Distributed neural activity supports a broad range of perceptual, motor, affective, and cognitive functions. Discovering how brain systems implement this broad behavioral repertoire is critical for elucidating the neural basis of human cognition. Past studies have revealed two organizational principles – low-dimensional architecture and multi-task hubs – that connect distributed neural activity with task performance across functional domains.

The application of dimensionality reduction techniques on whole-brain imaging data has revealed a relatively low-dimensional organization of cortical neural activity. Specifically, the number of variables required to explain a large amount of variance in distributed neural activity is far lower than the total number of variables in the data (MacDowell and Buschman, 2020; Nakai and Nishimoto, 2020; Shine et al., 2019b). The spatiotemporal patterns of these variables are commonly referred to as intrinsic networks in task-free conditions, or manifolds, motifs, and latent components in task contexts. This low-dimensional organization may reflect an elementary set of information processes implemented by distributed brain systems (Cunningham and Yu, 2014; Yeo et al., 2016), in which different tasks selectively engaged selective latent activity patterns depending on the specific processing requirements of individual tasks.

The anatomical overlap between spatiotemporal components predicts that some brain regions broadly participate in multiple tasks. In support of this prediction, studies have found that associative regions in frontal and parietal cortices are involved in executing a wide array of tasks (Cole et al., 2013; Duncan, 2010). These task-flexible regions, also commonly referred to as brain hubs (Gratton et al., 2018; van den Heuvel and Sporns, 2013), have diverging connectivity with multiple brain systems, and are thought to perform integrative functions that allow perceptual inputs to interact with contextual task representations for adaptive task control (Bertolero et al., 2018; Bertolero et al., 2015; Ito et al., 2022; Nee, 2021). The behavioral significance of brain hubs is affirmed by lesion studies demonstrating that lesions to hub regions are associated with task impairments across multiple functional domains (Hwang et al., 2021; Reber et al., 2021; Warren et al., 2014).

However, a prevailing assumption is that diverse human behavior depends on the organization of cortical activity, and the contribution from subcortical regions, particularly the thalamus, is not well understood. Our previous studies demonstrated that the human thalamus contains a complete representation of intrinsic cortical functional networks, and additionally, exhibits a hub-like connectivity profile interlinking multiple cortical systems (Hwang et al., 2021; Hwang et al., 2017). Furthermore, lesions to the anterior and the medial thalamus in human patients are associated with behavioral impairments across functional domains (Hwang et al., 2020a; Hwang et al., 2021). Given that every cortical region receives projections from multiple thalamic nuclei and that the thalamus mediates striatal and cerebellar influences on cortical (Shine, 2020), the thalamocortical system is ideally suited to shape cortex-wide activity patterns that instantiate cognitive representations. The relationship between thalamic task activity and cortical cognitive representations, however, are not well understood.

Given the observations described above, we expect that task functional magnetic resonance imaging (fMRI) data obtained from the human thalamus to exhibit a similar low-dimensional organization like the cortical system. Furthermore, there should be task hub regions within the human thalamus that broadly participate in multiple cognitive tasks. The current study applied a dimension reduction technique to determine how thalamic task-evoked activity and thalamic hubs are organized to support cognitive functions across diverse functional domains. To this end, we analyzed fMRI data where human subjects performed a rich battery of tasks designed to elicit neural activity for a wide range of cognitive functions. Our study addressed the following two specific questions. First, are there thalamic hub regions that exhibit task-evoked activity by multiple tasks across domains? Second, can thalamic task-evoked activity be used to predict cortical task-evoked activity patterns? Answering these questions would reveal a general principle of how the human thalamocortical system contributes to higher-order, multi-domain cognitive task activity.

To determine the organization of task-evoked thalamic activity, we analyzed two fMRI datasets, both of which utilized rich batteries of tasks designed to elicit a wide range of processes encompassing perceptual, affective, memory, social, motor, language, and cognitive domains (Figure 1A). In the first dataset, the multi-domain task battery (MDTB) dataset, 21 subjects performed 25 behavioral tasks (King et al., 2019). In the second dataset, the Nakai and Nishimoto (N&N) dataset, 6 subjects performed 103 tasks (Nakai and Nishimoto, 2020).

Figure 1 with 2 supplements see all

Download asset Open asset

Hub organization of thalamic task-evoked activity

Examining the spatial patterns of thalamic components (Figure 1B) revealed that several thalamic subregions showed spatial overlap across multiple components, which suggests that these thalamic subregions may support multiple tasks. We reasoned that if a thalamic voxel is broadly participating in multiple tasks, it will express stronger weights for top thalamic activity components that explained a large amount of variance in evoked activity patterns across tasks. Therefore, to further map ‘task hub’ regions in the human thalamus, we calculated a metric by summing each voxel’s absolute component weight for the top 10 components. Results showed that for the MDTB dataset, the anterior, medial, medio-posterior, and dorsal thalamus exhibited strong task hub properties (Figure 2A). For the N&N dataset, the posterior thalamus did not exhibit strong task hub property, and the spatial correlation between task hub metrics from the two datasets was r=0.36. We then recalculated the task hub metric using lower ranked components (components number 11th to 20th, explained less than 2% of variances), and results revealed a different anatomical pattern (Figure 2—figure supplement 1). These results suggest that the anterior, medial, and posterior thalamic task hub patterns we identified were primarily driven by component weights summed across top components that explained large portion of variance in task-evoked activity.

Figure 2 with 2 supplements see all

Download asset Open asset

Network hubs in the human thalamus exhibit diverse functional connectivity (FC) patterns interlinking multiple cortical systems (Greene et al., 2020; Hwang et al., 2017). To determine whether the tasks hubs we identified correspond to network hub regions previously identified in the thalamus, we compared the spatial similarity of task hubs with FC hubs. We calculated a network hub metric, the participation coefficient (Gratton et al., 2012; Guimerà and Nunes Amaral, 2005; Figure 2—figure supplement 2), using every thalamic voxel’s FC matrix with 400 cortical regions (Schaefer et al., 2018). When calculating FC matrices, we used timeseries after removing task-evoked variances from the preprocessed fMRI data, to reduce shared variances in task-evoked activations biasing FC estimates (see Methods for details). The task hubs showed significant spatial correspondence with FC hubs (Figure 2A; spatial correlation: MDTB mean = 0.17, SD = 0.083, p<0.001; N&N mean = 0.17, SD = 0.05, p<0.001; spatial correlation between MDTB and N&N: r=0.55). One discrepancy was the posterior thalamus, which showed strong task hub but not network hub property for the MDTB dataset. We then projected the task hubs estimates onto the cortex by calculating the dot product between each thalamic voxel’s task hub estimate and its thalamocortical FC matrix. We found that tasks hubs were most strongly coupled with associate regions in the frontal and parietal cortex, for example, the lateral frontal, the insula, the dorsal medial prefrontal, and the intraparietal cortices (Figure 2B; spatial correlation between MDTB and N&N: r=0.46).

Given that the thalamus is not a homogenous structure and can be divided into different subdivisions or nuclei based on its functional and histological properties, we summarized the spatial distribution of task hub estimates (CompW) using two different thalamic atlases (Figure 3). First, the Morel atlas was used to determine the location of different thalamic nuclei Krauth et al., 2010; this atlas maps human thalamic nuclei based on cyto- and myelo-architecture information in stained slices from five postmortem human brains, and further transformed to the template space. Second, for the network atlas, we assigned each thalamic voxel to one of the seven canonical cortical function networks (Schaefer et al., 2018) that we used for calculating network hubs. We then averaged the task hub estimates within each thalamic nucleus or functional parcellation. In both datasets, we found the task hub values to be highest in the anterior and mediodorsal nuclei. These nuclei groups are known to have strong reciprocal connectivity with frontal regions (Giguere and Goldman-Rakic, 1988; Selemon and Goldman-Rakic, 1988). We also found the medial pulvinar to show high task hub value but only in the MDTB dataset, likely because the N&N dataset did not show high task hub value in the posterior thalamus. For the functional network parcellation atlas, between datasets discrepancies were also found in posterior thalamus. For example, in the MDTB dataset, the DF parcellation (which overlapped with the anterior medial, medial, and posterior medial thalamus) was found to show the highest task hub value. For the N&N dataset, the posterior thalamus did not exhibit high task hub value, but the FP parcellation, which covered the dorsal bank of the medial thalamus, showed the highest task hub value. Overall, these results suggest that thalamic task hubs do not overlap with one functional network parcellation or thalamic nucleus, instead overlap with multiple thalamic parcellations and nuclei.

