Progressively shifting patterns of co-modulation among premotor cortex neurons carry dynamically similar signals during action execution and observation

Reviewer #1 (Public Review):

Summary and strengths. This paper starts with an exceptionally fair and balanced introduction to a topic, the mirror neuron literature, which is often debated and prone to controversies even in the choice of the terminology. In my opinion, the authors made an excellent job in this regard, and I really appreciated it. Then, they propose a novel method to look at population dynamics to compare neural selectivity and alignment between execution and observation of actions performed with different types of grip.

Weakness. Unfortunately, the goal and findings within this well-described framework are less clear to me. The authors aimed to investigate, using a novel analytic approach, whether and to what extent a match exists between population codes and neural dynamics when a monkey performs an action or observes it performed by an experimenter. This motivation stems from the fact that the general evidence in the literature is that the match between visual and motor selectivity of mirror neuron responses is essentially at a chance level. While the approach devised by the author is generally well-described and understandable, the main result obtained confirms this general finding of a lack of matching between the two contexts in 2 out of the three monkeys. Nevertheless, the authors claim that the patterns associated with execution and observation can be re-aligned with canonical correlation, indicating that these distinct neural representations show dynamical similarity that may enable the nervous system to recognize particular actions. This final conclusion is hardly acceptable to me, and constitutes my major concern, at least without a more explicit explanation: how do we know that this additional operation can be performed by the brain? Is this a computational trick to artificially align something that is naturally non-aligned, or can it capture something real and useful?
Based on the accumulated evidence on space-constrained coding of others' actions by mirror neurons (e.g., Caggiano et al. 2009; Maranesi et al. 2017), recent evidence also cited by the authors (Pomper et al. 2023), and the most recent views supported even by the first author of the original discovery (i.e., Vittorio Gallese, see Bonini et al. 2022 on TICS), it seems that one of the main functions of these cells, especially in monkeys, might be to prepare actions and motor responses during social interaction rather than recognizing the actions of others - something that visual brain areas could easily do better than motor ones in most situations. In this perspective, and given the absence of causal evidence so far, the lack of visuo-motor congruence is a potentially relevant feature of the mechanism rather than something to be computationally cracked at all costs.

Specific comments on Results/Methods:
I can understand, based on the authors' hypothesis, that they employed an ANOVA to preliminarily test whether and which of the recorded neurons fit their definition of "mirror neurons". However, given the emphasis on the population level, and the consolidated finding of highly different execution and observation responses, I think it could be interesting to apply the same analysis on (at least also) the whole recorded neuronal population, without any preselection-based on a single neuron statistic. Such preselection of mirror neurons could influence the results of EXE-OBS comparisons since all the neurons activated only during EXE or OBS are excluded. Related to this point, the authors could report the total number of recorded neurons per monkey/session, so that also the fraction of neurons fitting their definition of mirror neuron is explicit.
Furthermore, the comparison of the dynamics of the classification accuracy in figures 4 and 5, and therefore the underlying assumption of subspaces shift in execution and observation, respectively, reveal substantial similarities between monkeys despite the different contexts, which are clearly greater than the similarities among neural subspaces shifts across task epochs: to me, this suggests that the main result is driven by the selected neural populations in different monkeys/implants rather than by an essential property of the neuronal dynamics valid across animals. Could the author comment on this issue? This could easily explain the "strange" result reported in figure 6 for monkey T.

Reviewer #2 (Public Review):

In this work, the authors set out to identify time-varying subspaces in the premotor cortical activity of monkeys as they executed/observed a reach-grasp-hold movement of 4 different objects. Then, they projected the neural activity to these subspaces and found evidence of shifting subspaces in the time course of a trial in both conditions, executing and observing. These shifting subspaces appear to be distinct in execution and observation trials. However, correlation analysis of neural dynamics reveals the similarity of dynamics in these distinct subspaces. Taken together, Zhao and Schieber speculate that the condition-dependent activity studied here provides a representation of movement that relies on the actor.
This work addresses an interesting question. The authors developed a novel approach to identify instantaneous subspaces and decoded the object type from the projected neural dynamics within these subspaces. As interesting as these results might be, I have a few suggestions and questions to improve the manuscript:
1- Repeating the analyses in the paper, e.g., in Fig5, using non-MN units only or the entire population, and demonstrating that the results are specific to MNs would make the whole study much more compelling.
2- The method presented here is similar and perhaps related to principal angles (https://doi.org/10.2307/2005662). It would be interesting to confirm these results with principal angles. For instance, instead of using the decoding performance as a proxy for shifting subspaces, principal angles could directly quantify the 'shift' (similar to Gallego et al, Nat Comm, 2018). Relatedly, why the decoding of the 'object type' is used to establish the progressive shifting of the subspaces? I would be interested to see the authors' argument. The object type should be much more decodable during movement or hold, than instruction, which is probably why the chance-level decoding performance (horizontal lines) is twice the instruction segment for the movement segment.
3- Why aren't execution and observation subspaces compared together directly? Especially given that there are both types of trials in the same session with the same recorded population of neurons. Using instantaneous subspaces, or the principal angles between manifolds during exec trials vs obs trials.
4- The definition of the instantaneous subspaces is a critical point in the manuscript. I think it is slightly unclear: based on the Methods section #715-722 and the main text #173-#181, I gather that the subspaces are based on trial averaged neural activity for each of the 4 objects, separately. So for each object and per timepoint, a vector of size (1, n) -n neurons- is reduced to a vector of (1, 2 or 3 -the main text says 2, methods say 3-) which would be a single point in the low-d space. Is this description accurate? This should be clarified in the manuscript.
5- Isn't the process of projecting segments of neural dynamics and comparing the results equivalent to comparing the projection matrices in the first place? If so, that might have been a more intuitive avenue to follow.
6- Lines #385-#389: This process seems unnecessarily complicated. Also, given the number of trials available, this sometimes doesn't make sense. E.g. Monkey R exec has only 8 trials of one of the objects, so bootstrapping 20 trials 500 times would be spurious. Why not, as per Gallego et al, Nat Neurosci 2020 and Safaie et al, Nat 2023 which are cited, concatenate the trials?
7- Related to the CCA analysis, what behavioural epoch has been used here, the same as the previous analyses, i.e. 100ms? how many datapoint is that in time? Given that CCA is essentially a correlation value, too few datapoints make it rather meaningless. If that's the case, I encourage using, let's say, one window combined of I and G until movement, and one window of movement and hold, such that they are both easier to interpret. Indeed low values of exec-exec in CC2 compared to Gallego et al, Nat Neurosci, 2020 might be a sign of a methodological error.

