Introduction
Contemporary accounts of brain organisation conceptualise cognition as reflecting interactions of large-scale networks of brain regions, organised in a systematic fashion along cortical gradients. These gradients capture similarities in connectivity patterns across disparate areas of the cortex (Bressler and Menon
2010; Margulies et al.
2016; Medaglia et al.
2015; Paquola et al.
2018; Yeo et al.
2011). Cortical gradients provide a new tool for understanding patterns of hemispheric specialisation, since networks with lateralised connectivity will occupy different positions along these gradients in the left (LH) and right hemispheres (RH). This study exploits the potential of cortical gradients to uncover hemispheric differences in patterns of intrinsic connectivity, (i) by assessing the position of canonical networks in the left and right hemisphere along gradients derived bilaterally, and (ii) by examining the functional significance of these hemispheric differences for the highly left-lateralised domain of semantic cognition, compared with other cognitive domains (working memory (WM) and visual reasoning) that are expected to show different patterns of lateralisation.
The principal gradient, which explains the most variance in whole-brain decompositions of intrinsic connectivity, captures the separation between sensory-motor cortex and heteromodal Default Mode Network (DMN) (Huntenburg et al.
2018; Margulies et al.
2016). In this way, it relates to previously described cortical hierarchies that extract progressively more complex or heteromodal information from sensory inputs, or that maintain more abstract goals for action, in lateral and medial temporal lobes, and lateral and medial prefrontal cortex (Badre
2008; Badre and D’Esposito
2007; Bajada et al.
2017,
2019; Fuster
2001; Jackson et al.
2017,
2019; Koechlin et al.
2003; Petrides
2005; Thiebaut De Schotten et al.
2017). The principal gradient goes beyond these observations to explain why similar hierarchies occur in multiple brain regions. The principal gradient is correlated with physical distance along the cortical surface from primary systems, with the DMN falling at a maximum distance from sensory and motor systems in multiple locations across the cortex. Since DMN is a highly distributed network, with multiple nodes located in distant brain regions, the functional transitions captured by the principal gradient are repeated across the cortex, and these are seen in both hemispheres. The principal gradient also captures the sequence of networks found along the cortical surface—from DMN, through frontoparietal control networks, to attention networks (Dorsal and Ventral, DAN and VAN) and finally primary somatomotor and visual networks. A recent study showed that when gradient decomposition is performed for the two hemispheres separately, both hemispheres contain a similar (but not identical) principal gradient (Liang et al.
2021). However, the functional relevance of these similarities and differences between the left and right hemisphere has not been established.
Patterns of intrinsic connectivity tend to be highly symmetrical, with the strongest time-series correlations seen between homotopic regions that occupy the same position in the two hemispheres (Jo et al.
2012). However, symmetrical patterns of connectivity are weaker within heteromodal networks towards the DMN apex of the gradient (Raemaekers et al.
2018). These increasing asymmetries are related to structural connectivity: primary cortices are connected across the hemispheres through fast fibres of the corpus callosum, while heteromodal cortices are connected by slower fibres that show less homotopic connectivity (Stark et al.
2008). A recent study using large-scale novel meta-analytic and voxel mirroring methods confirmed that areas with less similar connectivity across hemispheres are associated with heteromodal functions, such as memory, language and executive control (Mancuso et al.
2019). Moreover, higher-order networks, including DMN, frontoparietal network (FPN) and dorsal attention network (DAN), show the highest degrees of interhemispheric differences in intrinsic connectivity (Karolis et al.
2019; Wang et al.
2014). These lateralised patterns of connectivity have functional significance, giving rise to lateralised functions like verbal semantics and other components of language (Joliot et al.
2016; Knecht et al.
2000) and aspects of attention (Bartolomeo and Seidel Malkinson
2019). For example, Gotts et al. (
2013) identified that a ‘segregation’ mode of lateralisation in the left hemisphere (i.e., heightened intrinsic connectivity with other left hemisphere regions), conferred behavioural advantages in a verbal semantic task (vocabulary). In contrast, cross-hemisphere connections for the right hemisphere were related to better visual reasoning (block design). Given that segregated connectivity is also associated with higher-order heteromodal networks, we would expect this left hemisphere semantic pattern to involve lateralised connectivity at the heteromodal end of the gradient.
