Author response:
The following is the authors’ response to the original reviews.
Reviewer #1 (Public review):
Summary:
The "multiple-demand" (MD) system is a well-known finding of human brain imaging and is thought to play a central role in cognitive control. To directly compare the MD system in humans and monkeys, Mione et al. used functional magnetic resonance imaging to measure whole-brain activation in a multi-step saccadic maze task. In humans, the authors found a distributed pattern of brain activity close match to the canonical MD network and extends to adjacent regions of dorsal attention and other networks. While there was good correspondence between monkey and human data, differences were also notable in the lateral frontal cortex, the dorsal parietal cortex, and the sensorimotor cortex.
Strengths:
Though previous data hint at a corresponding network in the macaque, there has been no direct comparison to human data. This study provides a direct cross-species comparison with whole-brain data from fMRI, and the findings suggest an extended and strongly interconnected brain network recruited by increased cognitive challenge.
Weaknesses:
In previous human imaging, the MD system is defined by overlapping activation for many kinds of cognitive demands. In the present work, however, the authors used just a single task. Although there is some overlap between the putative monkey MD network and the canonical MD network identified in human imaging, there should be caution in linking current findings to the MD system based on limited task events.
In the Discussion, we acknowledge the limitation of using a single task. With this one task, however, the canonical MD network is clearly shown in our human data. Accompanying activation, especially of the canonical dorsal attention network, likely reflects the specific spatial demands of the maze task. In the monkey data, much of this dorsal attention activity is not seen (e.g. superior and medial parietal), likely reflecting limited power. Instead, there is distributed overlap with the previous limited monkey studies than have contrasted higher with lower cognitive demand. Though we agree that task-specific activations may contribute to our results, these arguments suggest that, in large part, our method does successfully identify a distributed set of multiple-demand regions. At the same time we acknowledge the desirability of further work to examine a wider range of task demands.
Reviewer #1 (Recommendations for the authors):
(1) Though the whole-brain data obtained by fMRI can provide a direct comparison between species, a single cognitive task might be insufficient to link the findings with the MD system. A cognitively challenging task likely activates multiple regions of the MD system; however, it may also recruit some task-specific regions, which do not belong to the canonical MD network. Furthermore, this is probably the reason that the dorsal attention network showed the strongest activation rather than the core MD for the current visuospatial maze task.
In the Discussion, we acknowledge the limitation of using a single task (p. 15-16), and the likely contribution of task-specific activations to our data, especially involving the dorsal attention network (p. 15).
(2) Ideally, a meta-analysis recruiting more fMRI studies on humans and monkeys when they perform various similar cognitive tasks may strengthen the evidence that there is a comparable MD system between species.
Many human meta-analyses, of course, show the common MD system. For monkeys, however, at least to our knowledge, there are insufficient studies for a similar meta-analysis. Instead we discuss overlaps between the current activation findings and two previous studies of respectively antisaccades and task switching (p. 15), suggesting that, in monkey as in human, there is convergence for different kinds of demand.
(3) Different from human subjects, monkeys usually require substantial training before fMRI scanning. More details about the training of the two monkeys should be given. In addition, training may reshape cognitive task activations. This potential impact should also be discussed.
We now address this point in the Discussion (p. 18). As we note, similar results for the two species apparently survive even large differences in protocol. A training summary has been added to Methods (p. 27).
(4) According to the description in the text, the two monkeys have obvious differences in cognitive task activations. It is necessary to show individual-level brain activations as well.
Individual results and a conjunction map are shown in Supplementary Figure 6 (see accompanying text on p. 13). As expected, the conjunction map had substantial similarity to the findings from the two animals combined. Similarities and differences between animals are addressed in the Discussion (p. 17).
(5) Based on the current research content, the title seems to be too general.
For the reasons given above (see point 2), we think our title is reasonable.
