Sensitizing PCs showed greater locomotion-related increases than depressing PCs across the overlapping range of initial response amplitudes. Although response increase was positively correlated with initial amplitude among sensitizers, this relationship was absent among depressors (Supplementary Fig. 2B). Thus, initial response amplitude influenced amplification within sensitizers but did not explain their greater amplification relative to depressors.

    Finally, ΔF/F was calculated from the raw fluorescence trace using the mode of the entire recording as F0, rather than the pre-stimulus fluorescence of individual trials. The same F0 was therefore applied throughout the recording. This normalization preserves differences in fluorescence baseline between rest and locomotion and cannot shift individual trials into different ranges according to their pre-stimulus activity.

    Spike inference using MLSpike explicitly models baseline noise and subtracts its effect from that of the spikes calculated from the calcium signal (Deneux et al., 2016). This transformation preserved the difference in amplitude triggered by locomotion between depressing and sensitizing PCs while revealing their similar baseline activity (see Figs. 3E-I). These results indicate that the distinct changes in response amplitude between sensitizing and depressing PCs during locomotion are not driven by baseline differences. We have now added these clarifications to the Results and Methods sections.

    Action taken

    - Adding Supplementary Fig. 2E, F.

    - Adding Supplementary Fig. 2B

    - At line ~1009, Methods: “ΔF/F0 was calculated from the raw fluorescence trace using the mode of the entire trace as F0”

    (1d) It would be informative to see per-neuron comparison for adaptive indices during rest and locomotion states. This could be visualized using a scatter plot with AI-rest vs. AI-locomotion for Figures 1D- 1F and 2J- 2L.

    (1e) Are neurons more strongly modulated between locomotion and rest, also more likely to experience a shift in AI indices (i.e. delta AI). Is there a correlation between the change in firing rate between behavioral states and Delta AI (Loco-Rest)? If so, is this present for all neuron subtypes (e.g. VIP, SST, and PV)?

    We have now added per-neuron comparisons of AI during rest and locomotion for PCs, VIP, SST and PV cells (Supplementary Fig. 2A,G-I). AI during rest and locomotion was significantly correlated in all cell types except SST cells, indicating that adaptation profiles were partially conserved across behavioural states.

    We also tested whether locomotion-related changes in response integral were correlated with ΔAI for each cell type (Supplementary Fig. 2C,J-L). In PCs, the change in response integral was negatively correlated with ΔAI, indicating that cells that become more sensitizing during locomotion experienced greater increases in response gain. Among interneurons, only PV cells showed a similar relationship between ΔIntegral and ΔAI, while these variables were not correlated in VIP and SST cells.

    Action taken

    - Adding Supplementary Fig. 2A, G-I

    - Adding Supplementary Fig. 2C, J-L

    (1f) Optogenetic inhibition of VIP neurons on average abolished the slow depressive effects of adaptation in SST (Figure 3). The strength and prevalence of this effect are unclear. Perhaps one can perform a bootstrap control and opto AI indices and calculate whether AI was significantly reduced following optogenetics inhibition, and if so, on average, how likely was this to occur for the recorded SST neurons? This is important in knowing that the average effects (Figure 3D) aren't driven by a portion of SST neurons, especially as this is later used to confirm the region of parameter space and affects the subsequent results in Figure 4.

    Because AI values with and without VIP inhibition were obtained from the same SST neurons, we assessed the effect using a paired linear model rather than bootstrapping the two conditions. VIP inhibition significantly reduced SST AI (Fig. 4E; Control AI = 0.44 ± 0.04; Opto AI = 0.12 ± 0.02; p < 0.001). We have also added a histogram of the within-cell change in AI (Fig. 4F), showing that this reduction occurred in 90% of the recorded SST neurons rather than being driven by a small subset. These analyses confirm that the average response reflects a loss of slow depression across the SST population.

    Action taken

    - Adding Fig. 4E,F.

    (2) Statistics for the effects. There is a mention of Liner mixed models, but no information is given on the actual models being used and tested. This is particularly for the case of Figure 1G, where there is a composition of effect sizes between different populations. What precise significance test is being used? Are the stats on paired cells when considering locomotion and rest?

    We used linear mixed-effects models to account for the hierarchical and paired structure of the data. For Fig. 1G, locomotion state, adaptive type, and their interaction were included as fixed effects, with mouse and cell identity as random effects. Thus, comparisons between rest and locomotion account for measurements obtained from the same cells. We tested differences among adaptive types separately within each behavioural state and differences between states separately within each adaptive type. The reported interaction (p < 0.001) indicates that the locomotion-related increase in amplitude differed between sensitizers and depressors. We have now revised the Methods section and figure legends to explicitly describe the model and statistical test used.

    Action taken

    - At line ~1312, Methods: “Locomotion state, adaptive type, and their interaction were included as fixed effects, with mouse and cell identity as random effects. We tested differences among adaptive types separately within each behavioural state, differences between states separately within each adaptive type, and finally the interaction between behavioural state and adaptive type which indicates whether the locomotion-related increase in a given variable differed between sensitizers and depressors…”

    (3) Model parameters: It is acknowledged that there is a large range of parameters that can model the responses effectively, up to 11% of initial conditions. At 9000 initial conditions, this is around 1000. The parameter estimates are then considered as the mean of each parameter. This seems like a strange choice for a few different reasons:

    (3a) A mean solution might not be one of the solutions. Let's say the parameters range over a large dimensional space. They could occupy non-overlapping / discontinuous subspaces. In that case, the mean parameters do not necessarily fall within the solution subspaces. Therefore, this reduction to means might not be valid.

    (3b) Compare distributions rather than means. There are multiple distributions of parameters between conditions. All stats should be on the comparison of distributions rather than just the means.

    We agree that a limitation of our previous presentation was that it did not reflect the full range of solutions produced by the model. We now report both the whole population of solutions and describe the differences associated with locomotion in term of their effect sizes rather than averages alone. Most parameter distributions were approximately Gaussian or showed a monotonic decay when distributed close to 0 (Supplementary Fig. 4). The only exception was the PC->SST connection, which was described by two overlapping Gaussian components. Importantly, both components were similarly displaced towards lower values during locomotion (Supplementary Fig. 4E, change in Gaussian centre: component 1 = 0.5, component 2 = 0.47). This bimodality could potentially reflect heterogeneity among SST interneurons and is highlighted in the Discussion.

    To test for meaningful differences, conventional significance testing would be strongly influenced by the number of solutions, which are not independent biological observations. We therefore quantified changes taking into account effect sizes, more specifically the normalized Mann–Whitney U statistic, Unorm = U/(n1𝑛2), equivalent to the Vargha–Delaney A effect size (Vargha and Delaney, 2000) (See Methods). We considered a substantial difference when 0 < Unorm < 0.3 or 0.7 < Unorm < 1.

    Our original use of the mean was based on the fact that introducing the average for each parameter in the overall model, rendered a solution that would be classified as a “good fit” by our criteria criteria (rest < 3, locomotion < 12). Given this validation and the predominantly unimodal parameter distributions, we have retained the means in the main figures as a concise summary while providing the complete distributions in Supplementary Fig. 4.

    Action taken

    - Adding Supplementary Fig. 4.

    - At line ~517, “Moreover, examination of the distribution of each parameter revealed that they were mostly unimodal (Supplementary Fig. 4)”.

    - At line ~844, Discussion: “In line with this heterogeneity, the PC→SST connection showed a bimodal distribution at rest, with both components shifting similarly towards lower values during locomotion.”

    (4) Visualizing weight matrices: It is very challenging to interpret the weight matrices. Furthermore, it appears that the stationary and locomotion conditions fit independently, and given the large parameter spaces, it is even harder to interpret. Can the fitting instead be done by fitting on one and using those at the initial conditions for the other state? Figure 7 shows an initiative cartoon, but it is not clear how the matrices in Figures 5 and 6 lead to the summary shown in Figure 7. It is also not clear why the connections between inhibitory neurons are not shown in Figure 7. One option is to perhaps run some kind of dimensionality reduction on the parameter space to better interpret the data. When showing delta Weights, was the model initialised with 'Rest' weights and allowed to change? It is not obvious what the difference is between 'relative change in connection weights' and 'relative change in synaptic weights'. This needs to be clarified.

    Thanks for raising these concerns, which prompted us to substantially revise the model presentation and analysis. We applied dimensionality reduction after the initial global sampling step by identifying weights that changed substantially from rest to locomotion. We quantified each change using normalized Mann–Whitney U statistic, Unorm = U/(n1n2) (See Methods) and then examined whether any additional weights could be fixed by performing principal component analysis on the original unrestricted solutions and examining the three weights with borderline effect size (w5, w6 and w12). PC1 explained approximately 30% of the variance across solutions. Most weights fixed according to their effect sizes contributed little to PC1, whereas the weights that were kept variable generally made larger contributions (Supplementary Table 2). Among the borderline weights, w12 made only a small contribution to PC1 and was therefore fixed, while w5 and w6 made substantial contributions and remained free to vary. Based on these criteria, we fixed a total of 10 weights showing no consistent change and achieved similar model performance ( values: Rest original = 2.9 ± 0.0; Rest reduced = 2.9 ± 0.0; Loco original = 9.2 ± 0.01; Loco reduced = 8.1 ± 0.03; Supplementary Table 2). We have added the selection criteria and the results of these analyses to the Results and Methods sections, simplifying the weight matrices and facilitating its interpretation.

    We previously tried an approach similar to that described by Dipoppa et al.(2018) by using the locomotion weights as initial conditions for the rest traces and introducing penalties at later optimization stages. However, the local optimization algorithms failed to reach distant regions of parameter space containing the good fitting solutions for the rest condition. Consequently, the outcome depended strongly on initialization and did not provide a reliable assessment of which changes could be constrained through penalties. We therefore fitted rest and locomotion using independent global searches and subsequently reduced the model using the effect-size, PC1, and model-refitting procedure described above. We believe that this approach provides a robust and interpretable reduction because it has both allowed us to describe circuit changes during internal-state transitions (the present paper) and, more recently, it has made a series of predictions about different learning states that have been subsequently supported by optogenetic experiments (Hinojosa et al., 2026).

