Movement errors during skilled motor performance engage distinct prediction error mechanisms

The brain detects deviations from intended behaviors by estimating the mismatch between predicted and actual outcomes. Axiomatic to these computations are salience and valence prediction error signals, which alert the brain to the occurrence and value of unexpected events. Despite the theoretical assertion of these prediction error signals, it is unknown whether and how brain mechanisms underlying their computations support error processing during skilled motor behavior. Here we demonstrate, with functional magnetic resonance imaging, that internal detection, i.e., without externally-provided feedback, of self-generated movement errors evokes instantaneous activity increases within the salience network and delayed lingering decreases within the nucleus accumbens – a key structure in the reward valuation pathway. A widespread suppression within the sensorimotor network was also observed. Our findings suggest that neural computations of salience and valence prediction errors during skilled motor behaviors operate on different time-scales and, therefore, may contribute differentially to immediate and longer-term adaptive processes.

This paper reports detailed analysis of behavioural and fMRI data in a sequence production task, gathered as a preliminary part of another previously published report.
The aim was to measure neural responses to rare errors in the key-stroke sequences, and to determine whether there are network differences in the error-free vs erroneous motor performance, in error detection and in salience and valence processing (the former driven by errors; the latter considering the sign of the error). Because the number of spontaneous errors was low, data from a sub-group of the experiment were also analysed, as a "replication".
It is an interesting approach, and in principle a good design, and I think the questions are of wide interest. But I am not especially enthusiastic for several reasons.
First the task requires sequences to be as fast as possible, and so individual keystrokes are well beyond the temporal resolution of the fMRI analysis. To some extend this has been mitigated by considering whole sequences of 5 keystrokes at a time, but this greatly reduces the interest of the results, as it seems critical to assess error onset at the level of single keystrokes, not sequences.
Second, the instructions to participants were that when they made errors they should effectively restart from the beginning of the sequence. The long increase in duration may well reflect this process of resetting and reinitiating the sequence. Hence it seems appropriate to compare posterror performance with initiation of the sequences at the start of each block, and not with performance mid-block.
Third, errors were rare -only 203 across nearly 50 participants, and so there must be considerable variability -some participants making only one, some possibly making many more. This may lead to bias in the results, towards both behavioral and fMRI effects driven by only a few participants. It may well affect the statistics, and the t-tests used to compare behavioral data may be compromised by uneven variance. At a minimum, so analysis of the distribution of errors across the population should be presented. The attempt to mitigate against these low numbers with a replication study is good -but a sub-sample of the original pool of participants is not a valid replication. An independent sample of new participants is needed.
Fourth I am puzzled but not sure I fully understand how the fMRI data is modeled. The authors state the errors are modeled a zero duration. I guess this will bias the fitting towards very short error sequences, and away from the longer multi-sequence events? Could this effect lead to the apparent reduction in fMRI signal in error sequences? I also am not sure I fully understand the behavioral measures. Error duration is from end of a correct sequence to start of a correct sequence, if I understand it. Hence one would expect a clear order effect with longer duration for later elements in the sequence, if participants successful stop and reset the sequence as instructed. But this was not seen. In addition, when comparing % change in transitions to mean % change in duration of pre and post sequences, it seems one is comparing a single keystroke change to mean change over 5 keystrokes, and so it is unsurprising the latter are smaller effects.
Finally, the design rests on spontaneous errors, all of which are bad. So, I cannot see how the data can support an argument on valence, which would seem to need separation of positive and negative prediction outcomes. There are no unexpected positive outcomes.
Reviewer #2 (Remarks to the Author): The authors set out to identify the neural correlates of 'salience' (unsigned) and 'valence' (signed) Prediction Errors related to reward. The contributions of this manuscript are twofold: first, although signed and unsigned PEs have been observed throughout the brain, there is ongoing debate as to how/where these signals may overlap/interact. In particular, the authors focus on dACC is apt; dACC is the subject of much debate along these lines, with various authors suggesting that the error signals in dACC are signed, unsigned, or a mixture of the two. Given the array of findings that can support any of these positions, it is a valuable to directly investigate the nature of error signals in dACC, especially in contrast to other areas that are more reliably implicated in signed PEs (nucleus accumbens).
