Recent work has shown that cortical function and intrinsic connectivity change systematically along a ‘principal gradient’, which has primary sensory and motor cortex at one end, and transmodal regions implicated in abstract and memory-based functions at the other (Margulies et al., 2016). This principal gradient is the dimension of intrinsic connectivity that typically explains the most variance in whole-brain decompositions of human resting-state fMRI data (Hong et al., 2020, Huntenburg et al., 2018). It is correlated with physical distance along the cortical mantle from primary sensory-motor landmarks. It captures the order of large-scale networks in frontal, temporal, and parietal regions – all of which show transitions from sensory-motor regions, through attention networks to frontoparietal control and then default mode network (DMN) regions. It also captures functional transitions: for instance, opposing ends of the principal gradient are implicated in sensory and memory-based decisions (Murphy et al., 2019; Murphy et al., 2018; Lanzoni et al., 2020). Theories suggest the principal gradient supports increasingly abstract levels of representation, allowing heteromodal regions in the DMN to take on roles that are less directly influenced by externally driven neural activity (Smallwood et al., 2021; Margulies et al., 2016; Huntenburg et al., 2018; Wang et al., 2020). In line with this account, the representations of events extracted from movie-watching data are thought to vary in length along this hierarchy, with short events represented in sensory areas and longer events in transmodal regions including within the default mode network (Baldassano et al., 2017).

Semantic representation is thought to draw on both ends of this principal gradient (Huth et al., 2016; Barsalou, 2008; Kiefer and Pulvermüller, 2012; Meteyard et al., 2012; Patterson et al., 2007; Wang et al., 2010; Visser and Lambon Ralph, 2011; Murphy et al., 2019). The graded hub and spoke model proposes that heteromodal concepts are formed at a distance from visual and auditory input regions: diverse sensory processing pathways gradually converge in transmodal areas to form abstract generalisable semantic representations (Patterson et al., 2007; Ralph et al., 2017). Moreover, we flexibly deploy conceptual information to suit our goals or context: this involves the recruitment of several heteromodal networks, including the DMN, semantic control network (SCN) and multiple-demand network (MDN), which lie at different points along the principal gradient (Wang et al., 2020; Gao et al., 2021). These networks show differential recruitment across trials depending on the semantic similarity of the items being linked, with higher conceptual overlap associated with stronger activation at the top end of the principal gradient (Wang et al., 2020). Anterior temporal lobe (ATL) and angular gyrus (AG), which are allied with DMN, contribute to the relatively automatic retrieval of semantic information strongly supported by long-term memory (Davey et al., 2015; Jefferies et al., 2020, Teige et al., 2019; Teige et al., 2018). In contrast, when unusual aspects of knowledge are required – for example, for weak associations – there is greater recruitment of the SCN to shape the meaning that is retrieved (Davey et al., 2015; Jackson, 2021; Jefferies, 2013; Jefferies et al., 2020), along with other control regions within the MDN, which responds to higher task demands across domains (Fedorenko et al., 2013; Erb et al., 2013; Assem et al., 2020).

While recruitment is known to vary along the principal gradient according to the demands of the task (Murphy et al., 2019; Wang et al., 2020), there have been few if any investigations of the way in which neural representation changes over the length of this whole-brain gradient depending on the type of information that is retrieved. Local transitions in the information that is represented neurally have been observed for specific domains. Visual object recognition is supported by a processing hierarchy extending from primary visual cortex, through higher visual regions in inferior temporal cortex, to ventral anterior temporal areas that capture the meaning of concepts across modalities (Murphy et al., 2017; Ju and Bassett, 2020). A similar functional transition is proposed in medial temporal cortex, which supports detailed, fine-grained spatial representations in posterior hippocampus and more gist or schema like, coarse-grained memory representations in anterior hippocampus (Poppenk et al., 2013). Lateral prefrontal cortex has also been proposed to support a rostral–caudal gradient, representing specific motor actions in posterior regions through to more abstract goals in anterior regions (Badre and D’Esposito, 2009; Badre and Nee, 2018). However, these studies have not examined changes in connectivity gradients as a function of task demands. Dimensionality reduction techniques like principal component analysis (PCA) can also quantify the number of components needed to explain a specific amount of variance in the neural response of a given brain region, providing an index of the complexity or dimensionality of the representational space over time (Ahlheim and Love, 2018). Neural representation in unimodal regions is thought to have many dimensions, allowing the cortex to discriminate between similar inputs, while the representational space may have fewer dimensions in transmodal DMN regions, allowing a variety of inputs to be encoded into common activity patterns that generalise knowledge across circumstances (Ju and Bassett, 2020). For example, we can easily tell the difference between ‘Dalmatian’ and ‘Bulldog’ yet understand their abstract similarity as ‘dogs’. Category learning based on an abstract rule is also associated with a prefrontal response characterised by fewer dimensions (Mack et al., 2020).

In the current study, we manipulated the degree to which ongoing semantic cognition was aligned with long-term semantic knowledge and quantified the similarity of the multivariate response to each trial along the principal gradient. Our task asked participants to generate links between pairs of words that varied in their associative strength, including strongly associated items which often co-occur, weakly linked items and unrelated pairs. We asked whether functional connectivity patterns during this task would mirror the principal gradient and vary flexibly according to the associative strength of the items in each trial. We also asked whether neural representations along the principal gradient differed in terms of their complexity. With less support from long-term memory, unusual patterns of conceptual retrieval might be supported by control regions that promote specific sensory-motor features, reducing the functional separation of the two ends of the principal gradient in terms of their patterns of connectivity. In contrast, frequently occurring patterns of conceptual retrieval may emerge from long-term memory in the absence of additional constraints from brain regions; this retrieval state might be associated with increased functional separation between DMN and sensory-motor regions, allowing memory to underpin cognition irrespective of ongoing sensory-motor experience. We might also expect the dimensionality of neural representations in a semantic task to decrease from unimodal to transmodal areas along the principal gradient, reflecting increasingly abstract and culturally shared representations towards the apex of the gradient, particularly when participants retrieve strong associations that link words together in specific well-established ways.

This study modulated the associative strength between word-pairs to delineate the neural patterns underpinning semantic retrieval along the principal gradient of cortical organisation. We combined dimensionality reduction methods (gradient analysis) with informational connectivity capturing the multivariate similarity between brain regions during the retrieval of semantic associations. This analysis showed that the presentation of strong associations increased the functional separation between unimodal and transmodal regions at opposite ends of the principal gradient. Stronger associations elicited higher gradient values in transmodal regions and lower gradient values in unimodal regions, suggesting that commonly occurring patterns of cognition mediated by long-term memory involved greater informational separation from sensory-motor features. In contrast, more idiosyncratic patterns of retrieval involved less separation between transmodal and unimodal representations, perhaps because participants fell back on specific features to generate a link when dominant aspects of long-term memory were insufficient for the task. This change was greatest in semantic control regions, which were close to the apex of the principal gradient along with DMN at rest and during the retrieval of strong associations, yet showed substantially lower gradient values during the retrieval of weak associations. We also found a larger decrease in the dimensionality of neural representations from unimodal to transmodal regions along the principal gradient when strong associations were retrieved; the response towards the DMN apex of the gradient was less complex or variable across participants for strong associations, reflecting the way that these trials relied on aspects of knowledge or retrieval states that could generalise across the specific experiences of individuals. The neural representation of weakly associated word-pairs was associated with higher dimensionality or heterogeneity in distributed transmodal areas, including in lateral DMN and semantic control regions. Finally, mediation analysis showed that the relevance of cortical gradients to individual differences in rated semantic similarity was partially explained by neural dimensionality differences between weak and strong associations. Therefore, the neural organisation of semantic cognition not only reflects global patterns of brain connectivity but changes in global connectivity patterns associated with variation in transmodal neural representation.

Notably, while most gradient studies generate connectivity matrices by correlating mean univariate signals between ROIs over time, normally using resting-state fMRI data (Huntenburg et al., 2018; Margulies et al., 2016), we employed a novel informational connectivity approach to measure the similarity of multivariate response patterns between ROIs during a task (Anzellotti and Coutanche, 2018; Coutanche and Thompson-Schill, 2013). The cortical gradient explaining the most variance in informational connectivity was highly correlated with the principal gradient estimated from resting data (Margulies et al., 2016) using the traditional univariate approach. [This relationship between task-based informational connectivity and resting-state fMRI analysis was also found for the gradient which explained the second most variance for the semantic task. The gradient which explained the third most variance showed a moderate correlation with the third gradient derived from resting-state data in the left hemisphere. However, the correlation was negative in the right hemisphere, suggesting an influence of the task: semantic cognition might be expected to separate the connectivity patterns of left and right hemispheres, since it is left lateralised.] Moreover, the variance explained by the principal gradient was higher for the decomposition of informational connectivity than for time-series correlation (40–50% vs. 25%). Therefore, the current study helps to validate a new approach to deriving cortical gradients from task data which was sensitive to the effects of associative strength in semantic cognition.

