The Ingestive Response Reflects Neural Dynamics in Gustatory Cortex

Reviewer #1 (Public review):

Summary:

This study investigates how ingestive behaviors are reflected in muscle activity and how these behaviors relate to neural dynamics in the brain. By combining muscle recordings with computational analysis, the authors identify patterns of mouth movements and show that these change over time and align with changes in brain activity. The work suggests that ingestion is not defined by a single action but by coordinated changes across multiple behaviors.

Strengths:

(1) Addresses an important and underexplored question about how ingestive behavior is organized.

(2) Combines behavioral, physiological, and computational approaches creatively.

(3) Provides a novel framework for quantifying complex ingestive movements.

(4) Demonstrates a clear temporal relationship between behavior and brain activity.

Weaknesses

(1) Behavioral labels rely on video-based scoring, which may not fully capture subtle or hidden movements.

(2) The relationship between brain activity and behavior is correlational, but sometimes interpreted more strongly.

(3) The manuscript could be clearer and more accessible to readers outside the field.

Reviewer #2 (Public review):

Summary:

In this study, Baas-Thomas et al. aim to study the neural mechanisms underlying ingestive versus rejection responses to taste stimuli by developing an EMG-based approach to identify ingestion-related orofacial movements. Whereas prior work has focused primarily on detecting rejection-related gapes, the authors introduce a machine-learning classifier that uses waveform features extracted from anterior digastric (AD) EMG signals to detect mouth- and tongue-movement (MTM) events associated with ingestion. Clustering analyses further suggest that ingestive behavior consists of multiple MTM subtypes whose relative frequencies vary across trial time and taste conditions. Finally, simultaneous recordings indicate that shifts in MTM expression follow transitions in gustatory cortex (GC) population dynamics into palatability-related firing states, supporting a role for cortical ensemble activity in coordinating ingestive motor responses.

Strengths:

Overall, the scientific question addressed in this study is well motivated. A mechanistic understanding of ingestive decision-making requires a precise characterization of the motor patterns that implement ingestion, and these behaviors have remained insufficiently resolved in prior work. The authors take a reasonable and technically innovative approach by leveraging AD EMG recordings to classify distinct orofacial movement patterns. The extracted waveform features appear effective in separating gapes from ingestion-related mouth-tongue movements, and clustering analyses further suggest the presence of distinguishable MTM subtypes that show meaningful temporal structure and neural correlates. Taken together, the work provides a potentially useful framework for linking gustatory cortical dynamics to the motor expression of taste-guided decisions.

A particularly valuable aspect of this work is the attempt to move beyond a binary characterization of ingestive behavior and instead identify multiple subtypes of ingestion-related movements. This finer behavioral resolution has the potential to provide a more realistic account of how complex consummatory actions are organized. More broadly, the effort to relate structured behavioral motifs to population-level neural dynamics is conceptually interesting and could prove useful for future studies seeking to connect circuit dynamics with the motor implementation of motivated behaviors.

Weaknesses:

(1) I have several concerns regarding the methodological comparisons used to establish the superiority of the proposed XGBoost classifier. In particular, the comparison between the XGBoost classifier and previously used QDA approaches (Figure 3) may not be entirely well-matched. The QDA framework was originally designed primarily to detect gape events and does not explicitly assign labels to MTM movements. As a result, the apparent advantage of XGBoost in identifying MTMs may partly reflect differences in task formulation rather than intrinsic differences in classification performance. From visual inspection, gape detection performance appears broadly comparable across methods.

A more informative benchmark would involve comparing XGBoost to an extended pipeline in which QDA-based gape detection is combined with a secondary movement-detection stage, distinguishing MTMs from periods of no movement. Such a comparison would better isolate the contribution of classifier architecture per se. Without this control analysis, the strength of the claim that XGBoost provides superior performance for behavioral decoding remains somewhat uncertain.

(2) The presentation of the neural ensemble analyses is considerably less comprehensive and intuitive than that of the behavioral analyses. The manuscript would benefit from more direct visualization of inferred neural state transitions. For example, plotting predicted neural states in a manner analogous to the behavioral states illustrated in Figure 6B would improve interpretability and help readers understand how neural dynamics relate temporally to behavioral changes.

In addition, the interpretation that GC ensemble dynamics drive behavioral state transitions may require further clarification. If GC activity plays a causal role in initiating behavioral changes, one might expect a consistent brain-to-behavior lag across changepoints. However, Figure 6 appears to show such lag primarily at the second transition but not at the first. This raises questions about how uniformly the proposed causal interpretation applies across state boundaries, and additional analysis or discussion is needed.

