2018-Evidence for a Functional Hierarchy of Association ... · Evidence for a Functional Hierarchy of Association Networks Eun Young Choi1, Garrett K. Drayna1, and David Badre2 Abstract

Evidence for a Functional Hierarchy ofAssociation Networks

Eun Young Choi1, Garrett K. Drayna1, and David Badre2

Abstract

■ Patient lesion and neuroimaging studies have identified arostral-to-caudal functional gradient in the lateral frontal cortex(LFC) corresponding to higher-order (complex or abstract) tolower-order (simple or concrete) cognitive control. At the sametime, monkey anatomical and human functional connectivitystudies show that frontal regions are reciprocally connectedwith parietal and temporal regions, forming parallel and dis-tributed association networks. Here, we investigated the linkbetween the functional gradient of LFC regions observed dur-ing control tasks and the parallel, distributed organization ofassociation networks. Whole-brain fMRI task activity corre-sponding to four orders of hierarchical control [Badre, D., &D’Esposito, M. Functional magnetic resonance imaging evi-dence for a hierarchical organization of the prefrontal cortex.Journal of Cognitive Neuroscience, 19, 2082–2099, 2007] wascompared with a resting-state functional connectivity MRI esti-mate of cortical networks [Yeo, B. T., Krienen, F. M., Sepulcre,J., Sabuncu, M. R., Lashkari, D., Hollinshead, M., et al. The orga-

nization of the human cerebral cortex estimated by intrinsic func-tional connectivity. Journal of Neurophysiology, 106, 1125–1165,2011]. Critically, at each order of control, activity in the LFC andparietal cortex overlapped onto a common association networkthat differed between orders. These results are consistent witha functional organization based on separable association net-works that are recruited during hierarchical control. Furthermore,corticostriatal functional connectivity MRI showed that, consistentwith their participation in functional networks, rostral-to-caudalLFC and caudal-to-rostral parietal regions had similar, order-specific corticostriatal connectivity that agreed with a striatalgating model of hierarchical rule use. Our results indicate thathierarchical cognitive control is subserved by parallel and distrib-uted association networks, together forming multiple localizedfunctional gradients in different parts of association cortex. Assuch, association networks, while connectionally organized inparallel, may be functionally organized in a hierarchy via dynamicinteraction with the striatum. ■

INTRODUCTION

The lateral frontal cortex (LFC) has an established role incognitive control function. However, the functional orga-nization of the frontal cortex remains widely debated(Fedorenko,Duncan, &Kanwisher, 2013; Curtis &D’Esposito,2004; Petrides & Pandya, 1994; Goldman-Rakic, 1988). Onehypothesis proposes that the LFC is organized hierarchi-cally to support behavior according to abstract, complex,or multistep rules (see Badre & D’Esposito, 2009; Badre,2008). This hypothesis has been motivated in part by thefrequent observation of an apparent functional gradientfrom caudal to rostral LFC in relation to progressively moreabstract and complex tasks (Badre&D’Esposito, 2007, 2009;Christoff, Keramatian, Gordon, Smith, & Mädler, 2009;Badre, 2008; Buckner, 2003; Koechlin, Ody, & Kouneiher,2003). Though there have been some notable exceptionsto this pattern observed in some experiments (Crittenden& Duncan, 2014; Reynolds, O’Reilly, Cohen, & Braver,2012), rostrocaudal functional distinctions have beenshown to be replicable across a variety of experimentsthat have controlled for variables like overall difficulty and

individual differences in the control structure of the task(Badre & Nee, 2018; Nee & D’Esposito, 2016; Bahlmann,Blumenfeld, & D’Esposito, 2015; Dixon, Fox, & Christoff,2014; Nee, Jahn, & Brown, 2014; Nee & Brown, 2012).Beyond the observation of a functional gradient, there

has been evidence from fMRI effective connectivity (Nee& D’Esposito, 2016; Koechlin et al., 2003) and studies ofpatients with focal neurological damage (Azuar et al.,2014; Badre & D’Esposito, 2009) that separate regionsof LFC along this gradient form a functional hierarchywhereby mostly rostral, higher-order regions differen-tially influence the activity of mostly caudal, lower-orderregions. Though there is some debate as to which regionis the ultimate “top” of this functional hierarchy (Badre &Nee, 2018; Nee & D’Esposito, 2016; Parkin, Hellyer, Leech,& Hampshire, 2015; Goulas, Uylings, & Stiers, 2014;Koechlin et al., 2003), there is strong evidence that inter-regional interactions increase along the rostrocaudal axisof LFC as rules become more complex, and they do sobased on a priority relationship among regions dependingon the task. For example, in a recent electrocorticographystudy, Voytek et al. (2015) showed that, during rule-guidedbehavior, rostral regions of LFC asymmetrically influencedgamma amplitude changes in caudal LFC via directional1Harvard University, 2Brown University

© 2018 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 30:5, pp. 722–736doi:10.1162/jocn_a_01229

theta phase-to-gamma amplitude coupling. Importantly,this hierarchical, interregional theta phase-to-gammaamplitude coupling only occurred when the neurosurgerypatients performed higher-order, abstract rules. Thus,there is evidence that regions along the rostrocaudal axisare not only functionally distinguishable but also, undercertain task conditions, can interact with each otherhierarchically to support higher-order rule use.What might support this functional hierarchy? One

class of hypotheses focuses on corticocortical connectiv-ity (reviewed in Badre & D’Esposito, 2009) with putativeasymmetry in connectivity among caudal-to-rostral frontalregions supporting increasingly higher-order control rep-resentations. Though early theory focused on cortico-cortical connections, recent hypotheses propose that thesehierarchical relationships are established functionally viadynamic interaction with other mediating regions, suchas the striatum. One such hypothesis proposes that thestriatum supports hierarchical relationships among sepa-rate LFC regions via a cortico-basal ganglia-thalamic gatingfunction (Chatham & Badre, 2015; Frank & Badre, 2012).Selection of a task-relevant LFC control representationmay be accomplished by a striatal gating mechanism inwhich the LFC transmits candidate rule information to alocal region of the striatum, which in turn selectively am-plifies cortical processing of adaptive rules in the LFC re-gion through its disinhibitory pallido-thalamo-corticalloop (Hazy, Frank, & O’Reilly, 2007; O’Reilly & Frank,2006; Frank, Loughry, & O’Reilly, 2001; Braver & Cohen,2000). Frank and Badre proposed that this striatal gatingmechanism could be elaborated to support hierarchicalrule use. Specifically, rostral, higher-order LFC regionscould send top–down contextual information to the stria-tum that, through cortico-basal ganglia-thalamic loops, gatecortical processing in caudal, lower-order LFC regions,thereby allowing contextual representations maintained inrostral LFC to influence the selection of contexts or actionsby caudal LFC (Badre & Frank, 2012; Frank & Badre, 2012).These properties may be reflected in the architecture ofcorticostriatal connections in two predictions. First, theremight be an ordered relationship of corticostriatal connec-tivity reflecting candidate rule maintenance loops betweenthe cortex and subregions of the striatum that maintainlevel-specific rule information. Second, there might be anasymmetry in the corticostriatal connectivity (Badre &Frank, 2012; Frank & Badre, 2012) such that higher-ordercortical (e.g., more rostral LFC) regions connect with abroader caudal extent than caudal LFC regions do rostrallyin the striatum, reflecting the influence of higher-order LFCregions on lower-order cortico-basal ganglia-thalamicloops. Thus, there may be an ordered but increasinglysymmetric corticostriatal connectivity from higher- tolower-order cortical regions. Some initial evidence forthese asymmetric connectional relationships between LFCand striatum has been observed using high angular resolu-tion diffusion tractography (Verstynen, Badre, Jarbo, &Schneider, 2012).