    As a summary to facilitate the interpretation of our results, Fig. 9 (former Fig. 7) puts together the inferred changes in connectivity from Fig. 6 and the differences between depressors and sensitizers (Fig. 7). The reduced model is now easier to summarize in this figure: the schematic focuses on the local connectivity changes from Fig. 6F, and added the change in SST→PC and PV→PC connectivity that differs between sensitizing and depressing PCs from Fig. 7. We have also added the inferred changes in interneuron-to-interneuron connectivity, namely the SST→VIP increase with locomotion, as well as the external connections. We revised the figure legend to link each element of the schematic more clearly to the corresponding quantitative results.

    Finally, relative weight changes were calculated using Equation 4, based on the rest and locomotion weights obtained from independent fits. “Connection weights” (Equation 2), describe the total strength of the connection from population j to population i. “Synaptic weights” (Equation 3) defines the strength of individual synapses by dividing the connection weight by the number of presynaptic cells and the connection probability. This distinction is necessary because the model was fitted to responses averaged across all imaged neurons, whereas estimating the strength of individual synapses requires accounting for both the number of presynaptic cells and their probability of connection. We have added this explanation to the Results and Methods sections, more clearly differentiating between both types of weights.

    Action taken

    - Adding Supplementary Table 2 showing the effect sizes and PC1 values.

    - At line ~441, “To simplify the model, we then examined which parameters differed consistently between behavioural states using effect sizes and principal component analysis. This allowed 10 parameters to be constrained to the same values at rest and during locomotion without substantially reducing the quality of either fit (Supplementary Table 2; see Methods).”

    - At line ~608, “The simplest model that maintained a good-quality fit kept excitatory drive from PCs and all external inputs fixed and equal across the three PC types, restricting their differences to their local inhibitory connections (Fig. 7D,E)”.

    - At line ~1202, In Methods “To reduce the number of free parameters and attempt to fit a simpler model, we measured the magnitude of the difference between rest and locomotion for each parameter…” - Updated Figure 9 and its legend.

    (4a) Model parameters were reduced differently for locomotion and rest (Figure 4). We suggest evaluating the results for locomotion and rest using the same chi-square value of 3 for both behavioral states (at least in controls).

    Setting the same chi-square threshold across conditions would not correspond to an equivalent goodness-of-fit criterion. As defined in equation 12, reduced chi-square is inversely proportional to the standard error of the data which is higher in the rest condition compared to locomotion. As a consequence, the same chi-square threshold would impose a disproportionately strict constraint on the condition with lower variability, where deviations between model and data are more heavily penalized. For this reason, we used condition-specific thresholds to ensure comparable fit quality relative to the noise level in each condition. In addition, to enable direct comparison across conditions independent of their noise levels, we used the RMSE as a complementary metric. We now explain these different metrics in more detail in the Methods section.

    Recommendations for the authors:

    Reviewer #1 (Recommendations for the authors):

    (1) The authors use high-contrast, very long stimuli. This should be specified early on in the results and discussed in terms of how it could influence the results.

    At line ~116, we have added: “The GCaMP6f signal in PCs responding to a prolonged, high-contrast stimulus (20º drifting grating at 1 Hz, 10 s duration) are shown in Fig. 1B”.

    At line ~813, we have added to the discussion “We used prolonged, high-contrast stimuli to provide a high signal-to-noise ratio and reliably evoke slow adaptation, whereas stimuli of different durations or contrasts may engage different patterns of locomotion-dependent modulation.”

    (2) In lines 24-25 (Abstract), the authors write: "These results demonstrate that behavioural state selectively shifts cortical computation toward distinct adaptive regimes". This statement is vague and hard to grasp. Consider rewriting.

    We have made this statement more precise in the new version of the abstract.

    (3) In line 96, the authors should include a brief description of the adaptive index metric.

    At line ~133, we have added: “This was calculated as AI = (R1 − R2)/ (R1 + R2) where R1 is the average response over the period 0.5 - 2.5 s after stimulus onset and R2 is the average at 8-10 s (AI; Methods).”

    (4) In line 85, the authors write: "Locomotion increased the response averaged across all PCs by a factor of 2.2 (plus minus) 0.9". Yet, in lines 99-101, they write "but the amplitude of the visual response increased by a factor of 4.5 (plus minus) 0.7 in sensitizing PCs, 2.5 (plus minus) 0.3 in intermediate group and 2.2 (plus minus) 0.2 in depressors". How can the average across all neurons be 2.2 if all the groups have an increase factor equal to or higher than that, and each group represent a third of the population?

    We thank the reviewer for identifying this inconsistency. After preserving cell identity by classifying PCs according to their AI during locomotion, we recalculated the response increases. The response averaged across all PCs increased by a factor of 1.9 ± 0.2, while the increase was 2.1 ± 0.4 in sensitizing PCs and 1.7 ± 0.3 and 1.7 ± 0.4 in intermediate and depressing PCs, respectively. These values have been corrected in the revised manuscript.

    (5) Figures 1F and 2C do not seem to agree. In Figure 1F, there is a clear increase in response across time in locomotion vs resting, while in Figure 2D, this increase is smaller and only appears in the very late phase. How do the authors explain this? Does it have to do with the spike inference and the different temporal resolution? (Figure 2C loco (calcium) is different from Figure 1F).

    These differences are indeed related to the spike-inference method (but not to a difference in temporal resolution – they are the same). Spike inference revealed the initial rapidly adapting phase of the visual response that was not evident in the calcium signal but is well-documented electrophysiologically. Better definition of this peak reduced the apparent difference between rest and locomotion during the early phase of the response in depressing PCs. Note, however, that the difference in rest and locomoting responses quantified by spike inference remained clear during the later, slower, phase of the visual response.

    (6) The authors observe that SST interneurons do not change their response magnitude during locomotion (Figure 2I, L). How do authors explain their results compared to Dipoppa et al., 2018and Pakan et al., 2016?

    There is a key potential difference between our analysis and that of Dipoppa et al., (2018) and Pakan et al., (2016): we focused throughout on mice that were naïve to the visual stimulus (i.e. the first imaging session) whereas these other studies did not report visual-exposure history or control for session number. We have observed greater locomotion-related modulation of SST responses in later sessions, consistent with experience-dependent changes, and these are now described in the context of a study focusing on the changes associated with simple forms of learning, which is now referenced (A.J. Hinojosa, S.E. Dominiak, Ye. Kosiachkin, L. Lagnado (2026). Distinct Disinhibitory Circuits Link Short-Term Adaptation to Familiarity and Reward Learning in Visual Cortex. bioRxiv 2026.03.24.713929; doi: https://doi.org/10.64898/2026.03.24.713929). The Methods now contains the sentence: “We only analyzed responses to the first day of imaging in mice naïve to the stimulus to limit the introduction of longer-term changes in circuit function reflecting simple forms of learning such as habituation (Hinojosa et al., 2026)”. Differences in stimulus duration and presentation protocols may also contribute to the contrasting results.

    (7) In line 250, the authors state they modelled the feedback of higher visual areas with a 380ms delay after stimulus onset. Yet there is no reference for previous literature supporting this. Why did they choose this delay specifically? In the main text, it says 380ms, whereas in the supplementary material, it is 370ms.

    We modelled feedback with a delay because top-down signals in V1 have been shown to emerge after the initial feedforward response (Kirchberger et al., 2023). However, the 370-ms delay in our model was not intended to represent the minimal axonal transmission delay reported in that study. Rather, it was selected from our experimental responses as the onset of the delayed SST response that the feedback input was introduced to reproduce. Because SST responses became approximately flat when VIP influence was removed, we inferred that their remaining delayed drive arose primarily from feedback, consistent with evidence that LM to V1 feedback strongly targets L2/3 SST interneurons. We have now clarified this rationale, added the relevant references (at line ~360), and corrected the inconsistent value of 380 ms to the correct one of 370 ms.

    (8) In lines 307-309 authors declare: "This third stage could not be usefully applied to data at rest because there were not enough recording episodes across all four optogenetic conditions and responses were smaller and more variable than during locomotion". Does this mean that the model fit better represents synaptic weights during locomotion than during the static state? How can this influence the conclusions drawn from the model?

    We agree that the absence of the third, optogenetic filtering stage at rest means that the inferred rest weights are less strongly constrained than those during locomotion. We did not apply this third step to the rest data because of the lower response amplitudes, smaller number of responsive cells, and smaller number of stimulus trials across all four optogenetic conditions. Nevertheless, the rest data provided sufficiently reliable population responses for the initial model fitting; only their subdivision by optogenetic condition produced traces that were too variable for meaningful additional filtering. We have therefore clarified that the inferred weights should not be interpreted as equally precise estimates of synaptic strength in the two states (at line ~451). The lack of optogenetic filtering at rest increases uncertainty in the magnitude of these changes but does not invalidate the model’s conclusions.

    (9) Are the results presented in Figure 5F a time integral of the connection weights changes?

    No. Figure 5F shows the relative difference between the mean connection weights inferred independently at rest and during locomotion, calculated as:

    Connection weights are time-independent parameters within each condition and were estimated by fitting the model to the full time course of the population responses; therefore, no temporal integration of the weights was performed. We have clarified this in the figure legend: “6F. Relative change between the time-independent connection weights inferred at rest and during locomotion (Equation 4).”

    (10) In lines 435-437, the authors state: "Notably, the model predicted almost zero net change in total input to SSTs, consistent with the negligible change in SST activity observed experimentally (Fig. 2I) but inconsistent with the simple VIP->SST->PC disinhibitory model". The authors should discuss how temporal patterns of activation, depending on nicotinic versus metabotropic receptors, could explain these results. E.g., SSTs inhibit VIPs, and the dynamic of this loop would only be revealed in the ms timescale.

    We thank the reviewer for this suggestion. Our revised analysis indicates that net input to SSTs decreases during locomotion, consistent with increased VIP inhibition, although average SST activity changes little. As the reviewer suggests, rapid nicotinic activation of VIPs could produce a brief SST suppression that is not resolved by our calcium recordings, while slower metabotropic or recurrent effects may subsequently compensate for it. Averaging across the stimulus could therefore obscure these dynamics. We have added this possibility and the temporal limitations of our model to the Discussion (at line ~823).