Previous studies have identified hemispheric differences in control networks, situated between DMN and sensory-motor cortex. In the left hemisphere, the frontoparietal control network couples preferentially to DMN and language regions, while in the right hemisphere, this network shows stronger connectivity to attentional regions (Wang et al.
2014). These findings suggest that control networks might be critical for the emergence of lateralised cognition. In line with this view, the most lateralised regions of the semantic network are associated with controlled semantic retrieval, as opposed to conceptual representation (Gonzalez Alam et al.
2019). Furthermore, the clustering of connectivity patterns within the FPN across hemispheres reveals a bipartite organisation, with one subnetwork showing more intrinsic connectivity to DMN, whilst the other shows more connectivity to DAN (Dixon et al.
2018). These subnetworks may support the capacity of the FPN to couple efficiently with the DAN and DMN, depending on the task (Niendam et al.
2012; Spreng et al.
2013; Vincent et al.
2008; Wang et al.
2014). These observations collectively give rise to the hypothesis that differences in network interactions between the hemispheres might be reflected in the location of control networks on the principal gradient, with left hemisphere control regions nearer to DMN, and right hemisphere control areas nearer to the sensory-motor end of the gradient. In line with this view, Davey et al. (
2016) suggested that left-lateralised semantic control processes reflect an interaction of heteromodal conceptual representations, associated with DMN, and control processes that can promote the retrieval of currently-relevant aspects of knowledge, even when these are not dominant in long-term memory. Semantic cognition may be left lateralised because these DMN and control networks interact more strongly in the left hemisphere.
This study contrasted patterns of lateralisation for semantic cognition with working memory (Studies 1 and 2) and visual reasoning using matrix problems (Study 1). Since semantic cognition is thought to draw on left-lateralised interactions between DMN and control regions (Davey et al.
2016), working memory tasks provide an interesting contrast: increased working memory demands are expected to increase reliance on a bilateral multiple-demand network that supports executive demands across tasks (Duncan
2001,
2010; Fedorenko et al.
2013; Hugdahl et al.
2015), with the differential engagement of left and right hemispheres when verbal and spatial working memory tasks are compared (Emch et al.
2019; Hong et al.
2000). The multiple-demand network is adjacent to but somewhat spatially distinct from the semantic control network (Davey et al.
2016; Gao et al.
2021; Jackson
2021), and its recruitment typically shows less lateralisation (Camilleri et al.
2018; Müller et al.
2015; Rottschy et al.
2012). Visual reasoning tasks are also expected to show a distinct pattern of lateralisation compared with semantic cognition. While semantic cognition is strongly left-lateralised, meta-analytic and patient evidence suggests a bilateral basis for a wide variety of reasoning tasks (Hobeika et al.
2016; Wertheim and Ragni
2018; Shin and Jeon
2021; Gläscher et al.
2010). Shin and Jeon (
2021) found common bilateral activation in multiple demand cortex across inductive and deductive tasks, with stronger responses in the right hemisphere for more complex tasks. In some studies, matrix reasoning is more right-lateralised than analogical reasoning, consistent with this task’s greater visual and spatial demands (Hobeika et al.
2016; but see also Wertheim and Ragni
2018). Visual reasoning tasks are expected to involve an interaction of control/attention networks with visual regions (Hearne et al.
2017), without strong engagement of memory processes in DMN; this pattern of network interaction might be bilateral or stronger in the right hemisphere. Structural equation modelling has shown that while executive and perceptual attention both contribute to the RAPM, executive control plays a larger role (Schweizer and Moosbrugger
2004; Ren et al.
2012;
2013). Although simple visual attention is thought to be right-lateralised (Kinsbourne
1987; Fink et al.
2000,
2001), the RAPM gives rise to bilateral responses with some evidence pointing to a right-lateralised bias under certain conditions (Prabhakaran et al.
1997; Bishop et al.
2008).
Study 1 examined the organisation of the principal gradient across the left and right human cerebral hemispheres in participants who took part in a resting-state scan (
N = 253) and behavioural tasks in a separate session (
N = 175). We considered how individual differences in semantic cognition related to the position of large-scale networks on the principal gradient of intrinsic connectivity defined by Margulies et al. (
2016), in the left versus right hemispheres, deriving a hemispheric difference gradient score per network for each participant. The semantic component that we examined was derived from a wide variety of semantic tasks, and is likely to reflect the capacity to access relevant conceptual knowledge in different contexts. We would expect this component to be left-lateralised as all elements of the semantic cognition network appear to show a left-hemisphere bias in meta-analytic evidence (the map for the term ‘semantic cognition’ from Neurosynth, for example, is highly left-lateralised), with the possible exception of the anterior temporal lobe (Rice et al.