Second, the authors attempt to disentangle these PEs in the area of motor control. dACC is generally acknowledged to be critical in cognitive control, to which motor control is related. By relating their motor control paradigm to processes and functions frequently studied in the context of cognitive control, the manuscript aims to generalize findings from one line of research (cognitive control) to a separate (although closely related) line.
To do this the authors adopt a finger-tapping task in which subjects are asked to memorize a fivedigit sequence in which they are to tap their fingers. During the task proper, subjects tap their finger in a continuous, self-paced fashion, and error sequences are compared to non-error sequences in both behavioral and fMRI analyses.
On the whole this study addresses an important and relevant issue in the field of cognitive/motor control, i.e., distinguishing the nature of different error signals in the brain, especially w/r/t salience and value. The approach taken to answering this question seems sensible, and the major results appear to be of sufficient interest for the readership. However, there are a few major concerns I would like to see addressed before this manuscript can be accepted. MAJOR 1)Some aspects of the experiment design are unclear or ambiguous. For example, I guess the term 'matching sequence' means instances when the subject correctly reproduced the learned sequence of key presses (?), but it is not clearly defined anywhere in the manuscript that I could find, and it seems from text on lines 298-299 errorless performance is not the same thing as the matching sequence... this is one example of the design needs to be made more clear.
2) The analyses that contribute to fig. 5 (A,B,C), pg. 15, seem to run afoul of one of the classical neuroimaging fallacies. Specifically, the authors claim that, using errorless performance as a baseline, 'salience' network regions are specifically engaged on error and NOT matching sequences. It appears they base this on a significant difference from the BOLD response in those regions to the errorless baseline. However, just because the BOLD signal is different from baseline in one condition, and not different from baseline in another condition, this does not indicate that the BOLD signal in condition 1 is different from the BOLD signal in condition 2. I note that they do not do this in the analyses supporting figure 3, so perhaps they could make it clear that they are not basing all their claims on the fig. 5. analyses.
3) The authors note increased activity in the salience network prior to the occurrence of an error; this does not seem to fit well with their story about the 'instantaneous' recruitment of the salience network in response to errors; if the salience network responds instantly to error, why should it be active during correct performance preceding an error. I think there are ways one could support an unsigned PE story for these periods, but it doesn't appear that the authors have attempted to do so. I would like a more thorough treatment of how increased pre-error activity relates to the underlying theory supporting the authors predictions. 4) Given the authors' explicit aim to dissociate value PEs from salience PEs, especially in the realm of control, it seems appropriate to acknowledge recent work that likewise attempts to dissociate value and surprise signals in control, specifically Vassena et al., 2020, Nature Human Behavior. Full disclosure: I am an author on that paper, but it legitimately seems relevant to this study. 5) From your behavioral analyses, it appears that there is some amount of response slowing in the PreE1 periods; dACC activity is frequently correlated with time on task (e.g., grinband et al., 2010). It appears that the fMRI analyses do not incorporate parametric modulators that might be used to model response time or sequence time. It could be possible that the increased salience network activity observed in preE periods might be related to increases in RT. They authors should probably re-rerun the analyses with RT included as a modulator to ensure that their data could not be explained as deriving from longer sequence durations. 6) This task doesn't appear to carry an explicit reward manipulation, i.e., subjects are not given any incentive (monetary or otherwise) to perform the sequences correctly. I don't necessarily think that such an incentive is absolutely required, but given the long history of studying reward/valence PEs with explicit rewards, the authors should spend some time justifying why the signals they observe in, e.g., NAcc are in fact reflecting reward PEs and not some other signal. Communications Biology Review The authors had subjects repeated a simple motor task involving repeating a sequence of key presses many times over, while being scanned in an fMRI. The subjects occasionally made errors in the sequence, but were not given related feedback. Thus, any knowledge of the error came from endogenous, not exogenous sources. During sequences in which the subject made an error, the speed of performance was slower. In addition, the dorsal anterior cingulate, pre-supplementary motor area, inferior frontal cortex, subthalamic nucleus, substantia nigra and subthalamic nucleus all showed activity increases, whereas paracentral lobule and sensorimotor cortex showed activity decreases. In addition, performance slowed even more following the error, and activity dropped in the medial thalamus and nucleus accumbens. The authors list many, many effects, but the takehome message is that regions associated with negative-valence prediction errors, such as dACC increase in activity during errors, whereas regions associated with task performance, such as sensorimotor cortex showed reduced activity. The task is interesting, the N is impressive, and the data is fairly compelling. I have only a few points.
