Part of the brain called the lateral prefrontal cortex has a critical role in many of the processes seen as hallmarks of human cognition, such as reasoning, planning and problem-solving. Individuals with damage to the lateral prefrontal cortex are disorganized and easily distracted, and may show behaviors that are inappropriate for their context. However, the involvement of the lateral prefrontal cortex in such a wide range of processes makes it difficult to study. This in turn presents a significant roadblock to a full understanding of cognition and human intelligence.

Of particular interest is whether the lateral prefrontal cortex has a hierarchical organization wherein a region coordinates the roles of other regions, much like the chief executive of a company. Therefore, Nee and D’Esposito set out to map how the lateral prefrontal cortex is organized, and how its different regions communicate with each other to support complex cognition.

Brain imaging revealed that the rear (posterior) part of the lateral prefrontal cortex processes an individual’s current situation, while the front (anterior) prepares for future situations. Areas in the middle process both types of information. These central areas appear to be highly influential as they have stronger connections to the anterior and posterior regions than vice versa.

In cognitively demanding situations, the middle areas receive input from both anterior regions (potentially about future needs) and posterior regions (potentially about current needs). By combining the two sets of information, the middle areas can select behaviors that take into account both present circumstances and longer-term goals. With this strategic overview, the middle areas of the lateral prefrontal cortex are well positioned to play the part of the brain’s chief executive.

Future experiments should test whether the interactions observed between the different regions of the lateral prefrontal cortex are essential for complex planning and thinking. Additional work in animals would improve our understanding of the mechanisms underlying these interactions. Finally, studying how these interactions are altered in disorders such as schizophrenia, where the lateral prefrontal cortex shows abnormal activity, might pave the way for more effective treatments.

The prefrontal cortex (PFC) is central to higher-level cognition. Damage to the PFC produces severe deficits in goal-directed cognition (Luria, 1966; Damasio and Anderson, 1993; Lhermitte, 1986) and neuroimaging studies have demonstrated the involvement of the PFC across a wide range of tasks (Duncan and Owen, 2000; Niendam et al., 2012). Despite its ubiquitous influence, mechanistic insight into PFC function has remained difficult.

Progress towards understanding PFC function has been made through the development of organizing principles that describe gradients of function across the full extent of PFC. A prominent organizing principle is that the rostral/caudal axis of the lateral PFC (LPFC) is structured according to the level of cognitive control (Fuster, 1997; Koechlin et al., 2003; Koechlin and Summerfield, 2007; Badre, 2008; Badre and D'Esposito, 2007; 2009). Mounting evidence from neuroimaging and lesion studies suggests that progressively rostral areas perform progressively higher levels of cognitive control through processing increasingly abstract representations (Koechlin et al., 2003; Badre, 2008; Badre and D'Esposito, 2007; 2009; Koechlin and Jubault, 2006; Nee and Brown, 2012; 2013; Koechlin et al., 1999; Bahlmann et al., 2014; Azuar et al., 2014). This organization enables the LPFC to control behavior across multiple timescales including adapting actions according to the prevailing context, and aligning behaviors with longer-term goals and plans (Fuster, 2001; 2008).

While differentiating the functional sensitivities of sub-regions of the LPFC is an important first step, mechanistic insight into how the LPFC supports goal-directed cognition requires understanding intercommunication among LPFC sub-regions. To this end, it has been proposed that the LPFC is organized hierarchically such that more rostral areas exert asymmetrically greater influence upon more caudal areas (Badre and D'Esposito, 2009). Hierarchical processing confers substantial computational efficiency and flexibility (Botvinick and Weinstein, 2014; Duncan, 2013; Solway et al., 2014), and such processing in the LPFC could explain the massive propensity for higher-level cognition found in primates. However, data demonstrating functional hierarchical processing in the LPFC have been limited (Koechlin et al., 2003) and controversial (Reynolds et al., 2012; Crittenden and Duncan, 2014). Furthermore, no data have linked hierarchical processing to its surmised cognitive advantages. Finally, although it has been widely assumed that the rostral-most areas of LPFC are situated at the apex of the hierarchy (Koechlin and Summerfield, 2007; Badre and D'Esposito, 2009), recent data call this assumption into question (Goulas et al., 2014). Hence, understanding whether and how hierarchical processing occurs in the LPFC remains a significant barrier to modeling goal-directed cognition.

Here, we examine how interactions within the LPFC support higher-level cognition. To do so, we devised a paradigm to engage the entire LPFC by orthogonally manipulating demands on stimulus domain and cognitive control (Figure 1). We began by confirming that the rostral/caudal axis of the LPFC operates according to abstraction principles by examining gradients of stimulus domain-generality/specificity, and temporal control over action. Next, we modeled interactions among LPFC sub-regions to examine whether and how cognitive control demands elicit asymmetrical influences indicative of hierarchical processing. Finally, we aimed to establish whether individual differences in hierarchical interactions predict trait-measured higher-level cognitive functioning, consistent with the presumed computational advantage of hierarchical processing.

Figure 1 with 1 supplement see all

Download asset Open asset

To engage the full extent of LPFC necessary to test our predictions, we adapted a task (Koechlin et al., 1999; Charron and Koechlin, 2010) that orthogonally manipulated demands on stimulus domain (i.e. verbal vs. spatial) and cognitive control (Figure 1). Participants began each block of trials by performing a basic sequencing task which required the determination of whether the current stimulus followed the previous stimulus in a trace of the points of a star (spatial task) or in the word 'tablet' (verbal task) with the appropriate stimulus feature cued by color. Hence, a critical aspect of the basic task was Feature Control – selecting the appropriate stimulus feature to integrate into ongoing cognition. These demands recruit caudal dorsal and ventral LPFC for spatial and verbal selection, respectively (Nee et al., 2013). Since there was no preceding stimulus at the start of each block, participants decided whether the first stimulus of a given block was the start of the sequence (i.e. top of the star for the spatial task, 't' for the verbal task).

After 2 to 5 trials, participants were cued to either perform one of three different sub-tasks (Restart, Delay, or Dual), or to continue with the basic task (Control). Tasks were cued by the shapes of frames with the basic/Control task always cued by square frames. After a 3 to 5 trial sub-task phase, participants were cued to return to the basic task, which we refer to as a return trial. Participants began the Restart task by starting the sequence anew, deciding whether the stimulus was the start of the respective sequence, and then performing the basic task for the remainder of the sub-task phase. When cued to return the basic task, participants once again started the sequence anew. Thus, the Restart task required Contextual Control – selecting the appropriate action rules (i.e. sequence-match vs sequence-start) for the given circumstance. This process is involved, for example, in preparing to switch tasks (Monsell, 2003), comprising readying a new rule set as distinct from a previous one. Previous work has demonstrated that this type of control process engages the mid LPFC (MacDonald et al., 2000; Sohn et al., 2000).

In the Delay task, participants kept in mind their place in the sequence across a distractor-filled delay. During the delay, participants rejected all stimuli with a keypress corresponding to a 'no' response. When cued to return to the basic task, participants evaluated the current stimulus with respect to the delay-maintained memorandum. As a result, the Delay task placed minimal demands on current stimulus processing, while requiring participants to prepare to integrate past information into future processing thereby requiring Temporal Control. This is an important component of working memory, supporting the ability to maintain information in the face of ongoing processing in order to inform future cognition (D'Esposito and Postle, 2015; Baddeley, 2003). Previous work has demonstrated that this type of control process engages rostral LPFC (De Pisapia et al., 2012; De Pisapia et al., 2007; Nee et al., 2014).

Finally, the Dual task contained the essential processes of both the Restart and Delay tasks. Here, participants started the sequence anew, while simultaneously remembering their place in the previous sequence. When cued to return to the basic task, participants disengaged from the current sequence and evaluated the current stimulus with respect to the delay-maintained memorandum. Hence, this task required a combination of Contextual Control and Temporal Control to select the appropriate action rules and temporal origin of information (i.e. sequence-match current sequence vs sequence-match previous sequence).