2015a,
b; Jackson et al.
2017; Gonzalez Alam et al.
2019). These effects for semantic judgements were compared with Raven’s Advanced Progressive Matrices (RAPM), a measure of non-verbal reasoning, and Digit Span, a measure of verbal working memory, allowing us to determine whether hemispheric differences on the principal gradient are related in distinct ways to left-lateralised semantic cognition (Gonzalez Alam et al.
2019; Jackson
2021; Noonan et al.
2013), compared with other types of demanding cognition. Using similar methods, Mckeown et al. (
2020) found associations between individual differences in gradient values and patterns of spontaneous thought, suggesting that variation in gradient organisation is reflected in people’s cognition and experience.
Having established that individual differences in semantic cognition were associated with the magnitude of hemispheric differences on the principal gradient in a specific control network in Study 1, we examined how semantic and non-semantic task demands modulated activation within this lateralised control network in Study 2. We re-analysed fMRI data examining parametric manipulations of difficulty in semantic and verbal working memory tasks (Gao et al.
2021). Controlled semantic retrieval demands were varied by presenting word pairs that were strongly or more weakly associated: previous studies have shown greater recruitment of the left-lateralised semantic control network when participants are required to identify weak associations that are not dominant within the semantic store (Jackson
2021; Noonan et al.
2013). Semantic control demands were compared with the effects of working memory load, since higher working memory demands increase the recruitment of bilateral multiple-demand cortex that is partially distinct from the semantic control network (Fedorenko et al.
2013). In this way, we assessed whether hemispheric differences in the position of large-scale networks on the principal gradient of intrinsic connectivity at rest corresponded with hemispheric differences in the recruitment of these networks during task performance.
Discussion
This study investigates the lateralisation of function along the principal gradient—a key topographical component of large-scale intrinsic connectivity that captures the separation of unimodal and heteromodal cortex (Margulies et al.
2016). We show that intrinsic connectivity patterns in the two hemispheres are situated at different points along the principal gradient: overall, left hemisphere parcels are closer to the heteromodal end of the principal gradient than right hemisphere parcels, consistent with the role of this hemisphere in key heteromodal functions, such as semantic cognition and language. This pattern was observed in many canonical heteromodal networks derived from a whole-brain parcellation of resting-state data (Yeo et al.
2011), including control, default, dorsal and ventral attention networks; however, this pattern was inverted for Limbic-A, centred on the ventral anterior temporal lobe (ATL). There was also no gradient difference between the hemispheres in sensorimotor networks. In Study 1, individual differences in the relative gradient positions of networks across the hemispheres were found to have functional associations with two cognitive processes with opposing patterns of lateralisation, semantic cognition and visual reasoning (there were no effects for working memory). Participants whose Control-B network was closer to the heteromodal DMN end of the principal gradient in the left hemisphere compared with the right showed more efficient semantic retrieval; in contrast when the DAN-B network was closer to the heteromodal end of the principal gradient in the right hemisphere compared with the left, participants showed better visual reasoning on a progressive matrices task. Finally, in Study 2, we established that Control-B dissociates from DAN-B in the effect of verbal task demands on task activation in the left and right hemispheres. Control-B shows a left-lateralised response to semantic control demands but not working memory load, consistent with the view that lateralised control regions near the DMN apex of the principal gradient support controlled semantic retrieval.
To date, only one previous study has attempted to describe hemispheric differences in the principal gradient (Liang et al.
2021). Despite important differences in methodology, our findings align with Liang et al.’s study: both investigations found higher gradient values in the left than right hemisphere for ventromedial prefrontal cortex, IFG and lateral ATL. However, Liang et al. extracted separate gradients for the left and right hemispheres and considered these patterns within a 7-network parcellation; consequently, they could not identify the sub-network hemispheric differences that we observed, or directly compare left and right hemisphere networks within the same decomposition. The study by Liang et al. also did not assess the functional significance of hemispheric differences on the principal gradient, which was the main focus of the current study.
We found that left hemisphere parcels were, in general, closer than right hemisphere parcels to the DMN apex of the principal gradient, helping to explain why key heteromodal functions—such as language and semantic cognition—are left-lateralised. Margulies et al. (
2016) found that the terms
language:syntax and
language:semantics were among the BrainMap behaviour terms closest to the heteromodal end of the principal gradient; similarly,
verbal semantics was towards the heteromodal apex in Neurosynth (a meta-analytic tool; Yarkoni et al.