Although much work has focused on the functionalgradient within the LFC and its relationship to the striatum,these structures also interact with other cortical regionsduring cognitive control. Monkey anatomical tract-tracingstudies show that the LFC has widespread connections toregions in frontal, parietal, and temporal association cor-tices (Saleem, Miller, & Price, 2014; Morecraft et al., 2012;Petrides & Pandya, 1999, 2006;Morecraft,Geula,&Mesulam,1992; Cavada & Goldman-Rakic, 1989a, 1989b; Vogt &Pandya, 1987). Based on anatomical connectivity patterns,Goldman-Rakic (1988) suggested that there are multipleparallel, distributed association networks, each consistingof a complementary set of regions in frontal, parietal, andtemporal association cortices and dedicated to a distinctfunction. Consistent with this model, recent network par-cellations of the human cerebral cortex using resting-statefunctional connectivity MRI (fcMRI) also yield parallel, dis-tributed association networks, recognizably associatedwith distinct functions (Lee et al., 2012; Power et al.,2011; Yeo et al., 2011) and functionally coupled to distinctregions of the striatum (Choi, Yeo, & Buckner, 2012). Thisparallel, distributed organization of association networkssuggests that cognitive control arises from the interactionsbetween the distributed regions of a network, rather than asingle region acting alone (Cole, Bassett, Power, Braver, &Petersen, 2014; Mesulam, 1981, 1990; Goldman-Rakic,1988). Indeed, numerous fMRI studies have reported activ-ity in parietal and temporal cortices in addition to frontalcortex during cognitive control tasks, including those in-volving complex hierarchical cognitive control (Farooqui,Mitchell, Thompson, & Duncan, 2012; Kouneiher, Charron,& Koechlin, 2009; Badre & D’Esposito, 2007; Jubault,Ody, & Koechlin, 2007; Koechlin & Jubault, 2006; Koechlinet al., 2003). Importantly, however, frontal regions in thesenetwork parcellations from fcMRI are not arrayed in anobvious rostrocaudal manner (Power et al., 2011), nor dothey show a clear hierarchical ordering in their restingcorrelation structure (Yeo et al., 2011). Thus, this raises afundamental question: How do the rostral-to-caudal func-tional distinctions observed in LFC relate to the widelyobserved parallel, distributed organization of associationnetworks?

In this study, we provide some initial answers to thisquestion by contrasting task-based fMRI activations foundduring a hierarchical control study with parcellationsfrom fcMRI. First, we assessed the overlap of fMRI taskactivity from a hierarchical cognitive control study (fromBadre & D’Esposito, 2007) with a comprehensive estimateof human cortical networks in 1000 healthy, young adultparticipants using fcMRI (from Yeo et al., 2011) to directlydetermine their relationship. This comparison revealedthat the separate regional activations identified with fMRIalong the rostrocaudal axis of LFC were systematically as-sociated with distinct functional networks, both in LFCand in parietal cortex. Next, given the striatal gating hy-pothesis and its importance for establishing hierarchicalfunctional relationships, we investigated the functional

Choi, Drayna, and Badre 723

connectivity between the LFC and parietal regions of theseassociation networks and the striatum for agreement withthe first of the predictions of the corticostriatal model.We note that we were unable to test the second predictionof asymmetric skew in the corticostriatal connectivity herewith fcMRI because of confounds arising from the physicalconstraints of the striatum (see Discussion).

METHODSHierarchical Rule Study

We reexamined fMRI task activity from a cognitively de-manding hierarchical rule use study conducted by Badreand D’Esposito (2007). The relevant methods are de-scribed briefly here; see their work for further details.

Participants

The study involved 19 paid participants (10 men) whowere 18–31 years, clinically normal, native Englishspeakers, and right-handed. Written informed consentwas obtained in accordance with procedures set by theCommittee for Protection of Human Subjects at the Uni-versity of California, Berkeley.

Task and Procedure

Conceptually, hierarchical rules consist of higher-orderrules that specify the use of differing lower-order rulesin different contexts. In Badre and D’Esposito’s (2007)study, participants completed four independent hierar-chical rule tasks that ranged from an easy first-order taskto a highly demanding fourth-order task. Tasks had one,two, three, or four hierarchical levels of rule complexity(see Badre & D’Esposito, 2007, their Figure 2; also termedpolicy abstraction, see Badre, Kayser, & D’Esposito, 2010;Badre, 2008; Botvinick, 2008). In a single trial of each task,participants were shown a visual contextual cue andmade a button press response according to the relevantrule. Trials had one, two, or four available rules at thehighest order of each task (termed rule competition,see Badre & D’Esposito, 2007, their Figure 1), whichserved as a parametric variable approximating cognitivedemand in the data analysis to identify regions specificallyinvolved in cognitive control at each rule order.

MRI Data Acquisition

MRI data were collected on a 4-T Varian/Inova (Palo Alto,CA) MRI system. Structural data included anatomical im-ages acquired by a T1-weighted gradient-echo multislicesequence and a high-resolution T1-weighted MP-FLASH3-D sequence. The functional imaging data were acquiredusing a two-shot gradient-echo EPI sequence sensitive toBOLD contrast. Functional imaging parameters were asfollows: repetition time (TR) = 2000 msec, echo time

(TE) = 28 msec, 3.5 × 3.5 × 5 mm, 18 axial slices(whole-brain coverage), and 0.5 mm interslice gap voxels.

fMRI Data Preprocessing

fMRI data were corrected for slice acquisition-dependenttime shifts and interpolated to a 1-sec resolution. Volumesthat were large outliers in global signal were removed byreplacing them with the global mean signal. Subsequentpreprocessing using SPM2 (Wellcome Department ofCognitive Neurology, London, United Kingdom) includedmotion correction, normalization to Montreal NeurologicalInstitute (MNI) space using a 12-parameter affine transfor-mation and a nonlinear transformation, and resamplinginto 3-mm3 voxels. Data were then spatially smoothedwith an 8-mm FWHM Gaussian kernel.

Data Analysis

Data for each task were independently analyzed in SPM2at the participant level using a general linear model,which included a regressor for rule competition. Theresulting beta maps reflected the degree to which signalat each voxel parametrically correlated with rule compe-tition at each order.