Figure 3

Download asset Open asset

Predict cortical task activity with thalamic task-evoked activity

To further test whether thalamic task-evoked activity is related to diverse cognitive functions putatively implemented by distributed cortical activity, we adapted the activity flow mapping procedure (Cole et al., 2016; Ito et al., 2017) to test whether thalamic task-evoked responses can be used to predict cortical task activity (Figure 4A). Briefly, for every thalamic voxel, we calculated the dot product between the thalamic-evoked response pattern of each task and the thalamocortical FC matrix, using a split-half cross-validation procedure (see Methods). This calculation yielded a predicted cortex-wide activity pattern for every task, and prediction accuracy was calculated by comparing the predicted pattern to the observed cortical activity pattern using Pearson correlation. We then compared the model performance against three null models. The first null model randomly shuffled thalamic-evoked responses (the ‘null model’), which assumed no spatial structure in thalamic task-evoked responses. The second null model set all thalamic-evoked responses to the same value (‘1’) across all voxels (the ‘uniformed evoked model’), which assumed that cortical activity patterns are determined only by the summated inputs from thalamocortical FC patterns. The third null model was constructed by averaging the evoked response amplitudes across tasks for every thalamic voxel, which assumed that there is no difference in thalamic activity patterns between tasks. Finally, activity flow model performance is limited by the reliability of the data used to calculate the predicted pattern. Therefore, we calculated the noise ceiling that accounted for the split-half reliability of thalamic-evoked responses, thalamocortical FC, and cortical-evoked responses (see Methods for details). All prediction accuracy was normalized by the noise ceiling. For the null models, prediction accuracy was normalized by the thalamocortical model’s noise ceiling.

Figure 4 with 1 supplement see all

Download asset Open asset

The mean prediction accuracy of the thalamocortical activity flow model was 0.29 for the MDTB dataset and 0.15 for the N&N dataset (Figure 4B). As described above, the prediction accuracy is limited by the reliability of the thalamic and cortical data (noise ceiling). After accounting for the noise ceiling (Figure 4—figure supplement 1), the normalized prediction accuracy was 0.5 for MDTB and 0.62 for N&N (Figure 4C), which indicates that the thalamocortical model was able to account for up to 38% of variance in task-evoked activity patterns in the cortex. We further found that thalamocortical activity flow model outperformed the null models, except for the N&N dataset when compared to the averaged evoked pattern model (Figure 4C; MDTB thalamocortical model vs. null model: t(20) = 11.22, p<0.001; MDTB thalamocortical model vs. uniformed evoked model: t(20) = 12.05, p<0.001; MDTB thalamocortical model vs. averaged pattern model: t(20) = 6.88, p<0.001; N&N thalamocortical model vs. null model: t(5) = 11.22, p=0.018; N&N thalamocortical model vs. uniformed evoked model: t(5) = 4.75, p=0.005; N&N thalamocortical model vs. averaged pattern model: t(5) = 1.61, p=0.17). These results indicate that thalamic task-evoked responses can predict cortical activity patterns.

We then sought to compare our thalamocortical activity flow model to other cortical and subcortical regions. We constructed a series of comparison models by applying the same split-half cross-validation activity flow mapping procedure to 100 cortical regions of interests (ROIs) that had similar size as the thalamus (Schaefer et al., 2018), as well as to other subcortical regions including the hippocampus, the caudate, the putamen, and the globus pallidus. We found that several cortical regions also showed strong prediction performance, including the insula, the inferior frontal cortex, the dorsal medial frontal cortex, the intraparietal sulcus, and the lateral occipital cortex (Figure 4D for unnormalized prediction accuracy, Figure 4E for noise normalized prediction accuracy). However, the thalamocortical model outperformed all comparison models. For example for both datasets, lateral ROIs in the visual network showed the highest normalized prediction accuracy, but still statistically weaker than the thalamocortical activity flow model (MDTB: t(20) = 2.91, p=0.008; N&N: t(5) = 3.3, p=0.021). These results indicate that the thalamocortical activity flow model is a strong predictor of cortical task activity.

Lesions to task hubs in the thalamus are associated with impaired prediction of task reorientations and representational geometry

The results presented in Figure 4 demonstrated that thalamic task-evoked responses can predict cortical activity patterns that putatively support cognitive representations. However, it is unclear to what degree task hubs in the thalamus contribute to these results. One possibility is that all thalamic voxels contribute equally to cortical task activity patterns. Another possibility is that, since thalamic task hubs are broadly engaged by multiple tasks and exhibited diverging connectivity patterns with multiple cortical systems, thalamic task hubs have a stronger contribution to predicting cortical task activity patterns. To evaluate this prediction, we performed virtual lesion simulations. Specifically, we systematically removed 20% of thalamic voxels based on their percentile rank of task hub estimates and calculated the percentage of reduction in prediction accuracy from the activity flow analysis. For both datasets, we found that removing voxels with strong task hub estimates decreased the prediction accuracy in cortical task-evoked activity patterns (Figure 5A). We fitted a regression model to test the relationship between reduction in prediction and percentile rank of task hub metrics removed, and found significant negative associations (MDTB activity flow prediction: b=–0.43, SE = 0.019, t=–22.63, p<0.001; N&N activity flow prediction: b=–0.1, SE = 0.028, t=–3.79, p<0.001). These results indicate that virtual lesioning of thalamic task hubs had a stronger impact on reducing the model’s ability to predict cortical activity patterns. Furthermore, thalamic voxels with the strongest impact on prediction performance were spatially located in the anterior, medial, mediodorsal, and medial posterior thalamus that we previously identified as task hubs (Figure 5B; spatial correlation between MDTB and N&N r=0.29).

Figure 5

Download asset Open asset

We compared these simulation results with neuropsychological impairments found in 20 human patients with focal thalamic lesions. Neuropsychological profiles of these patients were described in detail in our previous publication (Hwang et al., 2021). Briefly, patients performed a battery of neuropsychological tests to assess executive, language, memory, learning, visuospatial, and construction functions (Lezak et al., 2012). Test performance was then compared to published, standardized norms and converted to z-scores quantify the severity of impairment (Figure 6A). Patients were grouped into two groups, those that showed impairment (z-score <–1.695, worse than 95 percentile of the normative population) in single or fewer domains (the SM group) vs. those showed impairment across multiple domains (the MM group; Figure 6B). There were no statistically significant differences in the lesion volumes between these two groups of patients (MM group: mean = 1559 mm³, SD = 1370 mm³; SM group: mean = 1073 mm³, SD = 925 mm³; group difference in lesion size, t(18) = 0.87, p=0.39).

Figure 6

Download asset Open asset

Lesion sites from each group were compared to the effects of the simulated lesions presented in Figure 5. Specifically, we compared the percentages of reduction in voxels from these two groups of lesions. We found that simulated lesions to voxels that overlapped with lesion sites in the MM patient group showed larger reductions in task activity predictions for the N&N dataset but not the MDTB dataset (Figure 6C; MDTB: Komogorov-Smirnov test D=0.048, p=0.065; N&N: Komogorov-Smirnov test D=0.25, p<0.001). These empirical results from human patients support our simulated lesion analyses, suggesting that lesioning task hubs in the human thalamus are associated with behavioral deficits across functional domains.

It was hypothesized that the thalamus influences cognitive representations beyond sensory and motor domains (Halassa and Sherman, 2019; Wolff and Vann, 2019). Anatomically, every cortical region receives inputs from one or many thalamic subregions, and most thalamic subregions send signals to one or many cortical systems (Jones, 2001; Sherman, 2007). Functionally, resting-state fMRI studies have found a complete representation of intrinsic cortical networks in the human thalamus (Greene et al., 2020; Hwang et al., 2017; Yuan et al., 2016; Zhang et al., 2008). Several of these networks, such as the FP, CO, and DF, have been implicated in multiple cognitive functions (Sadaghiani et al., 2010; Seeley et al., 2007). These observations raised a broader question of if and how the human thalamus contributes to cognitive task activity across functional domains. The current study addressed this question by investigating how thalamic task-evoked activity are organized and how it can predict cortical cognitive task activity.

We identified several key organizational characteristics. First, the anterior, medial, and posterior-medial thalamus participated in multiple tasks, and these ‘task hubs’ spatially overlapped with FC hubs mapped in our previous studies (Hwang et al., 2021; Hwang et al., 2017). Second, thalamic task-evoked activity can predict task-specific activity patterns in the cortex via a linear thalamocortical activity flow model. Critically, when compared to comparison models and models built on other brain structures, this thalamocortical activity flow model performed the best in predicting cortical task activity, highlighting the capacity of the thalamus in influencing distributed activity patterns that may instantiate cognitive representations. Finally, findings from the thalamocortical activity flow model were further corroborated by simulated lesions and neuropsychological impairments observed in patients with focal thalamic lesions, affirming the behavioral significance of thalamocortical interactions. Collectively, these findings highlight a general and critical role the human thalamocortical system supporting cognitive functions.