(3) The neural ensemble analyses primarily focus on constructing higher-level behavioral state variables rather than directly testing how individual movement subtypes relate to neural activity. The behavioral interpretation of the inferred state structure, therefore, remains somewhat unclear. While this approach is consistent with previous work from the authors and with broader state-transition frameworks of gustatory processing, it is not immediately obvious that this is the most informative level of analysis for the present dataset.

In particular, it would strengthen the manuscript to examine whether GC neurons or ensembles also encode lower-level motor structure, such as the occurrence of gapes or specific MTM subtypes. Demonstrating selective or mixed encoding across hierarchical levels (movement motifs versus abstract behavioral states) would help clarify the functional interpretation of the reported neural dynamics. At present, the manuscript largely assumes that GC activity reflects higher-order behavioral states without directly testing alternative representational possibilities.

(4) Because direct behavioral ground truth for intra-oral ingestive movements is difficult to obtain, MTM subtypes are inferred primarily through clustering of EMG waveform features. Although the authors demonstrate statistical separability and cross-session stability of these clusters, it remains unclear whether they correspond to discrete motor programs or instead reflect a structured partitioning of a continuous behavioral space shaped by feature selection and preprocessing choices. Perhaps some additional robustness analyses or convergent validation (e.g., alternative clustering methods, feature perturbation tests, or stronger neural and behavioral dissociations) would help clarify the biological significance of the inferred subtype structure.

Reviewer #3 (Public review):

Summary:

This study examines how ingestive-related orofacial movements relate to ensemble dynamics in gustatory cortex (GC) during taste processing. Previous work has shown that GC activity evolves through a sequence of population states following taste delivery, culminating in a transition to palatability-related firing that precedes rejection-related orofacial movements (e.g., gaping). Importantly, perturbing GC activity around the time of this transition alters the timing of gaping, suggesting that these ensemble dynamics play a functional role in linking taste evaluation to behavioral responses. The present study asks whether similar neural dynamics are also associated with ingestive-related orofacial movements that occur during the consumption of palatable stimuli.

To address this question, the authors develop a machine-learning classifier to identify distinct orofacial movements from anterior digastric EMG recordings. Using a set of labeled EMG waveforms obtained from video-scored trials, a gradient-boosted (XGBoost) classifier is trained to detect gapes, mouth/tongue movements (MTMs), and periods of no movement. Applying this classifier to a larger EMG dataset reveals that ingestive-related MTMs cluster into three distinct movement subtypes whose frequencies change systematically within trials.

The authors then relate these behavioral dynamics to previously described GC ensemble transitions identified using changepoint modeling. They report that changes in MTM subtype frequencies tend to occur shortly after the transition to palatability-related activity in GC. These results suggest that GC population dynamics are temporally associated not only with rejection-related behaviors but also with ingestive motor patterns that occur as animals prepare to consume palatable tastants.

Strengths:

The study introduces an innovative framework for extracting intricate orofacial movement information from EMG recordings. The machine-learning classifier provides a scalable method for identifying specific orofacial movements and performs better than previously published algorithms designed for gape detection. This approach allows the authors to examine movement microstructure at a temporal resolution that cannot be achieved with video scoring in freely moving animals.

A second strength is the integration of orofacial movement analysis with neural population dynamics. By relating EMG-derived movement subtypes to ensemble state transitions in GC, the study builds on a substantial body of work examining the temporal evolution of taste responses in cortex. The finding that ingestive-related movement dynamics occur shortly after the emergence of palatability-related firing provides an interesting extension of previous observations linking GC state transitions to rejection behavior.

The manuscript is also commendable in its commitment to data accessibility. By providing clear information about how the datasets can be accessed and making training data for the classifier publicly available, the authors make it possible for other researchers to examine the analytical pipeline and apply similar approaches to their own datasets. This transparency provides a useful framework for extending and building upon the methods presented here.

Weaknesses:

Some aspects of the EMG-based movement classification pipeline warrant careful interpretation. The training dataset used for classifier development is relatively small and is derived from a subset of trials in which mouth movements were clearly visible in video recordings. While the classifier performs well on this labeled dataset, it is not entirely clear how representative these labeled examples are of the full range of EMG signals present in the larger dataset.

The interpretation of the three identified MTM subtypes also remains somewhat tentative. The study convincingly demonstrates that distinct waveform-defined clusters exist in the EMG data, but the functional significance of these clusters as ingestive "behaviors" is less clear. As acknowledged by the authors, the specific roles of these movement patterns in the ingestion process remain speculative.