Cortical Network Parcellation and Resting-stateFunctional Connectivity Preprocessing

We used the 17-network cortical parcellation created byYeo et al. (2011) and fully preprocessed fcMRI data fromChoi et al. (2012) for the corticostriatal fcMRI analyses.Their relevant methods for creating the parcellationand preprocessing the fcMRI data are described brieflyhere; see their works for further details.

Participants

The study involved 1000 paid participants (427 men) whowere aged 18–35 years (mean age = 21.3 years), clinicallynormal, native English speakers, and majority right-handed(91%). A subset (Discovery sample set, see Choi et al.,2012; Yeo et al., 2011) was used for the fcMRI analysis:500 paid participants (213 male) were ages 18–35 years(mean age = 21.3 years), clinically normal, native Englishspeakers, majority right-handed (91%), and with normal orcorrected-to-normal vision. Participants were excludedfrom the study if their slice-based fMRI signal-to-noise ratiowas low (<100; Van Dijk, Sabuncu, & Buckner, 2012) orartifacts were detected in the MRI data. Participants wereprescreened for prior neurological or psychiatric illness,psychoactive medications, and MRI contraindications.Written informed consent was obtained in accordance withprocedures set by institutional review boards of HarvardUniversity or Partners Healthcare.

724 Journal of Cognitive Neuroscience Volume 30, Number 5

MRI Data Acquisition

MRI data were collected on matched 3-T Tim Trio scanners(Siemens, Erlangen, Germany) using the vendor-supplied12-channel phased-array head coil. Structural data includeda high-resolution multiecho T1-weighted magnetization-prepared gradient-echo image (multiecho MPRAGE).Structural scan (multiecho MPRAGE) parameters were asfollows: TR = 2200 msec, inversion time = 1100 msec,TE = 1.54 msec for Image 1 to 7.01 msec for Image 4, flipangle= 7°, 1.2× 1.2× 1.2mm, and field of view=230. Thefunctional imaging data were acquired using a gradient-echo EPI sequence sensitive to BOLD contrast. Functionalimaging parameters were as follows: TR= 3000msec, TE=30 msec, flip angle = 85°, 3 × 3 × 3 mm voxels, field ofview = 216, and 47 slices aligned to the AC–PC plane(whole-brain coverage). During the functional scans, par-ticipants were instructed to stay still, stay awake, and keeptheir eyes open.

fMRI Data Preprocessing

First four volumes of each run were discarded, sliceacquisition-dependent time shifts were compensatedper volume using SPM2, and head motion was correctedusing rigid body translation and rotation using FMRIBSoftware Library (FSL). The data underwent further pre-processing specific to fcMRI analysis, including low-passtemporal filtering and regression for head motion, whole-brain signal, and ventricular and white matter signal.

Structural MRI Data Preprocessing and SurfaceFunctional–Structural Data Alignment

For each participant, the cerebral cortex was extractedfrom the structural MRI image, modeled as a 2-D flat sur-face, and registered to a common spherical coordinatesystem using the FreeSurfer version 4.5.0 (surfer.nmr.mgh.harvard.edu). Functional images were aligned to thestructural image using FsFast (surfer.nmr.mgh.harvard.edu/fswiki/FsFast). A 6-mm FWHM Gaussian smoothingkernel was applied to the fMRI data in the surface space,and data were downsampled to a 4-mm mesh.

Hybrid Surface- and Volume-based Alignment, Mappingbetween Surface- and Volume-based Coordinates,and Visualization

To perform the fcMRI analyses between the FreeSurfercortical surface and volume, the surface and volume werealigned using a nonlinear volumetric registration algo-rithm (see Choi et al., 2012). Briefly, the striatum wasmodeled as a volume and aligned to the cortical surface.The participant’s fMRI data were transformed into thecommon FreeSurfer nonlinear volumetric space andsmoothed with a 6-mm FWHM smoothing kernel. Spatialcorrespondence between the FreeSurfer surface and vol-

umetric coordinate systems was established by averagingover 1000 participants the transformation from each par-ticipant’s native space to the FreeSurfer surface spaceand the transformation from the FreeSurfer nonlinearvolumetric space. For figure images, the normalized Free-Surfer nonlinear volumetric data were then transformedinto FSL MNI space. The striatum was defined using aFreeSurfer template mask of the striatum (Note that themask does not include the tail of the caudate because ofresolution limitations). All analyses were performed inFreeSurfer surface and volumetric spaces (fsaverage52 mm isotropic resolution). Results were displayed inMNI atlas space for the volume and the left and right in-flated PALS cortical surfaces using Caret software (VanEssen, 2005) for the surface.

Regression of Cerebral Cortex Signal from Striatum

As described by Choi et al. (2012), the mean signal of cor-tical voxels adjacent to the left (within 4 voxels or 8 mm)or right (within 4.5 voxels or 9 mm) putamen was re-gressed from the entire left or right striatum at the partic-ipant level. This signal was determined to likely be spuriousbleeding because (1) it was part of a continuous region ofcorrelations primarily in the insula, but also extending intothe claustrum and putamen; (2) the correlations in thestriatum did not agree with the pattern of monkey ana-tomical projections from the insula to the striatum; and(3) regression of this signal revealed topographic motorcorrelations with the putamen as shown by monkey anat-omy (see Choi et al., 2012, their Figure 13).

We noted another likely region of spurious signal cov-ering the rostral ventral edge of the caudate and adjacentventricles and spread into the septal wall in the fcMRImap for premotor dorsal (PMd). This signal was presenteven after regression of the mean ventricular signal dur-ing preprocessing. These correlations are contradictorywith monkey anatomical projections from premotor cor-tex, which target the dorsal most part of the putamen andcaudate, but not the rostral ventral caudate. Seeding therostral ventral edge of the caudate, adjacent ventricularregion, or septal wall yielded maps with broad, weak cor-relations in these regions, much of the entire medial wallof the cortex, and the lateral motor and premotor corti-ces. Because regression of this spurious signal from thedata did not show qualitative changes to the resultingfcMRI maps, we decided not to perform this step. Wenote their presence with asterisks in Figures 4 and 5.

Creation of Network Parcellation

Yeo et al. (2011) divided the cerebral cortex into 18,715uniformly distributed regions. At the individual participantlevel, for each region, a connectivity profile was computedof its Pearson’s product moment correlation to 1175uniform ROI vertices (∼16 mm apart). The resulting18,715 × 1175 matrix was thresholded to the top 10% of


correlations and binarized. Group matrix was obtained byaveraging over 1000 participants. A k-means clusteringmethod was used to empirically group regions into adefined number of networks based on their connectivityprofiles. Clustering was repeatedly run with random splitsof participants into two groups of n = 500 to assess thestability of different network number solutions. The 17-network parcellation was selected as a relatively stablesolution with fine-grained detail.