The human capacity to retrieve abstract meaning from sounds, visual objects, words, and phrases exceeds the capacities of other animal species; however, the development of this ability is poorly understood (Enge et al., 2021). One recent study Dong et al., 2021 found that, in children, the overarching organisational gradient is anchored within the unimodal cortex, between somatosensory/motor and visual territories, while in adolescence, the principal gradient of connectivity transitions into an adult-like spatial framework, with the default network at the opposite end of a spectrum from primary sensory and motor regions. Whether the development of abstract semantic representations is related to this gradual emergence of the principal gradient in the brain remains an open question which can be addressed in future studies.

Our findings are broadly consistent with the controlled semantic cognition framework (Jefferies et al., 2020, Ralph et al., 2017), which proposes that heteromodal semantic cognition emerges from the distillation of sensory-motor features into more abstract concepts; semantic control processes then shape the responses across this network to suit the demands of the task. There is strong overlap between the semantic network recruited during tasks and DMN regions identified from the parcellation of resting-state data, for example in inferior and lateral temporal and lateral parietal cortices (Davey et al., 2016; Jackson et al., 2016; Seghier et al., 2010; Humphreys and Lambon Ralph, 2015). When semantic retrieval is focused on dominant features of conceptual knowledge, automatic retrieval might emerge from the unconstrained behaviour of these semantic ‘hub’ regions, including ATL and potentially other sites with similar patterns of connectivity, such as left AG (Humphreys and Lambon Ralph, 2015). Gradient analysis of both task and rest data shows that these transmodal regions have a high degree of pattern similarity to each other, and weaker similarity to sensory-motor cortex. Long-term knowledge amplifies the informational distance between the opposite ends of the principal gradient, suggesting that when strong associations are retrieved, areas associated with abstract and heteromodal conceptual representation separate from specific sensory-motor features and attentional networks at the opposite end of the gradient, as these elements are not needed for task-relevant patterns of retrieval. In contrast, when retrieval is focused on unusual aspects of semantic knowledge, semantic control processes shape retrieval to suit the demands of the task – for example, inhibiting dominant associative links and retrieving specific features that support specific and unusual connections between words (Jackson, 2021; Jefferies et al., 2020, Gao et al., 2021, Noonan et al., 2013). In these circumstances, stronger activation is seen within the ‘semantic control network’ comprising inferior frontal gyrus, posterior middle temporal gyrus, and pre-supplementary cortex (Jackson, 2021; Noonan et al., 2013, Gao et al., 2021).

We found a positive linear relationship between the difference of gradient values across weak and strong associations and principal gradient location, providing evidence that regions towards the DMN apex can flex their patterns of connectivity depending on the demands of the task. Semantic control regions showed the largest changes in gradient values and dimensionality between strong and weak associations, suggesting that these regions separate from other heteromodal areas towards the DMN apex of the gradient and interact more strongly with attentional networks and unimodal regions to generate unusual connections between words. While our results are in line with the controlled semantic cognition framework in general, multiple cognitive processes might contribute to the differences between stronger and weaker associations in our task, including the requirement to generate a novel link, as opposed to recognising an existing link. There are also likely to be differing demands on more automatic versus more controlled retrieval, imagery and processes associated with creativity.

Our gradient analyses are highly consistent with inter-voxel similarity analysis across participants and trials. Using this approach to index the complexity of neural representations, researchers have revealed a graded transition from coarse to fine-grained representations along the long axis of the hippocampus from posterior to anterior regions (Brunec et al., 2018). Lower inter-voxel similarity over time in the hippocampus has been linked to more differentiated memory representations (Chanales et al., 2017). Motivated by these studies, we measured the heterogeneity of the voxels’ responses in local areas to assess the dimensionality of representational space during semantic retrieval. We found a dimensionality decrease from unimodal to transmodal regions along the principal gradient which was larger for strong associations. This finding is consistent with recent studies showing that multivariate patterns within transmodal areas of ventral ATL represent meaning irrespective of presentation modality, feature input, or language (Murphy et al., 2017; Coutanche and Thompson-Schill, 2015; Correia et al., 2014) – this abstraction of conceptual representations away from surface features of the input is likely to be critical for the generalisation of knowledge across modalities and diverse contexts and to support shared understanding across individuals. In line with this view, higher-dimensional spaces are thought to enable the discrimination of similar inputs but consume more resources to embed the information, while lower-dimensional spaces facilitate generalisation and novel learning across stimuli or situations (Badre et al., 2021; Ju and Bassett, 2020). DMN regions near the apex of the principal gradient may support more abstract representations of meaning when the retrieval demands of a task are well aligned with the information stored in long-term semantic memory.

Consistent with this hypothesis, neural representations showed higher dimensionality for weak associations in transmodal regions, including the semantic ‘hub’ regions in ATL and AG, and semantic control regions in lateral and medial prefrontal, parietal, and temporal cortex (Humphreys and Lambon Ralph, 2015, Jackson et al., 2016; Jefferies, 2013; Jefferies et al., 2020; Seghier et al., 2010). Notably, the dimensionality difference peaked in semantic control regions, suggesting the SCN is important for promoting flexible semantic cognition (Jefferies et al., 2020). In this way, our results extended prior observations from univariate analyses by revealing that more complex representations and/or processes were implemented in transmodal areas during more controlled patterns of retrieval. Although regions towards the DMN apex of the principal gradient were associated with lower dimensionality of neural representation, we also observed higher sensitivity to the manipulation of associative strength in these areas. This suggests that transmodal regions adopt low-dimensional states when retrieval is adequately constrained by semantic long-term memory and higher-dimensional states when more idiosyncratic patterns of retrieval are generated.

It is worth noting that not all brain regions showed the expected pattern in the dimensionality analysis – especially when considering the global dimensionality of all semantic trials, as opposed to the influence of strength of association in the semantic task. In particular, the limbic network, including regions of ventral ATL thought to support a heteromodal semantic hub, showed significantly higher dimensionality than sensory-motor areas – these higher-order regions are expected to show lower dimensionality corresponding to more abstract representations. However, this analysis does not assess the psychological significance of data dimensionality differences (unlike our contrast of strong and weak associations, which are more interpretable in terms of semantic cognition). Limbic regions are subject to severe distortion and signal loss in functional MRI, which might strongly influence this metric. Future studies using data acquisition and analysis techniques that are less susceptible to this problem are required to fully characterise global dimensionality and its relation to the principal gradient.

In conclusion, we found that when strong associations were retrieved, and therefore cognition was well aligned with semantic information in long-term memory, the principal gradient was strengthened, suggesting greater separation of heteromodal and unimodal connectivity. When less related concepts were linked, this dimension of connectivity was reduced in strength as semantic control regions near the apex of the gradient separated from DMN to generate more flexible and original responses. In addition, semantic regions near the apex of the principal gradient showed higher dimensionality for weak associations, with activation patterns in these regions predicting individual differences in the similarity of semantic ratings. These results reveal that the principal gradient provides an organising principle for semantic cognition in the cerebral cortex.

Experiment materials, behavioural data, source data for producing the figures, brain parcellation template, and group-level neuroimaging data (gradient-relevant analysis) are accessible in the Open Science Framework at https://osf.io/mkgcy/. The Neurovault collection provides the processed version of the dataset for the other analyses, including neural dimensionality and second-order representational analysis: https://neurovault.org/collections/12539/. All analysis codes and software used in this study have been uploaded onto osf: https://osf.io/mkgcy/, which include but are not limited to the gradient analysis (Matlab), dimensionality analysis (python), and second-order RSA analysis (Matlab). The conditions of our ethical approval do not permit public archiving of the data because participants did not provide sufficient consent for the release of their biomedical data. Researchers who wish to access the data should contact the Research Ethics and Governance Committee of the York Neuroimaging Centre, University of York, or the corresponding authors. Data will be released to researchers when this is possible under the terms of the GDPR (General Data Protection Regulation). The decision as to whether the data can be reused and how access can be provided will be taken by the Research Ethics and Governance Committee of the York Neuroimaging Centre; data access arrangements are likely to exclude commercial use of the data.