Neural systems support a rich and diverse behavioral repertoire. Human functional neuroimaging and studies of animal models found a low-dimensional organization scheme in cortical neural activity (Beam et al., 2021; Karolis et al., 2019; MacDowell and Buschman, 2020; Nakai and Nishimoto, 2020; Shine et al., 2019b). Given this finding, multi-task fMRI data obtained from the human thalamus should exhibit similar property. In our study, we applied a dimension reduction procedure to decompose multi-task thalamic activity into low-dimensional task components. Notably, we found thalamic task components that explained large amount of variances across tasks expressed strongest weights in anterior, medial, and posterior thalamus, but not in lateral thalamic regions that overlap with first-order thalamic relays such as the lateral geniculate nucleus and the ventrolateral nucleus.

Because of its unique connectivity profile, the thalamus may be in an ideal position to influence cortex-wide task activity and the diverse cognitive functions cortical activity patterns instantiate. Because every cortical region receives inputs from one or multiple thalamic regions, thalamocortical connectivity may be more effective in pushing cortical activity patterns to the desired task state. We speculate that the thalamus may be at the nexus where specific activity patterns in the thalamus can be selectively down- or up-regulated to meet the specific information processing demands of different tasks. Functions that may be influenced by the thalamocortical system suggested by previous studies include selection and gain control (Halassa and Kastner, 2017), adjustment of inter-regional communication (Hwang et al., 2020b; Saalmann et al., 2012), and modulation of cortical excitability (Kosciessa et al., 2021). The small size of the thalamus further suggests a potential role in modulating the dimensionality of task-evoked cortical activity – for instance, previous studies have found that thalamic activity correlates with the dimensionality and the strength of cross-system coupling between cortical regions (Garrett et al., 2018; Shine et al., 2019b), and thalamic lesions disrupt the low-dimensional cortical network organization (Hwang et al., 2017).

The studies described above suggest a strong relationship between thalamic activity and cortical cognitive activity. We further tested whether we can predict cortical activity patterns based on thalamic task-evoked activity and thalamocortical FC pattern with cortical systems. We found that our data-driven thalamocortical activity flow model can indeed successfully predict task-specific activity patterns observed in the cortex, better than null models that assumed that the thalamocortical system does not carry task-specific information. While the observed prediction accuracy was lower than previous studies that focused on cortico-cortical activity flow models (Cole et al., 2016), most of those studies included multiple cortical regions to build the prediction model. Our study tested a different question to evaluate the predictive power of a single region, the thalamus, relative to other brain regions. We found that after accounting for the noise ceiling, thalamic activity alone was able to predict approximately 38% of the systematic variance of cortical activity patterns across tasks. Critically, we found that the thalamocortical model outperformed comparison models constructed with other cortical and subcortical structures, including frontal (the insula and the inferior frontal sulcus), parietal, and striatal regions previously implicated in adaptive control of tasks (Badre and Nee, 2018; Ito et al., 2022). This result suggests that the thalamocortical system exerts a strong influence on cortical activity that instantiates cognitive representations.

Placing the thalamus in a central position for influencing cortex-wide cognitive activity has several benefits. For instance, the thalamus consists of two classes of thalamocortical projection cells, ‘matrix’ and ‘core’ cells, where core cells target specific granular cortical structures, and matrix cells diffusely project to multiple brain regions (Jones, 2001). Given that different brain regions may be encoding different perceptual and cognitive information (Christophel et al., 2017), selectively modulating thalamic activity may facilitate targeted activations of cortical representations via associated thalamocortical interactions. The diffuse projection pattern of matrix cells in the thalamus may simultaneously activate multiple cortical regions and promote inter-system integration when required (Hwang and D’Esposito, 2022; Jones, 2001). Furthermore, the thalamus is modulated by inputs from other subcortical structures, such as the reticular nucleus, the basal ganglia, the superior colliculus, and the cerebellum (Bostan and Strick, 2018; Hwang et al., 2020b). This anatomical feature suggests that the thalamus may be an efficient target for exerting neuromodulatory (e.g., dopaminergic) influences on cognitive processes when required (Castro-Alamancos and Gulati, 2014; Garrett et al., 2022). Several computation models have leveraged these characteristics to explain the functions of basal-ganglia-thalamic circuits for higher-order cognitive functions (Jaramillo et al., 2019; O’Reilly and Frank, 2006).

In our previously study (Hwang et al., 2017), we found that the anterior, medial, and posterior thalamus to have strong network hub properties, exhibiting strong FC with multiple cortical systems in frontal and parietal cortices. Network hubs in the frontal and parietal cortices have also been found to broadly participate in cognitive tasks across different functional domains (Bertolero et al., 2015; Yeo et al., 2016), likely reflect its role in instantiating task representations (Schumacher and Hazeltine, 2016) and support cross-domain cognitive functions that arise from brain-wide network interactions (Gratton et al., 2018; van den Heuvel and Sporns, 2013). We hypothesized that network hubs in the thalamus may share this property with cortical network hubs. However, our previous study utilized resting-state FC, and multi-task fMRI data are needed to test task participation. In the current study, we analyzed task-evoked activity patterns in the thalamus and found task hub regions overlapped with network hubs in the anterior, medial, and posterior thalamic regions.

We found that thalamic task hubs were most strongly coupled with multiple regions across frontal and parietal associative cortices, including subregions that found to exhibit functional specialization for specific tasks as well as task-flexible regions (Yeo et al., 2016). Several selective regions within the frontal and parietal cortices have been hypothesized to be part of a flexible multiple-demand system (Assem et al., 2020) and thus, these thalamic subregions may be part of a core system whose functions are required to perform specific computations that are common across many different task contexts. One strong test of the behavioral relevance of these thalamic task hubs is to examine the effects of lesions. If thalamic task hubs are part of a domain-general core, lesioning task hubs should be associated with broad behavioral deficits not confined to a specific behavioral domain. Indeed, we found that simulated lesions to thalamic task hubs impaired prediction accuracy in cortical activity pattern across tasks, and stroke lesions to the anterior-medial thalamus in human patients were associated with behavioral deficits across functional. At this point we are agnostic to the specific function these thalamic task hubs perform. One hypothesis, suggested by findings from rodent models, is that the medial thalamus excites task-relevant and inhibits task-irrelevant cortical activity that encodes working memory representations (Mukherjee et al., 2021; Rikhye et al., 2018). Thus, in human subjects, thalamocortical interactions between thalamic task hubs and frontoparietal systems may control task representations that bind sensory, motor, and contextual information that are necessary for adaptive behavior across tasks (Schumacher and Hazeltine, 2016). This is an open question that should be evaluated in future studies.

On important limitation of our study is that our thalamocortical activity flow model cannot establish the directionality of thalamic activity influencing cortical activity. It is more likely that brain systems instantiate cognitive representations via recurrent connectivity among cortical regions and the thalamus, as well as integrating inputs from other brain structures. As discussed above, including all brain regions into the activity flow model can substantially improve model performance in predicting task-specific activity patterns (Cole et al., 2016). Nevertheless, our findings suggest that, across the whole brain, the thalamus likely has one of the strongest contributions to cortical task activity. Another important limitation is that the posterior thalamus exhibited strong task hub property in the MDTB dataset but not the N&N dataset. We suspect this could be related to differences in behavioral tasks across the two datasets, and speculatively, to the increased noise in the N&N dataset, given the lower number of observations per task and lower subcortical signal-to-noise ratio. One strong test of mapping task hubs is the lesion methods, and in our previous study (Hwang et al., 2021) we did not have good lesion coverage on the pulvinar nucleus. Future studies should test whether patients with pulvinar lesions have cognitive deficits beyond visual attention tasks (Snow et al., 2009). To conclude, our study reported important organizational principles of how thalamic task-evoked activity and thalamocortical connectivity supports cognitive tasks. These results highlight how the thalamus is in a central position for supporting cognitive task representations.

Raw data are available at OpenNeuro.org (https://openneuro.org/datasets/ds002105/ and https://openneuro.org/datasets/ds002306/). Code and data are available at (https://github.com/HwangLabNeuroCogDynamics/ThalamicTaskHubs; copy archived at swh:1:rev:d876ea93f1174ffdfe35c60f1f8b0ba8b637b037).