    (11) Based on the results shown in Figure 6F, the authors conclude that sensitizers are dominantly inhibited by SSTs and depressors by PVs as previously reported. However, Figure 6F only shows differences during SST inhibition and PV activation but not the rest of the manipulations. How do the authors interpret this? How does this relate to previous reports?

    We agree that our original conclusion was too strong. Significant differences between sensitizers and depressors were observed during SST inhibition and PV activation, whereas SST activation and PV inhibition showed tendencies in the predicted direction but were more variable. We therefore interpret these results as qualitative support, rather than confirmation, that sensitizers are more strongly influenced by SSTs and depressors by PVs, and have revised the text accordingly. The previous report referred to Heintz et al. (2022), which showed that SST and PV interneurons promote sensitizing and depressing adaptation, respectively, leading to the proposal that PC adaptation depends on the balance between these inhibitory inputs. Our present results extend this model by suggesting that sensitizing and depressing PC subpopulations differ in their relative sensitivity to SST and PV manipulation. We have now toned down this statement (at line ~638).

    (12) Figure 6G shows a preferential reduction of the dominant source of inhibition (SSTs for sensitizers, PV for depressors). Can this be explained by a regression-to-the-mean kind of phenomenon? The higher the inhibition from one source is, the more it can be reduced.

    The shifts are not attributable to regression to the mean across model solutions, because no solutions were selected according to extreme weight values and the complete distributions, rather than only their extremes, shifted between conditions. We have added Supplementary Fig. 5 to show this. We agree, however, that a stronger inhibitory input has greater scope for reduction, such that a broad weakening of PV and SST transmission can produce a larger effect in the PC population that depends more strongly on that input. Differences in the preexisting PV balance make depressors more sensitive to reduced PV inhibition and sensitizers more sensitive to reduced SST inhibition. Therefore, the baseline connection strength likely contributes to the magnitude of the reduction. We have clarified this possibility in the Discussion section (at line ~867: “More generally, a broad weakening of inhibitory transmission would have a greater functional impact on PCs receiving stronger baseline input from that pathway, without requiring selective modulation of those connections”).

    (13) Discussion lines 500-502 should be moved to the beginning of the results section, when describing the two populations of excitatory neurons with different response patterns.

    The first paragraph now ends “Previous work has shown that receptive field location and orientation tuning are not significantly different between depressing and sensitizing PCs (Heintz et al., 2022)”.

    (14) Figure 1: overlap between sensitisers, intermediate and depressors: choose a bin size that matches the cut-off for the different categories? Do neurons switch from rest to locomotion?

    Figure 1C shows the same matched population of neurons, but ranks their AI values separately within each state to compare the state-specific distributions of adaptive properties. We have now adjusted the bins for each category as suggested. We have also performed the same binning for spike inferred traces (Supplementary Fig. 3A). The paired comparison in Supplementary Fig. 3A directly shows how individual AI values change between rest and locomotion. In the remaining Figure 1 analyses assessing locomotion-related gain differences, PCs are classified according to their AI during locomotion, and the same cell identities and classifications are retained when analysing their responses at rest.

    (15) Row 405: "25-50% reduction", but in Figure 5G, it is between 22% to 51%.

    We have now changed the values to accurately match the new Figure 6 (18-60%).

    (16) Statistics for Figure 6F and the number of cells accounted for were not well described.

    This has now been reported in the legend of Figure 7F.

    (17) Figure 7 is a bit misleading to read. The figure shows a dramatic difference between Depressors at rest and in locomotion, but the actual results did not show such a big change in PV inputs.

    We thank the reviewer for pointing out that the schematic could give the impression that line thickness represents the magnitude of the inferred changes quantitatively. Figure 7 (now 9) is intended as a qualitative summary of the main circuit differences rather than a scaled representation of connection weights. We have now clarified this explicitly in the figure legend. The quantitative changes in PV and SST inputs are shown in Fig. 6.

    Reviewer #2 (Recommendations for the authors):

    (1) Lines 68-69, "Contrary to the prevailing model ... we find that visual responses in SSTs are not inhibited by locomotion but instead weakly enhanced." That was clearly established in Dipoppa et al. (2018) and probably in other papers as well.

    We agree that “prevailing model” overstated the contrast with previous literature. We have therefore revised the sentence to refer more specifically to “a simple model of gain control through the VIP→SST→PC pathway”.

    (2) You call the change in response to locomotion an increase in "gain". Gain usually refers to a scaling, i.e. a multiplicative change, whereas from your Figure 1D-F it looks far more additive, i.e. a roughly constant factor added to the time-varying response. It would be clearer to, for example, just talk about response increases rather than gain increases.

    A related problem is that you use various language to describe responses -- the 'amplitude', which in line 93 is defined as the integral over 10 sec of the dF/F response, but is used in different ways, sometimes without any definition (for example in Figure 6F), throughout the text; you show both dF/F response, and the inferred spike rate response, which are qualitatively different in some ways, without commenting on these differences, or telling us what we should really believe. Re differences: the numbers in lines 100-101 (for dF/F) are different from those in 164-166 (for spikes), particularly for depressing E cells, but the numbers are just presented without any comparison or reference to one another; Figures 2D-E (for spikes) are pretty different from Figures 1D,F (for dF/F), but again, no comment on this, no comparison, no reference from one to the other. While it's good to see the raw dF/F as well as the inferred spikes, it would help to note the differences and say what you believe best represents the biology, e.g. the spikes, and stick to that thereafter (unless there is some specific reason to show the other, in which case you should explain that reason).

    We agree that “gain” was used too broadly in the manuscript. We have therefore replaced “gain increase” with “response increase” or “response amplification” where referring to our experimental measurements. We have also clearly distinguished between amplitude and integral, while using “response” as a general term where applicable and clearly defining the specific measure used beforehand.

    We agree that the rationale for presenting both calcium responses and inferred spike rates was not clear enough and have corrected this. Inferred spike rates provide a better representation of the temporal dynamics of electrical activity, particularly the initial fast component of adaptation that is obscured by the kinetics of GCaMP6f. We therefore use inferred spikes when analysing these temporal dynamics. However, spike inference results in quantized responses and consequently distorts the distribution of AI values towards the extremes (Supplementary Fig. 3A). We therefore used the directly measured calcium signals for population-level analyses of adaptive properties and their dependence on locomotion speed (Fig. 1 and 2). Importantly, both measures support the same central conclusion that locomotion preferentially increases responses in sensitizing compared with depressing PCs (Supplementary Fig. 3B and D).

    (3) Lines 95-97: "we split the population into three groups of equal size" according to adaptive index (AI), i.e. into tertiles; but then "These groups were sensitizing (AI less than zero), intermediate and depressing (AI more than zero)" -- the parenthetical AI<0 and AI>0 are wrong, and confusing, make it sound like the division is by sign and there is no intermediate group. This problem also shows up in Figure 2A legend, where you say "averaged across all sensitizing PCs" when it's actually averaged over the most sensitizing tertile.

    This has now been corrected. We have changed the definition of the three groups so that sensitizing cells clearly means the most sensitizing PCs with the most negative AI. We have adjusted the legend to Fig. 3 accordingly.

    Interneurons have more consistent adaptive properties so we kept the distinction between sensitizers and depressors based simply on the sign of their adaptive index. We have explained this distinction accordingly. Line ~284: “Unlike PCs, interneurons were not divided into tertiles but were classified as sensitizing or depressing according to whether their AI was negative or positive, respectively.”

    Finally, we have removed this calculation to avoid confusion: “The probabilities of a PC being sensitizing (AI <0, Psens) or depressing (AI > 0, Pdep) were both ~0.5”.

    (4) Equation 1 and thereafter, your notation is very awkward and non-standard. It is standard to call the connection from j to i w_ij, which makes sense since the entry of the weight matrix that multiples j's rate to determine the input to i is the matrix entry labeled w_ij, i.e. the total input to neuron i is \sum_j w_ijr_j. But you label it w_ji. You're entitled to your own notation if you insist, but it is very non-standard and so confusing. Similarly, the expression w_inIn is awkward, "In" looks like I times n, and it doesn't indicate there are multiple sources of input. You have three different sources of input, so why not write this term as \sum_i w^in_i I_i where i runs from 1 to 3 and I_i (or h_i or some other letter of your choice) represents the activity of the i^th external input?

    We have now made this adjustment in line with standard matrix notation. Subscripts follow the convention w_ij, with the postsynaptic population listed first and the presynaptic population second (e.g. w_(PC,SST) denotes the SST→PC connection). We have also revised the notation for external inputs, which are now explicitly represented as a sum over , where hk is the activity of external input k and its connection weight onto postsynaptic population i.

    (5) On line 217 you describe C_ji as the total number of synaptic connections from cells of type j to those of type i, but just a little ways down in Equation 3 you define it as N_j p_ji which is the expected number of ‘connections’ from ‘responsive’ cells of type j to cells of type i. Please define it correctly to begin with; this is very confusing.

    We thank the reviewer for identifying this ambiguity. We have now defined Cij at its first appearance as the number of connections from responsive presynaptic cells and explicitly stated that Cij = Njpij, where Nj is the number of responsive presynaptic cells and pij the connection probability.

    (6) In lines 217-218 you write "average strength of an individual synapse between neuron type j and i", 'between' should be 'from'.

    This has now been changed.

    (7) Lines 225-227, for "a PC and a nearby PV interneuron the probability of them being connected is ..." -- you need to specify the direction, PC to PV or PV to PC. "... while the average probability of an SST interneuron is ..." -- probably of what -- of the SST neuron getting an input from a nearby PC? Projecting to it? Something else?

    This has now been corrected to specify the direction of each connection from PV and SST to PCs.

    (8) Figure 3A: You should explain some of your model assumptions. For example, why do SS inputs not go to SST cells? You say (line 276) these include inputs from higher visual areas (HVAs), and SST cells definitely get inputs from HVAs (Shen, ..., & Tolias, 2022). Why no PV- >SST connections (these are seen with comparable strengths to other connections in some studies, not in others). Figures 3C, D: Are these locomoting or at rest? Figure 3E, why is the SS input modeled on VIP response in locomotion, rather than at rest, or the difference between locomotion and rest, or something else?