Collectively, the task was a 2 x 2 x 2 design with factors of Stimulus Domain (verbal, spatial), Contextual Control (low – Control, Delay; high – Restart, Dual), and Temporal Control (low – Control, Restart; high – Delay, Dual). Since the Delay task required negligible demands on Feature Control due to the minimization of current processing, it was expected that the statistical interaction between the cognitive control main effects would reveal demands on Feature Control. Collectively, these manipulations were designed to engage the entire LPFC, providing an opportunity to examine the hypothesized hierarchical interactions among LPFC areas that support cognitive control. We focus our analyses on the sub-task trials, unless otherwise specified.

Behavioral results

To examine behavioral manifestation of cognitive control demands, separate 2 x 2 x 2 ANOVAs were computed on error-rate (ER), and reaction times (RT) on correct trials (Figure 2). These analyses revealed significant effects of cognitive control demands with main effects of Temporal Control (ER: F(1,23) = 5.72, p<0.05; RT: F(1,23) = 28.72, p<0.0001), Contextual Control (ER: F(1,23) = 16.21, p=0.0005; RT: F(1,23) = 147.79, p<10^–11), and their interaction (ER: F(1,23) = 23.78, p=0.0001; RT: F(1,23) = 87.16, p<10^–8). As anticipated, the interaction was largely driven by fast and accurate performance on the Delay condition, which placed minimal demands on current processing during the sub-task phase. By contrast, there were no main effects of Stimulus Domain or interactions of Stimulus Domain with cognitive control demands (ER and RT all p>0.05).

Figure 2

Download asset Open asset

Following each sub-task, there was a return trial that required either the integration of an item held across the sub-task (Delay and Dual conditions) or a switch to a new sequence (Restart and Dual conditions). These were analyzed with separate 2 x 2 x 2 ANOVAs on ER and RTs on correct trials (Figure 2). These analyses revealed a significant effects of cognitive control demands with main effects of Temporal Control in ER, but not RT (ER: F(1,23) = 11.44, p<0.005; RT: F(1,23) = 1.89, p>0.15), Contextual Control (ER: F(1,23) = 20.00, p<0.0005; RT: F(1,23) = 87.05, p<10^–8), and their interaction (ER: F(1,23) = 19.18, p<0.0005; RT: F(1,23) = 10.24, p<0.005). Again, no effects of Stimulus Domain were observed (ER and RT all p>0.05). These analyses confirm the effective manipulation of cognitive control demands, but equivalent loadings of control demands across stimulus domain.

Evaluating the rostral/caudal axis and cognitive control

To verify the expected rostral/caudal gradient by cognitive control in the LPFC, fMRI data from the epoch of the sub-tasks (see Materials and methods for full modeling details) were analyzed with a 2 x 2 x 2 ANOVA. Here, we focus on the effects of cognitive control (Figure 3), while the effects of Stimulus Domain are depicted in Figure 3—figure supplement 1. The main effect of Temporal Control produced activations across the rostral-most aspects of the LPFC, mostly within the lateral frontopolar cortex (FPl; putative area 10) and mid-middle frontal gyrus (MFG; putative area 9/46), also spanning orbitofrontal cortex and pars orbitalis ventrally, through rostral-MFG (rMFG; putative area 46) laterally, up to the superior frontal gyrus dorsally. Previous studies have suggested that rostral LPFC is engaged as difficulty is increased regardless of functional demands (Crittenden and Duncan, 2014; Fedorenko et al., 2013). Notably, the main effect of Temporal Control in the present design includes both the easiest condition (Delay) and hardest condition (Dual). The rostral LPFC was engaged considering each condition in isolation (Figure 3—figure supplement 2), ruling out the possibility that activations in this area are attributable to difficulty. The main effect of Contextual Control was most pronounced in mid-lateral areas including the inferior frontal sulcus (IFS; putative area 45 or 9/46v) extending into pars triangularis ventrally, and the caudal middle frontal gyrus (cMFG; putative area 8Ar or 9/46d) dorsally. Significant effects extended both rostrally into the FPl, and caudally into the inferior frontal junction (IFJ; putative area 44 or 8Av) and pars opercularis ventrally, and the caudal superior frontal sulcus (SFS; putative area 8Ad) dorsally. Finally, the statistical interaction of the effects of cognitive control produced the strongest foci in caudal LPFC areas including the IFJ and pars opercularis ventrally, and the caudal SFS dorsally, spreading into the mid-lateral PFC. As indicated in Figure 3—figure supplement 3, this interaction was primarily due to reduced activation during the Delay task, which had minimal demands on current processing. Thus, the effects are consistent with Feature Control. Collectively, these results demonstrate a rostral/caudal gradient as a function of cognitive control processes with an orderly progression of Temporal Control, Contextual Control, and Feature Control running from rostral to caudal spanning the entire LPFC (Figure 3; Figure 3—figure supplement 4). These effects were replicable when the dataset was split into two, confirming their robustness (Figure 3—figure supplement 5).

Figure 3 with 5 supplements see all

Download asset Open asset

Figure 3—source data 1

Statistical parametric maps of the univariate whole-brain results.

Maps contain t-statistics and have been thresholded to correct for multiple comparisons as detailed in the Materials and methods.

https://doi.org/10.7554/eLife.12112.007

Download elife-12112-fig3-data1-v1.zip

Abstraction: stimulus domain-sensitivity and generality

The previous analyses demonstrated the predicted rostral/caudal functional gradient as a function of cognitive control processes. Next, we sought to verify that this gradient reflected increasingly abstract processing in increasingly rostral areas. If processing in the LPFC becomes more abstract in progressively rostral areas, caudal areas should process concrete features such as stimulus-domain, whereas more rostral areas should instantiate control processes that act across domains (e.g. a plan that can be applied to spatial or verbal stimuli). Thus, we reasoned that abstract versus concrete processing would reveal itself through stimulus domain-generality versus specificity.

To test this hypothesis, we used activation peaks of the LPFC ROIs defined by the cognitive control contrasts above. Notably, each contrast produced a distinct dorsal and ventral activation peak spanning rostral areas (FPl, MFG), mid areas (IFS, cMFG), and caudal areas (IFJ, SFS). Hence, this analysis enabled the examination of whether stimulus domain-sensitivity is present throughout the rostral/caudal axis, or whether domain-generality emerges at a particular rostra/caudal level. First, we performed a 2 (dorsal, ventral) x 3 (rostral, mid, caudal) ANOVA on the main effect of Stimulus Domain. This analysis revealed a significant effect of the dorsal/ventral axis (F(1,23) = 100.21, p<10^–9), a significant effect of the rostral/caudal axis (F(2,46) = 10.63, p<0.0005), and critically, an interaction between the dorsal/ventral and rostral/caudal axes (F(2,46) = 56.12, p<10^–12). Follow-up t-tests revealed that caudal and mid ventral areas demonstrated verbal-sensitivity (SFS: t(23) = 5.93, p_corrected<0.0001; cMFG: t(23) = 5.08, p_corrected<0.0005), but not rostral areas (MFG: t(23) = 1.54, p_corrected>0.8). Symmetrically, caudal and mid dorsal areas demonstrated spatial-sensitivity (IFJ: t(23) = 7.83, p_corrected<10^–6; IFS: t(23) = 5.33, p_corrected<0.0005), but not rostral areas (FPl: t(23) = −2.25, p_corrected >0.2). A similar lack of stimulus domain-sensitivity was observed in a third rostral peak (rMFG: t(23) = 0.83, p>0.4). This region is not considered further due to lack of resting connectivity with the other LPFC regions (see Materials and methods; Figure 6—figure supplement 4). Collectively, these data indicate caudal and mid LPFC are sensitive to stimulus domain, whereas rostral LPFC areas are domain-general, consistent with a gradient of abstraction (Figure 4). These effects were replicable when the dataset was split into two, confirming their robustness (Figure 4—figure supplement 1). Given the sensitivity of caudal areas to stimulus domain, and the close proximity of the peaks of effects of Stimulus Domain and Feature Control (see Materials and methods; Figure 3—figure supplement 1), we use the contrast of Stimulus Domain to define caudal areas in future analyses since they enable the most straight-forward separation of spatial and verbal aspects of Feature Control.

