Thalamocortical contributions to cognitive task activity

Article activity feed

1. 

   Version published to 10.7554/elife.81282 on eLife

   Dec 20, 2022
2. 

   eLife

   Nov 28, 2022

   Author Response

   Reviewer 1 (Public Review):

   1. The finding that thalamic activity exhibits a low dimension structure is in my opinion less of a finding, but rather an assumption that motivates the use of dimensionality reduction techniques. When the authors ask (line 101) "whether thalamic task activity exhibits similar low dimensional structure", what is the alternative hypothesis? I think it is a foregone conclusion that with a restricted number of tasks, and the intrinsic smoothness of fMRI activity data, there are always K<<N components that capture 50,75, 90% of the variance. If you had measured the spiking of the entire population of thalamic neurons or increased the threshold to 99%, the structure of activity would be more high dimensional. So I believe you can either frame this as an assumption going in, or you build carefully …

   Author Response

   Reviewer 1 (Public Review):

   1. The finding that thalamic activity exhibits a low dimension structure is in my opinion less of a finding, but rather an assumption that motivates the use of dimensionality reduction techniques. When the authors ask (line 101) "whether thalamic task activity exhibits similar low dimensional structure", what is the alternative hypothesis? I think it is a foregone conclusion that with a restricted number of tasks, and the intrinsic smoothness of fMRI activity data, there are always K<<N components that capture 50,75, 90% of the variance. If you had measured the spiking of the entire population of thalamic neurons or increased the threshold to 99%, the structure of activity would be more high dimensional. So I believe you can either frame this as an assumption going in, or you build carefully an alternative hypothesis of what a "high-dimensional" structure would look like. Generating activity data i.i.d would be the simplest case, but given that both signal and measurement noise in fMRI are reasonably smooth, this would be a VERY trivial null hypothesis.

   We thank the reviewer for pointing out this inherent assumption in our analysis. We agree that given the smoothed nature of BOLD signal and the restricted task design we likely cannot effectively test an alternative high dimensional organization hypothesis. We have revised our introduction accordingly and clarify that we use a dimensionality reduction technique with the assumption that we will observe a low dimension structure of thalamic task fMRI data, similar to past fMRI studies that focused on cortical ROIs (line 102). Furthermore, we have revised the discussion section to remove discussion highlighting the low-dimension organization as a novel finding (line 404).

   1. The measure of "task hub" properties that is central to the paper would need to be much better explained and justified. You motivate the measure to be designed to find voxels that are "more flexibly recruited by multiple thalamic activity components", but it is not clear to me at this point that the measure defined on line 634 does this. First, sum_n w_i^2 is constrained to be the variance of the voxel across tasks, correct? Would sum_n abs(w) be higher when the weights are distributed across components? Given that each w is weighted by the variance (eigenvalue) of the component across the thalamus, would the score not be maximal if the voxel only loaded on the most important eigenvector, rather than being involved in a number of components? Also, the measure is clearly not rotational invariant - so would this result change after some rotation PCA solution? Some toy examples and further demonstrations that show why this measure makes sense (and what it really captures) would be essential. The same holds for the participation index for the resting state analysis.

   Please see our response to essential revision point #1.

   1. For the activity flow analysis, the null models (which need to be explained better) appear weak (i.e. no differences across tasks?), and it is no small wonder that the thalamus does significantly better. The Pearson correlations are not overwhelmingly impressive either. To give the reader a feel for how good/bad the prediction actually is, it would be essential that the authors would report noise ceilings - i.e. based on the reliability of the cortical activity patterns and thalamic activity patterns, what correlation would the best model achieve (see King et al., 2022, BioRxiv, as an example).

   Please see our response to essential revision point #4.

   1. Overall it has not been made clear what the RDM analysis adds to the prediction of the actual activity patterns. If you predicted the activity patterns themselves up to the noise ceiling, you would also hit the RDM correctly. The opposite is not the case, you could predict the correct RDM, but not the spatial location of the activity. However, the two prediction performances are never related to each other and it remains unclear what is learned from the latter (less specific) analysis.

   We agree that the utility of the RDM analysis is not clear, and we have removed it from the manuscript.

   Read the original source
3. 

   eLife

   Nov 28, 2022

   eLife assessment

   This valuable study examines a largely ignored brain structure (the thalamus) in functional brain imaging studies. In general, the study shows convincing evidence from the reanalysis of two task-based MRI studies that localized thalamic regions show hub properties in terms of their activation properties and connectivity to cortical regions. While the strength of the study is that converging evidence was shown across two large data sets, the empirical support for some of the claims in the current version remains incomplete.

   Read the original source
4. 

   eLife

   Nov 28, 2022

   Reviewer #1 (Public Review):

   Using two openly available multi-task fMRI datasets, the authors decompose thalamic activity into a smaller set of components. They show that voxels with higher loadings on the main components (high task hub property) also have a high participation coefficient as derived from resting state data. Cortical activity patterns can be predicted to some degree from thalamic activity patterns, and generally better than from a number of other cortical areas. This prediction relies mainly on the voxels with high task hub scores. The results are valuable and methodological generally solid, with some aspects being incomplete.