    We thank the reviewer for pointing out that these model assumptions were not sufficiently explained. The SS input was introduced to reproduce the strong sensitizing dynamics observed in VIPs during locomotion, which were greatly reduced at rest. We did not include a direct SS→SST connection because optogenetic silencing of VIPs largely abolished slow adaptation in SSTs indicating that a separate direct SS input to SSTs was not required to reproduce these dynamics (Figs. 4C-F). This does not imply that SSTs do not receive inputs from HVAs; accordingly, the model includes a separate FB input to SSTs, we have clarified that the anatomical origin of the SS signal remains uncertain (At line ~406). We have also clarified that traces in Figs. 4C-F correspond to the locomotion state.

    We excluded PV→SST connection because it was reported to be very weak in mouse V1 (Pfeffer et al., 2013), although we acknowledge that other studies have reported stronger connectivity (Jiang et al., 2015). We have now clarified that its exclusion represents a simplifying assumption of the model (at line ~423).

    (9) Model: Please tell us in the main text how many solutions you averaged over for rest and for locomotion. From lines 328-329, it sounds like there were about 990 for locomotion? What about at rest? And tell us if the fits for rest and for locomotion are done completely independently or have some dependence - you may say that in the Methods and I have missed it, but in any case, it would be nice to have that info in the main text. As I mentioned above, in addition to the mean weights, it would be nice to show the first few PC's of the weights, if they capture most of the variance across models.

    The final sets comprised 1007 solutions for locomotion and 11657 solutions for rest, fitted independently to the average visual responses of the four cell types. We have now clarified this in the main text.

    We also performed PCA on the fitted weights; however, the first PCs did not provide a low dimensional representation of the solution space, with PC1 explaining ~30% of the variance and the first three PCs together only ~50%. We therefore did not consider visualisation of the first few PCs particularly informative.

    (10) Figure 6: G, H are relative change in locomotion vs. rest; but I says "Change in total input during presentation of the stimulus". This is ambiguous. Is this the change with locomotion in the presence of a stimulus - or the change in response induced by presenting a stimulus, perhaps at rest? Is this any different from the conditions being compared in G and H - are they without a stimulus? If they're all comparing the same two things, please use the same description for them. If 6I is different in what it's comparing than G & H, explain so and explain why. More generally, please make what you're comparing clear and unambiguous.

    We apologise for the ambiguity. Panels G and H show relative changes between locomotion and rest, whereas panel I shows the absolute change (locomotion − rest), as absolute values are required to sum the individual inputs and calculate the net change in total synaptic input. We have revised the figure legend to make this distinction explicit. We have also removed the reference to stimulus presentation, as that is valid for all the results presented in the paper.

    (11) But, if Figure 6I is the change with locomotion vs rest: then how is the change in net input being similar, but even larger, for depressing E-cells vs sensitising ones, consistent with the sensitizing ones showing a much larger increase in response with locomotion?? Please explain.

    Following dimensionality reduction and refitting, the net input now shows a slightly larger locomotion-associated increase in sensitizing than depressing PCs, although this difference is not significant relative to the variability across solutions (Unorm = 0.4). We have clarified this point and discuss a potential explanation for the remaining discrepancy: our rate model treats inhibition as subtractive, whereas SST inhibition can act divisively. The model may therefore underestimate the effect that a reduction in SST inhibition has on PC responses, particularly in sensitizing PCs.

    (12) Also, here and elsewhere, you give RMSE's. But this is the square root of the sum of the squared differences of all parameters; the number doesn't have any intrinsic meaning, for example, since it depends on how many parameters there are, and weights all differences equally, no matter how big the relative difference and no matter how big the standard error. So the number tells us nothing, except that it's bigger in some cases and smaller in others. Why not use numbers that have some meaning, for example, the Chi-squared reduced, which divides by N, roughly, and also by the standard error; perhaps even better, divide by the data value to get the relative error, and again, per parameter, i.e. divided by N.

    We agree that RMSE does not account for the uncertainty of individual data points and therefore should not be interpreted as a statistical goodness-of-fit measure. We used reduced χ² for model fitting and selection, where weighting by experimental uncertainty is appropriate. As already defined in the Methods, RMSE was additionally used as a descriptive measure of absolute fit error independent of data uncertainty and the number of varied parameters, but dependent on the number of data points N. This allowed comparison across analyses with different uncertainty structures. We have clarified this distinction in the revised text (At line ~1163).

    (11) Line 490-491: Locomotion has "minimal effects on depressing PCs (Figures 1 and 2) - well, that's true for spikes (Figure 2), but not for dF/F (Figure 1) - say this more precisely.

    We have toned down this line to say “smaller effects on depressing PCs”.

    (12) In Figure 7, Summary, it would be helpful to show the external input to each cell type.

    We have now included in the summary figure the external input connections that showed meaningful changes in the model, while omitting unchanged inputs for simplicity.

    (13) Lines 573-576 assert that given low numbers of PV and SST cells and low probabilities of their connection to PCs, simple random connectivity can give rise to the distribution of PV:SST input ratios on PCs inferred from the model - this is a hypothesis, and it would be simple to test it in simulations.

    Thank you very much for this prompt, which has triggered a major addition to the paper. We find that variations in PV:SST input ratio predicted by the model are largely reproduced by simulations in which PV and SST connections are assigned randomly according to experimentally measured cell densities, connection probabilities and synaptic strengths. We now include this point in the Abstract and Introduction and evidence it in a new section of the Results, “Random local connectivity as a driver of functional heterogeneity”, and the accompanying Fig. 8. This random connectivity rule provides a simple and parsimonious explanation for variations in PV and SST inputs, although in the Discussion we point out that we cannot rule out the possibility that more specific local connectivity rules also contribute to variations in the PV:SST input ratios that cause variations in adaptation across the PC population.

    (14) Line 579: "the number of" should be omitted.

    Done.

    (15) Methods: Line 645, what was the spatial frequency of the gratings? Lines 659-660: How many/what percentage of cells of each type were excluded? Line 696, omit 'linear'. Fitting the model, lines 763 and following: you are fitting the average time traces - how many & which time points are you fitting? As you mention somewhere, most of the average traces show some signs of the 1Hz grating frequency. Are the models penalized for not fitting these wiggles, or do you use some smoothed version of the traces, or sample time in a way to avoid the 1Hz oscillations? Line 776, give the values of N_v and N. Lines 772 and 779-780, repeat the same information about standard errors. Line 836, omit 'in'.

    We have now added information about spatial and temporal frequency (line ~996), the percentage of excluded non-responsive cells (line ~1019) and the time points being fitted (line ~1164).

    As described in the Methods, the 1 Hz oscillations are prominent in individual PC responses but are strongly reduced in the population averages because cells with receptive fields at different locations respond at different phases of the grating. Modulation at 1 Hz remains visible in average response of the SST population because these have larger receptive fields and phase differences are smaller. We therefore fitted the population-average traces without explicitly modelling the 1 Hz component. We have also added the values of N and Nv for the unrestricted four-population fit and made the rest of corrections pointed out by the reviewer in this question.

    (16) Supplementary Figure 4: I can't figure out what these sentences are saying "... and their corresponding fits from the solution with lower (RMSE=0.14) calculated by a model without slow-sensitising (SS) input" and "Same as in A for missing feedback input (FB) (RMSE=0.14)". A solution with lower what? Does the RMSE=0.14 refer to the model before removing the given input, or after? What's the significance of RMSE=0.14? Was that the lowest RMSE across models before omitting inputs, or does it have some other significance?

    RMSE = 0.14 refers to the best-fitting solution obtained after removal of each respective input. We have revised the legend to make this clearer.

    Reviewer #3 (Recommendations for the authors):

    (1) Supplementary Figure 1 - 'AI Rest' is compared to 'Delta AI'. It would be useful to see 'AI Rest' versus AI Locomotion'. In addition, how are these responses related to the baseline rates of the neurons? It would be good to plot 'Delta AI' versus 'Delta mean-rate' (i.e. difference in mean rates between locomotion and stationary conditions).

    As also requested in the public review, we have now added per-neuron comparisons of AI during rest and locomotion for PCs, VIP, SST and PV cells (Supplementary Fig. 2A, G–I). AI during rest and locomotion was significantly correlated in all cell types except SST cells, indicating that adaptation profiles were partially conserved across behavioural states.

    We also examined whether locomotion-related changes in response magnitude were associated with changes in AI (Supplementary Fig. 2C, J–L). We plotted ΔAI against the change in response integral during the stimulus, which is directly proportional to the change in mean response rate requested by the reviewer. In PCs, the change in response was negatively correlated with ΔAI, indicating that cells that became more sensitizing during locomotion showed greater increases in response. Among interneurons, PV cells showed a similar relationship, whereas no significant relationship was observed in VIP or SST cells.

    Finally, potential differences in baseline activity are addressed in Supplementary Fig. 2E, F, where baseline activity was similar between rest and locomotion.

    (2) 'Sensitize' is a term that is not well-established and hence would be worth defining early on.

    We agree this term is not as well stablished as depression. We have now added the definition of sensitizing adaptation early in the introduction of the paper. Later on, we explain in more detail the dynamics of sensitizing PCs in contrast to depressing PCs at line ~119.