Figure 4 with 1 supplement see all

Download asset Open asset

Abstraction: temporal control over action

Abstraction along the rostral/caudal axis of the LPFC is hypothesized to relate to the temporal control over action (Fuster, 1997; Koechlin et al., 2003; Koechlin and Summerfield, 2007). By these ideas, concrete processing in caudal LPFC provides control over current processing (e.g. translating stimuli to actions), while abstract processing in rostral LPFC organizes behavior according to future considerations (e.g. goals or plans). To test this proposal, we examined the degree to which different rostral/caudal levels of the LPFC were related to current processing demands and future processing demands, as indexed by RT. Current demands were operationalized as RT during sub-task trials (i.e. RTs corresponding directly to the epochs at which fMRI signal was measured). Future demands were operationalized as RT during return trials (i.e. RTs corresponding to demands the trial after the fMRI signal was measured). Regressing these measures onto activation revealed a clear gradient such that caudal LPFC was related to current, but not future processing, while rostral LPFC was related to future, but not current processing, with mid LPFC showing activity related to both current and future processing (Figure 5A). To quantify these relationships, we examined partial correlations with the ROIs described above. This revealed significant relationships between caudal LPFC and current (r = 0.69, p_corrected<10^–22), but not future processing (r = 0.04, p>0.5). Conversely, rostral LPFC was related future (r = 0.31, p_corrected<0.0005), but not current processing (r = 0.06, p>0.45). Finally, activity in mid LPFC was related to both current (r = 0.62, p_corrected<10^–17) and future processing (r = 0.31, p_corrected<0.0005; Figure 5B).

Figure 5 with 1 supplement see all

Download asset Open asset

Figure 5—source data 1

Statistical parametric maps of the activation-behavior correlations.

Maps contain r-values and have been thresholded to correct for multiple comparisons as detailed in the Materials and methods.

https://doi.org/10.7554/eLife.12112.016

Download elife-12112-fig5-data1-v1.zip

To bolster these findings, we regressed current and future RT on activations at each LPFC level in each subject. Then inferences on the resultant slope parameters were performed at the group-level (i.e. summary-statistic approach). These parameters were submitted to a 3 x 2 ANOVA with factors of LPFC level (caudal, mid, rostral) and time (current, future). This analysis revealed a significant LPFC level x time interaction (F(2,46) = 30.85, p<10^–8; Figure 5C). T-tests on the individual parameter estimates corroborated the same pattern (caudal-current: t(23) = 8.11, p_corrected<10^–6; caudal-future: t(23) = 0.80, p>0.4; mid-current: t(23) = 7.21, p_corrected<10^–5; mid-future: t(23) = 3.14, p_corrected<0.05; rostral-current: t(23) = −0.61, p>0.5; rostral-future: t(23) = 3.60, p_corrected<0.01). Collectively, these data indicate a temporal abstraction gradient such that caudal LPFC areas reflect current processing demands, while rostral LPFC areas reflect future processing demands. Mid LPFC areas were related to both current and future demands, suggesting a potential zone for integrating rostral and caudal influences. These effects were replicable when the dataset was split into two, confirming their robustness (Figure 5—figure supplement 1).

Evaluating hierarchical interactions

The above analyses demonstrate that the rostral/caudal LPFC axis is sensitive to abstraction with caudal areas showing concrete properties including stimulus domain-sensitivity and present-oriented processing, and rostral areas showing abstract properties including stimulus domain-generality and future-oriented processing. However, the question remains regarding how these areas interact to support goal-directed cognition and whether such interactions are hierarchical.

Hierarchical interactions are presumed to be reflected by asymmetries in the influence of one region upon another (Badre and D'Esposito, 2009). Regions that are high in the hierarchy are thought to influence lower regions more than vice versa. Since directed influences cannot be inferred on the basis of correlations alone, we performed dynamic causal modeling (DCM) which describes directed influences between brain regions (effective connectivity) by estimating parameters in a series of differential equations (Friston et al., 2003, see Stephan et al., 2010 for an accessible review). These differential equations include a forward model linking the hemodynamic response to underlying neural activity in order to describe how activity changes in a region as a function of inputs to that region. The ability of DCM to uncover underlying directed neuronal interactions has been validated using large-scale biophysical models (Lee et al., 2006) and invasive recordings in rodents (David et al., 2008). However, since computational complexity precludes modeling all sources of inputs that drive changes in a particular region, DCM can only be applied to simplified neural circuits (i.e. a small number of regions). Such simplifications also require including experimental manipulations such as stimulus delivery and experimental condition as direct drivers and modulators of activity. Despite these simplifications, DCM has been demonstrated to accurately identify the appropriate connections and modulations, even when the modeled regions are only a subset of the regions contributing to a process (Lee et al., 2006). Finally, DCM leverages model comparison to determine the presence and strength of directed regional couplings. This requires specifying multiple (plausible) models and comparing model evidence to reveal which model provides the best account of the data. Hence, DCM can only make relative claims with respect to the explored model space.

To examine interactions among LPFC sub-regions, we modeled the dynamics of activity in the six LPFC ROIs as a function of three sets of parameters: 1) inputs into the system from stimuli received by caudal areas; 2) connectivity among regions along which activity is propagated (fixed connectivity); and 3) changes in connectivity induced by demands of stimulus domain and cognitive control (connectivity modulations). An example is schematized in Figure 6—figure supplement 1. We then searched for asymmetries in the fixed connectivity and their modulations for evidence of hierarchical control.

Group-level Bayesian model comparison adjudicated between models of effective connectivity within the LPFC (Stephan et al., 2009a) (see Methods for full details of the model procedure). Significant parameter estimates resulting from the best model are depicted in Figure 6. We began by examining asymmetries in fixed connectivity (Figure 6A). Fixed connectivity is constrained by structural connectivity (Stephan et al., 2009b), and thus provides a window into the relative strength of efferent and afferent connections. To evaluate whether fixed connections demonstrate hierarchical dependencies, we calculated the relative strength of efferent versus afferent connections with a simple contrast (average efferent strength – average afferent strength) based on estimated fixed connectivity of the model. Higher hierarchical areas would be expected to show a positive value on this metric. Similar explorations in the monkey have provided evidence for hierarchical structure, but with controversy surrounding whether the rostral LPFC is the apex of the hierarchy (Badre and D'Esposito, 2009) or the mid LPFC (Goulas et al., 2014). In our dataset, hierarchical strength increased monotonically from caudal areas to mid areas to MFG, but FPl demonstrated the lowest value on this metric resulting in an inverted-U pattern (Figure 7). These effects were replicable when the dataset was split into two, confirming their robustness (Figure 7—figure supplement 1). This pattern was not due to the average connectivity strength or total connectivity strength, which showed markedly different patterns (Figure 7—figure supplement 2), indicating that hierarchical strength is independent of the general magnitude of connectivity.

Figure 6 with 5 supplements see all

Download asset Open asset

Figure 7 with 2 supplements see all

Download asset Open asset

To better quantify the hierarchical organization of the LPFC, we fit a quadratic (e.g. inverted-U) function to each participants’ hierarchical strength metrics. Here, the height of the vertex of the quadratic function estimates the hierarchical strength of the apex of the hierarchy, assuming it is positive. The position of the vertex provides an estimate of the rostral/caudal location of apex of the hierarchy. Across participants, the height of the vertex was significantly positive (t(23) = 2.34, p<0.05), indicating that the vertex tended to represent the apex of the hierarchy. Furthermore, the vertex tended to be localized to the mid LPFC (mean y-coordinate: 24.7, standard error: 6.1) and differed significantly from both the rostral-most (i.e. FPl; t(23) = -4.35, p<0.0005), and caudal-most ROIs (i.e. SFS; t(23) = 4.18, p<0.0005). These results are inconsistent with a strict rostral-to-caudal hierarchy (Badre and D'Esposito, 2009) and instead suggest that the apex of the LPFC hierarchy is in mid areas (Goulas et al., 2014).