1. The paper is difficult to understand. I have to take responsibility for some of that -it is likely that part of my difficulty is merely due to me not being smart enough, or well-versed enough in some of the subject matter. It is also partly due to the fact that English may not be the first language of the authors. Here and there are some tell-tale mistakes in word choice. For example, page 4, line 67, "upon such conditions"; page 17, line 336, "theoretical acclaim showing that movement…"; and page 18 line 363, "caution is imposed because of the low temporal resolution…" These are small language errors that I will not dwell on. I do suggest the authors have a native English speaker proofread the work. One identifiable, general problem in the writing is that the authors often use overly long sentences, trying to cram too many pieces of information into one point. For example, the first two sentences on page 12 (lines 235-243): "Next, we assessed temporal characteristics of error-specific changes in activity by estimating changes in BOLD responses immediately before, during and immediately after errors relative to those evoked during continuous periods of errorless performance immediately before, during and after matching sequences, respectively." And then "Thus, we conducted both region of interest (ROI) and whole-brain analysis applying repeated measures ANOVA approach on individual contrast values and activation maps, respectively, obtained from pairwise comparisons between estimated BOLD responses during (1) penultimate sequences before errors and before matching sequences, (2) last sequences immediately before errors and before matching sequences, (3) errors and matching sequences, (4) first sequences immediately after errors and after matching sequences, and (5) second sequences after errors and after matching sequences." I understand that in the later sentence, lists tend to make long sentences, but I still argue that both sentences are made harder to follow by being too long. I would advise/request that the authors go through the paper carefully with the goal of breaking the longest sentences into two or more shorter, more succinct sentences. I realize that some of the points the authors are trying to make are complex and nuanced, but still, breaking such points into smaller sections with a clear progression of what needs to be understood in what order could greatly improve the readability of this work.
I also have to note that most of my points have to do with not understanding the description, so in each case, re-writing portions of the text may be the solution.
2. Exactly what point in an error was the fMRI time-locked to? It seems that when the participant got a sequence wrong at any point, that the entire sequence as labelled an error. Does that mean that the participant could get item 1 through 4 right, press the wrong key for item 5, and then the onset of the error is still locked to the onset of the first (correct) key press? 3. Related to determining errors, what happened if the participant made a single keypress error on the first item, and then instead of going to what would normally be the second item, repeated the first item correctly and then continued on. Would that make all following key presses an error? That is, how did the authors handle errors that were merely cases where the sequence was offset by one? 4. Page 8 line 170: "we further reaffirmed that slowing associated with error commission occurred as early as at the error onset, persisted during the error and reached its peak at the error offset" Is "at error onset" here the transition from the first erroneous press to the next press, or the transition from the last correct press to the first incorrect press? 5. The "matching sequences" were matched according to their within-block position, meaning an error on block 1 could be matched to an errorless sequence on block 14, s that right? There appears to be an effect across blocks on, transition duration. Why would the authors match to across blocks instead of just matching to periods of errorless performance within the same block?
(a) Activation map with main effect of the first-block sequence (p < 0.05, FWE-corrected). Only blocks with successful task initiation starting with correctly performed and completed sequence (S1) were included in the analysis. For each subject, the number of events in the regressor used to assess the effect of the firstblock sequence was equal to the number of errors analyzed for that subject, so that the statistical power to detect changes was comparable to the statistical power to detect changes during the post-error sequence (PostE1). (b) Activation map with main effect of the first post-error sequence (PostE1) (p < 0.001, unc). (c-e) Parameter estimates extracted from the dACC, NAc and SMC. Green, gray and yellow columns represent group means. Error bars represent s.e.m.