2011). Language and semantics both depend on the retrieval of heteromodal representations—extracted from diverse sensory-motor features when we acquire concepts and words; moreover, they both require retrieval to be controlled to fit rapidly changing goals and contexts. These different components of semantic cognition—conceptual representations plus control processes—are lateralised to different degrees (Gonzalez Alam et al.
2019). Semantic control processes are supported by a strongly left-lateralised network, encompassing left inferior frontal gyrus, and left posterior middle and inferior temporal cortex (Jackson
2021; Noonan et al.
2013). The resting-state functional connectivity between these semantic control sites is stronger in the left hemisphere compared with the right (Gonzalez Alam et al.
2019). In contrast, heteromodal conceptual representation is thought to be supported by bilateral ventral ATL (Ding et al.
2020; Lambon Ralph et al.
2017; Patterson et al.
2007). Evidence for bilateral conceptual representation in ventral ATL is provided by neuroimaging studies (Bright et al.
2004; Tranel et al.
2005; Vandenberghe et al.
1996; Visser et al.
2009,
2011) and neuropsychology; patients with bilateral ventrolateral ATL damage show severe semantic impairment (for example, in semantic dementia), while patients with unilateral lesions have milder deficits (Rice et al.
2018).
This difference between strongly lateralised semantic control processes and bilateral conceptual representations may help to explain why Limbic-A, centred on the ventral anterior temporal lobe, was situated closer to the DMN end of the principal gradient in the right hemisphere compared with the left. Gonzalez Alam et al. (
2019) found that right ATL was more connected to core DMN regions, including angular gyrus and dorsomedial prefrontal cortex; in contrast, left ATL was more connected to left-lateralised sites implicated in semantic control, including left intraparietal sulcus and left anterior insula bordering ventral parts of inferior frontal gyrus. In the left hemisphere, the principal gradient captures the order of networks from DMN, through the semantic control network, to executive regions (Wang et al.
2020). As a consequence, this proximity (and shared connectivity) of left ATL to semantic control regions might explain the unique gradient difference in Limbic-A. Right hemisphere components of this network might be closer to the heteromodal apex of the principle gradient because they are further from left-lateralised control networks situated towards the middle of the gradient.
The left-lateralised semantic control network is thought to be partially distinct from multiple demand cortex that responds to executive demands across domains: for example, effects of semantic but not non-semantic control demands are observed in anterior aspects of inferior frontal gyrus and posterior middle temporal gyrus (Davey et al.
2015,
2016; Hoffman et al.
2010; Jackson
2021; Noonan et al.
2013; Whitney et al.
2011,
2012). Similarly, the frontoparietal control network, defined through analyses of intrinsic functional connectivity, shows a bipartite organisation (Dixon et al.
2018), overlapping with Control-A and Control-B networks within the Yeo et al. (
2011) parcellation used in this study. Dixon et al.’s (
2018) control subnetwork including more anterior parts of both inferior prefrontal cortex and middle temporal gyrus has a topographical distribution that is similar to the functionally defined semantic control network (Jackson
2021; Noonan et al.
2013), and shows stronger interactions with DMN regions than the other control subnetwork. Similarly, the functionally defined semantic control network shows relatively strong intrinsic connectivity to both DMN, associated with heteromodal integration or abstraction, and domain-general executive and attention networks (Davey et al.
2016). This pattern of connectivity may allow states of controlled semantic cognition in which ongoing activation within DMN regions is shaped through the application of goal representations within the executive cortex to promote more weakly encoded aspects of knowledge (Wang et al.
2020). This finding is consistent with our observation of more efficient semantic cognition when the Control-B network was closer on the principal gradient to DMN in the left hemisphere as opposed to the right. Gradient differences between the two hemispheres might allow one control subnetwork to connect more strongly with DMN, supporting semantic control in the left hemisphere, while the other control subnetwork in the right hemisphere connects more strongly with sensory-motor regions, with advantages for demanding tasks that are oriented towards external sensory-motor features. This possibility is consistent with Wang et al. (
2014) who found that control network regions in the left hemisphere have stronger connectivity with DMN, while right hemisphere control sites are closer in connectivity to attentional networks.