Next, we searched for asymmetries in the modulations of effective connectivity by cognitive demands. Modulatory dynamics (Figure 6B) had the following properties: first, Stimulus Domain demands increased connectivity from caudal to mid LPFC areas (hereafter, caudal-to-rostral modulations are referred to as 'bottom-up'). That is, when processing spatial stimuli, connectivity was increased from SFS to cMFG (t(23) = 3.70, q<0.005), while processing verbal stimuli increased connectivity from IFJ to IFS (t(23) = 3.89, q<0.001). Second, cognitive control demands modulated connectivity from rostral to mid LPFC areas (hereafter, rostral-to-caudal modulations are referred to as 'top-down'). Temporal Control increased connectivity from FPl to MFG (t(23) = 2.40, q<0.05). Contextual Control decreased connectivity from FPl to MFG (t(23) = -5.45, q<0.0001), IFS (t23) = -4.76, q<0.0005), and cMFG (t(23) = -4.15, q<0.001), while increasing connectivity from MFG to IFS (t(23) = 8.18, q<10^–6) and cMFG (t(23) = 7.70, q<10^–6). Collectively, these data indicate that demands on stimulus domain influence bottom-up processing, while demands on cognitive control influence top-down processing. These results suggest that demands on cognitive control result in a convergence of influences at the mid LPFC where top-down and bottom-up information meet. These effects were replicable when the dataset was split into two, confirming their robustness (Figure 6—figure supplement 2).

Relationship between LPFC dynamics and cognitive ability

Hierarchical processing is important due to high degree of flexibility and computational efficiency it confers (Botvinick and Weinstein, 2014; Duncan, 2013; Solway et al., 2014). Therefore, we predicted that the modeled LPFC hierarchical interactions should be related to cognitive abilities. To explore this possibility, we examined the relationship between the modeled estimates of effective connectivity and trait-measured higher-level cognitive ability.

First, we observed that individual differences in the magnitude of top-down modulations by cognitive control (as measured by estimates of modulations of effective connectivity; Figure 6B) tended to be inter-related (Figure 8—figure supplement 1A). This suggests that top-down control is an individual difference metric. Using principle components analysis, we distilled the cognitive control-induced changes in top-down effective connectivity into a single factor (Figure 8—figure supplement 2). Similarly, we combined the stimulus domain-induced changes in bottom-up effective connectivity into a single factor. We found that these measures were inversely related, such that individuals that tended to have strong top-down influences tended to have weaker bottom-up influences and vice versa (r = −0.52, p<0.01; robust regression t(22) = −2.79, p<0.05; Figure 8A). This was not due to the model fitting procedure as the model covariance matrix indicated a negligible relationship between these factors (Figure 8—figure supplement 1B).

Figure 8 with 3 supplements see all

Download asset Open asset

Next, we examined several independent measures of higher-level cognitive ability which included tests of working memory and fluid intelligence. Performance on these measures tended to be inter-correlated (Figure 8—figure supplement 1C), so we combined them into a single factor using principle components analysis. We examined the relationship between model-derived metrics of top-down and bottom-up processing, and trait-measured higher-level cognitive ability. While there was no relationship between bottom-up processing and higher-level cognitive ability (r = −0.01, p>0.95; robust regression t(22) = 0.09, p>0.9; Figure 8C), there was a significant positive relationship between top-down processing and higher-level cognitive ability (r = 0.47, p<0.05; robust regression t(22) = 2.25, p<0.05; Figure 8B). This relationship remained after removing 2 participants with higher-level cognitive abilities greater than 2 standard deviations below the mean (r = 0.43, p<0.05; robust regression t(20) = 2.21, p<0.05). Hence, those individuals with greater top-down processing demonstrate greater higher-level cognitive ability.

Subsequently, we explored the relationship between the hierarchical strength of the modeled fixed connectivity and trait-measured higher-level cognitive ability. In particular, we were interested in whether the degree to which mid LPFC demonstrated hierarchically organized fixed connectivity predicted higher-level cognitive ability. We related the height of the vertex fit described above – summarizing the hierarchical strength of mid LPFC – to trait-measured higher-level cognitive ability. This relationship was significantly positive (r = 0.61, p<0.005; robust regression t(22) = 4.68, p=0.0001; Figure 8D) indicating that those individuals with stronger mid LPFC hierarchical fixed connectivity demonstrate greater higher-level cognitive ability. This relationship largely remained after removing 2 participants with higher-level cognitive abilities greater than 2 standard deviations below the mean (r = 0.39, p=0.07; robust regression t(20) = 3.00, p<0.01). Furthermore, this relationship remained after controlling for contributions of average connectivity of mid LPFC (cMFG, IFS, and MFG), and global connectivity of the LPFC (ρ = 0.47, p<0.05; robust regression t(20) = 2.66, p<0.05), indicating that hierarchical connectivity makes important contributions to higher-level cognitive ability over-and-above simple connectivity magnitude. This suggests that the hierarchical strength of the mid LPFC is beneficial to higher-level cognition.

The height of the vertex summarizes the hierarchical strength of mid LPFC. However, it is obtained by fitting the hierarchical strengths of all areas, thereby incorporating some information from other regions, as well. One might be interested in the degree to which the relationship between higher-level cognitive ability and hierarchical strength is selective to mid LPFC. This is difficult to disentangle due to the fact that hierarchical strengths of different regions are interdependent. Nevertheless, a simple average of the mid LPFC hierarchical strength (including cMFG, IFS, and MFG) produced similar results to the height of the vertex method described above (r = 0.48, p<0.05; robust regression t(22) = 2.38, p<0.05). Opposite results were obtained by relating higher-level cognitive ability to the hierarchical strength of caudal LPFC (r = −0.43, p<0.05; robust regression t(22) = −2.00, p=0.057) and FPl (r = −0.42, p<0.05; robust regression t(22) = −2.41, p<0.05), which is to be expected given the strong interdependencies between the hierarchical strength of mid LPFC to both caudal and rostral LPFC (mid to caudal, r = −0.84, p<10^–6; mid to FPl, r = −0.87, p<10^–7). Partial correlations were unable to convincingly determine whether the relationships between higher-level cognitive ability and hierarchical strength were primarily due to the hierarchical strength of mid LPFC or other areas (ρ_mid·caudal = 0.27, p>0.2; ρ_mid·FPl = 0.29, p>0.15; ρ_caudal·mid = −0.03, p>0.85; ρ_FPl·mid = 0.03, p>0.85). However, that the relationships between caudal/rostral LPFC and higher-level cognitive ability were negligible after partialing out the contributions of mid LPFC and that the reverse was not true is suggestive that the main driver is the hierarchical strength of the mid LPFC.

Finally, we assessed the unique contribution of each LPFC dynamic factor described above to higher-level cognitive ability. The strength of top-down modulations by cognitive control, bottom-up modulations by stimulus domain, and hierarchical fixed connectivity of the mid LPFC were simultaneously regressed on trait-measured higher-level cognitive ability. This analysis revealed significant effects of top-down modulations (ordinary least-squares regression t(20) = 2.45, p<0.05; robust regression t(20) = 2.21, p<0.05), and mid LPFC hierarchical fixed connectivity (ordinary least-squares regression t(20) = 3.13, p<0.01; robust regression t(20) = 3.58, p<0.005), but no effect of bottom-up modulations (ordinary least-squares regression t(20) = 1.81, p>0.05; robust regression t(20) = 1.48, p>0.15). These results indicate that both the hierarchical orientation of the mid LPFC at baseline, and the top-down modulation from rostral LPFC to mid LPFC during cognitive control make separable and positive contributions to higher-level cognitive ability.