1. 1. King M
   2. Hernandez-Castillo CR
   3. Poldrack RA
   4. Ivry RB
   5. Diedrichsen J

   (2019) OpenNeuro

   multi-domain task battery.

   https://doi.org/10.18112/openneuro.ds002105.v1.1.0
2. 1. Nakai T
   2. Nishimoto S

   (2020) OpenNeuro

   Over 100 Task fMRI Dataset.

   https://doi.org/10.18112/openneuro.ds002306.v1.0.3

1. 1. Assem M
   2. Glasser MF
   3. Van Essen DC
   4. Duncan J

   (2020) A domain-general cognitive core defined in multimodally parcellated human cortex

   Cerebral Cortex 30:4361–4380.

   https://doi.org/10.1093/cercor/bhaa023

   • PubMed
   • Google Scholar
2. 1. Badre D
   2. Nee DE

   (2018) Frontal cortex and the hierarchical control of behavior

   Trends in Cognitive Sciences 22:170–188.

   https://doi.org/10.1016/j.tics.2017.11.005

   • PubMed
   • Google Scholar
3. 1. Beam E
   2. Potts C
   3. Poldrack RA
   4. Etkin A

   (2021) A data-driven framework for mapping domains of human neurobiology

   Nature Neuroscience 24:1733–1744.

   https://doi.org/10.1038/s41593-021-00948-9

   • PubMed
   • Google Scholar
4. 1. Behzadi Y
   2. Restom K
   3. Liau J
   4. Liu TT

   (2007) A component based noise correction method (compcor) for BOLD and perfusion based fMRI

   NeuroImage 37:90–101.

   https://doi.org/10.1016/j.neuroimage.2007.04.042

   • PubMed
   • Google Scholar
5. 1. Bertolero MA
   2. Yeo BTT
   3. D’Esposito M

   (2015) The modular and integrative functional architecture of the human brain

   PNAS 112:E6798–E6807.

   https://doi.org/10.1073/pnas.1510619112

   • PubMed
   • Google Scholar
6. 1. Bertolero MA
   2. Yeo BTT
   3. Bassett DS
   4. D’Esposito M

   (2018) A mechanistic model of connector hubs, modularity and cognition

   Nature Human Behaviour 2:765–777.

   https://doi.org/10.1038/s41562-018-0420-6

   • PubMed
   • Google Scholar
7. 1. Bostan AC
   2. Strick PL

   (2018) The basal ganglia and the cerebellum: nodes in an integrated network

   Nature Reviews. Neuroscience 19:338–350.

   https://doi.org/10.1038/s41583-018-0002-7

   • PubMed
   • Google Scholar
8. 1. Castro-Alamancos MA
   2. Gulati T

   (2014) Neuromodulators produce distinct activated states in neocortex

   The Journal of Neuroscience 34:12353–12367.

   https://doi.org/10.1523/JNEUROSCI.1858-14.2014

   • PubMed
   • Google Scholar
9. 1. Christophel TB
   2. Klink PC
   3. Spitzer B
   4. Roelfsema PR
   5. Haynes JD

   (2017) The distributed nature of working memory

   Trends in Cognitive Sciences 21:111–124.

   https://doi.org/10.1016/j.tics.2016.12.007

   • PubMed
   • Google Scholar
10. 1. Cole MW
    2. Reynolds JR
    3. Power JD
    4. Repovs G
    5. Anticevic A
    6. Braver TS

    (2013) Multi-task connectivity reveals flexible hubs for adaptive task control

    Nature Neuroscience 16:1348–1355.

    https://doi.org/10.1038/nn.3470

    • PubMed
    • Google Scholar
11. 1. Cole MW
    2. Ito T
    3. Bassett DS
    4. Schultz DH

    (2016) Activity flow over resting-state networks shapes cognitive task activations

    Nature Neuroscience 19:1718–1726.

    https://doi.org/10.1038/nn.4406

    • PubMed
    • Google Scholar
12. 1. Cole M.W
    2. Ito T
    3. Schultz D
    4. Mill R
    5. Chen R
    6. Cocuzza C

    (2019) Task activations produce spurious but systematic inflation of task functional connectivity estimates

    NeuroImage 189:1–18.

    https://doi.org/10.1016/j.neuroimage.2018.12.054

    • PubMed
    • Google Scholar
13. 1. Cox RW

    (1996) AFNI: software for analysis and visualization of functional magnetic resonance NeuroImages

    Computers and Biomedical Research, an International Journal 29:162–173.

    https://doi.org/10.1006/cbmr.1996.0014

    • PubMed
    • Google Scholar
14. 1. Cunningham JP
    2. Yu BM

    (2014) Dimensionality reduction for large-scale neural recordings

    Nature Neuroscience 17:1500–1509.

    https://doi.org/10.1038/nn.3776

    • PubMed
    • Google Scholar
15. 1. Duncan J

    (2010) The multiple-demand (MD) system of the primate brain: mental programs for intelligent behaviour

    Trends in Cognitive Sciences 14:172–179.

    https://doi.org/10.1016/j.tics.2010.01.004

    • PubMed
    • Google Scholar
16. 1. Esteban O
    2. Markiewicz CJ
    3. Blair RW
    4. Moodie CA
    5. Isik AI
    6. Erramuzpe A
    7. Kent JD
    8. Goncalves M
    9. DuPre E
    10. Snyder M
    11. Oya H
    12. Ghosh SS
    13. Wright J
    14. Durnez J
    15. Poldrack RA
    16. Gorgolewski KJ

    (2019) FMRIPrep: a robust preprocessing pipeline for functional MRI

    Nature Methods 16:111–116.

    https://doi.org/10.1038/s41592-018-0235-4

    • PubMed
    • Google Scholar
17. 1. Fonov V
    2. Evans AC
    3. Botteron K
    4. Almli CR
    5. McKinstry RC
    6. Collins DL
    7. Brain Development Cooperative Group

    (2011) Unbiased average age-appropriate atlases for pediatric studies

    NeuroImage 54:313–327.

    https://doi.org/10.1016/j.neuroimage.2010.07.033

    • PubMed
    • Google Scholar
18. 1. Garrett DD
    2. Epp SM
    3. Perry A
    4. Lindenberger U

    (2018) Local temporal variability reflects functional integration in the human brain

    NeuroImage 183:776–787.

    https://doi.org/10.1016/j.neuroimage.2018.08.019

    • PubMed
    • Google Scholar
19. Preprint

    1. Garrett DD
    2. Kloosterman NA
    3. Epp S
    4. Chopurian V
    5. Kosciessa JQ
    6. Waschke L
    7. Skowron A
    8. Shine JM
    9. Perry A
    10. Salami A
    11. Rieckmann A
    12. Papenberg G
    13. Wåhlin A
    14. Karalija N
    15. Andersson M
    16. Riklund K
    17. Lövdén M
    18. Bäckman L
    19. Nyberg L
    20. Lindenberger U

    (2022) Dynamic regulation of neural variability during working memory reflects dopamine, functional integration, and decision-making

    bioRxiv.

    https://doi.org/10.1101/2022.05.05.490687

    • Google Scholar
20. 1. Giguere M
    2. Goldman-Rakic PS

    (1988) Mediodorsal nucleus: areal, laminar, and tangential distribution of afferents and efferents in the frontal lobe of rhesus monkeys

    The Journal of Comparative Neurology 277:195–213.

    https://doi.org/10.1002/cne.902770204

    • PubMed
    • Google Scholar
21. 1. Gratton C
    2. Nomura EM
    3. Pérez F
    4. D’Esposito M

    (2012) Focal brain lesions to critical locations cause widespread disruption of the modular organization of the brain

    Journal of Cognitive Neuroscience 24:1275–1285.

    https://doi.org/10.1162/jocn_a_00222

    • PubMed
    • Google Scholar
22. 1. Gratton C
    2. Sun H
    3. Petersen SE

    (2018) Control networks and hubs

    Psychophysiology 55:13032.

    https://doi.org/10.1111/psyp.13032

    • PubMed
    • Google Scholar
23. 1. Greene DJ
    2. Marek S
    3. Gordon EM
    4. Siegel JS
    5. Gratton C
    6. Laumann TO
    7. Gilmore AW
    8. Berg JJ
    9. Nguyen AL
    10. Dierker D
    11. Van AN
    12. Ortega M
    13. Newbold DJ
    14. Hampton JM
    15. Nielsen AN
    16. McDermott KB
    17. Roland JL
    18. Norris SA
    19. Nelson SM
    20. Snyder AZ
    21. Schlaggar BL
    22. Petersen SE
    23. Dosenbach NUF

    (2020) Integrative and network-specific connectivity of the basal ganglia and thalamus defined in individuals

    Neuron 105:742–758.

    https://doi.org/10.1016/j.neuron.2019.11.012

    • PubMed
    • Google Scholar
24. 1. Guimerà R
    2. Nunes Amaral LA

    (2005) Functional cartography of complex metabolic networks

    Nature 433:895–900.

    https://doi.org/10.1038/nature03288

    • PubMed
    • Google Scholar
25. 1. Halassa MM
    2. Kastner S

    (2017) Thalamic functions in distributed cognitive control

    Nature Neuroscience 20:1669–1679.

    https://doi.org/10.1038/s41593-017-0020-1

    • PubMed
    • Google Scholar
26. 1. Halassa MM
    2. Sherman SM

    (2019) Thalamocortical circuit motifs: a general framework

    Neuron 103:762–770.