Quantification of Task Activity Overlap onNetwork Parcellation

For each participant, SPM beta maps of activity paramet-rically modulating with rule competition were obtainedfor each of the four tasks. SPM beta maps were non-linearly transformed from standardMNI space to FreeSurfervolumetric space and then projected to the FreeSurfercortical surface for overlay with the cortical parcellation.Activity was thresholded at a very lenient threshold (0.05)to maximize sensitivity to overlap at the acceptable cost ofbiasing toward the null hypothesis (i.e., failing to detectdifferences in overlap). Activity was quantified by findingthe mean beta value within the top three networks in theLFC that overlapped the most with the activity over thefour orders (Figure 2) and contained regions that wereconsistent with prior knowledge of the location of theLFC functional gradient. These were a somatomotor-related network (green; Figure 1, top) and two associationnetworks covering the middle (orange) or rostral lateral(maroon) LFC. Mean beta values were also obtained forparietal regions of the same networks. Statistical signifi-cance was determined by an Order× ROI (4× 3) repeated-measures ANOVA. Mauchley’s test of sphericity (i.e., thevariances in the differences of all possible pairs of groupsare the same) showed sphericity in the data for the parietalcortex but not the LFC. To address this, the Greenhouse–

Geisser correction was applied to the LFC ANOVA to adjustthe degrees of freedom and obtain a corrected F ratio.To determine whether network or rostrocaudal posi-

tion (estimated by y coordinate) significantly predictsthe order of task activity, multiple and single linear regres-sions of these factors were performed on group mean taskactivity within the three networks in the left frontal orparietal cortex that were normalized by network and taskand thresholded at 0.05. The statistical significance ofthe difference of model R2 values (R2

network− R2y coordinate)

was assessed using statistical bootstrapping to obtain itsdistribution. This difference distribution was created byresampling 1000 times from the set of single voxel values toobtain R2network, R

2y coordinate, and their difference R2network −

R2y coordinate. The 95% confidence interval was obtained

from the difference distribution, and statistical significancewas determined by examining if the confidence intervalcontained 0.

Resting-state Functional Connectivity Analysis

Analyses were conducted on fully preprocessed fcMRI datafrom the n = 500 subset (Discovery set) described above.

Cortical Seed Region Selection

In the LFC, fcMRI analyses were performed on two setsof cortical vertices. The first set consisted of the closestcortical surface vertices to the regions that were mostparametrically correlated to rule competition in thefirst-, second-, third-, and fourth-order hierarchical ruletasks. The MNI coordinates for these cortical regionsare as follows: −30, −10, 68 (first order; PMd); −38, 10,34 (second order; pre-premotor dorsal [prePMd]); −50,26, 24 (third order; inferior frontal sulcus [IFS]); and−36, 50, 6 (fourth order; frontal polar cortex [FPC]). Thesecond set of cortical vertices was based on the first set but

Figure 1. Comparison ofhierarchical cognitive controlactivity and cortical networkboundaries. Whole-brain grouptask activity (mean beta values)parametrically varying with rulecompetition (approximatingcognitive control) from fourorders of hierarchical rule use(from Badre & D’Esposito,2007) was overlaid on a17-network estimate of corticalnetworks based on resting-statefcMRI in 1000 participants(from Yeo et al., 2011). Lateral(top row), dorsal (middle row),and medial (bottom row) viewsof left hemisphere activity areshown. Black lines denotenetwork boundaries. Networksare shown in unique colors inthe parcellation at the top.


moved by hand to be approximately equidistant fromnearby cortical network boundaries to avoid signal mixingwith neighboring networks, which is a potential confoundin the results. Seeding adjacent cortical vertices showedqualitatively similar fcMRI results. The MNI coordinatesfor these adjusted LFC regions are as follows: −23,−10, 56 (first order); −38, 4, 29 (second order); −54,19, 22 (third order); and −37, 59, −2 (fourth order).In the parietal cortex, seed regions were selected to belocated away from network boundaries and have repre-sentative fcMRI maps of their respective parietal parcel-lations. The MNI coordinates for parietal seed regionsare as follows: −53, −26, 37 (somatomotor-related net-work); −52, −38, 51 (Association Network 1); and −52,−50, 47 (Association Network 2).

Resting-state Functional Connectivity Analysis

Striatal fcMRI maps for specific cerebral seed regions wereobtained by computing the Pearson’s product–momentcorrelation between the cortical region’s fcMRI timecourse and the time courses of striatal voxels. Each corticalregion consisted of a single surface vertex (∼4× 4mm) butshould be considered spatially more extensive due tospatial smoothing. Group mean correlation z maps wereobtained by converting the correlation (r) maps of indi-vidual participants to z maps using Fisher’s r-to-z transfor-mation and averaging across all participants in the group.For participants with multiple runs, the individual par-ticipant z maps were first averaged within participant andsubmitted to the group average. An inverse Fisher’s r-to-z transformation was then applied to the group-averagedcorrelation z map, yielding a group mean correlation (r)map.

Corticostriatal Correlation Plots

To comprehensively evaluate the fcMRI correlationsthroughout the striatum, for each cortical region, the stron-gest correlation value at each coronal slice spanning thestriatum in the group-level maps was plotted to create cor-ticostriatal correlation curves. There were 21 coronal slicesintersecting the striatum in FreeSurfer fsaverage5 2 mmspace (spanning y = 24 to −20 in MNI space). Visual in-spection showed that the correlation curves accurately re-flected the overall pattern of correlations through thestriatum. Each correlation curve was assessed for its meanby considering the corticostriatal correlation curves as afunction r( y), where y is the y coordinate of the peak voxelin each of the 21 coronal slices through the striatum. Wethen calculated the means of the group mean correlationcurves, defined as mean = Σi r yið Þ # yi

Σi r yið Þ. However, we

could not assess the statistical significance using thesemean values because some individual participant cor-

relation curves contained negative correlation values, whichlead to nonsensical statistics. Specifically, with σ =ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

Σi r yið Þ # yi−!yð Þ2q

, a negative correlation value will leadto σ =

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffinegative number

p, that is, an imaginary number,

which cannot be used in statistical tests. Because we wereonly interested in assessing the relative statistical differ-ences of these curves, not an inherent statistical property,we performed hypothesis tests on the nonnormalizedmeans, defined as nonnormalized mean = first moment ="i r( yi) × yi, from the participants. Hypothesis testsconsisted of Order×Mean (4× 1) and Order× Skewness(4 × 1) repeated-measures ANOVAs for the LFC andNetwork × Mean (3× 1) and Network × Skewness (3× 1)repeated-measures ANOVAs for the parietal cortex. Therelationship between Order and the x, y, and z coordinatesof the peak corticostriatal correlation was estimated withsingle and multiple linear regressions.

Figure 2. Percent overlap of hierarchical cognitive control activity withcortical networks. Mean task activity was quantified within networksin the LFC and parietal cortex. Colors correspond to network colorsshown in the parcellation at the bottom. For both LFC and parietalcortex, the top three overlapping networks are the somatomotor-related network (SMN), Association Network 1 (AN1), and AssociationNetwork 2 (AN2). The greatest overlap of activity occurs with SMNin the first order, followed by AN1 in the second and third orders,and AN2 in the fourth order, with a gradient of change.