Reviewer
The analyses that contribute to fig. 5 (A,B,C), pg. 15, seem to run afoul of one of the classical neuroimaging fallacies. Specifically, the authors claim that, using errorless performance as a baseline, 'salience' network regions are specifically engaged on error and NOT matching sequences. It appears they base this on a significant difference from the BOLD response in those regions to the errorless baseline. However, just because the BOL D signal is different from baseline in one condition, and not different from baseline in another condition, this does not indicate that the BOLD signal in condition 1 is different from the BOLD signal in condition 2. I note that they do not do this in the analyses supporting figure 3, so perhaps they could make it clear that they are not basing all their claims on the fig. 5. analyses.

Response
The reviewer is right, and we regret this confusion. Before addressing this issue, we would like to bring to the r when revising the manuscript, we added new figures. Therefore, figure numbers in the current version of the manuscript are different from the previous version. Thus, previous Fig. 3 and 5 are now Fig. 4 and 6, respectively.
We approached the analysis of the fMRI data in the following order. First, we assessed the main effect of errors and their position-matched sequences, versus rest, separately ( Fig. 4a and Fig 4c upper panels, respectively). Next, we directly contrasted between the errors and position-matched sequences to test error-specific changes (Fig. 4b). Then, we also analyzed pre-and post-error trials to assess error-related activity over time. Our main results are shown in Fig 5 and Table 3. These analyses were conducted to characterize how neural changes associated with error processing evolve over time. To do so, activity levels during the pre-error trials, errors and post-error trials were estimated against the corresponding trials of the control error-free task condition. Thus, zero contrast values (the reference line marked by the x axis) correspond to activity levels during position-matched periods of errorless performance. We also estimated changes in activity during errors and their position-matched sequences, versus rest, separately (Fig. 6). This analysis, however, was done as a post-hoc/supplementary step to visualize the differences between the two conditions and to validate their dissociation. In the figure, we include asterisk symbols to indicate significant differences not only for errors and their position-matched sequences versus rest (red and blue asterisks, respectively) but also between these two conditions (black asterisks).
We believe that given the definition of each trial type and condition this information is currently provided in the Results section it is now clearer that position-matched sequences are, in fact, part of the error-free condition/errorless performance. It also seems that the term was previously used to refer to either errorless performance or rest, also introduced unclarity. Therefore, instead of using this term, we now specify what condition was used as a reference.
s section making sure that all necessary details are provided and that the description of the fMRI data analysis is comprehensive and clear.

Reviewer
The authors note increased activity in the salience network prior to the occurrence of an error; this does not seem to fit well with their story about the 'instantaneous' recruitment of the salience network in response to errors; if the salience network responds instantly to error, why should it be active during correct performance preceding an error. I think there are ways one could support an unsigned PE story for these periods, but it doesn't appear that the authors have attempted to do so. I would like a more thorough treatment of how increased pre-error activity relates to the underlying theory supporting the authors predictions.

Response
We suggestion and we acknowledge that remark is accurate. We did not attempt to elaborate on the idea of proactive salience PES in the Introduction section because it was not part of our hypothesis. Nevertheless, we respectfully disagree with the reviewer that the idea of the instantaneous recruitment of the salience network during error commission presented in the Introduction contradicts the suggestive interpretation of our findings in the Discussion, where we propose that salience prediction error signals (PES) may be generated proactively even before the error is actually expressed behaviorally.