    https://doi.org/10.1016/j.neuron.2019.06.005

    • PubMed
    • Google Scholar
27. 1. Hwang Kai
    2. Bertolero MA
    3. Liu WB
    4. D’Esposito M

    (2017) The human thalamus is an integrative hub for functional brain networks

    The Journal of Neuroscience 37:5594–5607.

    https://doi.org/10.1523/JNEUROSCI.0067-17.2017

    • PubMed
    • Google Scholar
28. 1. Hwang K
    2. Bruss J
    3. Tranel D
    4. Boes AD

    (2020a) Network localization of executive function deficits in patients with focal thalamic lesions

    Journal of Cognitive Neuroscience 32:2303–2319.

    https://doi.org/10.1162/jocn_a_01628

    • PubMed
    • Google Scholar
29. 1. Hwang K
    2. Shine JM
    3. Cellier D
    4. D’Esposito M

    (2020b) The human intraparietal sulcus modulates task-evoked functional connectivity

    Cerebral Cortex 30:875–887.

    https://doi.org/10.1093/cercor/bhz133

    • PubMed
    • Google Scholar
30. 1. Hwang K
    2. Shine JM
    3. Bruss J
    4. Tranel D
    5. Boes A

    (2021) Neuropsychological evidence of multi-domain network hubs in the human thalamus

    eLife 10:e69480.

    https://doi.org/10.7554/eLife.69480

    • PubMed
    • Google Scholar
31. Software

    1. Hwang K

    (2022) Thalamocortical contribution to cognitive task activity, version swh:1:rev:d876ea93f1174ffdfe35c60f1f8b0ba8b637b037

    Software Heritage.

    https://archive.softwareheritage.org/swh:1:dir:192178129b4458e3587eae8e0a629abc176bd8ea;origin=https://github.com/HwangLabNeuroCogDynamics/ThalamicTaskHubs;visit=swh:1:snp:95f89e7b5c34e4265b78df160571de038dcb3820;anchor=swh:1:rev:d876ea93f1174ffdfe35c60f1f8b0ba8b637b037
32. Book

    1. Hwang K
    2. D’Esposito M

    (2022)

    Cognitive control functions of the human thalamus

    In: Halassa M, editors. The Thalamus. New York: Oxford University Press. pp. 307–323.

    • Google Scholar
33. 1. Ito T
    2. Kulkarni KR
    3. Schultz DH
    4. Mill RD
    5. Chen RH
    6. Solomyak LI
    7. Cole MW

    (2017) Cognitive task information is transferred between brain regions via resting-state network topology

    Nature Communications 8:1027.

    https://doi.org/10.1038/s41467-017-01000-w

    • PubMed
    • Google Scholar
34. 1. Ito T
    2. Yang GR
    3. Laurent P
    4. Schultz DH
    5. Cole MW

    (2022) Constructing neural network models from brain data reveals representational transformations linked to adaptive behavior

    Nature Communications 13:673.

    https://doi.org/10.1038/s41467-022-28323-7

    • PubMed
    • Google Scholar
35. 1. Jaramillo J
    2. Mejias JF
    3. Wang XJ

    (2019) Engagement of pulvino-cortical feedforward and feedback pathways in cognitive computations

    Neuron 101:321–336.

    https://doi.org/10.1016/j.neuron.2018.11.023

    • PubMed
    • Google Scholar
36. 1. Jones EG

    (2001) The thalamic matrix and thalamocortical synchrony

    Trends in Neurosciences 24:595–601.

    https://doi.org/10.1016/s0166-2236(00)01922-6

    • PubMed
    • Google Scholar
37. 1. Karolis VR
    2. Corbetta M
    3. Thiebaut de Schotten M

    (2019) The architecture of functional lateralisation and its relationship to callosal connectivity in the human brain

    Nature Communications 10:1417.

    https://doi.org/10.1038/s41467-019-09344-1

    • PubMed
    • Google Scholar
38. 1. King M
    2. Hernandez-Castillo CR
    3. Poldrack RA
    4. Ivry RB
    5. Diedrichsen J

    (2019) Functional boundaries in the human cerebellum revealed by a multi-domain task battery

    Nature Neuroscience 22:1371–1378.

    https://doi.org/10.1038/s41593-019-0436-x

    • PubMed
    • Google Scholar
39. Preprint

    1. King M
    2. Shahshahani L
    3. Ivry R
    4. Diedrichsen J

    (2022) A Task-General Connectivity Model Reveals Variation in Convergence of Cortical Inputs to Functional Regions of the Cerebellum

    bioRxiv.

    https://doi.org/10.1101/2022.05.07.490946

    • Google Scholar
40. 1. Kosciessa JQ
    2. Lindenberger U
    3. Garrett DD

    (2021) Thalamocortical excitability modulation guides human perception under uncertainty

    Nature Communications 12:2430.

    https://doi.org/10.1038/s41467-021-22511-7

    • PubMed
    • Google Scholar
41. 1. Krauth A
    2. Blanc R
    3. Poveda A
    4. Jeanmonod D
    5. Morel A
    6. Székely G

    (2010) A mean three-dimensional atlas of the human thalamus: generation from multiple histological data

    NeuroImage 49:2053–2062.

    https://doi.org/10.1016/j.neuroimage.2009.10.042

    • PubMed
    • Google Scholar
42. Book

    1. Lezak MD
    2. Howieson DB
    3. Bigler ED
    4. Tranel D

    (2012)

    Neuropsychological Assessment

    Oxford University Press.

    • Google Scholar
43. 1. MacDowell CJ
    2. Buschman TJ

    (2020) Low-Dimensional spatiotemporal dynamics underlie cortex-wide neural activity

    Current Biology 30:2665–2680.

    https://doi.org/10.1016/j.cub.2020.04.090

    • PubMed
    • Google Scholar
44. 1. Mukherjee A
    2. Lam NH
    3. Wimmer RD
    4. Halassa MM

    (2021) Thalamic circuits for independent control of prefrontal signal and noise

    Nature 600:100–104.

    https://doi.org/10.1038/s41586-021-04056-3

    • PubMed
    • Google Scholar
45. 1. Nakai T
    2. Nishimoto S

    (2020) Quantitative models reveal the organization of diverse cognitive functions in the brain

    Nature Communications 11:1142.

    https://doi.org/10.1038/s41467-020-14913-w

    • PubMed
    • Google Scholar
46. 1. Nee DE

    (2021) Integrative frontal-parietal dynamics supporting cognitive control

    eLife 10:e57244.

    https://doi.org/10.7554/eLife.57244

    • PubMed
    • Google Scholar
47. 1. O’Reilly RC
    2. Frank MJ

    (2006) Making working memory work: a computational model of learning in the prefrontal cortex and basal ganglia

    Neural Computation 18:283–328.

    https://doi.org/10.1162/089976606775093909

    • PubMed
    • Google Scholar
48. 1. Reber J
    2. Hwang K
    3. Bowren M
    4. Bruss J
    5. Mukherjee P
    6. Tranel D
    7. Boes AD

    (2021) Cognitive impairment after focal brain lesions is better predicted by damage to structural than functional network hubs

    PNAS 118:e2018784118.

    https://doi.org/10.1073/pnas.2018784118

    • PubMed
    • Google Scholar
49. 1. Rikhye RV
    2. Gilra A
    3. Halassa MM

    (2018) Thalamic regulation of switching between cortical representations enables cognitive flexibility

    Nature Neuroscience 21:1753–1763.

    https://doi.org/10.1038/s41593-018-0269-z

    • PubMed
    • Google Scholar
50. 1. Saalmann YB
    2. Pinsk MA
    3. Wang L
    4. Li X
    5. Kastner S

    (2012) The pulvinar regulates information transmission between cortical areas based on attention demands

    Science 337:753–756.

    https://doi.org/10.1126/science.1223082

    • PubMed
    • Google Scholar
51. 1. Sadaghiani S
    2. Scheeringa R
    3. Lehongre K
    4. Morillon B
    5. Giraud AL
    6. Kleinschmidt A

    (2010) Intrinsic connectivity networks, alpha oscillations, and tonic alertness: a simultaneous electroencephalography/functional magnetic resonance imaging study

    The Journal of Neuroscience 30:10243–10250.

    https://doi.org/10.1523/JNEUROSCI.1004-10.2010

    • PubMed
    • Google Scholar
52. 1. Schaefer A
    2. Kong R
    3. Gordon EM
    4. Laumann TO
    5. Zuo XN
    6. Holmes AJ
    7. Eickhoff SB
    8. Yeo BTT

    (2018) Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI

    Cerebral Cortex 28:3095–3114.

    https://doi.org/10.1093/cercor/bhx179

    • PubMed
    • Google Scholar
53. 1. Schumacher EH
    2. Hazeltine E

    (2016) Hierarchical task representation: task files and response selection

    Current Directions in Psychological Science 25:449–454.