1. 1. Gao Z

   (2021) Open Science Framework

   ID mkgcy. Flexing the principal gradient of the cerebral cortex to suit changing semantic task demands.

   https://osf.io/mkgcy/
2. 1. Gao Z

   (2022) NeuroVault

   ID collection:12539. Flexing the principal gradient of the cerebral cortex to suit changing semantic task demands.

   https://identifiers.org/neurovault.collection:12539

1. 1. Ahlheim C
   2. Love BC

   (2018) Estimating the functional dimensionality of neural representations

   NeuroImage 179:51–62.

   https://doi.org/10.1016/j.neuroimage.2018.06.015

   • PubMed
   • Google Scholar
2. 1. Alexander-Bloch AF
   2. Shou H
   3. Liu S
   4. Satterthwaite TD
   5. Glahn DC
   6. Shinohara RT
   7. Vandekar SN
   8. Raznahan A

   (2018) On testing for spatial correspondence between maps of human brain structure and function

   NeuroImage 178:540–551.

   https://doi.org/10.1016/j.neuroimage.2018.05.070

   • PubMed
   • Google Scholar
3. Report

   1. Anderson JL
   2. Jenkinson M
   3. Smith S

   (2007) Non-linear registration aka Spatial normalisation FMRIB

   FMRIB Analysis Group of the University of Oxford.

   https://www.fmrib.ox.ac.uk/datasets/techrep/tr07ja2/tr07ja2.pdf

   • Google Scholar
4. 1. Anzellotti S
   2. Coutanche MN

   (2018) Beyond functional connectivity: investigating networks of multivariate representations

   Trends in Cognitive Sciences 22:258–269.

   https://doi.org/10.1016/j.tics.2017.12.002

   • PubMed
   • Google Scholar
5. 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

   Neuroscience 30:4361–4380.

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

   • Google Scholar
6. 1. Badre D
   2. D’Esposito M

   (2009) Is the rostro-caudal axis of the frontal lobe hierarchical?

   Nature Reviews. Neuroscience 10:659–669.

   https://doi.org/10.1038/nrn2667

   • PubMed
   • Google Scholar
7. 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
8. 1. Badre D
   2. Bhandari A
   3. Keglovits H
   4. Kikumoto A

   (2021) The dimensionality of neural representations for control

   Current Opinion in Behavioral Sciences 38:20–28.

   https://doi.org/10.1016/j.cobeha.2020.07.002

   • PubMed
   • Google Scholar
9. 1. Baldassano C
   2. Chen J
   3. Zadbood A
   4. Pillow JW
   5. Hasson U
   6. Norman KA

   (2017) Discovering event structure in continuous narrative perception and memory

   Neuron 95:709–721.

   https://doi.org/10.1016/j.neuron.2017.06.041

   • PubMed
   • Google Scholar
10. 1. Barsalou LW

    (2008) Grounded cognition

    Annual Review of Psychology 59:617–645.

    https://doi.org/10.1146/annurev.psych.59.103006.093639

    • PubMed
    • Google Scholar
11. 1. Brunec IK
    2. Bellana B
    3. Ozubko JD
    4. Man V
    5. Robin J
    6. Liu Z-X
    7. Grady C
    8. Rosenbaum RS
    9. Winocur G
    10. Barense MD
    11. Moscovitch M

    (2018) Multiple scales of representation along the hippocampal anteroposterior axis in humans

    Current Biology 28:2129–2135.

    https://doi.org/10.1016/j.cub.2018.05.016

    • PubMed
    • Google Scholar
12. 1. Chanales AJH
    2. Oza A
    3. Favila SE
    4. Kuhl BA

    (2017) Overlap among spatial memories triggers repulsion of hippocampal representations

    Current Biology 27:2307–2317.

    https://doi.org/10.1016/j.cub.2017.06.057

    • PubMed
    • Google Scholar
13. 1. Correia J
    2. Formisano E
    3. Valente G
    4. Hausfeld L
    5. Jansma B
    6. Bonte M

    (2014) Brain-based translation: fMRI decoding of spoken words in bilinguals reveals language-independent semantic representations in anterior temporal lobe

    The Journal of Neuroscience 34:332–338.

    https://doi.org/10.1523/JNEUROSCI.1302-13.2014

    • PubMed
    • Google Scholar
14. 1. Coutanche MN
    2. Thompson-Schill SL

    (2013) Informational connectivity: identifying synchronized discriminability of multi-voxel patterns across the brain

    Frontiers in Human Neuroscience 7:15.

    https://doi.org/10.3389/fnhum.2013.00015

    • PubMed
    • Google Scholar
15. 1. Coutanche MN
    2. Thompson-Schill SL

    (2015) Creating concepts from converging features in human cortex

    Cerebral Cortex 25:2584–2593.

    https://doi.org/10.1093/cercor/bhu057

    • PubMed
    • Google Scholar
16. 1. Davey J
    2. Cornelissen PL
    3. Thompson HE
    4. Sonkusare S
    5. Hallam G
    6. Smallwood J
    7. Jefferies E

    (2015) Automatic and controlled semantic retrieval: TMS reveals distinct contributions of posterior middle temporal gyrus and angular gyrus

    The Journal of Neuroscience 35:15230–15239.

    https://doi.org/10.1523/JNEUROSCI.4705-14.2015

    • PubMed
    • Google Scholar
17. 1. Davey J
    2. Thompson HE
    3. Hallam G
    4. Karapanagiotidis T
    5. Murphy C
    6. De Caso I
    7. Krieger-Redwood K
    8. Bernhardt BC
    9. Smallwood J
    10. Jefferies E

    (2016) Exploring the role of the posterior middle temporal gyrus in semantic cognition: integration of anterior temporal lobe with executive processes

    NeuroImage 137:165–177.

    https://doi.org/10.1016/j.neuroimage.2016.05.051

    • PubMed
    • Google Scholar
18. 1. Diedrichsen J
    2. Wiestler T
    3. Ejaz N

    (2013) A multivariate method to determine the dimensionality of neural representation from population activity

    NeuroImage 76:225–235.

    https://doi.org/10.1016/j.neuroimage.2013.02.062

    • PubMed
    • Google Scholar
19. 1. Dong H-M
    2. Margulies DS
    3. Zuo X-N
    4. Holmes AJ

    (2021) Shifting gradients of macroscale cortical organization mark the transition from childhood to adolescence

    PNAS 118:e2024448118.

    https://doi.org/10.1073/pnas.2024448118

    • PubMed
    • Google Scholar
20. 1. DuPre E
    2. Salo T
    3. Ahmed Z
    4. Bandettini P
    5. Bottenhorn K
    6. Caballero-Gaudes C
    7. Dowdle L
    8. Gonzalez-Castillo J
    9. Heunis S
    10. Kundu P
    11. Laird A
    12. Markello R
    13. Markiewicz C
    14. Moia S
    15. Staden I
    16. Teves J
    17. Uruñuela E
    18. Vaziri-Pashkam M
    19. Whitaker K
    20. Handwerker D

    (2021) TE-dependent analysis of multi-echo fMRI with tedana

    Journal of Open Source Software 6:3669.

    https://doi.org/10.21105/joss.03669

    • Google Scholar
21. 1. Enge A
    2. Abdel Rahman R
    3. Skeide MA

    (2021) A meta-analysis of fMRI studies of semantic cognition in children

    NeuroImage 241:118436.

    https://doi.org/10.1016/j.neuroimage.2021.118436

    • PubMed
    • Google Scholar
22. 1. Erb J
    2. Henry MJ
    3. Eisner F
    4. Obleser J

    (2013) The brain dynamics of rapid perceptual adaptation to adverse listening conditions

    The Journal of Neuroscience 33:10688–10697.

    https://doi.org/10.1523/JNEUROSCI.4596-12.2013

    • PubMed
    • Google Scholar
23. 1. Fedorenko E
    2. Duncan J
    3. Kanwisher N

    (2013) Broad domain generality in focal regions of frontal and parietal cortex

    PNAS 110:16616–16621.

    https://doi.org/10.1073/pnas.1315235110

    • PubMed
    • Google Scholar
24. 1. Gao Z
    2. Zheng L
    3. Chiou R
    4. Gouws A
    5. Krieger-Redwood K
    6. Wang X
    7. Varga D
    8. Ralph MAL
    9. Smallwood J
    10. Jefferies E

    (2021) Distinct and common neural coding of semantic and non-semantic control demands

    NeuroImage 236:118230.

    https://doi.org/10.1016/j.neuroimage.2021.118230

    • PubMed
    • Google Scholar
25. 1. Hong S-J
    2. Xu T
    3. Nikolaidis A
    4. Smallwood J
    5. Margulies DS
    6. Bernhardt B
    7. Vogelstein J
    8. Milham MP

    (2020) Toward a connectivity gradient-based framework for reproducible biomarker discovery

    NeuroImage 223:117322.

    https://doi.org/10.1016/j.neuroimage.2020.117322

    • PubMed
    • Google Scholar
26. 1. Humphreys GF
    2. Lambon Ralph MA

    (2015) Fusion and fission of cognitive functions in the human parietal cortex

    Cerebral Cortex 25:3547–3560.

    https://doi.org/10.1093/cercor/bhu198

    • PubMed
    • Google Scholar
27. 1. Huntenburg JM
    2. Bazin P-L
    3. Margulies DS

    (2018) Large-Scale gradients in human cortical organization

    Trends in Cognitive Sciences 22:21–31.