    References

    Bellardita, C., & Kiehn, O. (2015). Phenotypic characterization of speed-associated gait changes in mice reveals modular organization of locomotor networks. Curr Biol, 25(11), 1426–1436. https://doi.org/10.1016/j.cub.2015.04.005

    Chance, F. S., Nelson, S. B., & Abbott, L. F. (1998). Synaptic Depression and the Temporal Response Characteristics of V1 Cells. The Journal of Neuroscience, 18(12), 4785–4799. https://doi.org/10.1523/JNEUROSCI.18-12-04785.1998

    Deneux, T., Kaszas, A., Szalay, G., Katona, G., Lakner, T., Grinvald, A., Rózsa, B., & Vanzetta, I. (2016). Accurate spike estimation from noisy calcium signals for ultrafast three-dimensional imaging of large neuronal populations in vivo. Nature Communications, 7(1), 12190. https://doi.org/10.1038/ncomms12190

    Dipoppa, M., Ranson, A., Krumin, M., Pachitariu, M., Carandini, M., & Harris, K. D. (2018). Vision and Locomotion Shape the Interactions between Neuron Types in Mouse Visual Cortex. Neuron, 98(3), 602–615.e608. https://doi.org/10.1016/j.neuron.2018.03.037

    Fairhall, A. L., Lewen, G. D., Bialek, W., & de Ruyter Van Steveninck, R. R. (2001). Efficiency and ambiguity in an adaptive neural code. Nature, 412(6849), 787–792. https://doi.org/10.1038/35090500

    Heintz, T. G., Hinojosa, A. J., Dominiak, S. E., & Lagnado, L. (2022). Opposite forms of adaptation in mouse visual cortex are controlled by distinct inhibitory microcircuits. Nature Communications, 13(1), 1031. https://doi.org/10.1038/s41467-022-28635-8

    Hinojosa, A. J., Dominiak, S. E., Kosiachkin, Y., & Lagnado, L. (2026). Distinct Disinhibitory Circuits Link Short-Term Adaptation to Familiarity and Reward Learning in Visual Cortex. bioRxiv, 2026.2003.2024.713929. https://doi.org/10.64898/2026.03.24.713929

    Jiang, X. L., Shen, S., Cadwell, C. R., Berens, P., Sinz, F., Ecker, A. S., Patel, S., & Tolias, A. S. (2015). Principles of connectivity among morphologically defined cell types in adult neocortex. Science, 350(6264). https://doi.org/10.1126/science.aac9462

    Jin, M., & Glickfeld, L. L. (2020). Magnitude, time course, and specificity of rapid adaptation across mouse visual areas. J Neurophysiol, 124(1), 245–258. https://doi.org/10.1152/jn.00758.2019

    Keller, G. B., Bonhoeffer, T., & Hubener, M. (2012). Sensorimotor mismatch signals in primary visual cortex of the behaving mouse. Neuron, 74(5), 809–815. https://doi.org/10.1016/j.neuron.2012.03.040

    Kirchberger, L., Mukherjee, S., Self, M. W., & Roelfsema, P. R. (2023). Contextual drive of neuronal responses in mouse V1 in the absence of feedforward input. Sci Adv, 9(3), eadd2498. https://doi.org/10.1126/sciadv.add2498

    Pfeffer, C. K., Xue, M., He, M., Huang, Z. J., & Scanziani, M. (2013). Inhibition of inhibition in visual cortex: the logic of connections between molecularly distinct interneurons. Nature Neuroscience, 16(8), 1068–1076. https://doi.org/10.1038/nn.3446

    Rupprecht, P., Carta, S., Hoffmann, A., Echizen, M., Blot, A., Kwan, A. C., Dan, Y., Hofer, S. B., Kitamura, K., Helmchen, F., & Friedrich, R. W. (2021). A database and deep learning toolbox for noise-optimized, generalized spike inference from calcium imaging. Nat Neurosci, 24(9), 1324– 1337. https://doi.org/10.1038/s41593-021-00895-5

    Saleem, A. B., Ayaz, A., Jeffery, K. J., Harris, K. D., & Carandini, M. (2013). Integration of visual motion and locomotion in mouse visual cortex. Nat Neurosci, 16(12), 1864–1869. https://doi.org/10.1038/nn.3567

    Varela, J. A., Sen, K., Gibson, J., Fost, J., Abbott, L. F., & Nelson, S. B. (1997). A Quantitative Description of Short-Term Plasticity at Excitatory Synapses in Layer 2/3 of Rat Primary Visual Cortex. The Journal of Neuroscience, 17(20), 7926–7940. https://doi.org/10.1523/JNEUROSCI.1720-07926.1997

    Vinck, M., Batista-Brito, R., Knoblich, U., & Cardin, J. A. (2015). Arousal and locomotion make distinct contributions to cortical activity patterns and visual encoding. Neuron, 86(3), 740–754. https://doi.org/10.1016/j.neuron.2015.03.028

  5. eLife Assessment

    This manuscript presents a valuable analysis of how locomotion modulates the activity of different subtypes of cortical neurons in the mouse primary visual cortex, showing that locomotion more strongly increases responses in sensitizing than in depressing excitatory cells. This data is then used to constrain a model of the responses. While the data are very interesting, the analyses remain incomplete, in particular due to concerns surrounding the modelling.

  6. Reviewer #1 (Public review):

    In this manuscript, Hinojosa and colleagues analysed the changes in V1 visual responses induced by locomotion in head-fixed mice using two-photon calcium imaging. The authors observe that locomotion strongly increases the visual responses of V1 excitatory neurons that exhibit sensitizing responses to visual stimuli. Also, there is an increased response in VIP interneurons, and to a lesser extent, PV interneurons and SST interneurons (non-significant). The authors used a model fitted with data presented in the manuscript, as well as previous knowledge on cortical connectivity among different neuron types. The model suggests that the major component of the increased responses during locomotion is an increase in excitatory drive from external inputs (feedforward, feedback and modulatory), most importantly onto VIP interneurons and excitatory neurons. However, the excitatory drive of local excitatory neurons onto other surrounding excitatory and inhibitory cells is reduced.

    The manuscript is well presented and represents a valuable analysis of how locomotion modulates the activity of different subtypes of cortical neurons. However, major issues should be addressed to strengthen the results.

    Major issues:

    (1) Speed and mismatch between locomotion and visual stimulation.

    The authors do not clearly describe the definition of locomotion versus the resting state. The speed should, by itself, have an impact on neuronal responses, especially at the onset of locomotion. Several published studies show that the mismatch between a visual stimulus and the speed of the animal induces specific responses in V1, both in excitatory and subtypes of inhibitory neurons. The authors should address these points upfront in the manuscript, since it is likely a major variable explaining their results

    (2) Use of deconvolution with MLSpike.

    Some results (Figure 2) exclusively depend on the deconvolution of calcium signals into spikes (since the initial peak is not seen in calcium transients). The authors should validate this result either with electrophysiological recordings or with the use of another deconvolution method (e.g. CASCADE), emphasising the limitations of this approach and the limitations of the time resolution of calcium imaging.

    (3) The manuscript is centred around a specific increase in visual responses in sensitizing neurons during locomotion, both in the fraction of responsive neurons and response magnitudes.

    It is hard to tell whether this difference is due to a greater scaling effect of locomotion, a difference in responses during the resting state, or both. The manuscript should further explore and discuss the differences in responses between sensitizing and depressing neurons, both during the resting state and locomotion. Adding metrics and direct comparisons of the magnitudes of fast responses, slow responses, and time integrals between sensitizing and depressing neurons in resting and locomotion states would help to clarify this. Same for fractions of responsive neurons of each type in each condition. E.g., the slow phase is harder to judge from the plots, but the DeltaF/F integral shown in Figure 1G seems to suggest the difference in response magnitude between sensitizing and depressing neurons is largest in locomotion state, rather than resting state. How do these integrals look for inferred firing rates shown in Figure 2?

    (4) There is something counterintuitive about how the changes in inhibition onto sensitizing and depressing neurons during locomotion explain the reported activity changes.

    Sensitizers receive reduced SST input and increased PV input during locomotion. If SSTs depress and PVs sensitize (and this is the main reason why sensitizers, which receive dominant input from SSTs sensitize, and vice-versa), how is it possible that this switch does not alter the sensitizing or depressing nature of these neurons' responses in locomotion? Are these changes insufficient to flip the dominant SST-PV drive? Figure 6D-E seems to show there is a flip, at least for sensitizers. How do authors explain this? Do authors think this is related to the narrowing of the adaptive index distribution shown in Figure 1C?

    (5) Presentation of the experimental data and the model.

    The manuscript introduces the results of interneuron recordings during the description of the model. Similarly, the results of optogenetic manipulations are presented inside the model's description. It would be clearer to present all experimental data first and introduce the model later, fitting it to all experimental evidence previously presented.

  7. Reviewer #2 (Public review):

    This is an interesting paper with important results. The authors, working in V1, have previously, in a 2022 paper, defined sensitizing and depressing excitatory (E) cells as those whose response increases or decreases, respectively, across the 10 seconds of showing a drifting grating stimulus. They showed that sensitizing E cells are dominantly inhibited by SST inhibitory cells, which are dominantly depressing, and that depressing E cells are dominantly inhibited by PV inhibitory cells, which are very largely sensitizing. It's been well established that locomotion greatly increases E-cell firing rates in V1 compared to rest, but much remains to be worked out as to the mechanism. Here, they find that locomotion increases the responses of the sensitizing E cells much more than depressing cells. They develop a model of changes in synaptic weights between rest and locomotion to account for the changes. One reason that sensitizers are increased more by locomotion than depressors is that PV cells, which more strongly inhibit depressors, have increased firing for locomotion, whereas SST cells, which more strongly inhibit sensitizers, don't change their firing rates with locomotion. However, in the mode,l a complex array of postulated changes in connection strengths is also involved.

    I have, though, a number of concerns: with the model, with the lack of proper discussion of connection to some previous works, and with an overall unclear and confusing presentation and certain controls that should be done.

    In the model, they postulate that synapses within the 6-cell-type network - sensitizing, intermediate, and depressing E cells, and PV, SST, and VIP I cells - and from three sources of external input to each of the six types all change between rest and locomotion (except that connections between the E cells don't depend on their types). There are a lot of degrees of freedom, and this makes interpretation of the results difficult. I would have liked to have seen more efforts to constrain the degrees of freedom. For example, there seems to be very little difference between the three E cell types in any of the three types of external input received. Why not constrain them all to get the same external input and see if it significantly affects model fit? Or what if synapses from the three types of external input are left unchanged, and only change their strengths between rest and locomotion? How well could this do? During optimization, why not constrain the changes between rest and locomotion, for example, by putting an L1 penalty on the changes or the relative changes, trying to force them to be sparse, and see whether there are roughly equally good fits? And then, if the main changes are in a small set of synapses, can the authors isolate changes to that small set and do roughly equally well? What about looking at the principal components of the weight changes across models, to isolate patterns of change that are most important?

    In terms of comparing to previous works, when optogenetic manipulations of SST and PV are done to test various hypotheses, I would like to see some discussion of what is already known from the authors' 2022 paper and what they are adding or testing that wasn't known or tested from that paper. And Dipoppa et al (2018) also found weight changes to account for the difference between rest and locomotion. They were looking at a fixed point of responses of neurons across retinotopic space to stimuli of various sizes with only one E-cell type, whereas they are accounting for trajectories across time considering 3 E-cell subtypes but without variation in stimuli or retinotopic position of neurons, so the efforts are somewhat different, but still, it would be good to see a bit more discussion of what is in agreement or in contradiction in the conclusions.