Like the frontoparietal regions linked to cognitive control, DMN also has subnetworks; this study provides some evidence that these subdivisions within control and DMN networks are functionally related. Just as we found a control network that was closer to the heteromodal end of the principal gradient in the left hemisphere, DMN-B (the adjacent network), showed the same pattern. DMN-B includes regions such as lateral ATL, angular gyrus, inferior frontal gyrus and dorsomedial prefrontal cortex that are associated with semantic processing in the left hemisphere (Jackson
2021; Jefferies
2013; Lambon Ralph et al.
2017; Noonan et al.
2013; Rice et al.
2015b), and this DMN variant has repeatedly shown functional dissociations with core DMN regions such as posterior cingulate cortex and more ventromedial prefrontal regions (Chiou et al.
2020; Zhang et al.
2020), referred to here as DMN-A. DMN-B is associated with lateralised cognitive processes, like language and semantics, as well as social cognition (Andrews-Hanna et al.
2014). This network shows responses to externally generated, conceptual tasks, including those that interface with perception. In contrast, DMN-A or core DMN is thought to be more detached from perception, and is engaged by internally generated, self-referential and autobiographical memory processing (Chiou et al.
2020). It is interesting to note that it is DMN-B, not core DMN, that shows a lateralised position on the principal gradient. This is consistent with the possibility that lateralisation reflects the need to sustain and/or control heteromodal semantic retrieval (as opposed to the need to support internally-generated mental states, which are also associated with the heteromodal end of the principal gradient).
We found evidence of significant differences in lateralisation patterns within attentional networks as well, with both DAN and VAN falling closer to the heteromodal end of the gradient in the left hemisphere. Although attention has been traditionally conceptualised as a right-lateralised cognitive function, contemporary neuroscientific research paints a more nuanced picture with complex patterns of lateralisation across the traditionally accepted ventral and dorsal attention networks (Corbetta and Shulman
2002; Jeong and Xu
2016; Szczepanski et al.
2010; Thiebaut de Schotten et al.
2011a,
b). Critically, the DAN also plays a role in the flexible coupling of the control network across hemispheres and subdivisions (Dixon et al.
2018; Wang et al.
2014). Both DAN and control networks showed significant but opposing behavioural associations in our individual differences’ analysis of the position of networks on the principal gradient across hemispheres. Hemispheric differences in DAN-B were related to Raven’s matrices performance, but in contrast to semantic cognition, participants whose DAN-B was closer to the heteromodal end of the gradient in the right hemisphere were better at the task. Performance in reasoning tasks relies on efficient interregional communication within the bilateral multiple-demand system, and between this control network and other regions, for example, areas that maintain visuo-spatial representations, to orchestrate complex cognition (Gläscher et al.
2010; Shin and Jeon
2021). Moreover, previous research has linked performance on progressive matrices to attentional capacity (Schweizer and Moosbrugger
2004), and performance can also be decomposed into in two components relating to perceptual and executive attention (Ren et al.
2012), with the latter corresponding more closely to the DAN (Corbetta et al.
2008; Corbetta and Shulman
2002) and accounting for more variance in visual reasoning tasks (Ren et al.
2013). The right hemisphere is particularly activated during the performance of this task in certain conditions (Bishop et al.
2008; Prabhakaran et al.
1997). Contrasting specific types of reasoning tasks, like matrix and analogical reasoning, reveals greater right-lateralised responses in fronto-parietal regions for matrix reasoning (Hobeika et al.
2016). Right frontal regions in this network also show a greater response as matrix tasks increase in complexity (Krawczyk et al.
2010a,
b,
2011). Consequently, higher DAN gradient values in the right hemisphere might reflect closer integration of DAN and control networks in the right hemisphere, which facilitates the efficient deployment of attention to solve spatial relational reasoning problems.
We also compared semantic cognition with verbal WM tasks and found no lateralisation effects within these networks for the latter. This suggests that the lateralisation of semantic cognition is related to the controlled retrieval of conceptual representations as opposed to the need to use language in the task, and is compatible with accounts of WM demands engaging bilateral multiple demand cortex (Duncan,
2001,
2010; Fedorenko et al.
2013; Hugdahl et al.
2015). However, we do not provide a full description of lateralisation in verbal working memory, since Study 1 did not find any associations between digit span and hemispheric gradient differences for any networks; consequently networks relevant to working memory were not selected for analysis in Study 2.