Reviewer #2 (Public review):
Summary:
Mione et al. aim to resolve a long-standing question in comparative neuroscience: whether the macaque brain contains a functional analogue to the distributed human multiple demand (MD) network. To address this, the authors employ a direct cross-species fMRI comparison using a multi-step saccadic maze task in humans and a simplified two-step version in macaques. By contrasting goal-directed navigation against a control condition that requires similar motor responses but no strategic planning, the study isolates the neural signatures of cognitive control across species.
Strengths:
The most compelling aspect of this work is its methodological alignment. Previous attempts to compare these systems often relied on comparisons of human BOLD signals and macaque single-unit recordings. By running parallel fMRI protocols, the authors establish a shared measurement basis that allows for a more direct comparison. The resulting activation maps clearly demonstrate conserved network topology across dorsomedial frontal, lateral, and medial parietal, and insula cortices. Combining these results with recent research on functional and structural connectivity further supports the idea that these networks evolved across species and provides a helpful starting point for future comparative studies. The findings will be highly useful for researchers investigating the evolutionary origins of domain-general cognitive control, as well as for neuroimaging methodologists developing cross-species alignment pipelines.
Weaknesses:
However, there are several differences in how the two groups were studied that make it harder to compare the results precisely. The human task mixed 2-, 4-, and 6-step trials within the same experimental blocks, whereas macaques performed only 2-step trials. This design difference likely places human participants in a state of sustained proactive cognitive control (Braver, 2012), as they must remain prepared for highly demanding trials at any moment. This elevated baseline arousal may artificially inflate MD network activation during the simpler 2-step trials in humans, making direct magnitude comparisons with the macaque data difficult.
This is a reasonable concern, which we note in the Discussion. Crucially, as we point out, similarities between species appear to survive this and other differences in procedure.
Additionally, the general linear model combined correct and error trials into a single regressor. Given that macaques exhibited substantially higher error rates, this approach risks diluting task-specific planning signals with activity related to error monitoring and reward prediction errors. The preprocessing pipeline also applied a 4 mm full-width half-maximum smoothing kernel to macaque data acquired at 1.5 mm resolution. Relative to the smaller size of the macaque brain, this kernel is quite large and likely blurs fine-grained topographical distinctions. This may partly explain why the macaque lateral frontal cortex shows a single dorsal activation patch rather than multiple discrete patches seen in humans.
These are also reasonable concerns. To address them, we have added supplementary analyses using (a) only correct trials, and (b) a smaller smoothing kernel (see Supplementary Figure 3 and 4). In both cases, there is some loss of power, but otherwise similar results.
Furthermore, there is concerning inter-individual variability in the macaque data. Normally, a functional network like the MD system is identified by consistent activation across all individuals. In this study, however, the two monkeys show substantially different activation maps and behavioral patterns. This lack of consistency renders the group-level results questionable, as it is unclear whether the group-level map represents a unified biological system or merely an average of disparate individual maps.
Results for individual animals have now been supplemented with a conjunction map (Supplementary Figure 6). As expected, the conjunction map is similar to the map obtained by pooling data from the two animals.
Finally, the subcortical activations shown in Figure 7 require more precise anatomical localization to confidently distinguish cerebellar nodes from adjacent brainstem structures.
The slices shown in Figure 7 have been amended to better show the detail of activation outside cerebral cortex, in particular in cerebellum.
The authors demonstrate a broad functional correspondence between human and macaque cognitive control networks, moving the field beyond speculative homology. The data suggest that an extended, interconnected network is recruited by cognitive challenge in both species; however, the strength of this claim is limited by the inter-individual variability and methodological constraints noted above. Assertions of precise topological equivalence should therefore be tempered. The absence of ventrolateral prefrontal and strong dorsal parietal activations in the macaque group analysis may reflect genuine biological differences, but could also stem from limited statistical power, excessive smoothing, or task design asymmetries. While the overall conclusions are plausible, they would be significantly strengthened by a more explicit discussion of these limitations and additional analytical clarifications regarding individual-level consistency.