    In terms of presentation and controls, I have many concerns, which include:

    (1) The main result is that sensitizers increase their responses with locomotion ~2X (for dF/F) or about 3.5X (for spikes) more than depressors. But there are other differences between sensitizers and depressors, for example sensitizers have smaller initial stimulus responses at rest, and depressors have larger. What if cells were divided into tertiles by initial stimulus response at rest? Would the authors see the same differences in the effects of locomotion? If so, can they establish whether the difference is really attached to the adaptation properties rather than to, for example, the initial responses, for example, by comparing the regression of response increase against AI vs the regression of response increase against initial resting response? And there might be other controls to be done for other features in which sensitizers and depressors differ.

    (2) Lines 103 and following: the authors refer to a "second notable change" which is the narrower distribution of adaptive effects, but I think this is trivial. The adaptive index is AI=(R1-R2)/(R1+R2), where R1 is response 0.5-2.5s after stimulus onset and R2 over 8-10s. But if the change is additive, as suggested by the dF/F figures (and I believe the distributions of AI here are based on dF/F measurements) -- adding the same constant to R1 and R2 will shrink |AI| without changing the sign of AI. So this would seem to just be a signature of a change that is primarily additive rather than multiplicative.

    Also, if the authors do decide that they are going to focus on spikes after showing the raw dF/F, then this analysis should be repeated for spikes.

    (3) Figure 2, F is supposed to be D minus E, but it doesn't look like it. For example, the initial response under locomotion is very similar in sensitizers and depressors, so the initial difference in F should be small, but it's not; and at rest, depressors initially have larger responses than sensitizers, whereas later depressors have smaller responses than sensitizers, yet the difference at rest is positive at all times. Something seems wrong here.

  8. Reviewer #3 (Public review):

    This study aimed to understand the depressing and sensitizing effects of adaptation in mice visual cortex during different behavioral states: locomotion and stationary. There is an impressive characterisation of the responses in different cortical cell types and with different optogenetic manipulations to the inhibitory populations. These form a very interesting dataset to understand the effects of the state on the circuits and gain insight into the mechanisms. This data is then used to constrain a model of the responses. Unfortunately, the model appears to be too flexible, and it was difficult to interpret the insights gained from the different model fits.

    Strengths:

    The data is impressive. There is a characterisation of responses of PCs and VIP, SST and PV interneurons. Additionally, there is the characterisation of some responses to specific optogenetic manipulations, VIP inactivation, SST or PV activation or inactivation. These data will help develop a good insight into the system. The principle of using the optigenetic manipulations to constrain model parameters is very interesting.

    Weaknesses:

    Many of the analyses have some concerns in the methodology used, which we list in detail below. Further, the model used to gain insight into the mechanism appears overly complicated and seems hard to gain clear insights from.

    Major concerns:

    (1) Key concern is the usage of dF/F signals for all analyses, especially when comparing responses.

    1a) Figure 1G: Comparison of sensitisers and depressors. It is important to consider what the baseline rates are when making these comparisons, especially when comparing the degree of effects between different cell types. For example, if baseline rates for sensitizers were overall higher, it would mean the difference in gain of response would be lower, and could affect the results in the opposite direction of what is claimed. One option to account for this would be to z-score the overall responses, using the same normalization for locomotion and rest. We also suggest plotting differences in sensitisers, intermediates, and depressors as a function of firing rate. Matching for firing rate across each PC categorization and calculating delta AI for each matched firing rate bin.

    1b) Figure 2A-F: The above is an even more significant issue when it comes to estimating spiking rates. The methods do not state how dF/F is calculated. If these are based on using the pre-stim as the reference, the algorithms for spike rate used might not be appropriate if this were used. Using pre-stimulus referencing could result in the estimate going into the wrong range in the calculation of the spike rate.

    1c) In both cases above, it could be a problem if baseline firing rates are different between cell types, or states (locomotion/stationary). The latter is established to have effects on many cell types measured, and so needs to be accounted for very carefully.

    1d) It would be informative to see per-neuron comparison for adaptive indices during rest and locomotion states. This could be visualized using a scatter plot with AI-rest vs. AI-locomotion for Figures 1D- 1F and 2J- 2L.

    1e) Are neurons more strongly modulated between locomotion and rest, also more likely to experience a shift in AI indices (i.e. delta AI). Is there a correlation between the change in firing rate between behavioral states and Delta AI (Loco-Rest)? If so, is this present for all neuron subtypes (e.g. VIP, SST, and PV)?

    1f) Optogenetic inhibition of VIP neurons on average abolished the slow depressive effects of adaptation in SST (Figure 3). The strength and prevalence of this effect are unclear. Perhaps one can perform a bootstrap control and opto AI indices and calculate whether AI was significantly reduced following optogenetics inhibition, and if so, on average, how likely was this to occur for the recorded SST neurons? This is important in knowing that the average effects (Figure 3D) aren't driven by a portion of SST neurons, especially as this is later used to confirm the region of parameter space and affects the subsequent results in Figure 4.

    (2) Statistics for the effects. There is a mention of Liner mixed models, but no information is given on the actual models being used and tested. This is particularly for the case of Figure 1G, where there is a composition of effect sizes between different populations. What precise significance test is being used? Are the stats on paired cells when considering locomotion and rest?

    (3) Model parameters: It is acknowledged that there is a large range of parameters that can model the responses effectively, up to 11% of initial conditions. At 9000 initial conditions, this is around 1000. The parameter estimates are then considered as the mean of each parameter. This seems like a strange choice for a few different reasons:

    3a) A mean solution might not be one of the solutions. Let's say the parameters range over a large dimensional space. They could occupy non-overlapping / discontinuous subspaces. In that case, the mean parameters do not necessarily fall within the solution subspaces. Therefore, this reduction to means might not be valid.

    3b) Compare distributions rather than means. There are multiple distributions of parameters between conditions. All stats should be on the comparison of distributions rather than just the means.

    (4) Visualizing weight matrices: It is very challenging to interpret the weight matrices. Furthermore, it appears that the stationary and locomotion conditions fit independently, and given the large parameter spaces, it is even harder to interpret. Can the fitting instead be done by fitting on one and using those at the initial conditions for the other state? Figure 7 shows an initiative cartoon, but it is not clear how the matrices in Figures 5 and 6 lead to the summary shown in Figure 7. It is also not clear why the connections between inhibitory neurons are not shown in Figure 7. One option is to perhaps run some kind of dimensionality deduction on the parameter space to better interpret the data. When showing deltaWeights, was the model initialised with 'Rest' weights and allowed to change? It is not obvious what the difference is between 'relative change in connection weights' and 'relative change in synaptic weights'.This needs to be clarified.

    4a) Model parameters were reduced differently for locomotion and rest (Figure 4). We suggest evaluating the results for locomotion and rest using the same chi-square value of 3 for both behavioral states (at least in controls).

  9. Author response:

    Public Reviews:

    Reviewer #1 (Public review):

    (1) Speed and mismatch between locomotion and visual stimulation.

    The authors do not clearly describe the definition of locomotion versus the resting state. The speed should, by itself, have an impact on neuronal responses, especially at the onset of locomotion. Several published studies show that the mismatch between a visual stimulus and the speed of the animal induces specific responses in V1, both in excitatory and subtypes of inhibitory neurons. The authors should address these points upfront in the manuscript, since it is likely a major variable explaining their results.

    We will clarify in the methods that a trial was considered as locomotion when an animal ran at a minimum of 3 cm/s for at least 80% of the 10 s stimulus presentation, and was considered rest when running under 3 cm/s during the same fraction of time. Trials with abrupt changes from locomotion to rest were rare and excluded following these criteria.

    Locomotion speed and visuomotor mismatch can influence neuronal responses in V1 but in the large majority of our trials mice either run continuously at a stable speed or remained still

    i.e locomotion onsets or offsets did not occur (see Hinojosa et al. 2026 for example running traces). Furthermore, sensitizing and depressing neurons were typically recorded simultaneously within the same field of view, experiencing identical locomotor behaviour. For these reasons, we think it is unlikely that differences in speed or mismatch alone can account for the different increase in amplitude observed between depressors and sensitizers.

    To directly address this point and further explore the role of speed on V1 neurons, we will quantify the relationship between running speed and amplitude increase in both PCs and interneurons, and include these analyses in the revised version of the manuscript.

    (2) Use of deconvolution with MLSpike.

    Some results (Figure 2) exclusively depend on the deconvolution of calcium signals into spikes (since the initial peak is not seen in calcium transients). The authors should validate this result either with electrophysiological recordings or with the use of another deconvolution method (e.g CASCADE), emphasising the limitations of this approach and the limitations of the time resolution of calcium imaging.

    A similar initial increase in amplitude followed by fast depression has been observed previously with electrophysiological recordings in V1 (Chance et al., 1998; Jin & Glickfeld, 2020; Varela et al., 1997). We will further validate our results using an alternative spike inference method like CASCADE (Rupprecht et al., 2021), as well as expanding on the limitations of our approach.

    (3) The manuscript is centred around a specific increase in visual responses in sensitizing neurons during locomotion, both in the fraction of responsive neurons and response magnitudes.

    It is hard to tell whether this difference is due to a greater scaling effect of locomotion, a difference in responses during the resting state, or both. The manuscript should further explore and discuss the differences in responses between sensitizing and depressing neurons, both during the resting state and locomotion. Adding metrics and direct comparisons of the magnitudes of fast responses, slow responses, and time integrals between sensitizing and depressing neurons in resting and locomotion states would help to clarify this. Same for fractions of responsive neurons of each type in each condition. E.g., the slow phase is harder to judge from the plots, but the DeltaF/F integral shown in Figure 1G seems to suggest the difference in response magnitude between sensitizing and depressing neurons is largest in locomotion state, rather than resting state. How do these integrals look for inferred firing rates shown in Figure 2?

    We will further explore the response dynamics of adaptive types within the locomotion and resting state, highlighting the differences between calcium signals and inferred spikes. We will then include our findings in the new version.

    (4) There is something counterintuitive about how the changes in inhibition onto sensitizing and depressing neurons during locomotion explain the reported activity changes.