Reviewer #3 (Public Review):

Summary:
In their study, Zhao et al. investigated the population activity of mirror neurons (MNs) in the premotor cortex of monkeys either executing or observing a task consisting of reaching to, grasping, and manipulating various objects. The authors proposed an innovative method for analyzing the population activity of MNs during both execution and observation trials. This method enabled to isolate the condition-dependent variance in neural data and to study its temporal evolution over the course of single trials. The method proposed by the authors consists of building a time series of "instantaneous" subspaces with single time step resolution, rather than a single subspace spanning the entire task duration. As these subspaces are computed on an instant time basis, projecting neural activity from a given task time into them results in latent trajectories that capture condition-dependent variance while minimizing the condition-independent one. The authors then analyzed the time evolution of these instantaneous subspaces and revealed that a progressive shift is present in subspaces of both execution and observation trials, with slower shifts during the grasping and manipulating phases compared to the initial preparation phase. Finally, they compared the instantaneous subspaces between execution and observation trials and observed that neural population activity did not traverse the same subspaces in these two conditions. However, they showed that these distinct neural representations can be aligned with Canonical Correlation Analysis, indicating dynamic similarities of neural data when executing and observing the task. The authors speculated that such similarities might facilitate the nervous system's ability to recognize actions performed by oneself or another individual.

Strengths:
Unlike other areas of the brain, the analysis of neural population dynamics of premotor cortex MNs is not well established. Furthermore, analyzing population activity recorded during non-trivial motor actions, distinct from the commonly used reaching tasks, serves as a valuable contribution to computational neuroscience. This study holds particular significance as it bridges both domains, shedding light on the temporal evolution of the shift in neural states when executing and observing actions. The results are moderately robust, and the proposed analytical method could potentially be used in other neuroscience contexts.

Weaknesses:
While the overall clarity is satisfactory, the paper falls short in providing a clear description of the mathematical formulas for the different methods used in the study. Moreover, it was not immediately clear why the authors did not consider a (relatively) straightforward metric to quantity the progressive shift of the instantaneous subspaces, such as computing the angle between consecutive subspaces, rather than choosing a (in my opinion) more cumbersome metric based on classification of trajectory segments representing different movements.

Specific comments:
In the methods, it is stated that instantaneous subspaces are found with 3 PCs. Why does it say 2 here? Another doubt on how instantaneous subspaces are computed: in the methods you state that you apply PCA on trial-averaged activity at each 50ms time step. From the next sentence, I gather that you apply PCA on an Nx4 data matrix (N being the number of neurons, and 4 being the trial-averaged activity of the four objects) every 50 ms. Is this right? It would help to explicitly specify the dimensions of the data matrix that goes into PCA computation.

It would help to include some equations in the methods section related to the LSTM decoding. Just to make sure I understood correctly: after having identified the instantaneous subspaces (every 50 ms), you projected the Instruction, Go, Movement, and Holding segments from individual trials (each containing 100 samples, since they are sampled from a 100ms window) onto each instantaneous subspace. So you have four trajectories for each subspace. In the methods, it is stated that a single LSTM classifier is trained for each subspace. Do you also have a separate classifier for each trajectory segment? What is used as input to the classifier? Each trajectory segment should be a 100x3 matrix once projected in an instantaneous subspace. Is that what (each of) the LSTMs take as input? And lastly, what is the LSTM trained to predict exactly? Just a label indicating the type of object that was manipulated in that trial? I apologize if I overlooked any detail, but I believe a clearer explanation of the LSTM, preferably with mathematical formulas, would greatly help readers understand this section.