Reviewers' comments:
Reviewer #1 (Remarks to the Author): The careful revisions to the paper appear to have addressed (but not solved) the various issues raised in my original review.
I think the paper is acceptable, but the study has weak points and, in the limit, these are because of the study design, low temporal resolution of the fMRI methods, and the subsequent "trial" level analysis of errors. This forces analysis that does not really correspond to the actuality -for fMRI errors are assumed to be coincident with the start of the trial (and the data in Fig 2 shows in reality errors are distributed throughout); and likewise, the duration of an error trial is assumed to be zero, whereas it will be variable, depending on where in the sequence it occurs. It seems unlikely that participants would predict errors in advance of a trial (unless the error is on the first keypress). So, the "predictive" components of the fMRI results are not especially convincing for me. Finally, as I mentioned previously, the design has only spontaneous errors of negative valence, so the separation of salience and valence relies -I think -entirely on assumptions about these neural circuits, rather than by experimental manipulation that would allow their dissociation.

Reviewer #2 (Remarks to the Author):
In their revised manuscript, the authors have adequately addressed most of my major concerns with their initial manuscript. Aspects of the design and analysis approaches that were confusing or ambiguous have been clarified. I appreciate their additional discussion on the possible roles of salience PEs in behavior and NAc as signaling valence PEs.
However, I do have lingering concerns regarding the correlation of time-on-task with ACC activity I raised in my previous review. The authors argue that time-on-task effects are not relevant to this study for a couple reasons.
First, the authors provide new analysis showing that transitions within the PreE1 sequence were not significantly longer than for errorless performance (presumably using only the 4 withinsequence transitions as indicated on lines 153-154). However, there is a significant increase in the length of the transition from the PreE1 sequence to the Error sequence, and an even more pronounced increase for transitions from Error to PostE1 (Fig 3c).
Due to the design of the study, which involves continuous, self-paced performance, it is ambiguous how to model these transition periods. The authors decided to model the onset of a trial as the first key-press of the sequence, and it doesn't appear that their fMRI analyses include a regressor specifically modeling the between-sequence transitions (e.g., a stick function at the time a sequence is completed). That is, in the fMRI analyses presented, the transition from one sequence to the next is included as part of the first sequence.
Judging from average transition duration (Fig 2A) and depending on the block, it requires 1200ms (later blocks) to 2000ms (earlier blocks) to complete a sequence (i.e., 4 within-sequence transitions from the 1st button in the sequence to the 5th), well within the TR of scanner (2650ms). In other words, effects that occur following the completion of a sequence and prior to the onset of the next could be driving BOLD activity in PreE1 -increases in sequence time ought to include the between-sequence transition that is modeled as being part of a sequence in your fMRI analyses.
Furthermore, although I focused on PreE1 in my previous comments, the time-on-task issue also pertains to the Error sequence (where longer within-sequence transitions are observed, as well as increased between-sequence transitions from Error to PostE1), and I suggested including RT modulators generally rather than specifically for the PreE1 period.
The authors additionally argue that time-on-task effects are mainly relevant when a task involves processing external cues and feedback, particularly when control demands are manipulated, and that the simple sequence task, which does not involve external cues, does not significantly involve active task maintenance. This view seems at odds with some current theory (e.g., Shenhav et al., 2013) suggesting that control networks could monitor any number of relevant signals, including those which are internally generated (for example: progress through, completion of, or deviations from a sequence), in order to support task performance.
If time-on-task is genuinely irrelevant to the sequence task in the manuscript, including sequence time (four within-sequence transitions as well as the between-sequence transition) as a parametric modulator should have minimal effect on the analyses. However, this analysis seems necessary in order to rule out this alternative explanation for the authors' findings.
Reviewer #3 (Remarks to the Author): The authors have satisfactorily addressed my comments.