    https://doi.org/10.1177/0963721416665085

    • Google Scholar
54. 1. Seeley WW
    2. Menon V
    3. Schatzberg AF
    4. Keller J
    5. Glover GH
    6. Kenna H
    7. Reiss AL
    8. Greicius MD

    (2007) Dissociable intrinsic connectivity networks for salience processing and executive control

    The Journal of Neuroscience 27:2349–2356.

    https://doi.org/10.1523/JNEUROSCI.5587-06.2007

    • PubMed
    • Google Scholar
55. 1. Selemon LD
    2. Goldman-Rakic PS

    (1988) Common cortical and subcortical targets of the dorsolateral prefrontal and posterior parietal cortices in the rhesus monkey: evidence for a distributed neural network subserving spatially guided behavior

    The Journal of Neuroscience 8:4049–4068.

    https://doi.org/10.1523/JNEUROSCI.08-11-04049.1988

    • PubMed
    • Google Scholar
56. 1. Sherman SM

    (2007) The thalamus is more than just a relay

    Current Opinion in Neurobiology 17:417–422.

    https://doi.org/10.1016/j.conb.2007.07.003

    • PubMed
    • Google Scholar
57. 1. Shine JM
    2. Breakspear M
    3. Bell PT
    4. Ehgoetz Martens KA
    5. Shine R
    6. Koyejo O
    7. Sporns O
    8. Poldrack RA

    (2019a) Human cognition involves the dynamic integration of neural activity and neuromodulatory systems

    Nature Neuroscience 22:289–296.

    https://doi.org/10.1038/s41593-018-0312-0

    • PubMed
    • Google Scholar
58. 1. Shine JM
    2. Hearne LJ
    3. Breakspear M
    4. Hwang K
    5. Müller EJ
    6. Sporns O
    7. Poldrack RA
    8. Mattingley JB
    9. Cocchi L

    (2019b) The low-dimensional neural architecture of cognitive complexity is related to activity in medial thalamic nuclei

    Neuron 104:849–855.

    https://doi.org/10.1016/j.neuron.2019.09.002

    • PubMed
    • Google Scholar
59. 1. Shine JM

    (2020) The thalamus integrates the macrosystems of the brain to facilitate complex, adaptive brain network dynamics

    Progress in Neurobiology 199:101951.

    https://doi.org/10.1016/j.pneurobio.2020.101951

    • PubMed
    • Google Scholar
60. 1. Snow JC
    2. Allen HA
    3. Rafal RD
    4. Humphreys GW

    (2009) Impaired attentional selection following lesions to human pulvinar: evidence for homology between human and monkey

    PNAS 106:4054–4059.

    https://doi.org/10.1073/pnas.0810086106

    • PubMed
    • Google Scholar
61. 1. van den Heuvel MP
    2. Sporns O

    (2013) Network hubs in the human brain

    Trends in Cognitive Sciences 17:683–696.

    https://doi.org/10.1016/j.tics.2013.09.012

    • PubMed
    • Google Scholar
62. 1. Warren DE
    2. Power JD
    3. Bruss J
    4. Denburg NL
    5. Waldron EJ
    6. Sun H
    7. Petersen SE
    8. Tranel D

    (2014) Network measures predict neuropsychological outcome after brain injury

    PNAS 111:14247–14252.

    https://doi.org/10.1073/pnas.1322173111

    • PubMed
    • Google Scholar
63. 1. Wolff M
    2. Vann SD

    (2019) The cognitive thalamus as a gateway to mental representations

    The Journal of Neuroscience 39:3–14.

    https://doi.org/10.1523/JNEUROSCI.0479-18.2018

    • PubMed
    • Google Scholar
64. 1. Yeo BTT
    2. Krienen FM
    3. Sepulcre J
    4. Sabuncu MR
    5. Lashkari D
    6. Hollinshead M
    7. Roffman JL
    8. Smoller JW
    9. Zöllei L
    10. Polimeni JR
    11. Fischl B
    12. Liu H
    13. Buckner RL

    (2011) The organization of the human cerebral cortex estimated by intrinsic functional connectivity

    Journal of Neurophysiology 106:1125–1165.

    https://doi.org/10.1152/jn.00338.2011

    • PubMed
    • Google Scholar
65. 1. Yeo BTT
    2. Krienen FM
    3. Eickhoff SB
    4. Yaakub SN
    5. Fox PT
    6. Buckner RL
    7. Asplund CL
    8. Chee MWL

    (2016) Functional specialization and flexibility in human association cortex

    Cerebral Cortex 26:465.

    https://doi.org/10.1093/cercor/bhv260

    • PubMed
    • Google Scholar
66. 1. Yuan R
    2. Di X
    3. Taylor PA
    4. Gohel S
    5. Tsai YH
    6. Biswal BB

    (2016) Functional topography of the thalamocortical system in human

    Brain Structure & Function 221:1971–1984.

    https://doi.org/10.1007/s00429-015-1018-7

    • PubMed
    • Google Scholar
67. 1. Zhang D
    2. Snyder AZ
    3. Fox MD
    4. Sansbury MW
    5. Shimony JS
    6. Raichle ME

    (2008) Intrinsic functional relations between human cerebral cortex and thalamus

    Journal of Neurophysiology 100:1740–1748.

    https://doi.org/10.1152/jn.90463.2008

    • PubMed
    • Google Scholar

Decision letter after peer review:

Thank you for submitting your article "Thalamocortical contribution to cognitive task activity" for consideration by eLife. Your article has been reviewed by 2 peer reviewers, including Jörn Diedrichsen as Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Timothy Behrens as the Senior Editor. The following individual involved in the review of your submission has agreed to reveal their identity: Moataz Assem (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

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_i² 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.

2. 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. The authors do not provide enough details on how task-evoked responses were regressed and how the residuals were assessed to be clean. A stronger test would utilize an FC matrix derived from independent data. Alternatively, I suggest using the FC matrix from dataset 2 to examine cortical projections of dataset 1 (and vice versa).

3. Throughout the manuscript, there is a general dependence on qualitative comparisons instead of quantifying similarities between findings. For example, (a) the spatial similarity of task hubs across the two datasets was not assessed (Figure 1d) (b) similarities between thalamic task hub projections to the cortex (Figure 2b). The comparison between the two datasets should be better quantified.

4. 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).

Reviewer #2 (Recommendations for the authors):

The findings presented here are important. The following recommendations would make the conclusions stronger:

1. The study would benefit from linking the thalamic task hubs to canonical resting-state networks defined in the thalamus (e.g. Ji et al. 2019 NeuroImage). Do task hubs mainly overlap with one functional network [e.g. the frontoparietal network (FPN)] or do they cross multiple functionally distinct networks? The latter would suggest some fractionation of the identified hubs (which is alluded to in the current results, due to the non-replication of the posterior thalamic task hub across datasets). It would be most interesting if task hubs ended up occupying voxels at the intersection of multiple RSNs (similar to how Power et al. 2013 Neuron defined hubs).

2. Could fractionating the thalamic task hubs reveal different cortical contributions? The cortical results in Figure 2B are actually not that similar across the datasets (e.g. ventro-medial frontal and posterior cingulate areas). Further, the cortical areas identified occupy much of the association cortices, which is inconsistent with more localized cortical hubs that the authors reference (Bertolero et al. as well as other studies). The posterior portion of the thalamic task hubs is already one potential contributing factor to these differences. There is also evidence that anterior thalamic regions are the most connected to a localized core subset of the FPN (Assem et al. 2020 Cerebral Cortex). Similarly, would simulating fractionated lesions to the thalamic task hubs show different contributions in Figure 5?

3. 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. The authors do not provide enough details on how task-evoked responses were regressed and how the residuals were assessed to be clean. Perhaps these details could quell my concerns. That said, if available, a stronger test would utilize an FC matrix derived from independent data. Alternatively, I suggest using the FC matrix from dataset 2 to examine cortical projections of dataset 1 (and vice versa).

4. Throughout the manuscript, there is a general dependence on qualitative comparisons instead of quantifying similarities between findings. For example, (a) the spatial similarity of task hubs across the two datasets was not assessed (Figure 1d) (b) similarities between thalamic task hub projections to the cortex (Figure 2b). Please quantify any comparison between the two datasets.

5. This is optional but I believe it would serve the study well if they link their thalamic task hubs to underlying thalamic nuclei using one of the many existing atlases (e.g. Najdenovska et al. nature scientific data 2018 https://www.nature.com/articles/sdata2018270). Even if the links with nuclei is at a coarse level, it could serve as a nice anatomical foundation for future explorations.

Essential revisions:

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_i² 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.