RESULTSCognitive Control Recruits Gradients of Activity inthe LFC and Parietal Cortex

We reexamined whole-brain fMRI maps from Badre andD’Esposito (2007) of activity parametrically covaryingwith rule competition (estimating cognitive control) spe-cific to four orders of hierarchical rule use. The thresholdof activity in these maps was lowered to compare in themost liberal way possible regional activation with thefcMRI cortical network parcellation. Given that our objec-tive was to look for differences in overlap with separatefunctional networks, using a liberal threshold for thefMRI activation map biases in favor of the null hypothesis,as randomly spatially distributed spurious or noisy activa-tion would reduce our sensitivity to locate differences.

Even at this highly liberal threshold, the rostral-to-caudal gradient of activity from the fourth to first orderswas still distinct in the LFC, as originally reported (Badre& D’Esposito, 2007; Figure 1, top row). Activity was alsopresent in distributed regions in the parietal cortex (Fig-ure 1, middle row), medial frontal cortex (Figure 1, bottomrow), and temporal cortex (Figure 1, top row). In par-ticular, the activity in the parietal cortex was arrayed in asmaller and more overlapping but nonetheless distinctcaudal-to-rostral functional gradient (Figure 1, middlerow). Overall, the global, whole-brain pattern is one oflocalized activity in somatomotor cortex at lower ordersof cognitive control that moves centrifugally in frontaland parietal cortices as higher orders of control are re-cruited. However, given the low statistical threshold usedhere, one cannot draw inferences about these patterns ordistinguish any individual voxel activation from chance.

Nevertheless, we report them, as they are the maps thatare compared with the network parcellation in the sub-sequent analyses.

Cognitive Control Recruits Higher- andLower-order Cortical Networks

We asked whether this broad patterning is random or if ithas a systematic relationship with the independently de-fined cortical networks. To address this, the group meanfMRI map of activity parametrically covarying with rulecompetition at each rule order was overlaid with a previ-ously reported 17-network fcMRI estimate of human corti-cal networks derived from 1000 healthy adult participants(Yeo et al., 2011; Figure 1, top). In the LFC, parametricactivation related to the first-order rule task overlapped pri-marily with regions belonging to a somatomotor-relatednetwork (green), for the second and third orders with anassociation network (orange; Association Network 1), andthe fourth order with a second, separate associationnetwork (maroon; Association Network 2; Figure 2, leftcolumn). Together, these three networks overlapped with64% or more of the parametric activation for each orderin the LFC.Figure 3A shows a quantification of this activity across

the four orders and three networks. Black bars corre-sponding to the first rule order show the greatest para-metric activation in the somatomotor-related networkand progressively less activation for Association Networks 1and 2 (Figure 3A, top). The opposite pattern holds for thefourth-order parametric activation shown in light blue bars,which is least in the somatomotor-related network and

Figure 3. Quantification of hierarchical cognitive control activity in somatomotor-related and association networks. The overlap of order-specificgroup task activity with LFC (A) and parietal (B) regions of a somatomotor-related network and two association networks was quantified. Darkerto lighter bars correspond to lower- to higher-order tasks. Regions are highlighted below (somatomotor-related network, green; AssociationNetwork 1, orange; Association Network 2, maroon). Order× Region repeated-measures ANOVA: F(2, 42) = 8.99, p< .001 for LFC; F(6, 108) = 9.74,p < .001 for parietal cortex.


progressively increases in Association Networks 1 and 2.The second and third rule orders show the greatest para-metric activation in Association Network 1 and less activa-tion in the somatomotor-related network and AssociationNetwork 2. A repeated-measures ANOVA (Greenhouse–Geisser corrected) showed a significant Order × ROIinteraction effect, F(2, 42) = 8.99, p < .001.To determine whether network or rostrocaudal posi-

tion was a significant predictor of the order of task activ-ity and, if so, which factor better explained the variance inorder, multiple and single linear regressions of these fac-tors were performed on the normalized group mean taskactivity in the three networks (Table 1). We sought to testwhether the frontal BOLD activity observed at each orderwas better described as progressively following a rostro-caudal gradient, such as along a single dimension of ab-straction, or, rather, followed the distributed network offrontal regions. These networks are rostrocaudally ar-rayed in one part of LFC, but there are exceptions thatallow us to distinguish this pattern from a basic rostro-caudal gradient. Thus, we took the BOLD activity in thethree networks of interest in the frontal cortex from thefour task orders and tested the hypotheses that the orderof the BOLD activity is better explained by a given voxel’snetwork membership or, alternatively, purely by rostro-caudal location (approximated by the voxel’s y coordi-nate). The multiple regression showed that both factorsare significant predictors of order (βNetwork = 0.6867, p<.001; βy = 0.0201, p < .001) and explained 31.6% of thevariance in order, F(2, 1805) = 417.9, p< .001. However,single linear regressions showed that network explainedmore of the variance in order (R2 = 0.301, F(1, 1806) =

776.9, p < .001) than y coordinate did (R2 = 0.213, F(1,1806) = 489.2, p < .001). The 95% confidence intervalfor the distribution of the difference between these twomodel R2 values is [0.03, 0.14], which does not contain 0.These results indicate that network membership, ratherthan rostrocaudal position, is a stronger predictor of theorder of a given voxel’s BOLD activity in the LFC duringhierarchical cognitive control.

We then tested how the caudal-to-rostral functionalgradient in the parietal cortex overlapped with the parietalregions of the same three cortical networks overlappingwith the LFC functional gradient (Figure 2B, bottom). Inparietal cortex, the parametric activation related to eachrule level overlapped with the corresponding network thatoverlapped that level in LFC. Specifically, the greatest activa-tion in the parietal cortex was located in the somatomotor-related network (green) for the first-order task, AssociationNetwork 1 (orange) for the second- and third-order tasks,and Association Network 2 (maroon) for the fourth-ordertask (Figure 2, right column). Together, these three net-works overlapped with 67% or more of the parametricactivation for each order in parietal cortex.

Figure 3B shows a quantification of this parametricactivation across the four orders and three networks.Again, first-order activity showed the greatest activationwith the somatomotor-related network and progressivelydecreases in Association Networks 1 (orange) and 2(maroon), the second- and third-order activities showthe greatest activation with Association Network 1, andthe fourth-order activity shows the least activation with thesomatomotor-related network and progressively in-creases in Association Networks 1 and 2 (Figure 3B). A

Table 1. Single and Multiple Linear Regressions of Network and Rostrocaudal Position ( y Coordinate) on the Order of Task Activityacross the Four Tasks

Region Factors df

R2 βNetwork βY-Coordinate y Intercept

F Statistic/p Value t Statistic/p Value t Statistic/p Value t Statistic/p Value

Left frontal cortex 2 1805 0.316a 0.6867 0.0201 −0.1769

417.9/<.001 16.523/<.001 6.441/<.001 −0.997/.319

Left frontal cortex 1 1806 0.301 0.8668 — 0.8865

776.9/.001 27.874/<.001 — 13.497/<.001

Left frontal cortex 1 1806 0.213 — 0.0547 −1.2613

489.2/<.001 — 22.117/<.001 −7.133/<.001

Left parietal cortex 2 1775 0.434a 0.5738 −0.0819 4.5647

682.3/<.001 14.646/<.001 −15.217/<.001 17.184/<.001

Left parietal cortex 1 1776 0.361 0.9752 — 0.6143

1003/<.001 31.668/<.001 — 10.262/<.001

Left parietal cortex 1 1776 0.366 — −0.1350 7.6950

1027/<.001 — −32.042/<.001 46.090/<.001

aAdjusted R2 for two factors.


repeated-measures ANOVA showed a significant Order ×ROI interaction effect, F(6, 108) = 9.74, p < .001.