Finally, several conclusions in the Discussion rely on relatively strong mechanistic language when describing the relationship between GC dynamics and ingestive behavior. The data clearly demonstrate a temporal association between GC state transitions and changes in the frequencies of the different MTM subtypes. However, the results primarily support the interpretation that similar cortical dynamics are associated with ingestive and rejection-related behaviors rather than definitively establishing that these behaviors "are governed by the same underlying neural mechanisms".

Author response:

Public Reviews:

Reviewer #1 (Public review):

(1) Behavioral labels rely on video-based scoring, which may not fully capture subtle or hidden movements.

This is very true; certainly, this work is only a starting point. But the techniques used for this manuscript, despite starting with video-based scoring, specifically did allow us to differentiate behaviors that were too subtle to recognize in the video. For the revision, we will describe how this work leads to future studies in which we will be able to explore other means of collecting behavioral labels, potentially directly from simultaneous recordings of multiple muscles.

(2) The relationship between brain activity and behavior is correlational, but sometimes interpreted more strongly.

We will comb through the manuscript and make edits to be more precise and technically correct in presenting this relationship, and clarify that our suggestion of a causal link is only indirect and related to previous work (Mukherjee et al. 2019).

(3) The manuscript could be clearer and more accessible to readers outside the field.

We will edit the manuscript in multiple places to make technical and field-specific aspects more accessible. As part of this, in appreciation of Reviewer 2’s comments, we will take additional care to elaborate on and clarify our need and interpretation of SHAP values and classifier structure.

Reviewer #2 (Public review):

(1) I have several concerns regarding the methodological comparisons used to establish the superiority of the proposed XGBoost classifier. In particular, the comparison between the XGBoost classifier and previously used QDA approaches (Figure 3) may not be entirely well-matched. The QDA framework was originally designed primarily to detect gape events and does not explicitly assign labels to MTM movements. As a result, the apparent advantage of XGBoost in identifying MTMs may partly reflect differences in task formulation rather than intrinsic differences in classification performance. From visual inspection, gape detection performance appears broadly comparable across methods.

A more informative benchmark would involve comparing XGBoost to an extended pipeline in which QDA-based gape detection is combined with a secondary movement-detection stage, distinguishing MTMs from periods of no movement. Such a comparison would better isolate the contribution of classifier architecture per se. Without this control analysis, the strength of the claim that XGBoost provides superior performance for behavioral decoding remains somewhat uncertain.

The revision will further clarify that, as the reviewer notes, the primary improvement in XGB classification compared to QDA (in multi-class aggregated metrics) comes specifically from its ability to classify MTMs, and that for gapes, both QDA and XGB perform on par. We will be more explicit about the fact that our goal in constructing the classifier is not to compare “classifier architecture”—not to find the very best classifier possible—but rather to take the next step by generating an instance of a classifier that performs demonstrably better on aggregated orofacial movements. We will update the manuscript to be more clear in our claims in this regard, and how the current XGB classifier can, once validated, be bootstrapped by future techniques (possibly using more informative data sources) to more fully characterize orofacial movements.

(2) The presentation of the neural ensemble analyses is considerably less comprehensive and intuitive than that of the behavioral analyses. The manuscript would benefit from more direct visualization of inferred neural state transitions. For example, plotting predicted neural states in a manner analogous to the behavioral states illustrated in Figure 6B would improve interpretability and help readers understand how neural dynamics relate temporally to behavioral changes.

In addition, the interpretation that GC ensemble dynamics drive behavioral state transitions may require further clarification. If GC activity plays a causal role in initiating behavioral changes, one might expect a consistent brain-to-behavior lag across changepoints. However, Figure 6 appears to show such lag primarily at the second transition but not at the first. This raises questions about how uniformly the proposed causal interpretation applies across state boundaries, and additional analysis or discussion is needed.

We are happy to update the figures (likely by adding another panel to Figure 6) to clearly show inference of neural state transitions, in a manner similar to how we have shown behavioral state transitions in Fig. 6B. In addition, we will do a more comprehensive job of describing and referencing earlier work in which we have unpacked these analyses in greater detail—work that makes it clear why we would predict a lag-relationship for one set of change points and not the other.

(3) The neural ensemble analyses primarily focus on constructing higher-level behavioral state variables rather than directly testing how individual movement subtypes relate to neural activity. The behavioral interpretation of the inferred state structure, therefore, remains somewhat unclear. While this approach is consistent with previous work from the authors and with broader state-transition frameworks of gustatory processing, it is not immediately obvious that this is the most informative level of analysis for the present dataset.