There are several limitations of the current study. The tasks compared in Study 1 (semantic battery versus digit span and Raven’s progressive matrices) varied in multiple ways, including mode of response, task instructions, time allowed to respond and level of demand. In the analysis examining the association between individual differences in gradient lateralisation and semantic performance, we statistically controlled for performance on a visual matching task with similar input and response characteristics to the semantic task, helping to ensure that semantic task effects reflected the requirement for conceptual retrieval. The fact that we observed an association with hemispheric differences in gradient position in DAN-B but not in Control-B with this perceptual task, while the association of hemispheric differences in Control-B with semantic performance remained significant, makes it unlikely that differences between the semantic and non-semantic tasks gave rise to our results. However, there were no control tasks for digit span or visual reasoning; moreover, hemispheric differences in the principal gradient might be related to other non-semantic cognitive domains not assessed here. Furthermore, the parametric manipulations of difficulty in WM and semantic judgements in Study 2 are not analogous since we manipulated working memory load (i.e., items to be maintained) and the strength of the semantic association (i.e., semantic distance as measured by word2vec). As noted by Gao et al. (
2021), the WM task was associated with faster responses than the semantic control task, perhaps because word reading takes longer than letter identification, but reading times are not necessarily relevant to the activation of control networks. Another difference among these tasks was that strength of association had a larger effect on RT than working memory load, although RT does not provide a direct measure of cognitive control demands. We selected these manipulations because the literature shows that they robustly vary the activation of control regions (Noonan et al.
2013; Jackson
2021; Fedorenko et al.
2013; Emch et al.
2019); however, it would also be possible to manipulate control demands in a more comparable way across these domains, for example, by varying the strength of the distractors in both tasks.
Our methods also did not allow us to investigate the source of the network asymmetries at the sub-network or parcel level, since the parcellation we used (Schaefer et al.
2018) did not provide homotopic regions that could be compared (see Popovych et al.
2021, for the effect of parcellation choice on resting-state results). The parcels we used from Schaefer et al. (
2018) are derived separately for each hemisphere, and there are an unequal number assigned to each network across hemispheres. For example, the DAN-B network has 13 parcels in the left hemisphere organised in three subdivisions (postcentral, frontal eye fields and precentral ventral region), while it has only 11 in the right hemisphere (lacking the precentral ventral region and sporting only 8 postcentral parcels, opposed to 9 in the left hemisphere). These differences might give rise to local resting-state functional connectivity gradients that are present in one hemisphere but not the other (Gordon et al.
2016), and these differences in local organisation could have functional significance. Future research could therefore seek to verify these patterns using symmetrised parcellations (Glasser et al.
2016; Joliot et al.
2015), or use methods that exploit voxel-level timeseries homotopy (Gotts et al.
2013; Jo et al.
2012).
Another limitation is shared by many studies that employ dimensionality reduction methods: the number of dimensions retained for analysis is somewhat arbitrary (Supplementary Figure S6 shows that there is no clear plateauing of the eigenvalues). Here we focussed on the asymmetry of the principal gradient, as it captures the most variance and is known to be important for cognition as well as the organisation of large-scale networks on the cortical surface (Murphy et al.
2018,
2019; Turnbull et al.
2020; Wang et al.
2020). We also provide a supplemental analysis of gradient asymmetries in Gradient 2, but we opted not to extend the analysis to gradients explaining less variance as their interpretability is expected to be lower; moreover, studies have shown functional associations with Gradients 1 and 2, but not with Gradient 3 and beyond (Hong et al.
2019; Murphy et al.
2018,
2019; Turnbull et al.
2020; Wang et al.
2020). Future research could take a different approach by extracting a very large number of gradients, and then identify lower-order gradients that specifically capture hemispheric differences in higher-order gradients (see Valk et al.
2020).
Finally, it remains unclear why attentional networks (DAN-A; DAN-B and VAN-A) were closer to the heteromodal end of the principal gradient in the left hemisphere, even when the opposite pattern for DAN-B (closer proximity to heteromodal cortex in the right hemisphere) was associated with better visual attention. One possibility is that these attention networks can also support controlled semantic cognition, to varying degrees across people, and that these patterns of left-lateralised and right-lateralised connectivity are in competition. Future research could test whether the position of networks along the principal gradient relates to their capacity for efficient interaction, and whether there are differences in physical distance along the cortical surface in the two hemispheres that reflect the connectivity gradient differences we described.