    https://doi.org/10.1016/j.tics.2017.11.002

    • PubMed
    • Google Scholar
28. 1. Huth AG
    2. de Heer WA
    3. Griffiths TL
    4. Theunissen FE
    5. Gallant JL

    (2016) Natural speech reveals the semantic maps that tile human cerebral cortex

    Nature 532:453–458.

    https://doi.org/10.1038/nature17637

    • PubMed
    • Google Scholar
29. 1. Jackson RL
    2. Hoffman P
    3. Pobric G
    4. Lambon Ralph MA

    (2016) The semantic network at work and rest: differential connectivity of anterior temporal lobe subregions

    The Journal of Neuroscience 36:1490–1501.

    https://doi.org/10.1523/JNEUROSCI.2999-15.2016

    • PubMed
    • Google Scholar
30. 1. Jackson RL

    (2021) The neural correlates of semantic control revisited

    NeuroImage 224:117444.

    https://doi.org/10.1016/j.neuroimage.2020.117444

    • PubMed
    • Google Scholar
31. 1. Jefferies E

    (2013) The neural basis of semantic cognition: converging evidence from neuropsychology, neuroimaging and TMS

    Cortex; a Journal Devoted to the Study of the Nervous System and Behavior 49:611–625.

    https://doi.org/10.1016/j.cortex.2012.10.008

    • PubMed
    • Google Scholar
32. 1. Jefferies E
    2. Thompson H
    3. Cornelissen P
    4. Smallwood J

    (2020) The neurocognitive basis of knowledge about object identity and events: dissociations reflect opposing effects of semantic coherence and control

    Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 375:20190300.

    https://doi.org/10.1098/rstb.2019.0300

    • PubMed
    • Google Scholar
33. 1. Jenkinson M
    2. Smith S

    (2001) A global optimisation method for robust affine registration of brain images

    Medical Image Analysis 5:143–156.

    https://doi.org/10.1016/s1361-8415(01)00036-6

    • PubMed
    • Google Scholar
34. 1. Ju H
    2. Bassett DS

    (2020) Dynamic representations in networked neural systems

    Nature Neuroscience 23:908–917.

    https://doi.org/10.1038/s41593-020-0653-3

    • PubMed
    • Google Scholar
35. 1. Kiefer M
    2. Pulvermüller F

    (2012) Conceptual representations in mind and brain: theoretical developments, current evidence and future directions

    Cortex; a Journal Devoted to the Study of the Nervous System and Behavior 48:805–825.

    https://doi.org/10.1016/j.cortex.2011.04.006

    • PubMed
    • Google Scholar
36. 1. Kivisaari SL
    2. van Vliet M
    3. Hultén A
    4. Lindh-Knuutila T
    5. Faisal A
    6. Salmelin R

    (2019) Reconstructing meaning from bits of information

    Nature Communications 10:927.

    https://doi.org/10.1038/s41467-019-08848-0

    • PubMed
    • Google Scholar
37. Preprint

    1. Krieger-Redwood K
    2. Steward A
    3. Gao Z
    4. Wang X
    5. Halai A
    6. Smallwood J
    7. Jefferies E

    (2022) Creativity in Verbal Associations Is Linked to Semantic Control

    bioRxiv.

    https://doi.org/10.1101/2022.02.08.479385

    • Google Scholar
38. 1. Kundu P
    2. Inati SJ
    3. Evans JW
    4. Luh W-M
    5. Bandettini PA

    (2012) Differentiating BOLD and non-BOLD signals in fMRI time series using multi-echo EPI

    NeuroImage 60:1759–1770.

    https://doi.org/10.1016/j.neuroimage.2011.12.028

    • PubMed
    • Google Scholar
39. 1. Kundu P
    2. Brenowitz ND
    3. Voon V
    4. Worbe Y
    5. Vértes PE
    6. Inati SJ
    7. Saad ZS
    8. Bandettini PA
    9. Bullmore ET

    (2013) Integrated strategy for improving functional connectivity mapping using multiecho fMRI

    PNAS 110:16187–16192.

    https://doi.org/10.1073/pnas.1301725110

    • PubMed
    • Google Scholar
40. 1. Lanzoni L
    2. Ravasio D
    3. Thompson H
    4. Vatansever D
    5. Margulies D
    6. Smallwood J
    7. Jefferies E

    (2020) The role of default mode network in semantic cue integration

    NeuroImage 219:117019.

    https://doi.org/10.1016/j.neuroimage.2020.117019

    • PubMed
    • Google Scholar
41. 1. Mack ML
    2. Preston AR
    3. Love BC

    (2020) Ventromedial prefrontal cortex compression during concept learning

    Nature Communications 11:46.

    https://doi.org/10.1038/s41467-019-13930-8

    • PubMed
    • Google Scholar
42. 1. Margulies DS
    2. Ghosh SS
    3. Goulas A
    4. Falkiewicz M
    5. Huntenburg JM
    6. Langs G
    7. Bezgin G
    8. Eickhoff SB
    9. Castellanos FX
    10. Petrides M
    11. Jefferies E
    12. Smallwood J

    (2016) Situating the Default-mode network along a principal gradient of macroscale cortical organization

    PNAS 113:12574–12579.

    https://doi.org/10.1073/pnas.1608282113

    • PubMed
    • Google Scholar
43. 1. Meteyard L
    2. Cuadrado SR
    3. Bahrami B
    4. Vigliocco G

    (2012) Coming of age: a review of embodiment and the neuroscience of semantics

    Cortex; a Journal Devoted to the Study of the Nervous System and Behavior 48:788–804.

    https://doi.org/10.1016/j.cortex.2010.11.002

    • PubMed
    • Google Scholar
44. 1. Mirman D
    2. Landrigan J-F
    3. Britt AE

    (2017) Taxonomic and thematic semantic systems

    Psychological Bulletin 143:499–520.

    https://doi.org/10.1037/bul0000092

    • PubMed
    • Google Scholar
45. 1. Mkrtychian N
    2. Blagovechtchenski E
    3. Kurmakaeva D
    4. Gnedykh D
    5. Kostromina S
    6. Shtyrov Y

    (2019) Concrete vs. Abstract semantics: from mental representations to functional brain mapping

    Frontiers in Human Neuroscience 13:267.

    https://doi.org/10.3389/fnhum.2019.00267

    • PubMed
    • Google Scholar
46. 1. Montefinese M

    (2019) Semantic representation of Abstract and concrete words: a minireview of neural evidence

    Journal of Neurophysiology 121:1585–1587.

    https://doi.org/10.1152/jn.00065.2019

    • PubMed
    • Google Scholar
47. 1. Murphy C
    2. Rueschemeyer S-A
    3. Watson D
    4. Karapanagiotidis T
    5. Smallwood J
    6. Jefferies E

    (2017) Fractionating the anterior temporal lobe: mvpA reveals differential responses to input and conceptual modality

    NeuroImage 147:19–31.

    https://doi.org/10.1016/j.neuroimage.2016.11.067

    • PubMed
    • Google Scholar
48. 1. Murphy C
    2. Jefferies E
    3. Rueschemeyer S-A
    4. Sormaz M
    5. Wang H-T
    6. Margulies DS
    7. Smallwood J

    (2018) Distant from input: evidence of regions within the default mode network supporting perceptually-decoupled and conceptually-guided cognition

    NeuroImage 171:393–401.

    https://doi.org/10.1016/j.neuroimage.2018.01.017

    • PubMed
    • Google Scholar
49. 1. Murphy C
    2. Rueschemeyer S-A
    3. Smallwood J
    4. Jefferies E

    (2019) Imagining sounds and images: decoding the contribution of unimodal and transmodal brain regions to semantic retrieval in the absence of meaningful input

    Journal of Cognitive Neuroscience 31:1599–1616.

    https://doi.org/10.1162/jocn_a_01330

    • PubMed
    • Google Scholar
50. 1. Noonan KA
    2. Jefferies E
    3. Visser M
    4. Lambon Ralph MA

    (2013) Going beyond inferior prefrontal involvement in semantic control: evidence for the additional contribution of dorsal angular gyrus and posterior middle temporal cortex

    Journal of Cognitive Neuroscience 25:1824–1850.

    https://doi.org/10.1162/jocn_a_00442

    • PubMed
    • Google Scholar
51. 1. Patterson K
    2. Nestor PJ
    3. Rogers TT

    (2007) Where do you know what you know? the representation of semantic knowledge in the human brain

    Nature Reviews Neuroscience 8:976–987.

    https://doi.org/10.1038/nrn2277

    • PubMed
    • Google Scholar
52. 1. Pereira F
    2. Gershman S
    3. Ritter S
    4. Botvinick M

    (2016) A comparative evaluation of off-the-shelf distributed semantic representations for modelling behavioural data

    Cognitive Neuropsychology 33:175–190.

    https://doi.org/10.1080/02643294.2016.1176907

    • PubMed
    • Google Scholar
53. 1. Poppenk J
    2. Evensmoen HR
    3. Moscovitch M
    4. Nadel L