    Sensitizers receive reduced SST input and increased PV input during locomotion. If SSTs depress and PVs sensitize (and this is the main reason why sensitizers, which receive dominant input from SSTs sensitize, and vice-versa), how is it possible that this switch does not alter the sensitizing or depressing nature of these neurons' responses in locomotion? Are these changes insufficient to flip the dominant SST-PV drive? Figure 6D-E seems to show there is a flip, at least for sensitizers. How do authors explain this? Do authors think this is related to the narrowing of the adaptive index distribution shown in Figure 1C?

    This result is only counterintuitive if we consider exclusively the internal connections within V1. The PV:SST ratio changes from 0.9 during rest, dominated by SST induced sensitization, to 1.2, dominated by PV depression. Although adaptation is strongly driven by the opposing inhibition of PV and SST in PCs during locomotion, its origin is more easily explained by an external input (SS) that targets VIPs, PVs and PCs. As a result, when locomotion increases the drive coming from SS input, it injects a source of sensitization that partly balances the decrease in PV:SST ratio, preventing a switch in their adaptive properties which, although reduced, remain sensitizing. We will include these calculations in the revised version.

    (5) Presentation of the experimental data and the model.

    The manuscript introduces the results of interneuron recordings during the description of the model. Similarly, the results of optogenetic manipulations are presented inside the model's description. It would be clearer to present all experimental data first and introduce the model later, fitting it to all experimental evidence previously presented.

    We understand that a clear separation between experimental and modelling results is often preferred in papers that combine these approaches but in our case modelling and experimental data are highly interdependent and we believe that an overlapping presentation make it easier for the reader to appreciate the links. One example is Fig. 2G-L that shows experimental results validating a key feature of the model - the use of average response dynamics for each population of interneuron. Similarly, the results in Fig. 3 validate the use of the VIP response dynamics as the template for the slow modulatory input to layer 2/3. Then the results of optogenetic experiments in Fig. 4 are used to narrow down fits to the model. For these reasons, we have chosen to present experimental results and the model in this more integrated manner.

    Reviewer #2 (Public review):

    In the model, they postulate that synapses within the 6-cell-type network - sensitizing, intermediate, and depressing E cells, and PV, SST, and VIP I cells - and from three sources of external input to each of the six types all change between rest and locomotion (except that connections between the E cells don't depend on their types). There are a lot of degrees of freedom, and this makes interpretation of the results difficult. I would have liked to have seen more efforts to constrain the degrees of freedom. For example, there seems to be very little difference between the three E cell types in any of the three types of external input received. Why not constrain them all to get the same external input and see if it significantly affects model fit? Or what if synapses from the three types of external input are left unchanged, and only change their strengths between rest and locomotion? How well could this do? During optimization, why not constrain the changes between rest and locomotion, for example, by putting an L1 penalty on the changes or the relative changes, trying to force them to be sparse, and see whether there are roughly equally good fits? And then, if the main changes are in a small set of synapses, can the authors isolate changes to that small set and do roughly equally well? What about looking at the principal components of the weight changes across models, to isolate patterns of change that are most important?

    To reduce the number of degrees of freedom and ease interpretation we did limit the model fitting for adaptive subtypes by fixing the PC-PC (𝑤𝑃𝐶_𝑃𝐶) and restricting the external inputs weights (𝑤𝐹𝐹_𝑃𝐶, 𝑤𝑆𝑆_𝑃𝐶, 𝑤𝐹𝐵_𝑃𝐶) to changes of ± 10 %. We will explicitly explain these constrains in the methods and discuss its limitations.

    We thank the reviewer for their suggestions of testing different conditions to find those providing the best fit for sensitizing and depressing PCs. We tried an approach similar to that described by Dipoppa et al. 2018 by using the locomotion weights as initial conditions for the rest traces and introducing penalties at later stages. However, the local optimization algorithms failed to reach distant regions of parameter space containing minimum solutions for the rest condition. We finally opted for repeating the same process of initial condition searching for locomotion and rest, making the L1 penalty approach impracticable in our case. We believe this approach is effective because it has both allowed us to describe circuit changes during internal-state transitions (the present paper) and, more recently, it has made a series of predictions about different learning states that have been confirmed by optogenetic tests (Hinojosa et al., 2026). We will nevertheless explore this and other of the reviewer suggestions to further optimize the fitting in the revised manuscript.

    In terms of comparing to previous works, when optogenetic manipulations of SST and PV are done to test various hypotheses, I would like to see some discussion of what is already known from the authors' 2022 paper and what they are adding or testing that wasn't known or tested from that paper. And Dipoppa et al (2018) also found weight changes to account for the difference between rest and locomotion. They were looking at a fixed point of responses of neurons across retinotopic space to stimuli of various sizes with only one E-cell type, whereas they are accounting for trajectories across time considering 3 E-cell subtypes but without variation in stimuli or retinotopic position of neurons, so the efforts are somewhat different, but still, it would be good to see a bit more discussion of what is in agreement or in contradiction in the conclusions.

    Thanks for this prompt. We will add further discussion of this work in light of the Heintz et al. (2022) and Dipoppa et al. (2018) papers.

    (1) The main result is that sensitizers increase their responses with locomotion ~2X (for dF/F) or about 3.5X (for spikes) more than depressors. But there are other differences between sensitizers and depressors, for example sensitizers have smaller initial stimulus responses at rest, and depressors have larger. What if cells were divided into tertiles by initial stimulus response at rest? Would the authors see the same differences in the effects of locomotion? If so, can they establish whether the difference is really attached to the adaptation properties rather than to, for example, the initial responses, for example, by comparing the regression of response increase against AI vs the regression of response increase against initial resting response? And there might be other controls to be done for other features in which sensitizers and depressors differ.

    We will explore the possibility that initial response influences the increase in amplitude. Preliminary data suggest that initial amplitude is higher in depressors than in sensitizers.

    (2) Lines 103 and following: the authors refer to a "second notable change" which is the narrower distribution of adaptive effects, but I think this is trivial. The adaptive index is AI=(R1-R2)/(R1+R2), where R1 is response 0.5-2.5s after stimulus onset and R2 over 8-10s. But if the change is additive, as suggested by the dF/F figures (and I believe the distributions of AI here are based on dF/F measurements) -- adding the same constant to R1 and R2 will shrink |AI| without changing the sign of AI. So this would seem to just be a signature of a change that is primarily additive rather than multiplicative.

    Also, if the authors do decide that they are going to focus on spikes after showing the raw dF/F, then this analysis should be repeated for spikes.

    We agree with the reviewer and will change the text accordingly to highlight the additive nature of the change in amplitude. We will also show the analysis with spikes (this shows similar results as the calcium data).

    (3) Figure 2, F is supposed to be D minus E, but it doesn't look like it. For example, the initial response under locomotion is very similar in sensitizers and depressors, so the initial difference in F should be small, but it's not; and at rest, depressors initially have larger responses than sensitizers, whereas later depressors have smaller responses than sensitizers, yet the difference at rest is positive at all times. Something seems wrong here.

    We apologize for the confusion this has caused. Figure 2F does not represent the difference between sensitizing and depressing PCs from panels D and E. Instead, it shows the time-varying difference between locomotion and rest states of sensitizers (blue, in figure 2D) and depressors (green, in figure 2E). Thus, panel F shows within-population modulation by behavioural state, rather than differences between sensitizing and depressing neurons. We will amend the figure legend and main text to explain this point and avoid misinterpretation.

    Reviewer #3 (Public review):

    (1) Key concern is the usage of dF/F signals for all analyses, especially when comparing responses.

    (1a) Figure 1G: Comparison of sensitisers and depressors. It is important to consider what the baseline rates are when making these comparisons, especially when comparing the degree of effects between different cell types. For example, if baseline rates for sensitizers were overall higher, it would mean the difference in gain of response would be lower, and could affect the results in the opposite direction of what is claimed. One option to account for this would be to z-score the overall responses, using the same normalization for locomotion and rest. We also suggest plotting differences in sensitisers, intermediates, and depressors as a function of firing rate. Matching for firing rate across each PC categorization and calculating delta AI for each matched firing rate bin.

    (1b) Figure 2A-F: The above is an even more significant issue when it comes to estimating spiking rates. The methods do not state how dF/F is calculated. If these are based on using the pre-stim as the reference, the algorithms for spike rate used might not be appropriate if this were used. Using pre-stimulus referencing could result in the estimate going into the wrong range in the calculation of the spike rate.

    (1c) In both cases above, it could be a problem if baseline firing rates are different between cell types, or states (locomotion/stationary). The latter is established to have effects on many cell types measured, and so needs to be account ted for very carefully.

    The DF/F0 trace was calculated using the mode of the whole trace as F0. While this approach is less sensitive to biases than subtracting the pre-stimulus, it does not consider noise levels like the z-score suggested by the reviewer. We will, therefore, normalize the calcium traces to z-score to further account for changes in the baseline. Spike inference using MLSpike, however, explicitly models baseline noise and subtracts its effect from that of the spikes calculated from the calcium signal (Deneux et al., 2016). This transformation preserved the difference in amplitude triggered by locomotion between depressing and sensitizing PCs while revealing their similar baseline activity (see Figs. 2D,E and F). These results indicate that the distinct changes in response amplitude between sensitizing and depressing PCs during locomotion are not driven by baseline differences. We will add this explanation to the methods section.

    We will also plot the changes in activity with locomotion across cell types as a function of firing rate and add these results to the revised manuscript.

    (1d) It would be informative to see per-neuron comparison for adaptive indices during rest and locomotion states. This could be visualized using a scatter plot with AI-rest vs. AI-locomotion for Figures 1D- 1F and 2J- 2L.

    (1e) Are neurons more strongly modulated between locomotion and rest, also more likely to experience a shift in AI indices (i.e. delta AI). Is there a correlation between the change in firing rate between behavioral states and Delta AI (Loco-Rest)? If so, is this present for all neuron subtypes (e.g. VIP, SST, and PV)?