Indeed, as we note in the Discussion, there is a strong possibility that ventrolateral frontal and dorsal parietal activation are missing in our monkey data because of limited statistical power. Further work would be needed to address these possible limitations.
Reviewer #2 (Recommendations for the authors):
(1) Please discuss how the mixed-difficulty block design in humans may induce a state of proactive cognitive control that elevates baseline MD activation compared to the fixed 2step macaque condition. A supplementary analysis comparing early versus late block 2-step trials in humans would help clarify whether activation magnitudes reflect sustained task-set maintenance or transient trial demands.
We now note the potential importance of this in the Discussion (p. 18), and point out that similarities between species appear to survive this difference in procedure.
(2) Please clarify the rationale for combining correct and error trials in a single regressor.
Given the higher macaque error rates, please provide a supplementary GLM restricted to correct trials only for the monkey data. This will demonstrate whether the core activation topography remains consistent when error-related signals are excluded.
A new analysis addresses this point (Supplementary Figure 3). As we note, restricting analysis to correctly-completed problems somewhat reduces power but leaves major features of the results intact.
(3) Please justify the use of a 4 mm FWHM smoothing kernel for macaque data, given cortical thickness and brain size differences. If feasible, re-run the analysis with a smaller kernel or surface-based smoothing to assess whether finer topographical distinctions emerge in lateral frontal and parietal cortices.
This new analysis has also been run (Supplementary Figure 4), again with reduced power but major features of the results intact. We also refer to previous work indicating choice of either 3 mm or 4 mm smoothing for macaque fMRI (p. 13), approximately matching the smoothing needed to align electrophysiological and fMRI maps (Issa et al., 2013, J.Neurosci.).
(4) Please detail how the human '2-step problems only' analysis in Supplementary Figure 1 was specified in the GLM. Explicitly state whether 4- and 6-step trials were modeled as separate regressors of no interest to prevent hemodynamic bleed-over from contaminating the 2-step beta weights.
Indeed, 4- and 6-step trials were removed using regressors of no interest, as now specified in Methods (p. 24).
(5) Please provide higher-resolution slices or probabilistic atlas overlays for the cerebellar and subcortical activations in Figure 7. This will help clearly distinguish cerebellar hemispheres from adjacent brainstem structures and ensure anatomical labeling is accurate.
To address this question, additional slices have been added to a revised Figure 7, in particular adding detail to cerebellar activation.
(6) Please address the substantial inter-individual variability in the macaque data, particularly regarding Monkey B's performance. Even after excluding five poor-performing sessions, Monkey B only reached about 65% accuracy on the first step of the maze task. This suggests that the animal was largely guessing rather than following a strategic, goal directed plan. The notably higher accuracy on step 2 (>80%) could simply be a selection bias artifact, as the task terminates immediately following step 1 errors, meaning step 2 trials only occur when step 1 was already correct. Including an animal that likely did not fully grasp the overarching task rule in a cohort of N=2 raises serious concerns about signal dilution and the reliability of the group-level maps. Please explicitly justify why Monkey B's data were retained despite these performance concerns, or consider excluding this animal and acquiring data from a third, behaviorally reliable subject. At minimum, report a conjunction map showing regions strictly active in both animals alongside the combined analysis, and present individual subject maps more prominently to improve transparency.
Though we appreciate this concern, we do not think the data indicate that monkey B failed to use the maze goal to constrain choices. As shown in Figure 3, the great majority of errors were timing errors (mostly not waiting for go signal). Excluding these, for step 1, of choices directed to one of the two available alternatives, 88% were correct (Figure 4, compare “correct” with “wrong open location”).
As noted above, we have added a conjunction map (Supplementary Figure 6) to our previous presentation of individual data. We note (p. 13) that, as expected, this map strongly resembles results from the combined-animal analysis.