    (2013) Long-axis specialization of the human hippocampus

    Trends in Cognitive Sciences 17:230–240.

    https://doi.org/10.1016/j.tics.2013.03.005

    • PubMed
    • Google Scholar
54. 1. Ralph MAL
    2. Jefferies E
    3. Patterson K
    4. Rogers TT

    (2017) The neural and computational bases of semantic cognition

    Nature Reviews. Neuroscience 18:42–55.

    https://doi.org/10.1038/nrn.2016.150

    • PubMed
    • Google Scholar
55. 1. Rigotti M
    2. Barak O
    3. Warden MR
    4. Wang X-J
    5. Daw ND
    6. Miller EK
    7. Fusi S

    (2013) The importance of mixed selectivity in complex cognitive tasks

    Nature 497:585–590.

    https://doi.org/10.1038/nature12160

    • PubMed
    • Google Scholar
56. 1. Schwartz MF
    2. Kimberg DY
    3. Walker GM
    4. Brecher A
    5. Faseyitan OK
    6. Dell GS
    7. Mirman D
    8. Coslett HB

    (2011) Neuroanatomical dissociation for taxonomic and thematic knowledge in the human brain

    PNAS 108:8520–8524.

    https://doi.org/10.1073/pnas.1014935108

    • PubMed
    • Google Scholar
57. 1. Seghier ML
    2. Fagan E
    3. Price CJ

    (2010) Functional subdivisions in the left angular gyrus where the semantic system meets and diverges from the default network

    The Journal of Neuroscience 30:16809–16817.

    https://doi.org/10.1523/JNEUROSCI.3377-10.2010

    • PubMed
    • Google Scholar
58. 1. Smallwood J
    2. Bernhardt BC
    3. Leech R
    4. Bzdok D
    5. Jefferies E
    6. Margulies DS

    (2021) The default mode network in cognition: a topographical perspective

    Nature Reviews. Neuroscience 22:503–513.

    https://doi.org/10.1038/s41583-021-00474-4

    • PubMed
    • Google Scholar
59. 1. Teige C
    2. Mollo G
    3. Millman R
    4. Savill N
    5. Smallwood J
    6. Cornelissen PL
    7. Jefferies E

    (2018) Dynamic semantic cognition: characterising coherent and controlled conceptual retrieval through time using magnetoencephalography and chronometric transcranial magnetic stimulation

    Cortex; a Journal Devoted to the Study of the Nervous System and Behavior 103:329–349.

    https://doi.org/10.1016/j.cortex.2018.03.024

    • PubMed
    • Google Scholar
60. 1. Teige C
    2. Cornelissen PL
    3. Mollo G
    4. Gonzalez Alam TRDJ
    5. McCarty K
    6. Smallwood J
    7. Jefferies E

    (2019) Dissociations in semantic cognition: oscillatory evidence for opposing effects of semantic control and type of semantic relation in anterior and posterior temporal cortex

    Cortex; a Journal Devoted to the Study of the Nervous System and Behavior 120:308–325.

    https://doi.org/10.1016/j.cortex.2019.07.002

    • PubMed
    • Google Scholar
61. 1. Visser M
    2. Lambon Ralph MA

    (2011) Differential contributions of bilateral ventral anterior temporal lobe and left anterior superior temporal gyrus to semantic processes

    Journal of Cognitive Neuroscience 23:3121–3131.

    https://doi.org/10.1162/jocn_a_00007

    • PubMed
    • Google Scholar
62. 1. Wang J
    2. Conder JA
    3. Blitzer DN
    4. Shinkareva SV

    (2010) Neural representation of Abstract and concrete concepts: a meta-analysis of neuroimaging studies

    Human Brain Mapping 31:1459–1468.

    https://doi.org/10.1002/hbm.20950

    • PubMed
    • Google Scholar
63. 1. Wang X
    2. Margulies DS
    3. Smallwood J
    4. Jefferies E

    (2020) A gradient from long-term memory to novel cognition: transitions through default mode and executive cortex

    NeuroImage 220:117074.

    https://doi.org/10.1016/j.neuroimage.2020.117074

    • 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

Decision letter after peer review:

Thank you for submitting your article "Flexing the principal gradient of the cerebral cortex to suit changing semantic task demands" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Chris Baker as the Senior Editor. The following individuals involved in the review of your submission have agreed to reveal their identity: Xi-Nian Zuo (Reviewer #1); Wei Wu (Reviewer #3).

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 level of analysis is not explained. For instance, for the cortical gradient analysis: are the gradients computed subject by subject and then averaged? Or is it the parcel-by-parcel matrices that are averaged? Or is it a fixed-effect analysis without considering the participant random factor? There are no error bars in Figure S1, so I assume the PCA is performed on the group average. It should be clearly stated and explained, not only for the gradient analysis but also for the dimensionality analysis and the RSA.

2. The performed statistical tests are not sufficiently explained. For example, what are the confidence intervals of their results? What is Pspin? What statistical test was performed to confirm the mediation analysis (if any)? What are the error bars in Figures 2 and 3?

3. Justification is needed for the selection of the time window for the analyses (TRs 4-6 in the semantic task and 9-10 in the chevron task). The authors indicate that the semantic trials are separated by an easy chevron task for 6 s (Line 619), but in Figure 1A this control task lasts for 4.5 s, which is a bit confusing. Also, the authors suggest 10 chevrons were presented one after another in each trial (Line 621). I wonder if 6 s was long enough for participants to look at so many chevrons and make judgements.

4. I wonder if the authors could provide more explanation about why they chose to exclude abstract nouns and pairs of items drawn from the same taxonomic category (Line 609).

5. How do response times vary according to the strength of association? How are response times taken into account in the analysis? For example, if trials with low association strength are longer, how is it taken into account?

6. How was the sample size (N=36) determined? Please clarify. Please provide statistical power, effect size, confidence interval, and corrections if multiple comparisons done for all the reports of statistics.

7. As a major manipulation in the experiment, it is not very clear how the authors split/define their stimuli into strong and weak semantic association conditions. If I understood correctly, word2vec was used to measure the association strength in each pair of words. Then the authors grouped the top 1/3 association strength trials as a "strong association" condition and the bottom 1/3 as "weak association" (Line 689), and all analyses comparing the effect of "strong vs. weak association" were conducted with data from these two subsets of stimuli. However, in multiple places, the authors indicate the association strength of their stimuli ranges from completely unrelated to weakly related to highly related (Line 612, Line 147, Line 690, and the examples in Figure 1B). This makes me wonder if the trials with bottom 1/3 association strength (i.e., those were used in the current study) are actually "unrelated/no association" trials (more like a baseline condition), instead of "weak association" trials as the authors claimed. These two situations could be different regarding how they engage semantic knowledge and control processing. Besides, I am very interested in what will the authors find if they compare all three conditions (i.e., unrelated vs. weak association vs. strong association).

8. Because the comparison between weak vs. strong association conditions is the key to the current study, I feel it might be better to introduce more about the stimuli in these two conditions. Specifically, the authors only suggested the word pairs fell in these two conditions varied in their association strength, but how about other psycholinguistic properties that could potentially confound their manipulation? For example, words with higher frequency and concreteness may engage more automatic/richer long-term semantic information and words with lower frequency and concreteness need more semantic control. I feel there may be a possibility that the effect of semantic association was partly driven by the differences in these measures in different conditions.

9. Behavior: Correlating the mean response and word2vec distance is not very informative (except for confirming that word2vec distance is meaningful for the mean population). Is this correlation accurate at the group level? i.e., if we compute the correlation coefficient per individual and then test the mean of the coefficient at the group level, is it significant?

10. I wonder if the uniqueness of response (Line 627) could be used to calculate a more straightforward measure of association strength for word pairs than word2vec. That is, for each trial, the more participants link the two words in the same way, the stronger the link is. And the fewer participants link the words in the same way, the weaker and more sparse the association is.

11. Some correlation coefficients are puzzlingly high (e.g., 0.84 in Figure 1C). Please discuss.

12. One strength of the gradient method is its continuous spatial variation in regard to the brain divisions, thus a spatially continuous brain parcellation. Thus, it would be more taking such advantage by using a voxel/vertex analysis than the 400-parcel analysis. What is the tenet for the current work to use the large-parcel gradient analysis? Is this something due to the limitation of the computational resource? If so, please clarify, and if not, please discuss.

13. It seems different cortical surface models are used for rendering the spatial maps (e.g., Figure 2ABC vs Figure 2E vs Figure 3A). Please use an identical surface model for better direct comparisons across different visualizations.

14. It might be good practice to add labels (e.g., t value, r value etc.) next to the colour bars in the activation maps, which could make the figures easier to understand.