REVI EWER #1
We would like to express our satisfaction bT_S _SP ]PaTPbP]k^NZXXPY_ _SL_ hIhe careful revisions to the paper appear to have addressed f the various issues raised in my %]PaTPbP]k^& original review.i We are also glad to read that the reviewer thinks hfthat _SP [L[P] T^LNNP[_LMWPi+ The reviewer also mentions, however, that hfthe study has weak points and, in the limit, these are because of the study design, low temporal resolution of the fMRI methods, and the subsequent h_]TLWi WPaPW LYLWd^T^ZQ P]]Z]^+i Whereas some of these limitations cannot be solved in our current work, we believe that it does not diminish the significance of our findings. Below are our point-by-point responses to the remaining concerns brought up by the reviewer.

Reviewer
This forces analysis that does not really correspond to the actuality g for fMRI errors are assumed to be coincident with the start of the trial (and the data in Fig 2 shows in reality errors are distributed throughout); and likewise, the duration of an error trial is assumed to be zero, whereas it will be variable, depending on where in the sequence it occurs.

Response
The issue of relatively low temporal resolution of the fMRI signal is present in all neuroimaging studies using this cutting-edge technology. In the context of the current study, this limitation did not allow us to estimate changes in neural activity at the level of single keypresses. However, it was not our objective and we did not seek to characterize error processing on a scale of milliseconds, but rather sought to determine what brain mechanisms underpin computations of prediction error signals during skilled motor behavior in humans h a question that, to the best of our knowledge, has not been directly addressed before. The acquisition of the whole-brain fMRI data and the subsequent set of analyses allowed us to capture changes in activity across the entire brain that occur on a slower timescale in a space-resolved manner.
Although the analysis of the imaging data was carried out at the trial level, it is sufficiently sensitive and precise to capture specific neural changes associated with spontaneous deviations from the expected behavior (i.e., errors). During our previous revision, we conduced additional analysis and showed that activity levels at the first error key, which may correspond to correctly initiated sequence, and at the first wrong key are almost identical not only across the entire group but also for each individual (Supplementary Fig. 4).
In the current revision, we also estimated the effect of tapping speed on the magnitude of errorrelated changes in brain activity in two ways: (1) by estimating trial-related changes in the BOLD signal using brief epoch-related design with trial duration, and (2) by directly assessing the effect of tapping speed using parameter modulators with transition durations within and between trials (separate models for each trial phase). We did not find any significant changes in activity associated with trial durations or tapping speed during periods with errors. The results of these analyses are now included in Supplementary Materials as Supplementary Fig. 6 & 7, respectively, and are mentioned in the main text as follows:

REVI EWER #2
We are pleased to read that in our revised manuscript, we hfhave adequately addressed most of my %]PaTPbP]k^& major concernsfi and that hfLspects of the design and analysis approaches that were confusing or ambiguous have been clarifiedi+ Yet, the reviewer does hfhave lingering concerns regarding the correlation of time-on-task with ACCi. We addressed these concerns by conducting additional analyses. Please, see our point-bypoint responses _Z _SP ]PaTPbP]k^NZXXPY_^ below.

Reviewer
First, the authors provide new analysis showing that transitions within the PreE1 sequence were not significantly longer than for errorless performance (presumably using only the 4 withinsequence transitions as indicated on lines 153-154). However, there is a significant increase in the length of the transition from the PreE1 sequence to the Error sequence, and an even more pronounced increase for transitions from Error to PostE1 (Fig 3c).
Due to the design of the study, which involves continuous, self-paced performance, it is ambiguous how to model these transition periods. The authors decided to model the onset of a trial as the first key-press of the sequence, and it doesn't appear that their fMRI analyses include a regressor specifically modeling the between-sequence transitions (e.g., a stick function at the time a sequence is completed). That is, in the fMRI analyses presented, the transition from one sequence to the next is included as part of the first sequence.
Judging from average transition duration (Fig 2A) and depending on the block, it requires 1200ms (later blocks) to 2000ms (earlier blocks) to complete a sequence (i.e., 4 within-sequence transitions from the 1st button in the sequence to the 5th), well within the TR of scanner (2650ms). In other words, effects that occur following the completion of a sequence and prior to the onset of the next could be driving BOLD activity in PreE1 -increases in sequence time ought to include the between-sequence transition that is modeled as being part of a sequence in your fMRI analyses.