To examine whether network or rostrocaudal positionof task activity better explains order, a multiple linear re-gression was performed that showed that both networkand rostrocaudal position ( y coordinate) are significantpredictors of order (βNetwork = 0.5738, p < .001; βy =−0.0819, p < .001) and explained 43.4% of the variancein order, F(2, 1775) = 682.3, p < .001 (Table 1). How-ever, unlike the LFC, single linear regressions showed asimilar amount of variance in order explained by network(R2 = 0.361, F(1, 1776) = 1003, p < .001) and y coordi-nate (R2 = 0.366, F(1, 1776) = 1027, p < .001). The 95%confidence interval for the distribution of the differencebetween these two model R2 values is [−0.06, 0.04],which does contain 0. These results indicate that networkmembership and rostrocaudal position are similarly strongpredictors of the order of a given voxel’s BOLD activity inthe parietal cortex during hierarchical cognitive control.This is consistent with the single, rostrocaudally arrayedset of network regions in the parietal cortex, in contrastto the multiple, distributed nature of regions in the LFC.

Functional Connectivity between Higher- andLower-order Cortical Networks and the Striatum

We next investigated whether the functional connectivitybetween those networks and the striatum is consistentwith the corticostriatal gating model of hierarchical ruleuse (Badre & Frank, 2012; Frank & Badre, 2012). Basedon this model, each LFC region should feature ordered,spatially distinct, and progressively asymmetric cortico-striatal fcMRI connectivity for lower- versus higher-orderregions. Specifically, we predicted that LFC regionswould have correlations through the caudate of the stri-atum with the peak correlation at a distinct rostrocaudalsite corresponding to the rostrocaudal location of theLFC region based on prior monkey and human cortico-striatal connectivity results (Verstynen et al., 2012; Kemp& Powell, 1970). Importantly, we expected to see a sim-ilar pattern for parietal regions that are members of thesame functional networks.

We first examined striatal functional connectivity withthe four LFC regions most correlated with rule competi-tion at each task order (PMd, prePMd, IFS, and FPC fromBadre & D’Esposito, 2007) located in the somatomotor-related network and Association Networks 1 and 2. (Notethat the exception was the third-order region [IFS],which was not located in Association Network 1, but ratherin an adjacent network. However, because the majorityof the covarying fMRI activity at the third order was locatedin Association Network 1, we nonetheless believe that thisis the involved network.) However, because these seedregions were close to the borders of the association net-works where there is a potential blurring of signals fromneighboring networks, we created a second set of seedregions placed away from the network borders (PMd0, pre-

PMd0, IFS0, and FPC0; Figure 4B, top). The first-order seedregion (PMd0) was primarily correlated with the caudalputamen, consistent with its close functional link to themotor cortex, and weakly correlated with the caudal bodyand tail of the caudate (Figure 4A; note that correlations inthe rostral ventral caudate are likely due to signal bleed-ing from adjacent ventricles and gray and white matter;see asterisks in Figure 4B–D and Methods). The second(prePMd0), third (IFS0), and fourth (FPC0) order regionsshowed longitudinal correlations throughout the rostro-caudal extent of the caudate and parts of the medial puta-men (Figure 4A).The strongest correlations for 21 evenly spaced coronal

slices of the caudate were plotted to create a corticostriatalcurve of correlation for each LFC region (Figure 4B). Thesecurves revealed that, consistent with our first prediction,the fourth-order map had the most rostral peak correla-tion, followed by successively caudal peaks for the third-and second-order maps. A linear regression showed adependence of peak correlation location in the y axis onorder (βy = 3.26, p < .001). This was consistent with therostral-to-caudal locations of the means of each curve,which were located at fourth order (z = 6.92), third order(z = 3.53), second order (z = −1.55), and first order (z=−14.99). A repeated-measures ANOVA showed a significantOrder × Mean interaction effect, F(3, 1996) = 101.43, p<.001. Follow-up paired t tests were significant between allorders ( p < .005).In addition to the rostral–caudal ( y) axis, the medial–

lateral (x), and dorsal–ventral (z) axes were examinedand found to have smaller shifts in mean peak correlationlocation from lower to higher order. From first to fourthorder, the mean peak correlation shifted from medial tolaterally for the x axis (Figure 4C) and from ventral todorsally for the z axis (Figure 4D). A single linear regres-sion showed a dependence of the peak correlation loca-tion in the x or z axis on order (βx = −0.50 and βz =1.91, both ps < .001). A multiple linear regression of allthree x, y, z axes on order indicated a statistically signif-icant rostrolateral shift in peak correlation from lower tohigher order (βx =−0.11, βy = 0.22, both ps < .001; βz =−0.06, p = .07).We next examined the striatal functional connectivity

for parietal seed regions in the somatomotor-relatednetwork and Association Networks 1 and 2 (Figure 5B,top). If parietal regions are recruited with LFC regions aspart of cortical networks, we predicted a similar rostral-to-caudal differential of striatal connectivity from caudal-to-rostral (higher- to lower-order) parietal regions. Indeed,similar to the first-order LFC region (PMd0), the seed regionin the somatomotor-related network was most stronglycorrelated with the putamen (Figure 5A) and weakly cor-related with the tail of the caudate (Figure 5B–D; note thatcorrelations in the rostral ventral caudate are likely due tosignal bleeding from adjacent ventricles and gray and whitematter; see asterisks in Figure 5B–D and Methods). Likethe second-, third-, and fourth-order LFC regions (prePMd0,


IFS0, FPC0), Association Networks 1 and 2 seed regions inthe parietal cortex also showed longitudinal correlationsin the caudate and medial putamen (Figure 5A).A plot of peak correlations from 21 coronal slices

through the caudate showed a rostrocaudal differencein the location of the peak correlation for each order(Figure 5B, bottom), consistent with the first prediction.A linear regression showed a dependence of peak corre-lation location in the y axis on order (βy = 4.58, p <.001). This was consistent with the means of the curves,which were located at Association Network 2 (z = 5.45),Association Network 1 (z = −2.18), and somatomotor-related network (z = −14.22). A repeated-measuresANOVA showed a significant Network× Mean interactioneffect, F(2, 1497) = 132.97, p < .001. Follow-up pairedt tests were significant between all networks ( p < .001).As with the LFC, the x (Figure 5C) and z (Figure 5D)

axes showed smaller medial to lateral and dorsal to ven-tral shifts, respectively, in the positions of the peak cor-

relations from lowest to highest order. A linear regressionshowed a dependence of the peak correlation locationin the x or z axis on order (βx = −0.69, βz = 2.50, bothps < .001). A multiple linear regression of all three x, y,z axes on order did not yield any statistically significantregression coefficients (βx = −0.10, p = .54; βy = −0.14,p = .71; βz = 0.31, p = .57). Altogether, we found thatparietal regions correlate with the striatum with similarconnectivity as the corresponding LFC regions. These re-gions correlate with primarily rostrocaudal differences inpeak correlation, consistent with the first model prediction.