15. The dimensionality analysis in the current study is novel and interesting. In this section, the authors linked decreasing dimensionality with more abstract and less variable representations. However, most results here were built based on the comparison between the dimensionality effects for strong and weak association conditions. I wonder if these conclusions can be generalised to results within each condition and across different regions (i.e., regions having lower dimensionality are doing more abstract and cross-modal processing). If so, I am curious why the ATL (a semantic "hub") in Figure 3A has higher dimensionality than the sensory-motor cortices (quite experiences related) and AG (another semantic "hub").

16. A more extensive discussion should be accompanied by the revision to highlight the implications and values of the present work for brain-mind development from a lifespan perspective.

17. Why do the authors use a binary separation of the distance while it is a continuous measure? If this is an actual modulation, the intermediary bin (intermediate values not included in the analyses) should present an intermediary profile.

Essential revisions:

1. The level of analysis is not explained. For instance, for the cortical gradient analysis: are the gradients computed subject by subject and then averaged? Or is it the parcel-by-parcel matrices that are averaged? Or is it a fixed-effect analysis without considering the participant random factor? There are no error bars in Figure S1, so I assume the PCA is performed on the group average. It should be clearly stated and explained, not only for the gradient analysis but also for the dimensionality analysis and the RSA.

Sorry that we did not make these details clear in our initial submission. We now include more details about the level of analysis in both the method and Results sections.

a. The cortical gradients were computed by using group averaged parcel-by-parcel matrices. Please see page 35:

‘We performed gradient decomposition on a group-averaged parcel-by-parcel matrix of informational connectivity, rather than extracting the simple Pearson time series correlation between pairs of regions, since informational connectivity corresponds to the similarity in the multivariate response of pairs of brain regions at different points in time (reflecting the multivariate response to trials in this experiment).’

Also see page 10:

‘PCA was used to decompose the group-averaged connectivity matrices, with the resulting components (or gradients) reflecting spatial patterns of connectivity across the cortex.’

b. We concatenated all participants’ neuroimaging data into one 4D file for one single trial and performed PCA on this to estimate the dimensionality map, then iterated the same procedure for all trials. The statistical test was conducted to test the association strength effect between conditions. In this way, the analysis considers group-level effects. Please see page 37.

‘Next, we examined the effect of strength of association on the dimensionality of the neural response (i.e., between strongly and weakly associated trials) in the semantic task, across the whole group of participants.’

c. The RSA analysis was conducted on the group level. We firstly constructed neural pattern similarity matrices between trials at the individual level, then correlated these matrices between participants using a searchlight approach. We also calculated the similarity of behavioral rating patterns between participants to construct a semantic rating similarity matrix. Lastly, we correlated these two group level matrices between neural and semantic levels. Please see page 36.

‘A second-order representational similarity analysis was then conducted by correlating the semantic and brain models to assess brain-behaviour alignment at the group level.’

2. The performed statistical tests are not sufficiently explained. For example, what are the confidence intervals of their results? What is Pspin? What statistical test was performed to confirm the mediation analysis (if any)? What are the error bars in Figures 2 and 3?

Thanks very much for identifying these points where more information is needed. We have significantly updated the statistical reporting throughout the manuscript according to the reviewer’s comments.

a. Confidence intervals (95%) were used and are now systematically reported in the manuscript. Please see page 13:

‘Simple t-tests revealed significant differences in principal gradient values for the semantic task across SCN and SCN+MDN compared with MDN regions (t(30) = 3.259, p = 0.0167, 95% CI = [0.279 ~ 1.215], Cohen’s d = 0.827; and t(30) = 3.904, p = 0.003, 95% CI = [0.334 ~ 1.066], Cohen’s d = 0.965, respectively).’

And see page 22:

‘Consistent with our prediction, we found the correlation between the principal gradient and semantic-brain alignment could be partially explained by this dimensionality difference (indirect effect: a x b = 0.0053, 95% CI = [0.0039 ~ 0.01], p < 0.001, explaining 51.7% variance; direct effect: c’ = 0.0049, 95% CI = [0.0026 ~ 0.01], p < 0.001).’

b. The error bars in Figure 2D and Figure 3C represent the standard error of gradient differences and of dimensionality differences across parcels in each brain network.

Please see page 20:

‘Error bars represent the standard error of dimensionality differences across parcels within each network.’

c. Pspin represents the significance of the spin permutation test. The significance of the correlation between gradients estimated by our task-based data and gradients estimated by using resting state data were assessed by using spin permutation to control the spatial autocorrelation in MRI data. Please see page 10:

‘This similarity between cortical gradients was established by using spin permutation tests, which preserve the spatial autocorrelation present within gradients. Statistical significance was determined by comparing with a null distribution using the permutation approach (5000 permutations), in which we randomly rotated the spatial patterns and measured the correlations between maps that preserve spatial autocorrelation (Alexander-Bloch et al., 2018). The first gradient, accounting for the greatest variance in informational connectivity, was highly overlapping with the principal gradient of intrinsic connectivity identified by prior work (Margulies et al., 2016), for both the semantic and chevron tasks (Spin permutation p value: P_spin < 0.001), see Figure 2 —figure supplement 1.’

d. The significance of mediation effect was tested by using a bootstrapping approach which was highly recommended recently (Shrout and Bolger, 2002).

Please see page 22:

‘To test this hypothesis, we performed a mediation analysis in which the principal gradient was the predictor, semantic-brain alignment was the dependent variable, and the effect of strength of association on the dimensionality of the neural response was the mediator. The significance of the mediation effect was tested by using the bootstrapping approach with 5000 samples. Consistent with our prediction, we found the correlation between the principal gradient and semantic-brain alignment could be partially explained by this dimensionality difference (indirect effect: a x b = 0.0053, 95% CI = [0.0039 ~ 0.01], p < 0.001, explaining 51.7% variance; direct effect: c’ = 0.0049, 95% CI = [0.0026 ~ 0.01], p < 0.001).’

3. Justification is needed for the selection of the time window for the analyses (TRs 4-6 in the semantic task and 9-10 in the chevron task). The authors indicate that the semantic trials are separated by an easy chevron task for 6 s (Line 619), but in Figure 1A this control task lasts for 4.5 s, which is a bit confusing. Also, the authors suggest 10 chevrons were presented one after another in each trial (Line 621). I wonder if 6 s was long enough for participants to look at so many chevrons and make judgements.

a. Thanks for pointing this out. The BOLD signal reaches its peak around 4 ~ 8s after the stimuli’s onset. Considering the TR is 1.5s, we chose TRs (4 ~ 6) which correspond to 4.5 ~ 9s after stimuli’s onset to represent the semantic neural responses: this is the period expected to correspond most strongly to the semantic response. The chevron task started at the 5^th TR after stimuli onset but considering task-switching effects may contaminate the first few chevron trials, and to avoid potential contamination of neural responses from the forthcoming semantic period, we excluded the first chevron TR (started at TR 9) and ended this period one TR earlier (at TR 10); this period covers the peak response period for the chevron task 6 ~ 9s after trial onset. Please see page 34:

‘Because the BOLD signal reaches its peak around 4 ~ 8s after the stimuli’s onset, the brain data were averaged within-trial across three time points for the semantic task (4.5~9s after trial onset; TRs 4-6). To avoid potential contamination of neural responses from task-switching processes and the forthcoming semantic trial, we selected two time points for the chevron task (12~15s after trial onset; TRs 9-10), resulting in one pattern of brain activity for the semantic and chevron tasks per trial.’

b. The chevron task lasted for 6s in each trial. There was a typo in Figure 1A, which has been corrected. Sorry for the confusion caused by this. Please see page 9:

c. The participants’ averaged reaction time was 0.336s which was lower than the 0.6s that participants had to perform each chevron decision. The accuracy was 72.6%. Therefore, the chevron task was challenging enough to engage the participant, but not so challenging that they could not do it. Please see page 8.

‘On the chevron task, the average reaction time was 0.336s and the accuracy was 72.6%; these data suggest the task was challenging enough to engage the participant but not so difficult that they could not do it. The fast pace of this task reduced participants’ ability to think about semantic links across trials.’

4. I wonder if the authors could provide more explanation about why they chose to exclude abstract nouns and pairs of items drawn from the same taxonomic category (Line 609).

There is a debate about the nature of abstract and concrete concepts, with previous studies suggesting that they recruit partially different brain mechanisms (Montefinese, 2019, Wang et al., 2010). Previous studies have also suggested that there are distinct differences between the processes supporting taxonomic and thematic knowledge: taxonomic knowledge is based on shared features (e.g., dog – cat), while thematic knowledge is based on co-occurrence in events or scenarios (Mirman et al., 2017). Again, studies have identified some differences in the neural response to these different types of semantic relationships (Mirman et al., 2017, Schwartz et al., 2011, Lewis et al., 2015). To avoid potential confounds caused by these differences, we excluded taxonomic noun pairs in the current study. We have now provided an explanation for this decision in the method of our manuscript. Please see page 31.