We examined interactions within the LPFC by varying demands on stimulus domain and cognitive control. Collectively, these demands engaged most of LPFC providing the opportunity for a comprehensive account of its function. We found that progressively rostral areas processed progressively abstract levels of cognitive control with demands on Feature Control, Contextual Control, and Temporal Control eliciting increasingly rostral activations. Moreover, stimulus domain-sensitivity was present in caudal and mid LPFC, but not in rostral LPFC consistent with abstract processing in rostral areas. Correlations between neural activation and behavioral performance revealed that caudal LPFC varied linearly with demands on current processing, while rostral LPFC varied linearly with demands on future processing. Both relationships were present in the mid LPFC, indicating a potential zone where current and future processing converges. Dynamic causal modeling revealed that hierarchical asymmetries in connectivity were strongest in mid LPFC. Furthermore, cognitive demands elicited both top-down influences from rostral to mid LPFC and bottom-up influences from caudal to mid LPFC. Thus, analyses of effective connectivity and its modulations suggest that mid LPFC forms a nexus where information converges to influence action. Finally, both the hierarchical orientation of the mid LPFC in fixed connectivity, and the top-down influences from rostral LPFC to mid LPFC during the engagement of cognitive control were positively related to higher-level cognitive ability. Hence, the functional organization of the LPFC is intimately tied to cognitive capacity.

A matter of debate is whether or not processing in the LPFC is hierarchical. From a computational perspective, hierarchical processing is advantageous (Botvinick and Weinstein, 2014; Solway et al., 2014), although relating such processing advantages to the LPFC has been challenging. Badre and D’Esposito (Badre and D'Esposito, 2009) suggested that the hallmark of hierarchies is asymmetrical influence. That is, those components that are higher in the hierarchy influence lower components more than vice versa. Based on patterns of efferent and afferent projections, they proposed a rostral-to-caudal hierarchy such that rostral areas are higher than caudal areas. Recently, however, the evidence for a rostral-to-caudal hierarchy based on asymmetry of anatomical projections has been called into question. Based on a meta-analysis of projections among areas of the LPFC in monkeys, hierarchy was determined by the asymmetry with which a given area sent projections to other areas compared to the converse (Goulas et al., 2014). The results did not support the notion that rostral LPFC (frontopolar cortex, putative area 10) is the top of the hierarchy. Instead, mid LPFC areas (putative areas 45 and 46) tended to be at the highest hierarchical levels, while area 10 was low in the hierarchy. However, homologies between human and monkey LPFC are still uncertain and there are marked differences between human and monkey frontopolar cortex (Semendeferi et al., 2001; Neubert et al., 2014). Nevertheless, our results examining fixed connectivity, a putative correlate of structural connections, are consistent with the findings that mid LPFC is higher and rostral LPFC is lower in the LPFC hierarchy. Collectively, these data suggest that mid LPFC forms the apex of the frontal hierarchy as measured by asymmetrical connectivity.

Other support for a rostral-to-caudal LPFC hierarchy comes from observed changes in functional connectivity by cognitive demands. In a landmark study, Koechlin et al (Koechlin et al., 2003) demonstrated a cascade of activations across LPFC such that the more temporally remote the control process, the more rostral the activation. Moreover, structural equation modeling revealed significant effective connectivity from rostral to caudal areas as a function of cognitive demands, consistent with a rostral-to-caudal hierarchy. While compelling, rostral-to-caudal models were not compared to other organizational schemes. Since a different organization may have produced superior fits to the data, it is difficult to draw conclusions from these connectivity data. In the present dataset, a strictly rostral-to-caudal model of modulations produced significant top-down modulatory effects consistent with Koechlin et al (Koechlin et al., 2003). However, this model produced a substantially worse fit to the data than the model we describe (see Figure 6—figure supplement 3), rendering rostral-to-caudal dynamics doubtful.

In the present data, we performed explicit model comparison and examined effective connectivity as the LPFC was engaged by a variety of cognitive demands. We found that influences as a function of cognitive demands did not follow a strict rostral-to-caudal progression. While cognitive control demands did modulate influences from rostral to mid areas, stimulus domain demands modulated influences from caudal to mid areas. This pattern of connectivity is consistent with the notion that mid areas are an integration zone where domain-general top-down influences and domain-specific bottom-up influences converge. This is corroborated by the activation-behavior relationships which demonstrated that whereas rostral and caudal areas were related to future and current processing, respectively, mid areas were related to both. This suggests that mid LPFC simultaneously represents information from both rostral and caudal areas. From a developmental perspective, it is interesting to note that mid LPFC areas develop later than caudal and rostral LPFC areas (Shaw et al., 2008; Gogtay et al., 2004). As a general pattern, integrative areas tend to develop later than the areas over which they integrate (Gogtay et al., 2004). Thus, the maturation of mid LPFC areas may be influenced by simultaneous top-down and bottom-up demands (Badre and D'Esposito, 2009) preparing the mid LPFC for an integrative role. These ideas are bolstered by evidence that the mid LPFC is a cortical hub with high global connectivity, the degree of which predicts intelligence (Cole et al., 2012). If mid LPFC is a convergence zone where top-down influences from rostral LPFC and bottom-up influences from caudal LPFC meet, it could serve to integrate abstract goals or intentions from rostral areas and concrete contextual information from caudal areas in order to guide action selection.

Perhaps the strongest data to date for rostral/caudal LPFC asymmetries comes from lesion studies. Badre et al (Badre et al., 2009) manipulated the level of cognitive control while testing patients with focal lesions to different LPFC regions. They found that those patients that had damage to mid LPFC demonstrated impaired performance at higher levels of the task, but not lower levels, while patients that had damage to a caudal LPFC demonstrated impaired performance on both higher and lower levels of the task. These results were recently replicated and extended with a different task whereby similar asymmetries were observed across three levels: caudal, mid, and rostral LPFC (Azuar et al., 2014). Collectively, these studies suggest a rostral/caudal hierarchy wherein higher level controllers depend upon intact lower level controllers, but not vice versa.

Both the model of the present data (Figure 6) and the lesion studies are consistent with the proposal that operations of the mid LPFC depend upon caudal LPFC, but not vice versa. As depicted in the model, this is due to bottom-up, but not top-down, modulations from caudal LPFC to mid LPFC as a function of cognitive demand. Of note is that mid LPFC lesions in the studies of Badre et al (Badre et al., 2009) and Azuar et al (Azuar et al., 2014) appeared to include the area we refer to as MFG (putative area 9/46). In our model, MFG processes sub-task cues, consistent with its role in rule processing (Nee and Brown, 2012; Bunge et al., 2005; Genovesio et al., 2005; Wallis et al., 2001). Given that all of the cognitive control demands require processing the sub-task cues, lesions to MFG would affect all of the cognitive control sub-tasks. In this way, operations of FPl are dependent on an intact MFG. Furthermore, in the model, modulations from FPl onto mid LPFC are inhibitory in nature, counteracting the baseline positive influence observed in the fixed connectivity. Thus, the cancellation of negative and positive influences could lead to a negligible impact of FPl lesions on mid LPFC processing. Collectively then, our model indicates that a similar rostral/caudal cascade of asymmetric impairments could be observed with the present task dynamics. Furthermore, our data highlight the need for task-based modeling of neural dynamics in the intact brain to unravel the mechanisms that can lead to lesion-based deficits.

The parameters estimated in our model indicated both positive (i.e. excitatory) and negative (i.e. inhibitory) modulations of effective connectivity. Interestingly, all inhibitory modulations arose from FPl as a function of Contextual Control demands. That is, Contextual Control reduced the influence of this region upon more caudal areas. This is in contrast to the excitatory modulation from FPl to MFG observed during Temporal Control demands. While Temporal Control required preserving past information in order to integrate it into future processing, Contextual Control required the isolation of current processing from previous and future processing. Our data suggest that the dynamics of FPl reflect these two forms of processing. That is, excitatory influence from FPl promotes integrative processing, while inhibitory influence from FPl promotes segregated processing. Such inhibitory influences during Contextual Control may have been necessary to overcome the positive fixed connectivity observed between FPl and other regions. Hence, although FPl was active for both Temporal Control and Contextual Control, this activity promoted different forms of processing under different conditions, realized by dynamic interactions between FPl and other areas of the LPFC.