In particular, it would strengthen the manuscript to examine whether GC neurons or ensembles also encode lower-level motor structure, such as the occurrence of gapes or specific MTM subtypes. Demonstrating selective or mixed encoding across hierarchical levels (movement motifs versus abstract behavioral states) would help clarify the functional interpretation of the reported neural dynamics. At present, the manuscript largely assumes that GC activity reflects higher-order behavioral states without directly testing alternative representational possibilities.

The reviewer makes a good point. While previous work from the lab (Li et al. 2016) has assessed the relationship of GC activity with both the onset of gaping (i.e., the behavioral state transition) and individual gapes and found only a relationship with onset of gaping (findings that we now explicitly describe in the revision), we have not performed a similar analysis for MTMs. We will do so and add it to the paper.

(4) Because direct behavioral ground truth for intra-oral ingestive movements is difficult to obtain, MTM subtypes are inferred primarily through clustering of EMG waveform features. Although the authors demonstrate statistical separability and cross-session stability of these clusters, it remains unclear whether they correspond to discrete motor programs or instead reflect a structured partitioning of a continuous behavioral space shaped by feature selection and preprocessing choices. Perhaps some additional robustness analyses or convergent validation (e.g., alternative clustering methods, feature perturbation tests, or stronger neural and behavioral dissociations) would help clarify the biological significance of the inferred subtype structure.

We admit (in fact, we have done so in the text) that we are not yet to the point of being able to “split hairs” to this degree (although we, like R2, see that as a goal). In the meantime, we will expand the section of Results text in which we describe the fact that the clustering of behaviors is observed both in “waveform space” (Fig. 4E was generated using standardized waveforms) and “feature space” (Fig. 4 B,C, and F), and that as such the clusters are NOT simply a partitioning of continuous, unimodal behavioral space. We will report convergent results from alternative (k-means) clustering methods to further support that conclusion. Finally, we will describe (in the Discussion section) ways to more rigorously test and extend this claim in future work.

Reviewer #3 (Public review):

Some aspects of the EMG-based movement classification pipeline warrant careful interpretation. The training dataset used for classifier development is relatively small and is derived from a subset of trials in which mouth movements were clearly visible in video recordings. While the classifier performs well on this labeled dataset, it is not entirely clear how representative these labeled examples are of the full range of EMG signals present in the larger dataset.

Very good point. We will update the text to note this qualification to the reader. We will also, however, highlight the fact that our focus on a highly reliable and representative (i.e., agreed upon by 2 independent, blind scorers) subset of labels allows us to perform more targeted analyses and make more targeted interpretation in our results. And we will also be more pointed in the revision, as we have noted above, about the fact that this work is only scratching the surface of what can be accomplished in this domain, and that future work will involve STARTING with the waveforms that aren't accounted for in terms of gapes and MTMs.

The interpretation of the three identified MTM subtypes also remains somewhat tentative. The study convincingly demonstrates that distinct waveform-defined clusters exist in the EMG data, but the functional significance of these clusters as ingestive "behaviors" is less clear. As acknowledged by the authors, the specific roles of these movement patterns in the ingestion process remain speculative.

We share R3’s desire for clarity on this point—we do not wish to imply that we understand more than we understand—and will be sure to fine-tune our language to make clearer and more explicit the fact that the distinction in the roles of the MTM subtypes in ingestion at this point remains speculative.

Finally, several conclusions in the Discussion rely on relatively strong mechanistic language when describing the relationship between GC dynamics and ingestive behavior. The data clearly demonstrate a temporal association between GC state transitions and changes in the frequencies of the different MTM subtypes. However, the results primarily support the interpretation that similar cortical dynamics are associated with ingestive and rejection-related behaviors rather than definitively establishing that these behaviors "are governed by the same underlying neural mechanisms".

We will soften our language to clarify which of our Discussion suggestions are speculation, highlighting for the reader the fact that our data, while consistent with evidence suggesting a causal link between the GC transition and gaping (Li et al., 2016; Mukherjee et al., 2019), do not prove a causal neural-behavioral link for MTMs.

References:

Li, Jennifer X., et al. “Sensory Cortical Activity Is Related to the Selection of a Rhythmic Motor Action Pattern.” The Journal of Neuroscience, vol. 36, no. 20, May 2016, pp. 5596–607. DOI.org (Crossref), https://doi.org/10.1523/JNEUROSCI.3949-15.2016.

Mukherjee, Narendra, et al. “Impact of Precisely-Timed Inhibition of Gustatory Cortex on Taste Behavior Depends on Single-Trial Ensemble Dynamics.” eLife, edited by Laura L. Colgin et al., vol. 8, June 2019, p. e45968. eLife, https://doi.org/10.7554/eLife.45968.