‘Because abstract and concrete meanings, and taxonomic and thematic semantic relations, at least partially recruit different brain mechanisms (Mkrtychian et al., 2019, Mirman et al., 2017, Montefinese, 2019, Wang et al., 2010, Schwartz et al., 2011), we excluded abstract nouns and pairs of items drawn from the same taxonomic category, so that only thematic links were evaluated.’

5. How do response times vary according to the strength of association? How are response times taken into account in the analysis? For example, if trials with low association strength are longer, how is it taken into account?

We agree with the reviewer that response times vary according to the strength of association; it usually takes a longer time to verify semantic links for trials that are weakly associated (Teige et al., 2019, Gao et al., 2021). When participants make easier and harder semantic decisions, they are expected to spend more time on the harder task, since after a response has been made, participants are often not thinking about the task. However, our task was different, since participants had to generate a meaningful link between all pairs of words, irrespective of strength of association – they could do this for strongly-related items but also for unrelated items. They had 4.5 seconds to think about the relationship between the words and then rated the strength of the link that they had generated. In this way, the task structure does not give rise to clear differences in time on task across trials, although we acknowledge that there might be differences in the time spent on the initial generation of ideas compared with their elaboration.

We selected TR 4-6 to represent semantic signals in the brain since this corresponded to 6 ~ 9s after stimuli onset, which covers the peak signal of BOLD response (4 ~ 8s), please see more details from our reply to the comment 3.

Please see page 32:

‘We used a slow event-related design, with trials lasting 13.5s. Each trial began with a visually presented word-pair for 4.5s; during this period, participants were tasked with identifying and encouraged to keep thinking about a link between the words. Following the link generation for each trial, there was 1.5s to rate the strength of the link they retrieved.’

6. How was the sample size (N=36) determined? Please clarify. Please provide statistical power, effect size, confidence interval, and corrections if multiple comparisons done for all the reports of statistics.

No explicit power analysis was performed to determine the sample size in this experiment, however, the present sample size (N=31) is comparable to, or significantly exceeds, that of prior task-based fMRI studies examining the neural basis of semantic cognition (e.g., Ubaldi et al., 2022, J.Neuroscience; Gao et al., 2022, Cerebral Cortex; Sormaz et al., 2018, PNAS).

We have extensively modified the manuscript to add reports of statistical power, effect sizes, confidence intervals and we have explicitly stated whether corrections are performed for multiple comparisons, with this correction applied whenever relevant.

Please see page 13:

‘Simple t-tests revealed significant differences in principal gradient values for the semantic task across SCN and SCN+MDN compared with MDN regions (t(30) = 3.259, p = 0.0167, 95% CI = [0.279 ~ 1.215], Cohen’s d = 0.827; and t(30) = 3.904, p = 0.003, 95% CI = [0.334 ~ 1.066], Cohen’s d = 0.965, respectively).’

‘We found a significant interaction between brain network and task (F(2.53, 75.887) = 8.345, p < 0.001, η2p = 0.218 ) and significant main effects of task (F(1, 30) = 21.947, p < 0.001, η2p = 0.414) and network (F(2.01, 60.288) = 3.338, p = 0.042, η2p = 0.100), see Figure 2D.’

And see page 16:

‘The difference in principal gradient values between strong and weak association trials showed a significant interaction effect between functional networks and tasks (F(2.53, 75.887) = 8.345, p < 0.001, η2p = 0.218 ). Error bars represent the standard error of gradient difference across parcels within each network.’

And see page 22:

‘Consistent with our prediction, we found the correlation between the principal gradient and semantic-brain alignment could be partially explained by this dimensionality difference (indirect effect: a x b = 0.0053, 95% CI = [0.0039 ~ 0.01], p < 0.001, explaining 51.7% variance; direct effect: c’ = 0.0049, 95% CI = [0.0026 ~ 0.01], p < 0.001).’

7. As a major manipulation in the experiment, it is not very clear how the authors split/define their stimuli into strong and weak semantic association conditions. If I understood correctly, word2vec was used to measure the association strength in each pair of words. Then the authors grouped the top 1/3 association strength trials as a "strong association" condition and the bottom 1/3 as "weak association" (Line 689), and all analyses comparing the effect of "strong vs. weak association" were conducted with data from these two subsets of stimuli. However, in multiple places, the authors indicate the association strength of their stimuli ranges from completely unrelated to weakly related to highly related (Line 612, Line 147, Line 690, and the examples in Figure 1B). This makes me wonder if the trials with bottom 1/3 association strength (i.e., those were used in the current study) are actually "unrelated/no association" trials (more like a baseline condition), instead of "weak association" trials as the authors claimed. These two situations could be different regarding how they engage semantic knowledge and control processing. Besides, I am very interested in what will the authors find if they compare all three conditions (i.e., unrelated vs. weak association vs. strong association).

We appreciate the reviewer pointing this out. The association strength of word pairs ranged from completely unrelated to highly related, reflected in both word2vec values (-0.05 to 0.72) and participants’ semantic ratings (from 0 ~ 4). In this way, the reviewer is right that the bottom third of trials largely consisted of words that might be thought of as ‘unrelated’ in other contexts, although the task still required participants to come up with a meaningful way of linking these items – and in this way, we did not include any “no association” trials.

When we divided our stimuli into three sets as requested by the reviewer, the average semantic rating ranged from 1.59 (bottom 1/3) through 2.82 (middle 1/3) to 3.51 (top 1/3), with extremely significant difference between all sets of items (all raw p values were lower than 2.83e-14). Therefore, we have rephrased our description, now referring to the top and bottom 1/3 stimuli as more strongly and less strongly associated.

Please see page 35:

‘We grouped the semantic judgements into strongly associated (top 1/3 trials, with a mean rating strength of 3.51) and weakly associated (bottom 1/3 trials, with a mean rating strength of 1.59) sets, according to the strength of association between the words within each trial (as measured by word2vec, widely used in the field (Kivisaari et al., 2019, Pereira et al., 2016)).’

We have conducted additional analysis for the intermediary bin and compare it against the bottom for the gradient analysis and against the top 1/3 for the dimensionality analysis (compared to the baseline condition for each analysis), which did show a similar patten like the contrast between strong and weak association but with a smaller effect, thus representing an intermediary profile as expected. The correlation between the principle gradient difference between middle and weak association with the principle gradient value derived from resting state was also significant, see Figure 2 —figure supplement 3C, but its magnitude was smaller than what we reported in the main body of manuscript (r = 0.235 vs. r = 0.369). Given that the expected strongest effect is between top and bottom 1/3, thus, we have now included these results in the supplementary materials. Please see Figure 2 —figure supplement 3 in page 44.

8. Because the comparison between weak vs. strong association conditions is the key to the current study, I feel it might be better to introduce more about the stimuli in these two conditions. Specifically, the authors only suggested the word pairs fell in these two conditions varied in their association strength, but how about other psycholinguistic properties that could potentially confound their manipulation? For example, words with higher frequency and concreteness may engage more automatic/richer long-term semantic information and words with lower frequency and concreteness need more semantic control. I feel there may be a possibility that the effect of semantic association was partly driven by the differences in these measures in different conditions.

Thanks for raising this point. We have performed additional control analysis to examine the relationship between association strength and psycholinguistic features according to the reviewer’s suggestion. The association strength did not show significant correlation with word frequency (r = -0.010, p = 0.392), concreteness (r = -0.092, p = 0.285) or imageability (r = 0.074, p = 0.377). Direction comparison of these psycholinguistic features between more strongly and les strongly associated word-pairs also did not any significant difference: frequency (t = 0.912, p = 0.364), concreteness (t = 1.576, p = 0.119), imageability (t = 1.451, p = 0.153). Please see page 32:

‘The association strength did not show significant correlation with word frequency (r = -0.010, p = 0.392), concreteness (r = -0.092, p = 0.285) or imageability (r = 0.074, p = 0.377).’

9. Behavior: Correlating the mean response and word2vec distance is not very informative (except for confirming that word2vec distance is meaningful for the mean population). Is this correlation accurate at the group level? i.e., if we compute the correlation coefficient per individual and then test the mean of the coefficient at the group level, is it significant?

We agree with the reviewer that the correlation analysis conducted at the group level could not reveal the relationship for the individual level, though it does provide us clear and strong evidence that word2vec was a reliable measure of association strength. Following the reviewer’s suggestion, we performed a linear mixed effect model analysis to test whether word2vec was a reliable predictor of semantic ratings at the individual level. Participants were included as a random effect and word2vec was treated as a predictor of semantic ratings. We used the likelihood ratio test to compare models with and without word2vec to determine whether the inclusion of word2vec significantly improved the model fit. In line with the correlation analysis between mean response and word2vec, the results revealed a similar significant relationship (χ2(1) = 2266.2, p < 2.3-e16). Considering that our key conclusions nearly were all based on the group-level analysis, thus we think the group-level correlation analysis is an informative and suitable measure here.