Our data highlight that directions of influence are dynamic rather than static. That is, whereas fixed connectivity demonstrated asymmetrically greater influence from mid LPFC to other regions, various task demands modulated connectivity towards, but not from, the mid LPFC. This suggests that asymmetries in the LPFC shift according to task demands. Recent work has demonstrated that the LPFC flexibly adapts its connectivity as a function of task demands (Cole et al., 2013), suggesting that dynamics in the LPFC may be of central importance for cognitive control. Whether and how such dynamics support hierarchical processing is a critical matter for future research.

Connectivity modulations that did follow a rostral-to-caudal direction of influence positively predicted independent measures of higher-level cognitive ability. However, modulations in the opposite direction demonstrated no such relationship. While it has been well-documented that top-down influences from the LPFC to posterior areas of cortex are important for working memory, sustained attention, mitigating distraction, and a variety of other cognitive abilities (D'Esposito and Postle, 2015; Gazzaley and Nobre, 2012), the relationship between top-down control within the LPFC and such higher-level cognitive processes has been less well studied. Our data suggest that top-down control within the LPFC is an important determinant of higher-level cognitive function. Furthermore, we found that more strongly asymmetrical hierarchical projections from mid LPFC to other LPFC regions were related to better higher-level cognitive ability. This suggests a beneficial cognitive effect of a hierarchical organization with mid LPFC at the apex.

The term 'hierarchy' is typically used to describe dominance relationships. Accordingly, we have attempted to operationalize such dominance relationships by reasoning that greater outward relative to inward connectivity reflects larger influence of one region over another. Such rationale has been used previously (Badre and D'Esposito, 2009; Goulas et al., 2014). However, different definitions of hierarchy may lead to different assignments of rank among the LPFC. For example, hierarchy may be defined by temporal control, with higher levels organizing behavior over broader timescales (Koechlin et al., 1999; 2003). Alternatively, hierarchy could be defined by modulations of influence as a result of cognitive demands. Such relationships are typified by the top-down biasing that LPFC exerts over posterior areas during difficult selections (Miller and Cohen, 2001). These various definitions lead to different rank orderings in our data. However, regardless of definition, the mid LPFC appeared to play a central role in our data showing the greatest outward relative to inward asymmetry, containing both future and present oriented signals that related to behavior, and providing a convergence zone of modulatory inputs. Hence, these data make it clear that the mid LPFC is a critical nexus for cognitive control.

How to operationally define the rostral/caudal axis of the LPFC has been a matter of considerable debate. Gradients of activation along the rostral/caudal axis of the LPFC have been observed by manipulating the temporal timescale of control processing and relational complexity (see Badre, 2008 for an in-depth review). However, activations spanning the rostral/caudal axis of the LPFC can be observed by tasks that do not clearly manipulate these processes (Crittenden and Duncan, 2014), and manipulations of these processes do not always lead to orderly gradients of activation (Reynolds et al., 2012). The purpose of the present study was not to resolve these issues (but see Nee et al., 2014 for a recent attempt). However, our results are most consistent with the idea that the LPFC is temporally organized (Koechlin et al., 2003; Koechlin and Summerfield, 2007; Fuster, 2001). In these data, caudal LPFC was involved in current processing, providing selective attention to visual stimulus features, while rostral LPFC was involved in future processing, enabling the retention of information for integration into future processing. The mid LPFC appeared to synthesize both current and future processing allowing the use of current and future informed contextual information to organize behavior. Hence, the temporal timescale of processing/representation increased from caudal to rostral areas.

A rostral-to-caudal hierarchy makes intuitive sense from the perspective of the temporal control of action. Long-lasting abstract representations such as goals or plans are thought to influence concrete representations such as actions. Mid-level representations such as contexts are in-between, governing the correct action tendencies given the prevailing situation. By a rostral-to-caudal hierarchy then, goals should govern contexts. However, it is often the case that contexts dictate whether a goal is achievable or not (e.g. a social situation is not a good context within which to write a manuscript). In such cases, the context may suggest a new goal to pursue (e.g. network instead of write). Hence, the ideal controller may take into account longer-term objectives and the current situation in order to arrive at the correct behavior for the moment. Our data indicate that the mid LPFC has the connectivity properties of such a controller, thereby forming a central nexus for cognitive control.

1. 1. Andersson JL
   2. Hutton C
   3. Ashburner J
   4. Turner R
   5. Friston K

   (2001) Modeling geometric deformations in EPI time series

   NeuroImage 13:903–919.

   https://doi.org/10.1006/nimg.2001.0746

   • Google Scholar
2. 1. Ashburner J
   2. Friston K

   (1997) Multimodal image coregistration and partitioning--a unified framework

   NeuroImage 6:209–217.

   https://doi.org/10.1006/nimg.1997.0290

   • Google Scholar
3. 1. Azuar C
   2. Reyes P
   3. Slachevsky A
   4. Volle E
   5. Kinkingnehun S
   6. Kouneiher F
   7. Bravo E
   8. Dubois B
   9. Koechlin E
   10. Levy R

   (2014) Testing the model of caudo-rostral organization of cognitive control in the human with frontal lesions

   NeuroImage 84:1053–1060.

   https://doi.org/10.1016/j.neuroimage.2013.09.031

   • Google Scholar
4. 1. Baddeley A

   (2003) Working memory: looking back and looking forward

   Nature Reviews. Neuroscience 4:829–839.

   https://doi.org/10.1038/nrn1201

   • Google Scholar
5. 1. Badre D
   2. D'Esposito M

   (2007) Functional magnetic resonance imaging evidence for a hierarchical organization of the prefrontal cortex

   Journal of Cognitive Neuroscience 19:2082–2099.

   https://doi.org/10.1162/jocn.2007.19.12.2082

   • 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

   • Google Scholar
7. 1. Badre D
   2. Hoffman J
   3. Cooney JW
   4. D'Esposito M

   (2009) Hierarchical cognitive control deficits following damage to the human frontal lobe

   Nature Neuroscience 12:515–522.

   https://doi.org/10.1038/nn.2277

   • Google Scholar
8. 1. Badre D

   (2008) Cognitive control, hierarchy, and the rostro-caudal organization of the frontal lobes

   Trends in Cognitive Sciences 12:193–200.

   https://doi.org/10.1016/j.tics.2008.02.004

   • Google Scholar
9. 1. Bahlmann J
   2. Aarts E
   3. D'Esposito M

   (2015) Influence of motivation on control hierarchy in the human frontal cortex

   Journal of Neuroscience 35:3207–3217.

   https://doi.org/10.1523/JNEUROSCI.2389-14.2015

   • Google Scholar
10. 1. Bahlmann J
    2. Blumenfeld RS
    3. D'Esposito M

    (2014) The rostro-caudal axis of frontal cortex is sensitive to the domain of stimulus information

    Cerebral Cortex 25:.

    https://doi.org/10.1093/cercor/bht419

    • Google Scholar
11. 1. Botvinick M
    2. Weinstein A

    (2014) Model-based hierarchical reinforcement learning and human action control

    Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences 369:.

    https://doi.org/10.1098/rstb.2013.0480

    • Google Scholar
12. 1. Bunge SA
    2. Wallis JD
    3. Parker A
    4. Brass M
    5. Crone EA
    6. Hoshi E
    7. Sakai K

    (2005) Neural circuitry underlying rule use in humans and nonhuman primates

    Journal of Neuroscience 25:10347–10350.

    https://doi.org/10.1523/JNEUROSCI.2937-05.2005

    • Google Scholar
13. 1. Charron S
    2. Koechlin E

    (2010) Divided representation of concurrent goals in the human frontal lobes

    Science 328:360–363.

    https://doi.org/10.1126/science.1183614

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

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

    Nature Neuroscience 16:1348–1355.