DISCUSSION

Hierarchical cognitive control has been associated with agradient of higher- to lower-order rostral-to-caudal re-gions in the LFC. At the same time, monkey anatomicaltract-tracing and human resting-state fcMRI shows thatthe LFC is part of parallel and distributed association

Figure 4. Functionalconnectivity between the LFCand the striatum. (A) Coronalslices display the striatal resting-state correlations of corticalregions chosen based onpeak covariation with rulecompetition (from Badre &D’Esposito, 2007) at the first(PMd0), second (prePMd0),third (IFS0), and fourth (FPC0)orders of rule use (see B, top)and adjusted from originallyreported locations to avoidconfounding signal bleedingat network boundaries. (B)Corticostriatal correlationplots show the peak striatalcorrelations (darker centerline) with SEM (surroundinglighter area) for each order at21 coronal slices through thecaudate. Cortical seed regionsare shown above the plot andoverlaid with the networkboundaries (black lines) of the1000-participant 17-networkcortical parcellation (from Yeoet al., 2011). Medial–lateral(C) and dorsal–ventral(D) locations of the peakcorrelations for each orderare shown for the caudate inthe same 21 coronal slices.Gray lines denote striatalborders. Asterisks indicatespurious first-order correlationsdue to signal bleeding fromadjacent ventricles, gray matter,and white matter. Order ×Mean repeated-measuresANOVA: F(3, 1946) = 88.46, p < .001. Follow-up paired t tests were significant between all orders ( p < .001), except for the first versus secondorders. Because of technical confounds, assessment of the asymmetry of the curves was inconclusive (see Results and Discussion).


networks. How do these regional versus network viewsof LFC functional organization relate to one another?We report the following observations addressing thisquestion. (1) Distributed activity at each order of cogni-tive control is not restricted to LFC but also includes re-gions of medial frontal, parietal, and temporal cortices. Inparticular, in addition to the previously observed rostral-to-caudal gradient of activity in the LFC, an approximatecaudal-to-rostral gradient is present in the parietal cortexfrom higher- to lower-order cognitive control. (2) Activityrelated to each order of cognitive control preferentiallyoverlaps with a particular network. The same networkpreferences are observed in both frontal and parietalcortex. Crucially, for at least LFC, network membershipbetter accounted for hierarchical order than did absoluterostrocaudal position. (3) The rostral-to-caudal LFC andcaudal-to-rostral parietal regions of the same cortical net-work both have similar patterns of striatal connectivitysuch that higher-order networks show peak correlationwith more rostral parts of the striatum. Thus, to answerthe question above, our results suggest that there arespecific higher- and lower-order association networks thatare differentially recruited during hierarchical cognitivecontrol tasks. The distributed regions of different net-

works form apparent local functional gradients locatedthroughout association cortex. The largest, most robustone of these is the LFC gradient, and at least one otherfunctional gradient located in the parietal cortex is alsoinvolved. These results suggest that it is network member-ship rather than rostrocaudal locus that determines rankin a hierarchy. Furthermore, the respective functionalinfluence of these networks on each other may partlyarise from the differing influence each network has overthe striatum. Thus, these data are consistent with a modelwhereby these association networks are connectionallyorganized mostly in parallel and broadly support cognitivecontrol but can be functionally organized hierarchicallyvia dynamic interaction with the striatum.Several lines of study have suggested a hierarchical

relationship among distributed association networks.Throughout the cortex, there are broad gradients of cyto-architectonic differentiation (Pandya & Yeterian, 1990)and anatomical connectivity (Mesulam, 1998; Pandya &Yeterian, 1990; Jones & Powell, 1970) from sensory tosensory association to high-order association cortex, in-cluding in the LFC and parietal cortex. Regions of higheror lower hierarchical level within these distributed gradi-ents are connected by long-range anatomical connections.

Figure 5. Functionalconnectivity between theparietal cortex and the striatum.(A) Coronal slices display thestriatal resting-state correlationsof cortical regions (see B, top)from the somatomotor-relatednetwork and AssociationNetworks 1 and 2. (B)Corticostriatal correlationplots show the peak striatalcorrelations (darker center line)with the SEM (surroundinglighter area) for each corticalregion at 21 coronal slicesthrough the caudate. Corticalregions are indicated abovethe plot and overlaid with thenetwork borders (black lines) ofthe 1000-participant 17-networkcortical parcellation (from Yeoet al., 2011). Medial–lateral (C)and dorsal–ventral (D) locationsof the peak correlations for eachnetwork are shown for thecaudate in the same 21 coronalslices. Gray lines denote striatalborders. Asterisks indicatespurious first-order correlationsdue to signal bleeding fromadjacent ventricles, gray matter,and white matter. Order ×Mean repeated-measuresANOVA: F(2, 1407) = 134.37,p < .001. Follow-up pairedt tests were significant between all orders ( p < .005). Because of technical confounds, assessment of the asymmetry of the curves was inconclusive(see Results and Discussion).


For example, there are connections betweenmotor-relatedregions in the LFC and parietal cortex and centrifugallyfarther pairs of LFC and parietal regions in associationcortex (Margulies & Petrides, 2013; Pandya & Yeterian,1990; Chavis & Pandya, 1976). Based on such observations,Fuster’s model proposes a hierarchical organization ofcognitive networks involving the LFC and parietal cortex(Fuster, 2009). Here, we identify three networks in par-ticular that overlap primarily with hierarchical functionalactivation. Across prior functional studies of hierarchical cog-nitive control, LFC regions similar to the PMd, IFS, and FPCare readily observed to dissociate and approximately maponto this tripartite network distinction (Nee & D’Esposito,2016; Bahlmann et al., 2015; Nee & Brown, 2013; Barbalatet al., 2011; Koechlin et al., 2003). We note, however, thatthere are cases in which regions belong to the same networkbut do not always coactivate during tasks. Notably, prePMdand the more rostral IFS both belong to Association Net-work 1 but have been observed in some studies to showseparate response patterns (Nee & D’Esposito, 2016;Chatham, Frank, & Badre, 2014; Badre & D’Esposito,2007; Koechlin et al., 2003). This suggests that the networkparcellations delineate regions that are overall more func-tionally associated than others but do not necessarily co-activate within network nor dissociate across networks forall tasks.A second key point is that, though there was clear pref-