10. I wonder if the uniqueness of response (Line 627) could be used to calculate a more straightforward measure of association strength for word pairs than word2vec. That is, for each trial, the more participants link the two words in the same way, the stronger the link is. And the fewer participants link the words in the same way, the weaker and more sparse the association is.

We agree with the reviewer that the uniqueness of responses is closely associated with association strength: fewer participants link the words in the same way when the association strength is weaker. Theoretically, with enough data, the uniqueness of responses could reflect association strength in a stable and precise way. However, the current sample was not large enough to fully reflect shared knowledge in the culture, unlike word2vec. Word2vec was trained on a super large corpus dataset which aims to measure the cooccurrence of words in similar scenarios; this not only captures the definition of association strength for thematic knowledge but has proved a reliable and excellent predictor of association strength in the field. This is why we used word2vec as the index of association strength in the current study.

11. Some correlation coefficients are puzzlingly high (e.g., 0.84 in Figure 1C). Please discuss.

Word2vec was trained on a large corpus dataset using a deep neural network, and other studies have already found it is an excellent predictor of association strength (Hoffman, 2018, Pereira et al., 2016): in this way, this strong correlation is not unexpected. In another study using similar design in our lab (Gao et al., 2021), we collected 192 word pairs and calculated the correlation coefficient between word2vec and semantic ratings, on average across the whole sample. We obtained similar results: Word2vec was strongly positively correlated with rating score (r = 0.773, p < 0.0001), showing that people were more likely to judge word pairs as related when they had high word2vec values.

12. One strength of the gradient method is its continuous spatial variation in regard to the brain divisions, thus a spatially continuous brain parcellation. Thus, it would be more taking such advantage by using a voxel/vertex analysis than the 400-parcel analysis. What is the tenet for the current work to use the large-parcel gradient analysis? Is this something due to the limitation of the computational resource? If so, please clarify, and if not, please discuss.

We agree with reviewer that it would be ideal to measure brain gradients using a voxel/vertex analysis. However, as the reviewer pointed out, the computational resource would be extremely high, as the time series of neural similarity matrices is far larger than traditional averaged univariate time series in each voxel/vertex-centered cubes (searchlight-like analysis). Moreover, 400-parcels is the default option within the BrainSpace package. We have clearly stated the reason for using this parcellation in the revised method, please see in page 35.

‘Due to limitation of computational resources, we chose to use the default Schafer’s 400 region parcellation in Brainspace.’

13. It seems different cortical surface models are used for rendering the spatial maps (e.g., Figure 2ABC vs Figure 2E vs Figure 3A). Please use an identical surface model for better direct comparisons across different visualizations.

Thank you very much for raising this. We have replotted all relevant figures according to the reviewer’s comment. Now all spatial maps are projected on the identical cortical surface model.

14. It might be good practice to add labels (e.g., t value, r value etc.) next to the colour bars in the activation maps, which could make the figures easier to understand.

Thanks for raising this, we have now updated all relevant figures according to the reviewer’s comment.

15. The dimensionality analysis in the current study is novel and interesting. In this section, the authors linked decreasing dimensionality with more abstract and less variable representations. However, most results here were built based on the comparison between the dimensionality effects for strong and weak association conditions. I wonder if these conclusions can be generalised to results within each condition and across different regions (i.e., regions having lower dimensionality are doing more abstract and cross-modal processing). If so, I am curious why the ATL (a semantic "hub") in Figure 3A has higher dimensionality than the sensory-motor cortices (quite experiences related) and AG (another semantic "hub").

The dimensionality and its relationship to the cortical gradient was also examined for each condition, indeed, we also assessed whether this relationship was influenced by associative strength, averaging estimates derived for sets of four trials with similar word2vec values using a ‘sliding window’ approach. There was a negative correlation between overall dimensionality (averaged across all trials) and principal gradient values. And the magnitude of this negative relationship increased as a function of the association strength (and importantly, all correlation values (in Figure 3D) are negative). So, we believe our conclusion could be generalized across conditions although it would be more difficult to observe this relationship if only a subset of trials was included in the analysis.

In our results, we observed higher dimensionality in ATL/frontal orbital cortex than sensory-motor cortices, which seems contradictory to our conclusion. However, these areas are subject to severe distortion and signal loss in functional MRI (Binney et al., 2010). This low tSNR may have produced a higher dimensionality estimation in PCA. We conducted a control analysis in which regions in the limbic network were removed due to their low tSNR, and the negative relationship between dimensionality and principal gradient remained unchanged (r = -0.346, p = 0.038).

Please see the Discussion on page 30.

‘It is worth noting that not all brain regions showed the expected pattern in the dimensionality analysis – especially when considering the global dimensionality of all semantic trials, as opposed to the influence of strength of association in the semantic task. In particular, the limbic network, including regions of ventral ATL thought to support a heteromodal semantic hub, showed significantly higher dimensionality than sensory-motor areas – these higher-order regions are expected to show lower dimensionality corresponding to more abstract representations. However, this analysis does not assess the psychological significance of data dimensionality differences (unlike our contrast of strong and weak associations, which are more interpretable in terms of semantic cognition). Limbic regions are subject to severe distortion and signal loss in functional MRI, which might strongly influence this metric. Future studies using data acquisition and analysis techniques that are less susceptible to this problem are required to fully characterize global dimensionality and its relation to the principal gradient.’

16. A more extensive discussion should be accompanied by the revision to highlight the implications and values of the present work for brain-mind development from a lifespan perspective.

Thanks for raising this important issue. We have included relevant discussion in the discussion part for this important question. Please see it on page 26.

‘The human capacity to retrieval meaning from sounds, visual objects, words and phrases far exceeds the capacities of other animal species, however, its development is poorly understood (Enge et al., 2021). One recent study (Dong et al., 2021) found that, in children, the overarching organizational gradient is anchored within the unimodal cortex, between somatosensory/motor and visual territories. While in adolescence, the principal gradient of connectivity transitions into an adult-like spatial framework, with the default network at the opposite end of a spectrum from primary sensory and motor regions. Whether the development of semantic system is paralleled with gradual transition of the principal gradient in the brain remain to be an open question and is required with future studies.’

17. Why do the authors use a binary separation of the distance while it is a continuous measure? If this is an actual modulation, the intermediary bin (intermediate values not included in the analyses) should present an intermediary profile.

Thanks very much for raising this point out. We have conducted additional analysis for the intermediary bin and compare it against the bottom for the gradient analysis and against the top 1/3 for the dimensionality analysis (compared to the baseline condition for each analysis), which did show a similar patten like the contrast between strong and weak association but with a smaller effect, thus representing an intermediary profile as expected. The correlation between the principle gradient difference between middle and weak association with the principle gradient value derived from resting state was also significant, see Figure 2 —figure supplement 3C, but its magnitude was smaller than what we reported in the main body of manuscript (r = 0.235 vs. r = 0.369). Given that the expected strongest effect is between top and bottom 1/3, thus, we have now included these results in the Figure 2 —figure supplement 3. Please see page 44.

Author details

1. Zhiyao Gao

   Department of Psychology, University of York, New York, United Kingdom

   Contribution

   Conceptualization, Resources, Data curation, Software, Formal analysis, Validation, Investigation, Visualization, Methodology, Writing - original draft, Project administration, Writing - review and editing

   For correspondence

   zhiyao.gao@york.ac.uk

   Competing interests

   No competing interests declared

   "This ORCID iD identifies the author of this article:" 0000-0002-8909-8096

2. Li Zheng

   Department of Psychology, University of Arizona, Tucson, United States

   Contribution

   Formal analysis, Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

3. Katya Krieger-Redwood

   Department of Psychology, University of York, New York, United Kingdom

   Contribution

   Data curation, Investigation, Writing - review and editing

   Competing interests

   No competing interests declared

4. Ajay Halai

   MRC Cognition and Brain Sciences Unit, University of Cambridge, Cambridge, United Kingdom

   Contribution

   Methodology, Writing - review and editing

   Competing interests

   No competing interests declared

5. Daniel S Margulies

   Integrative Neuroscience and Cognition Center (UMR 8002), Centre National de la Recherche Scientifique, Paris, France

   Contribution

   Writing - review and editing

   Competing interests

   No competing interests declared

6. Jonathan Smallwood

   Department of Psychology, Queens University, Kingston, Canada

   Contribution

   Writing - review and editing

   Competing interests

   No competing interests declared

7. Elizabeth Jefferies

   Department of Psychology, University of York, New York, United Kingdom

   Contribution

   Conceptualization, Supervision, Funding acquisition, Writing - original draft, Writing - review and editing

   For correspondence

   beth.jefferies@york.ac.uk

   Competing interests

   No competing interests declared

• 1,323

  Page views

• 148

  Downloads

• 2

  Citations

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