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

    • Google Scholar
15. 1. Cole MW
    2. Yarkoni T
    3. Repovs G
    4. Anticevic A
    5. Braver TS

    (2012) Global connectivity of prefrontal cortex predicts cognitive control and intelligence

    Journal of Neuroscience 32:8988–8999.

    https://doi.org/10.1523/JNEUROSCI.0536-12.2012

    • Google Scholar
16. 1. Crittenden BM
    2. Duncan J

    (2014) Task difficulty manipulation reveals multiple demand activity but no frontal lobe hierarchy

    Cerebral Cortex 24:532–540.

    https://doi.org/10.1093/cercor/bhs333

    • Google Scholar
17. 1. D'Esposito M
    2. Postle BR

    (2015) The cognitive neuroscience of working memory

    Annual Review of Psychology 66:115–142.

    https://doi.org/10.1146/annurev-psych-010814-015031

    • Google Scholar
18. 1. Damasio AR
    2. Anderson SW

    (1993)

    Clinical Neuropsychology

    409–460, The frontal lobes, Clinical Neuropsychology, 3rd Ed, New York, Heilman KM & Valenstein E Oxford University Press.

    • Google Scholar
19. 1. David O
    2. Guillemain I
    3. Saillet S
    4. Reyt S
    5. Deransart C
    6. Segebarth C
    7. Depaulis A

    (2008) Identifying neural drivers with functional MRI: an electrophysiological validation

    PLoS Biology 6:2683–2697.

    https://doi.org/10.1371/journal.pbio.0060315

    • Google Scholar
20. 1. De Pisapia N
    2. Sandrini M
    3. Braver TS
    4. Cattaneo L

    (2012) Integration in working memory: a magnetic stimulation study on the role of left anterior prefrontal cortex

    PloS One 7:e43731.

    https://doi.org/10.1371/journal.pone.0043731

    • Google Scholar
21. 1. De Pisapia N
    2. Slomski JA
    3. Braver TS

    (2007) Functional specializations in lateral prefrontal cortex associated with the integration and segregation of information in working memory

    Cerebral Cortex 17:993–1006.

    https://doi.org/10.1093/cercor/bhl010

    • Google Scholar
22. 1. Duncan J
    2. Owen AM

    (2000)

    Common regions of the human frontal lobe recruited by diverse cognitive demands

    Trends in Neurosciences 23:475–483.

    • Google Scholar
23. 1. Duncan J

    (2013) The structure of cognition: attentional episodes in mind and brain

    Neuron 80:35–50.

    https://doi.org/10.1016/j.neuron.2013.09.015

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

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

    Proceedings of the National Academy of Sciences of the United States of America 110:16616–16621.

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

    • Google Scholar
25. 1. Friston KJ
    2. Harrison L
    3. Penny W

    (2003)

    Dynamic causal modelling

    NeuroImage 19:1273–1302.

    • Google Scholar
26. 1. Fuster JM

    (1997)

    pp xvi, 333 p, The prefrontal cortex: anatomy, physiology, and neuropsychology of the frontal lobe, (3rd ed), Philadelphia, Lippincott-Raven

    pp xvi, 333 p, The prefrontal cortex: anatomy, physiology, and neuropsychology of the frontal lobe, (3rd ed), Philadelphia, Lippincott-Raven.

    • Google Scholar
27. 1. Fuster JM

    (2001)

    The prefrontal cortex--an update: time is of the essence

    Neuron 30:319–333.

    • Google Scholar
28. Book

    1. Fuster JM

    (2008)

    The Prefrontal Cortex ((4th ed))

    London: Academic Press.

    • Google Scholar
29. 1. Gazzaley A
    2. Nobre AC

    (2012) Top-down modulation: bridging selective attention and working memory

    Trends in Cognitive Sciences 16:129–135.

    https://doi.org/10.1016/j.tics.2011.11.014

    • Google Scholar
30. 1. Genovesio A
    2. Brasted PJ
    3. Mitz AR
    4. Wise SP

    (2005) Prefrontal cortex activity related to abstract response strategies

    Neuron 47:307–320.

    https://doi.org/10.1016/j.neuron.2005.06.006

    • Google Scholar
31. 1. Gogtay N
    2. Giedd JN
    3. Lusk L
    4. Hayashi KM
    5. Greenstein D
    6. Vaituzis AC
    7. Nugent TF
    8. Herman DH
    9. Clasen LS
    10. Toga AW
    11. Rapoport JL
    12. Thompson PM

    (2004) Dynamic mapping of human cortical development during childhood through early adulthood

    Proceedings of the National Academy of Sciences of the United States of America 101:8174–8179.

    https://doi.org/10.1073/pnas.0402680101

    • Google Scholar
32. Book

    1. Goldman-Rakic PS

    (1987)

    Circuitry of the primate prefrontal cortex and the regulation of behavior by representational memory

    In: Plum F, Mountcastle V B, editors. Handbook of Physiology. Bethseda: American Physiological Society,. 373––417.

    • Google Scholar
33. 1. Goulas A
    2. Uylings HB
    3. Stiers P

    (2014) Mapping the hierarchical layout of the structural network of the macaque prefrontal cortex

    Cerebral Cortex 24:1178–1194.

    https://doi.org/10.1093/cercor/bhs399

    • Google Scholar
34. 1. Gratton C
    2. Nomura EM
    3. Pérez F
    4. D'Esposito M

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

    Journal of Cognitive Neuroscience 24:1275–1285.

    https://doi.org/10.1162/jocn_a_00222

    • Google Scholar
35. 1. Jeon HA
    2. Anwander A
    3. Friederici AD

    (2014) Functional network mirrored in the prefrontal cortex, caudate nucleus, and thalamus: high-resolution functional imaging and structural connectivity

    Journal of Neuroscience 34:9202–9212.

    https://doi.org/10.1523/JNEUROSCI.0228-14.2014

    • Google Scholar
36. 1. Jeon HA
    2. Friederici AD

    (2013) Two principles of organization in the prefrontal cortex are cognitive hierarchy and degree of automaticity

    Nature Communications 4:2041.

    https://doi.org/10.1038/ncomms3041

    • Google Scholar
37. 1. Koechlin E
    2. Basso G
    3. Pietrini P
    4. Panzer S
    5. Grafman J

    (1999) The role of the anterior prefrontal cortex in human cognition

    Nature 399:148–151.

    https://doi.org/10.1038/20178

    • Google Scholar
38. 1. Koechlin E
    2. Jubault T

    (2006) Broca's area and the hierarchical organization of human behavior

    Neuron 50:963–974.

    https://doi.org/10.1016/j.neuron.2006.05.017

    • Google Scholar
39. 1. Koechlin E
    2. Ody C
    3. Kouneiher F

    (2003) The architecture of cognitive control in the human prefrontal cortex

    Science 302:1181–1185.

    https://doi.org/10.1126/science.1088545

    • Google Scholar
40. 1. Koechlin E
    2. Summerfield C

    (2007) An information theoretical approach to prefrontal executive function

    Trends in Cognitive Sciences 11:229–235.

    https://doi.org/10.1016/j.tics.2007.04.005

    • Google Scholar
41. 1. Lee L
    2. Friston K
    3. Horwitz B

    (2006) Large-scale neural models and dynamic causal modelling

    NeuroImage 30:1243–1254.

    https://doi.org/10.1016/j.neuroimage.2005.11.007

    • Google Scholar
42. 1. Lhermitte F

    (1986) Human autonomy and the frontal lobes. part II: patient behavior in complex and social situations: the "environmental dependency syndrome"

    Annals of Neurology 19:335–343.

    https://doi.org/10.1002/ana.410190405

    • Google Scholar
43. 1. Lund TE
    2. Nørgaard MD
    3. Rostrup E
    4. Rowe JB
    5. Paulson OB

    (2005) Motion or activity: their role in intra- and inter-subject variation in fMRI

    NeuroImage 26:960–964.

    https://doi.org/10.1016/j.neuroimage.2005.02.021

    • Google Scholar
44. Book

    1. Luria AR

    (1966)

    Higher cortical functions in man

    New York: Basic Books.

    • Google Scholar
45. 1. MacDonald AW
    2. Cohen JD
    3. Stenger VA
    4. Carter CS

    (2000)

    Dissociating the role of the dorsolateral prefrontal and anterior cingulate cortex in cognitive control

    Science 288:1835–1838.