We thank the reviewer for raising these questions and would like to make a few clarifications. First, the V term in the original task hub definition refers to the percentage (%) of variance explained in evoked activity across tasks by a given component, not the variance within the voxel. We have renamed the variance explained term as EV to avoid this confusion. Second, we have performed a control analysis to determine factors that contribute to the task hub metric. Specifically, we compared the following two metrics:

The original task hub metric WxEV = ${\displaystyle \sum _{i=1}^n\left|W_i\right|\times {\text{EV}}_i}$

A revised task hub metric CompW = ${\displaystyle \sum _{i=1}^n\left|W_i\right|}$

The difference between these two metrics is that the revised metric is not weighted by the percentage variance explained of each component (EV), so the resulting spatial pattern should only reflect the “sum of component weights (sum(abs(w))).” We found the spatial map of CompW to be remarkably similar to the original metric WxEV (spatial correlation was 0.95 for the MDTB dataset and 0.86 for the N and N dataset). This suggests that the EV (variance explained) term had relatively minor contribution to the spatial topography when compared to the abs(W) term. In other words, the component weights more strongly contributed to the observed maps (Author response image 1).

Author response image 1

Download asset Open asset

In addition, we recalculated CompW using lower ranked components (components with lower EV, total summed variance <2%, components #11^th to 20^th). The result revealed a different spatial topography, suggesting that top components with higher explained variances do have a distinct spatial topography (spatial correlation for MDTB = -0.03, for N and N = 0.02). We have included this additional plot in Figure 2-supplement 1.Given these findings, for the revised manuscript we now only use the CompW metric as the hub metric, and we updated its definition with the following text in the revised manuscript:

“We reasoned that if a thalamic voxel is broadly participating in multiple tasks, it will express stronger weights for top thalamic activity components that explained a large amount of variance in evoked activity across tasks. Therefore, to further map “task hub” regions in the human thalamus, we calculated a metric by summing each voxel’s absolute component weight for the top ten components.” (line 165)

This revised description is a more accurate description of what the metric is measuring, and we have updated the language throughout the manuscript.

As suggested by the reviewer, we have also included a supplemental figure to describe the intuition of the participation coefficient (PC) metric for measuring network hubs (Figure 2—figure supplement 2). We also provided additional details on this measurement in the methods section (line 643). Briefly, PC is a measure of the weight of between-network connectivity (estimated from functional connectivity) for each thalamus voxel, normalized by their expected functional connectivity weight:

${\displaystyle PC=1-\sum _{s=1}^N{\left(\frac{K_{\text{is}}}{K_i}\right)}^2}$ where ${\displaystyle K_{\text{is}}}$ is the total functional connectivity weight between thalamus voxel i and cortical ROIs within network s, and ${\displaystyle K_i}$ is the sum of functional connectivity weight between voxel i and all cortical ROIs. N is the total number of cortical networks. If a voxel has connectivity uniformly distributed to all cortical networks, then its PC value will be close to 1; on the other hand, if connectivity is concentrated to a specific network, its PC value will be close to 0. Note that it is necessary to threshold the FC matrix for this calculation to remove negative functional connectivity weights. In our analyses, we calculated PC by thresholding the thalamocortical functional connectivity to preserve the top 1% to top 15% of functional connectivity weights. The results were then averaged across thresholds. As suggested by the reviewer, a toy example is how provided in Figure 2—figure supplement 2.

For the example on the left of Figure 2—figure supplement 2, the black ROI only has connectivity concentrated with ROIs in network B, therefore the ${\displaystyle \sum _{s=1}^N{\left(\frac{K_{\text{is}}}{K_i}\right)}^2}$ term will be 1, resulting in a PC value of 0. For the example on the right, its connectivity is distributed to both network A and B, therefore the ${\displaystyle \sum _{s=1}^N{\left(\frac{K_{\text{is}}}{K_i}\right)}^2}$ term will be 0.5, resulting in a higher PC value of 0.5

2. 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. The authors do not provide enough details on how task-evoked responses were regressed and how the residuals were assessed to be clean. A stronger test would utilize an FC matrix derived from independent data. Alternatively, I suggest using the FC matrix from dataset 2 to examine cortical projections of dataset 1 (and vice versa).

We agree that it is important to verify whether task-evoked activations have been removed from the timeseries data prior to calculating the FC matrix. We ran a control analysis to verify this. Specifically, we correlated the task regressor (stimulus timing convolved with the γ response function) with the thalamus and cortical timeseries data that we used for the activity flow analyses. We found the correlations to be very close to zero (correlation with cortical ROI timeseries, mean r = -2.02 * 10^-10, SD = 1.06* 10^-8; correlation with thalamic voxel timeseries, mean r = 4.37 * 10^-10, SD = 1.27* 10^-10). These weak correlations suggest that task-evoked activations have been successfully cleaned prior to FC calculations, and thus our activity flow analysis is likely not biased by the shared variance between task-evoked activations and timeseries used for calculating the FC matrix. We have included this information on line 679.

3. Throughout the manuscript, there is a general dependence on qualitative comparisons instead of quantifying similarities between findings. For example, (a) the spatial similarity of task hubs across the two datasets was not assessed (Figure 1d) (b) similarities between thalamic task hub projections to the cortex (Figure 2b). The comparison between the two datasets should be better quantified.

We appreciate the reviewer for pointing this out. We now report the following spatial correlations in the revised manuscript.

– The spatial correlation between task hubs from the two datasets across 2445 thalamus voxels was r = 0.36 (Figure 2A, line 174).

– The spatial correlation between FC hub (PC) from the two datasets was r = 0.55 (Figure 2A, Line 193).

– The spatial correlation of cortical projection of task hub between the two datasets is r = 0.46 (Figure 2B, Line 200).

– The spatial correlation between effects of simulated lesions on activity flow predictions between the two datasets was r = 0.29 (Figure 5B, Line 331).

Please note, we only performed statistical test on the spatial correlation between task hub and FC hub, as that was the only inferential statement we made in the manuscript (Line 191). We note that all other observed spatial correlations were stronger than the spatial correlations between FC and task hubs.

4. 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).

We thank the reviewer for pointing out this limitation. Following methods proposed in King et al., 2022, we calculated the noise ceiling for each region, task, model, and subject, using split-half reliability estimates from the activity flow model (${\displaystyle {Modeled\mathrm{\_}A}_{\text{ctx}}}$) and observed cortical responses (${\displaystyle {Observed\mathrm{\_}A}_{\text{ctx}}}$):${\displaystyle NoiseCeiling=\sqrt{\text{Reliability}\left({Modeled\mathrm{\_}A}_{\text{ctx}}\right)\times Reliability({Observed\mathrm{\_}A}_{\text{ctx}})}}$ ${\displaystyle \text{Reliability}\left({Modeled\mathrm{\_}A}_{\text{ctx}}\right)}$ is the split half reliability of activity flow prediction from the source ROI, and ${\displaystyle Reliability({Observed\mathrm{\_}A}_{\text{ctx}})}$ is the split half reliability of the observed cortical evoked responses. Details on how noise ceiling estimates were calculated are now reported in line 703. The noise ceiling level is reported in the manuscript (Figure 4—figure supplement 1), and results in figures 4 and 5 are updated with the noise normalized value (divided by the noise ceiling).

We also included a third null model, as suggested by the reviewer. This null model averaged the evoked response amplitudes across all conditions for every voxel, which assumed there is no difference in the voxel-wise evoked response patterns between conditions (line 254). Furthermore, as requested by the reviewer, we included additional details to better explain the assumption of each null model:

“The first null model randomly shuffled thalamic evoked responses (the “null model”), which assumed no spatial structure in thalamic task-evoked responses. The second null model set all thalamic evoked responses to the same value (“1”) across all voxels (the “uniformed” evoked model”), which assumed that cortical activity patterns are determined only by the summated inputs from thalamocortical FC patterns. The third null model was constructed by averaging the evoked response amplitudes across tasks for every thalamic voxel, which assumed that there is no difference in voxel-wise thalamic activity patterns between tasks.”

While the raw Pearson correlation values from the activity flow model were weaker than some of the previous activity flow studies (i.e., Cole et al., 2016), we would like to note that there is one important difference between our approach and previous studies. Previous studies included “multiple” cortical regions to build the activity flow prediction model, whereas our study tested a different question to evaluate the predictive power of a “single” subcortical region, the thalamus. We found that after accounting for the noise ceiling, thalamic activity alone was able to predict approximately 38% of the systematic variance across 400 cortical ROI activity patterns.

Reviewer #2 (Recommendations for the authors):

The findings presented here are important. The following recommendations would make the conclusions stronger:

1. The study would benefit from linking the thalamic task hubs to canonical resting-state networks defined in the thalamus (e.g. Ji et al. 2019 NeuroImage). Do task hubs mainly overlap with one functional network [e.g. the frontoparietal network (FPN)] or do they cross multiple functionally distinct networks? The latter would suggest some fractionation of the identified hubs (which is alluded to in the current results, due to the non-replication of the posterior thalamic task hub across datasets). It would be most interesting if task hubs ended up occupying voxels at the intersection of multiple RSNs (similar to how Power et al. 2013 Neuron defined hubs).