Response
The reviewer is correct. Activity levels associated with each trial type (i.e., correct trials immediately before and after errors, errors, and trials during position-matched periods of continuous errorless performance) were estimated using event-related design (i.e., zero duration) with onsets set to the first keypresses of these trials. To preserve the variance explained by each covariate, BOLD signal changes associated with adjacent trials were estimated using separate models. The continuous nature of the task and the sampling resolution of the fMRI data (one data point every 2.65 seconds), which is longer than the average duration of a single trial, did not allow us to isolate BOLD changes associated with a single keypress/transition or with a single trial. Thus, despite modeling events of interest with zero durations, changes in the BOLD signal estimated for a given trial could also capture, to a small extent, some of the effects associated with between-trial transitions and with adjacent trials. We acknowledge this limitation in the first paragraph of the Results section (page 7, lines 105-107) and in the Discussion (page 20, lines 360-361). We also elaborate on our analytic approach in the Methods section (pages 26-28).
Nevertheless, fZWWZbTYR _SP ]PaTPbP]k^NZXXPY_' be conducted additional analyses to estimate changes in the BOLD signal time-locked to the last keypress of the trial. In the detailed information presented below we refer to this time-point as the offset of this trial, which in turn is also the onset of the subsequent trial. Parameter estimates for each trial type at its first and last keypress, versus rest, and contrast values between periods with and without errors (i.e., error and control condition, respectively) are shown in the figure below.
@Q bP`YOP]^_LYO NZ]]PN_Wd' _SP ]PaTPbP]k^NZYNP]Y T^_SL_ TYN]PL^PO LN_TaT_d bT_STY _SP O8:: XLd be associated with a longer time to process and press the keys, either in the wrong or correct sequential order, rather than with the salience PES triggered by a failure to generate the sequence. However, the direct comparison between the periods with and without errors shows that activity levels within the dACC are already significantly higher at the onset of the correct trial preceding the error(t(48) = 3.35, p = 0.002) (the upper panel in the figure below). The magnitude of this change, versus the control condition, is comparable to that of the error onset (t(48) = 0.17, p = 0.865), although the error onset itself and the subsequent transitions within the error are significantly longer than transitions within and between correct trials ( Fig. 3b & 3c). The longest transition with the greatest slowing was observed at the error offset h the transition from the last error key to the first key of the first post-error sequence ( Fig. 3b & 3c). However, by the error offset, the dACC activity drops down dramatically and does not differ significantly from the control condition (t(48) = 0.19, p = 0.848). Thus, these results provide no evidence that increased activity within the dACC is directly linked to longer tapping time, not only during the trials themselves, but also between trials. These findings are, however, in line with the role of the dACC in generation of the salience PES as we hypothesized in the Introduction section. Note that the direction and the time-scale of changes in activity within the dACC are dissociable from changes within the nucleus accumbens (NAc) (the middle panel in the figure below). Finally, during periods with errors, both the dACC and NAc show changes in the opposite direction from the ones exhibited by the primary sensorimotor regions (the lower panel in the figure below). Figure. Temporal characteristics of changes in activity during error processing in the key regions of the salience, valence/reward-related and task-related networks (upper, middle and lower panel, respectively; left and right plots show changes in the left and right hemisphere, respectively). Activity levels during periods with and without errors versus rest (red and blue lines, respectively) and their contrast values (dark gray lines; values along the x-axis correspond to periods without errors, i.e., the control condition) were extracted from models with event-related design (i.e., zero durations) with onsets at the first (highlighted data-points) and the last trial keys. The first error-key and the first wrong key within the error (marked with the red oval) were compared to the first key of the position-matched sequence. The ROI within the dorsal anterior cingulate cortex (dACC) overlaid on the mean structural image of all participants is also shown. Black asterisks indicate significant differences in activity between corresponding trials and phases during periods with and without errors. * /* */* * * g significant results at 0.05/0.01/0.001 level; in the latter case, the results are significant after Bonferroni correction for 25 ROIs included in the analysis (p < 0.002).