   1. The finding that thalamic activity exhibits a low dimension structure is in my opinion less of a finding, but rather an assumption that motivates the use of dimensionality reduction techniques. …

   Reviewer #1 (Public Review):

   Using two openly available multi-task fMRI datasets, the authors decompose thalamic activity into a smaller set of components. They show that voxels with higher loadings on the main components (high task hub property) also have a high participation coefficient as derived from resting state data. Cortical activity patterns can be predicted to some degree from thalamic activity patterns, and generally better than from a number of other cortical areas. This prediction relies mainly on the voxels with high task hub scores. The results are valuable and methodological generally solid, with some aspects being incomplete.

   1. The finding that thalamic activity exhibits a low dimension structure is in my opinion less of a finding, but rather an assumption that motivates the use of dimensionality reduction techniques. When the authors ask (line 101) "whether thalamic task activity exhibits similar low dimensional structure", what is the alternative hypothesis? I think it is a foregone conclusion that with a restricted number of tasks, and the intrinsic smoothness of fMRI activity data, there are always K<<N components that capture 50,75, 90% of the variance. If you had measured the spiking of the entire population of thalamic neurons or increased the threshold to 99%, the structure of activity would be more high dimensional. So I believe you can either frame this as an assumption going in, or you build carefully an alternative hypothesis of what a "high-dimensional" structure would look like. Generating activity data i.i.d would be the simplest case, but given that both signal and measurement noise in fMRI are reasonably smooth, this would be a VERY trivial null hypothesis.

   2. The measure of "task hub" properties that is central to the paper would need to be much better explained and justified. You motivate the measure to be designed to find voxels that are "more flexibly recruited by multiple thalamic activity components", but it is not clear to me at this point that the measure defined on line 634 does this. First, sum_n w_i^2 is constrained to be the variance of the voxel across tasks, correct? Would sum_n abs(w) be higher when the weights are distributed across components? Given that each w is weighted by the variance (eigenvalue) of the component across the thalamus, would the score not be maximal if the voxel only loaded on the most important eigenvector, rather than being involved in a number of components? Also, the measure is clearly not rotational invariant - so would this result change after some rotation PCA solution? Some toy examples and further demonstrations that show why this measure makes sense (and what it really captures) would be essential. The same holds for the participation index for the resting state analysis.
   3. For the activity flow analysis, the null models (which need to be explained better) appear weak (i.e. no differences across tasks?), and it is no small wonder that the thalamus does significantly better. The Pearson correlations are not overwhelmingly impressive either. To give the reader a feel for how good/bad the prediction actually is, it would be essential that the authors would report noise ceilings - i.e. based on the reliability of the cortical activity patterns and thalamic activity patterns, what correlation would the best model achieve (see King et al., 2022, BioRxiv, as an example).
   4. Overall it has not been made clear what the RDM analysis adds to the prediction of the actual activity patterns. If you predicted the activity patterns themselves up to the noise ceiling, you would also hit the RDM correctly. The opposite is not the case, you could predict the correct RDM, but not the spatial location of the activity. However, the two prediction performances are never related to each other and it remains unclear what is learned from the latter (less specific) analysis.

   Read the original source
5. 

   eLife

   Nov 28, 2022

   Reviewer #2 (Public Review):

   This study investigates how thalamic functional MRI activations change across subjects performing many cognitive tasks. The results reveal localised regions in anterior, medial (and potentially posterior) portions of the thalamus that co-activate most consistently across multiple tasks. The authors then try to link these task hubs to cortical association cortices, first by showing that association cortices are most connected to thalamic task hubs. Second, by showing that thalamic activations can predict

   The findings are important, mainly because thalamic fMRI activations are largely ignored by the current literature. The major strengths of the study lie in examining thalamic activations under many cognitive tasks and replicating results across two independent datasets.

   The findings of thalamic hubs are …

   Reviewer #2 (Public Review):

   This study investigates how thalamic functional MRI activations change across subjects performing many cognitive tasks. The results reveal localised regions in anterior, medial (and potentially posterior) portions of the thalamus that co-activate most consistently across multiple tasks. The authors then try to link these task hubs to cortical association cortices, first by showing that association cortices are most connected to thalamic task hubs. Second, by showing that thalamic activations can predict

   The findings are important, mainly because thalamic fMRI activations are largely ignored by the current literature. The major strengths of the study lie in examining thalamic activations under many cognitive tasks and replicating results across two independent datasets.

   The findings of thalamic hubs are compelling. However, this current version of the manuscript could be strengthened by providing better links with the wider literature (e.g. with thalamic resting-state networks). The study also falls short in properly quantifying the similarity of findings across the two independent datasets. The subtle discrepancies between the results of the two datasets throughout the manuscript could point to finer-grained fractionations of the identified thalamic hubs. The least compelling set of results (though not necessarily wrong) is the thalamic prediction of cortical activations. This is because the functional connectivity (FC) matrix used to link the thalamus and cortex was derived from the same data after regressing out task-related variance. However, this process might not be clean enough. A stronger test would utilize an FC matrix derived from an independent dataset.

   Read the original source
6. 

   Version published to 10.1101/2022.06.28.497905v1 on bioRxiv

   Jul 1, 2022