    • Google Scholar
46. 1. Miller EK
    2. Cohen JD

    (2001) An integrative theory of prefrontal cortex function

    Annual Review of Neuroscience 24:167–202.

    https://doi.org/10.1146/annurev.neuro.24.1.167

    • Google Scholar
47. 1. Monsell S

    (2003)

    Task switching

    Trends in Cognitive Sciences 7:134–140.

    • Google Scholar
48. 1. Nee DE
    2. Brown JW
    3. Askren MK
    4. Berman MG
    5. Demiralp E
    6. Krawitz A
    7. Jonides J

    (2013) A meta-analysis of executive components of working memory

    Cerebral Cortex 23:264–282.

    https://doi.org/10.1093/cercor/bhs007

    • Google Scholar
49. 1. Nee DE
    2. Brown JW

    (2012) Rostral-caudal gradients of abstraction revealed by multi-variate pattern analysis of working memory

    NeuroImage 63:1285–1294.

    https://doi.org/10.1016/j.neuroimage.2012.08.034

    • Google Scholar
50. 1. Nee DE
    2. Brown JW

    (2013) Dissociable frontal-striatal and frontal-parietal networks involved in updating hierarchical contexts in working memory

    Cerebral Cortex 23:2146–2158.

    https://doi.org/10.1093/cercor/bhs194

    • Google Scholar
51. 1. Nee DE
    2. Jahn A
    3. Brown JW

    (2014) Prefrontal cortex organization: dissociating effects of temporal abstraction, relational abstraction, and integration with FMRI

    Cerebral Cortex 24:2377–2387.

    https://doi.org/10.1093/cercor/bht091

    • Google Scholar
52. 1. Neubert FX
    2. Mars RB
    3. Thomas AG
    4. Sallet J
    5. Rushworth MF

    (2014) Comparison of human ventral frontal cortex areas for cognitive control and language with areas in monkey frontal cortex

    Neuron 81:700–713.

    https://doi.org/10.1016/j.neuron.2013.11.012

    • Google Scholar
53. 1. Niendam TA
    2. Laird AR
    3. Ray KL
    4. Dean YM
    5. Glahn DC
    6. Carter CS

    (2012) Meta-analytic evidence for a superordinate cognitive control network subserving diverse executive functions

    Cognitive, Affective & Behavioral Neuroscience 12:241–268.

    https://doi.org/10.3758/s13415-011-0083-5

    • Google Scholar
54. 1. Petrides M

    (2005) Lateral prefrontal cortex: architectonic and functional organization

    Philosophical Transactions of the Royal Society B: Biological Sciences 360:781–795.

    https://doi.org/10.1098/rstb.2005.1631

    • Google Scholar
55. 1. Reynolds JR
    2. O'Reilly RC
    3. Cohen JD
    4. Braver TS

    (2012) The function and organization of lateral prefrontal cortex: a test of competing hypotheses

    PloS One 7:e30284.

    https://doi.org/10.1371/journal.pone.0030284

    • Google Scholar
56. 1. Romanski LM

    (2004)

    Domain specificity in the primate prefrontal cortex

    Cognitive, Affective & Behavioral Neuroscience 4:421–429.

    • Google Scholar
57. 1. Satterthwaite TD
    2. Elliott MA
    3. Gerraty RT
    4. Ruparel K
    5. Loughead J
    6. Calkins ME
    7. Eickhoff SB
    8. Hakonarson H
    9. Gur RC
    10. Gur RE
    11. Wolf DH

    (2013) An improved framework for confound regression and filtering for control of motion artifact in the preprocessing of resting-state functional connectivity data

    NeuroImage 64:240–256.

    https://doi.org/10.1016/j.neuroimage.2012.08.052

    • Google Scholar
58. 1. Semendeferi K
    2. Armstrong E
    3. Schleicher A
    4. Zilles K
    5. Van Hoesen GW

    (2001) Prefrontal cortex in humans and apes: a comparative study of area 10

    American Journal of Physical Anthropology 114:224–241.

    https://doi.org/10.1002/1096-8644(200103)114:3<224::AID-AJPA1022>3.0.CO;2-I

    • Google Scholar
59. 1. Shaw P
    2. Kabani NJ
    3. Lerch JP
    4. Eckstrand K
    5. Lenroot R
    6. Gogtay N
    7. Greenstein D
    8. Clasen L
    9. Evans A
    10. Rapoport JL
    11. Giedd JN
    12. Wise SP

    (2008) Neurodevelopmental trajectories of the human cerebral cortex

    Journal of Neuroscience 28:3586–3594.

    https://doi.org/10.1523/JNEUROSCI.5309-07.2008

    • Google Scholar
60. 1. Sohn MH
    2. Ursu S
    3. Anderson JR
    4. Stenger VA
    5. Carter CS

    (2000) The role of prefrontal cortex and posterior parietal cortex in task switching

    Proceedings of the National Academy of Sciences of the United States of America 97:13448–13453.

    https://doi.org/10.1073/pnas.240460497

    • Google Scholar
61. 1. Solway A
    2. Diuk C
    3. Córdova N
    4. Yee D
    5. Barto AG
    6. Niv Y
    7. Botvinick MM

    (2014) Optimal behavioral hierarchy

    PLoS Computational Biology 10:e1003779.

    https://doi.org/10.1371/journal.pcbi.1003779

    • Google Scholar
62. 1. Stephan KE
    2. Penny WD
    3. Daunizeau J
    4. Moran RJ
    5. Friston KJ

    (2009a) Bayesian model selection for group studies

    NeuroImage 46:1004–1017.

    https://doi.org/10.1016/j.neuroimage.2009.03.025

    • Google Scholar
63. 1. Stephan KE
    2. Penny WD
    3. Moran RJ
    4. den Ouden HE
    5. Daunizeau J
    6. Friston KJ

    (2010) Ten simple rules for dynamic causal modeling

    NeuroImage 49:3099–3109.

    https://doi.org/10.1016/j.neuroimage.2009.11.015

    • Google Scholar
64. 1. Stephan KE
    2. Tittgemeyer M
    3. Knösche TR
    4. Moran RJ
    5. Friston KJ

    (2009b) Tractography-based priors for dynamic causal models

    NeuroImage 47:1628–1638.

    https://doi.org/10.1016/j.neuroimage.2009.05.096

    • Google Scholar
65. 1. Thirion B
    2. Pinel P
    3. Mériaux S
    4. Roche A
    5. Dehaene S
    6. Poline JB

    (2007) Analysis of a large fMRI cohort: statistical and methodological issues for group analyses

    NeuroImage 35:105–120.

    https://doi.org/10.1016/j.neuroimage.2006.11.054

    • Google Scholar
66. 1. Turner ML
    2. Engle RW

    (1989) Is working memory capacity task dependent?

    Journal of Memory and Language 28:127–154.

    https://doi.org/10.1016/0749-596X(89)90040-5

    • Google Scholar
67. 1. Wallis JD
    2. Anderson KC
    3. Miller EK

    (2001) Single neurons in prefrontal cortex encode abstract rules

    Nature 411:953–956.

    https://doi.org/10.1038/35082081

    • Google Scholar

Author details

1. Derek Evan Nee

   1. Helen Wills Neuroscience Institute, University of California, Berkeley, United States
   2. Department of Psychology, University of California, Berkeley, United States

   Contribution

   DEN, Conception and design, Acquisition of data, Analysis and interpretation of data, Drafting or revising the article

   For correspondence

   denee@berkeley.edu

   Competing interests

   The authors declare that no competing interests exist.

   "This ORCID iD identifies the author of this article:" 0000-0001-7858-6871

2. Mark D'Esposito

   1. Helen Wills Neuroscience Institute, University of California, Berkeley, United States
   2. Department of Psychology, University of California, Berkeley, United States

   Contribution

   MD, Conception and design, Drafting or revising the article

   Competing interests

   The authors declare that no competing interests exist.

• 7,188

  views

• 1,061

  downloads

• 142

  citations

Views, downloads and citations are aggregated across all versions of this paper published by eLife.