Previous imaging studies, conducted by our and other research groups, constantly showed that successful performance in this type of task primarily relies on sensorimotor cortical and subcortical circuits (e.g., Debas et al., 2010;Orban et al., 2010;Wiestler and Diedrichsen, 2013; 8WMZ`d P_ LW)' -+,06 >LMT_Za P_ LW)' -+,0' -+,1' -+,46 :ZSPY LYO ;k<^[Z^T_Z' -+,16 KZVZT P_ LW)' 2018; Pinsard et al., 2019). Thus, as opposed to conflict-inducing reaction time tasks, the maintenance of correct performance during the motor sequence task is supported by iL`_ZXL_TNj processing with minimal demands from costly attentional and cognitive control circuits (James, 1890). However, in the case of occasional errors, these circuits are rapidly engaged to catch and correct deviations from the intended behavior. Optimization of such dynamic interplay between automatic and controlled behavior is presumably achieved through performance monitoring system with the dACC as its central node (Gehring et al., 1993;Botvinick et al., 2001).
Thus, we did not claim that task maintenance is irrelevant to the production of self-guided skilled movements, but rather that, under such conditions, successful task performance does not rely heavily on attentional systems, including the dACC. In support of this viewpoint, our results suggest that, whereas the dACC is not significantly activated during errorless periods (Fig. 6), this region, together with other closely affiliated functional areas, is rapidly engaged once participants fail to maintain the correct performance of the task and commit an error. The salience account further suggests that the dACC supports task performance not by being actively involved in task maintenance or in post-error performance recovery per se. Instead, the dACC role is to iTYQZ]Xj _SP ]P^_ ZQ _SP M]LTY about situations that may require increased attention and rapid adjustments in the ongoing behavior.
In the manuscript, we use the term i_TXP-on-_L^Vj to refer to time passage between block initiation and error commission. The potential differences in activity levels due to thT^i_TXP-on-_L^Vj intervals, changes in activity during periods with errors were assessed against control condition comprising errorless periods at the same within-block positions as errors.
As mentioned above, we now understand that by expressing his concern about the i_TXP-on-_L^Vj' the reviewer may have in mind the possibility that activity increases within the dACC are related to a longer processing time to generate the keypresses themselves, rather than to salience PES. We conducted additional analyses and found no evidence for a link between activity levels within the dACC and longer transition durations.

Reviewer
If time-on-task is genuinely irrelevant to the sequence task in the manuscript, including sequence time (four within-sequence transitions as well as the between-sequence transition) as a parametric modulator should have minimal effect on the analyses. However, this analysis seems necessary in order to rule out this alternative explanation for the authors' findings.
IZ^L_T^Qd _SP ]PaTPbP]k^]P\`P^_' be generated statistical models including transition durations as parametric modulators. These parametric predictors capture the variance related to transition durations that are not captured by the trial regressors. Note that the trial regressors were modeled with the stick function (i.e., zero duration) time-locked to the first trial keys or, in a case of between-trial transitions, to the last trial keys. However, the inclusion of parametric modulators REVIEWER #2 8 in the model does not control for the potential effect of various transition durations on the changes in the BOLD signal captured by the trial regressors (Büchel et al., 1998). Grinband and colleagues (2008) have shown that the variable epoch model with actual trial durations is a more physiologically plausible representation of changes in activity associated with response time. Therefore, we conducted an additional analysis estimating trial-related changes in the BOLD signal using brief epoch-related design with trial duration. The results derived from