    Sorting was carried out separately on locomotion and rest data sets to capture the adaptive properties of the network under each condition. When assessing the change in adaptive index in individual cells there was a weak but significant correlation (r = 0.10, p<0.05), probably due to trial to trial stochasticity in the network which has been shown to be present in V1 (Carandini, 2004; Lee et al., 2010). Although adaptation profiles of individual PCs are not fully conserved across rest and locomotion, the observed overlap exceeds that expected by chance, suggesting that stochastic fluctuations modulate an underlying, stable circuit organization. Despite including the stochastic component of the responses, the conclusions hold: sensitizers undergo a larger gain modulation than that of depressors. We will include this analysis and the correlation between change in firing rate and Delta AI in the revised version of the paper.

    (1f) Optogenetic inhibition of VIP neurons on average abolished the slow depressive effects of adaptation in SST (Figure 3). The strength and prevalence of this effect are unclear. Perhaps one can perform a bootstrap control and opto AI indices and calculate whether AI was significantly reduced following optogenetics inhibition, and if so, on average, how likely was this to occur for the recorded SST neurons? This is important in knowing that the average effects (Figure 3D) aren't driven by a portion of SST neurons, especially as this is later used to confirm the region of parameter space and affects the subsequent results in Figure 4.

    The strength and prevalence of the effect are reflected in the distribution of AI changes across SST neurons, which is centred at AI = -0.3 ± 0.3, indicating a consistent reduction in AI across the population instead of being driven by a small portion of SST neurons. To further clarify this, we will report the proportion of SST neurons showing a reduction in AI and include statistical analyses on the changes.

    (2) Statistics for the effects. There is a mention of Liner mixed models, but no information is given on the actual models being used and tested. This is particularly for the case of Figure 1G, where there is a composition of effect sizes between different populations. What precise significance test is being used? Are the stats on paired cells when considering locomotion and rest?

    We used Linear mixed models to test for statistical significance between different conditions composed of hundreds of cells from several mice, i.e. nested analysis (cells nested within mice; see (Judd et al., 2017)). For analyses such as Fig. 1G, we considered locomotion state, adaptive type and their interaction (loco’adap) as fixed effects and mouse number as the random effect. The p-values depicted in the legend indicates the interaction between locomotion and adaptive type, i.e. the increase in amplitude during locomotion is significantly different in sensitizers compared to depressors with p < 0.0001. We will revise the method section and figure legends to explicitly describe the model and statistical test used.

    (3) Model parameters: It is acknowledged that there is a large range of parameters that can model the responses effectively, up to 11% of initial conditions. At 9000 initial conditions, this is around 1000. The parameter estimates are then considered as the mean of each parameter. This seems like a strange choice for a few different reasons:

    (3a) A mean solution might not be one of the solutions. Let's say the parameters range over a large dimensional space. They could occupy non-overlapping / discontinuous subspaces. In that case, the mean parameters do not necessarily fall within the solution subspaces. Therefore, this reduction to means might not be valid.

    (3b) Compare distributions rather than means. There are multiple distributions of parameters between conditions. All stats should be on the comparison of distributions rather than just the means.

    To test for the presence of subsets of solutions grouped around different parameter values we plotted the distribution of each parameter across all the good solutions found. Most of the weights were a gaussian distribution centred around the mean and, most importantly, none of them had two peaks. Furthermore, after computing the mean weight values we plotted the solutions given by them in the model, and it rendered a good fit as shown in the figures. We will include those distributions in the new version and base the overall comparison on these distributions.

    (4) Visualizing weight matrices: It is very challenging to interpret the weight matrices. Furthermore, it appears that the stationary and locomotion conditions fit independently, and given the large parameter spaces, it is even harder to interpret. Can the fitting instead be done by fitting on one and using those at the initial conditions for the other state? Figure 7 shows an initiative cartoon, but it is not clear how the matrices in Figures 5 and 6 lead to the summary shown in Figure 7. It is also not clear why the connections between inhibitory neurons are not shown in Figure 7. One option is to perhaps run some kind of dimensionality deduction on the parameter space to better interpret the data. When showing deltaWeights, was the model initialised with 'Rest' weights and allowed to change? It is not obvious what the difference is between 'relative change in connection weights' and 'relative change in synaptic weights'.This needs to be clarified.

    Thanks for raising this concern. We will firstly try to make the weight matrices clearer to interpret.

    Regarding the fitting of rest and locomotion conditions, we fitted the locomotion traces first and used those solutions as initial conditions for the rest traces. However, this rendered no good solutions as minimums in the parameter space were too far from the initial starting points. We opted, therefore, for repeating the same process of initial condition searching for locomotion and rest. This approach is less biased in satisfying our aim of finding solutions that fit the data and can explain their dynamics, which are different for each condition. We believe this approach is effective, as not only has it allowed us to describe circuit changes during internal-state transitions but has also made a series of predictions under different learning states that were confirmed by optogenetic tests (Hinojosa et al., 2026).

    We simplified Fig. 7 for clarity but we will make it more accurate and explain it more in detail in the legend, including connections between interneurons.

    Interpreting high-dimensional parameter spaces can be challenging. In this study, we focused on low-dimensional summaries of the parameter space (e.g., average connection weights and their distributions across populations), which revealed consistent and interpretable differences between sensitizing and depressing neurons. Importantly, our conclusions do not rely on individual parameter values, but rather on systematic differences across populations that are robust across solutions. Additionally, we ran clustering analysis and found that there is no parameter that can be removed. We focused, therefore, on the larger and more robust differences. We will explore additional dimensionality reduction approaches and include these results if they provide further insight beyond the current analyses.

    Finally, the change in weights was calculated with equation 4, in which the weight from locomotion and rest, obtained through independent fits, were used to calculate the relative change from rest to locomotion. These were either connection weights (equation 2) which consider the strength of the connection between cell j and i, or synaptic weights (equation 3) which express the weight of individual synapses by dividing connection weights by the number of presynaptic cells and probability of connection. This distinction arises because we used average traces from all the neurons imaged to fit the model, requiring considering the number of cells to know the strength of individual synapses. We will add this explanation in the results and methods sections.

    (4a) Model parameters were reduced differently for locomotion and rest (Figure 4). We suggest evaluating the results for locomotion and rest using the same chi-square value of 3 for both behavioral states (at least in controls).

    Thank you for this prompt, this is an important point that we tried to resolve during our analysis. We used the reduced chi-square () to evaluate model fits within locomotion and rest condition independently. As defined in equation 12, reduced chi-square is inversely proportional to the standard error of the data which is higher in the rest dataset. As a consequence, setting the same threshold across conditions would not correspond to an equivalent goodness-of-fit criterion, and would impose a disproportionately strict constraint on the condition with lower variability, where deviations between model and data are more heavily penalized. For this reason, we used condition specific thresholds to ensure comparable fit quality relative to the noise level in each condition. In addition, to enable direct comparison across conditions independent of their noise levels, we used the RMSE as a complementary metric.

    References

    Carandini, M. (2004). Amplification of trial-to-trial response variability by neurons in visual cortex. PLoS Biol, 2(9), E264. https://doi.org/10.1371/journal.pbio.0020264

    Chance, F. S., Nelson, S. B., & Abbott, L. F. (1998). Synaptic Depression and the Temporal Response Characteristics of V1 Cells. The Journal of Neuroscience, 18(12), 4785–4799. https://doi.org/10.1523/JNEUROSCI.18-12-04785.1998

    Deneux, T., Kaszas, A., Szalay, G., Katona, G., Lakner, T., Grinvald, A., Rózsa, B., & Vanzetta, I. (2016). Accurate spike estimation from noisy calcium signals for ultrafast three-dimensional imaging of large neuronal populations in vivo. Nature Communications, 7(1), 12190. https://doi.org/10.1038/ncomms12190

    Dipoppa, M., Ranson, A., Krumin, M., Pachitariu, M., Carandini, M., & Harris, K. D. (2018). Vision and Locomotion Shape the Interactions between Neuron Types in Mouse Visual Cortex. Neuron, 98(3), 602–615.e608. https://doi.org/10.1016/j.neuron.2018.03.037

    Heintz, T. G., Hinojosa, A. J., Dominiak, S. E., & Lagnado, L. (2022). Opposite forms of adaptation in mouse visual cortex are controlled by distinct inhibitory microcircuits. Nature Communications, 13(1), 1031. https://doi.org/10.1038/s41467-022-28635-8

    Hinojosa, A. J., Dominiak, S. E., Kosiachkin, Y., & Lagnado, L. (2026). Distinct Disinhibitory Circuits Link Short-Term Adaptation to Familiarity and Reward Learning in Visual Cortex. bioRxiv, 2026.2003.2024.713929. https://doi.org/10.64898/2026.03.24.713929

    Jin, M., & Glickfeld, L. L. (2020). Magnitude, time course, and specificity of rapid adaptation across mouse visual areas. J Neurophysiol, 124(1), 245–258. https://doi.org/10.1152/jn.00758.2019

    Judd, C. M., Westfall, J., & Kenny, D. A. (2017). Experiments with More Than One Random Factor: Designs, Analytic Models, and Statistical Power. Annu Rev Psychol, 68, 601–625. https://doi.org/10.1146/annurev-psych-122414-033702

    Lee, J., Kim, H. R., & Lee, C. (2010). Trial-to-trial variability of spike response of V1 and saccadic response time. J Neurophysiol, 104(5), 2556–2572. https://doi.org/10.1152/jn.01040.2009

    Rupprecht, P., Carta, S., Hoffmann, A., Echizen, M., Blot, A., Kwan, A. C., Dan, Y., Hofer, S. B., Kitamura, K., Helmchen, F., & Friedrich, R. W. (2021). A database and deep learning toolbox for noise-optimized, generalized spike inference from calcium imaging. Nat Neurosci, 24(9), 1324–1337. https://doi.org/10.1038/s41593-021-00895-5

    Varela, J. A., Sen, K., Gibson, J., Fost, J., Abbott, L. F., & Nelson, S. B. (1997). A Quantitative Description of Short-Term Plasticity at Excitatory Synapses in Layer 2/3 of Rat Primary Visual Cortex. The Journal of Neuroscience, 17(20), 7926–7940. https://doi.org/10.1523/JNEUROSCI.17-20-07926.1997

  10. Locomotion Selectively Amplifies Sensitizing Neurons in Primary Visual Cortex https://www.biorxiv.org/content/10.1101/2025.07.24.666602v1

  11. Locomotion Selectively Amplifies Sensitizing Neurons in Primary Visual Cortex https://www.biorxiv.org/content/10.1101/2025.07.24.666602v1