We thank the reviewer for raising this question. We do not think that task hub in the thalamus is a monolithic structure, instead it is biologically more plausible that different subdivisions within the thalamus, either defined histologically or functionally, can exhibit strong task hub properties. This interpretation is more in line with our previous studies demonstrating that network hubs are distributed across nuclei and functional parcellations within the human thalamus (Hwang et al., 2017 JN; Hwang et al. 2021 eLife).

To further describe the anatomical distribution of thalamic task hubs, we summarized the spatial distribution of compW using two thalamic atlases: (1) The Morel atlas (Krauth et al., 2010 NeuroImage), which defined thalamic nuclei using histological information from postmortem human brains, and (2) a functional network parcellation atlas developed by Thomas Yeo (https://twitter.com/bttyeo/status/1248992927830781952), which defined thalamic parcels based on its thalamocortical functional connectivity with the 7 canonical resting cortical networks that we used for the network hub analysis (Figure 2A). We excluded the “limbic network” parcellation from this analysis because only 4 thalamic voxels were assigned to it. These results are presented in the new Figure 3.

In both datasets, we found the task hub value to be higher in the anterior and mediodorsal nuclei. These two thalamic nuclei are known to have strong reciprocal connectivity with the medial frontal and lateral frontal regions. We also found the medial pulvinar to show high task hub value but only in the MDTB dataset, likely because the N and N dataset did not show high task hub value in the posterior thalamus.

For the functional network parcellation atlas, between datasets discrepancies were also found in posterior thalamus. For example, in the MDTB dataset, the DF parcellation, which overlapped with the anterior medial, medial, and posterior medial thalamus, was found to show the highest task hub value. For the N and N dataset, the posterior thalamus did not exhibit high task hub value, but the FP parcellation, which covered the dorsal bank of the medial thalamus, showed the highest task hub value. Overall, these results suggest that thalamic task hubs do not mainly overlap with one functional network parcellation or thalamic nucleus, instead cover multiple parcellations/nuclei.

Krauth, A., Blanc, R., Poveda, A., Jeanmonod, D., Morel, A., and Székely, G. (2010). A mean three-dimensional atlas of the human thalamus: generation from multiple histological data. Neuroimage, 49(3), 2053-2062.

Schaefer, A., Kong, R., Gordon, E. M., Laumann, T. O., Zuo, X. N., Holmes, A. J., … and Yeo, B. T. (2018). Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI. Cerebral cortex, 28(9), 3095-3114.

Reber, J., Hwang, K., Bowren, M., Bruss, J., Mukherjee, P., Tranel, D., and Boes, A. D. (2021). Cognitive impairment after focal brain lesions is better predicted by damage to structural than functional network hubs. Proceedings of the National Academy of Sciences, 118(19), e2018784118.

2. Could fractionating the thalamic task hubs reveal different cortical contributions? The cortical results in Figure 2B are actually not that similar across the datasets (e.g. ventro-medial frontal and posterior cingulate areas). Further, the cortical areas identified occupy much of the association cortices, which is inconsistent with more localized cortical hubs that the authors reference (Bertolero et al. as well as other studies). The posterior portion of the thalamic task hubs is already one potential contributing factor to these differences. There is also evidence that anterior thalamic regions are the most connected to a localized core subset of the FPN (Assem et al. 2020 Cerebral Cortex). Similarly, would simulating fractionated lesions to the thalamic task hubs show different contributions in Figure 5?

Given that we found that task hubs are spatially distributed across different thalamic sub-regions, it is more likely that different subregions will have different coupling patterns with cortical regions. This is perhaps to be expected, especially given the known anatomical thalamocortical projection profile of different thalamic nuclei. We demonstrated this point by performing a control analysis, where we projected the task hub estimates (compW) from the following three thalamic nuclei (anterior nucleus [AN], mediodorsal nucleus [MD], and medial pulvinar [PuM]) to the cortex via the activity flow mapping procedure. The results are presented in the Author response imahe 2. As expected, the anterior nucleus showed strong coupling with medial and lateral frontal regions, the mediodorsal thalamus with the lateral frontal cortex and the insula, and the medial pulvinar with the occipitotemporal cortex for the MDTB dataset (less clear pattern for the N and N dataset). Given these are expected from the known thalamocortical anatomy, we opted to not include this control analysis in the main manuscript.

Author response image 2

Download asset Open asset

We agree with the reviewer that the frontoparietal association areas found to show strong coupling with thalamic task hubs were more extensive than the more selective task-flexible regions found in previous papers. We have revised the Discussion section to make it clear that our results showed thalamic coupling with multiple regions across the broad frontoparietal cortices, and among the overlapped regions some selective frontoparietal regions have been shown to be part of a flexible multiple-demand system (line 491).We also simulated lesions to different thalamic nuclei that we found to exhibit high task hub values. Specifically, we simulated lesions separately for the anterior nucleus (AN), the mediodorsal nucleus (MD), and the medial pulvinar (PuM), which corresponded to the anterior, medial, and posterior medial portion of the thalamus that we found to show strong task hub values. We found that simulated lesions to these nuclei significantly impaired activity flow model performance (MDTB: AN mean reduction = 17.33%, SE = 5.71%, MD mean reduction = 16.98%, SE = 6.06%, PuM mean reduction = 17.45%, SE = 6.42%; N and N AN: mean reduction = 13.59%, SE = 9.24%, MD: mean reduction = 14.29%, SE=8.83%, PuM mean reduction = 6.7%, SE = 9.26%), with the exception that weaker reduction in activity flow model performance was observed for PuM in the N and N dataset. Critically, we did not observe a statistically significant difference in the effects of simulating lesions to these nuclei. Thus, the main thalamic task hub regions in the anterior, medial, and posterior medial thalamus appear to have similar contributions to activity flow model prediction. Given that we did not observe significant differences, we opted to not include these control analyses to the main manuscript.

3. 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. The authors do not provide enough details on how task-evoked responses were regressed and how the residuals were assessed to be clean. Perhaps these details could quell my concerns. That said, if available, a stronger test would utilize an FC matrix derived from independent data. Alternatively, I suggest using the FC matrix from dataset 2 to examine cortical projections of dataset 1 (and vice versa).

Please see our response to essential revisions point #2.

4. Throughout the manuscript, there is a general dependence on qualitative comparisons instead of quantifying similarities between findings. For example, (a) the spatial similarity of task hubs across the two datasets was not assessed (Figure 1d) (b) similarities between thalamic task hub projections to the cortex (Figure 2b). Please quantify any comparison between the two datasets.

Please see our response to essential revisions point #3.

5. This is optional but I believe it would serve the study well if they link their thalamic task hubs to underlying thalamic nuclei using one of the many existing atlases (e.g. Najdenovska et al. nature scientific data 2018 https://www.nature.com/articles/sdata2018270). Even if the links with nuclei is at a coarse level, it could serve as a nice anatomical foundation for future explorations.

Please see our response to reviewer 2 comment #1. We have included this suggested analysis in the new figure 3.

Author details

1. Kai Hwang

   1. Department of Psychological and Brain Sciences, University of Iowa, Iowa City, United States
   2. Cognitive Control Collaborative, University of Iowa, Iowa City, United States
   3. Iowa Neuroscience Institute, University of Iowa, Iowa City, United States
   4. Department of Psychiatry, University of Iowa, Iowa City, United States

   Contribution

   Conceptualization, Resources, Software, Formal analysis, Supervision, Funding acquisition, Validation, Investigation, Visualization, Methodology, Writing – original draft, Project administration, Writing – review and editing

   For correspondence

   kai-hwang@uiowa.edu

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-1064-7815

2. James M Shine

   Brain and Mind Center, University of Sydney, Sydney, Australia

   Contribution

   Conceptualization, Methodology, Writing – review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-1762-5499

3. Michael W Cole

   Center for Molecular and Behavioral Neuroscience, Rutgers University-Newark, Newark, United States

   Contribution

   Software, Methodology, Writing – review and editing

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0003-4329-438X

4. Evan Sorenson

   1. Department of Psychological and Brain Sciences, University of Iowa, Iowa City, United States
   2. Cognitive Control Collaborative, University of Iowa, Iowa City, United States
   3. Department of Psychiatry, University of Iowa, Iowa City, United States

   Contribution

   Resources, Data curation, Software, Formal analysis, Investigation, Visualization, Methodology, Writing – review and editing

   Competing interests

   No competing interests declared

• 1,423

  Page views

• 199

  Downloads

• 3

  Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.