erence for a given network for each level of control, theoverlap also was not complete, instead reflecting a differ-ential mixing of the networks. For example, as can be seenin Figure 2, activity related to the highest level of control ismost associated with Association Network 2. However, italso has some overlap with Association Network 1. Con-versely, the third-order control activity overlaps primarilywith Association Network 1 but includes some overlap withAssociation Network 2, as well. What accounts for this mix-ing? It is possible that this gradient of network involvementreflects noise. Another possibility is that the networkboundaries themselves are estimated from fcMRI data inseparate participants. As both the task and individuals dif-fer, slight differences in these network boundaries mightblur sharper functional bounds. Finally, each networkmay not correspond to a particular single level of controlper se, but solving each control problem requires inter-action among the networks to differing degrees.Consistent with the hypothesis that hierarchical rela-

tionships are established, at least in part, through inter-action with the striatum, we found for both LFC andparietal components of each network that the overallpattern of striatal connectivity was primarily a higher- tolower-order rostrocaudal differential in peak correlationlocus. This pattern is consistent with the first predictionof the corticostriatal gating model that these separatenetworks are regulated by distinct, local cortico-basalganglia-thalamic loops. These rostral-to-caudal LFC andcaudal-to-rostral striatal correlations agree with diffusionMRI studies showing a rostral-to-caudal correspondence

in fibers from the LFC to the striatum (Verstynen et al.,2012; Draganski et al., 2008). This is also consistent withanatomical tract-tracing observations that cortical regionsproject longitudinally through the striatum (Selemon &Goldman-Rakic, 1985; Goldman & Nauta, 1977; Künzle,1975) with, it is thought, the densest projections froma cortical region terminating at its most proximal striatalregion (Kemp & Powell, 1970). Interestingly, however, thereverse organization is true for parietal cortex: caudal-to-rostral parietal regions have a rostral-to-caudal differentialin striatal connectivity. A direct comparison of specificprefrontal and parietal monkey anatomical tract-tracingcases from the literature shows evidence for the organiza-tion seen here. Figure 6 shows striatal projection patternsfor three concentric pairs of anatomically connected LFCand parietal regions, each of which has a unique patternof striatal connectivity that is shared by the LFC and pari-etal region of each pair. Furthermore, these projectionpatterns resemble the lower- to higher-order humanfcMRI correlation patterns for the LFC and parietal cortex(Figures 4A and 5A). Although it is difficult to discernwhere the densest projection sites are without directlyexamining the anatomical tissue, these cases show thatprefrontal and parietal regions share similar patterns ofanatomical projections to the striatum and that these pro-jection patterns bear close resemblance to those seen withfcMRI here. In addition, these fcMRI patterns are consistentwith overlapping prefrontal and parietal connection zonesobserved in the striatum frommonkey and human connec-tivity studies (Jarbo&Verstynen, 2015; Selemon&Goldman-Rakic, 1985), including evidence that these cortical regionsform interconnected networks with distinct patterns ofassociated striatal connectivity (Choi, Tanimura, Vage, Yates,& Haber, 2017; Choi et al., 2012; Yeterian & Van Hoesen,1978; but see Selemon & Goldman-Rakic, 1985).

A second prediction of the corticostriatal gating modelis that there may be an asymmetric skew of the cortico-striatal correlation curves that is reflective of the influ-ence of higher-order regions on caudal regions. In thepresent study, we were unable to test this prediction withfcMRI due to confounds arising from the physical con-straints of the striatum, which create a ceiling-floor effectthat artificially affects assessments of asymmetric connec-tivity across the striatum. Any curves that peak at theends of the striatum (e.g., fourth-order curve) artificiallyappear to have greater asymmetry simply from being “cutoff” before it resolves to 0. Thus, for technical reasons,we are unable to test for an asymmetric, hierarchicalrelationship between the curves with fcMRI. We note thatevidence for an asymmetric structure of LFC to striatalconnections was found previously with diffusion trac-tography (Verstynen et al., 2012). Future study usingneurophysiological methods that can observe dynamic,real-time activity would be a next step to answering thisquestion.

What is the role of the parietal cortex within thesehierarchical networks? By virtue of the parietal cortex’s


high-order, multimodal role in processing and integrat-ing complex visual and somatomotor properties, theparietal cortex likely performs this processing in anorder-specific manner during hierarchical cognitive con-trol. Collins and Frank (2013) have suggested and pro-vided evidence for a more involved role of the parietalcortex that represents the visual properties of the stimu-lus, combines this with context-related information fromthe pFC, and sends this multiplexed context–stimulus in-formation to the premotor corticostriatal loop for actionselection. Within the multiple, hierarchical networksschema reported here, a similar process may occur ateach order, where a parietal region processes the visualproperties of the relevant, order-specific features of thecontext and stimulus and creates a multiplex of context–stimulus information that is sent to the pFC, parietal, andstriatal regions of the downstream, lower-order cortico-striatal network.

Altogether, this study broadens a regional view of thefunctional organization of the LFC by demonstrating therelationship between the functional gradient observed inthe LFC during hierarchical cognitive control and anfcMRI-based estimate of distributed cortical networks.We observe evidence that higher- and lower-order asso-

ciation networks are differentially recruited across levelsof hierarchical cognitive control and that the distributedregions of these networks form local functional gradientsthat have been observed throughout association cortex.Furthermore, the LFC and parietal regions of these net-works have similar correlation patterns within the stria-tum that are consistent with cases reported in theanatomical literature. The present identification of spe-cific hierarchically organized networks serves as a basisfor further functional, anatomical, and neurocomputationalinvestigations of between-network organization andinteractions.

Acknowledgments

The authors thank Drs. Randy Buckner, Christopher Chatham,and Derek Nee for valuable discussion and Drs. Thomas Yeo,Avram Holmes, Fenna Krienen, and Julia Lehman for technicalassistance. This work was supported by a National ScienceFoundation Graduate Research Fellowship (E. Y. C.). Theauthors declare no competing financial interests.

Reprint requests should be sent to Eun Young Choi, Departmentof Neurosurgery, Stanford University, 318 Campus Drive, W1103, Stanford, CA 94305, or via e-mail: [email protected].

Figure 6. Anatomicalprojections from the LFC orparietal cortex to the striatum.Top: Red circles indicateapproximate injection siteslocated in (A) area 46d, (A0) area7a/Opt, (B) area 6d, (B0) area5b/PEc, (C) area 4, and (C0)area 5a/PE. Pairs of LFC andparietal regions are anatomicallyconnected (Cavada & Goldman-Rakic, 1989b, Case 2; Petrides &Pandya, 1994, Case 17 Chavis& Pandya, 1976, Case 15;Petrides & Pandya, 1994, Case 6;Burman, Bakola, Richardson,Reser, & Rosa, 2014b, Cases7-11; Burman, Bakola,Richardson, Reser, & Rosa,2014a, Case 7). A and A0 projectto the dorsal and medial edgesof the caudate. B and B0 projectto regions in and around thedorsal internal capsule andlateral edge of the putamen.C and C0 project to the dorsalputamen. Anatomical tracingsshow dense corticostriatalprojections and were redrawnfor conformity with permissionfrom the following: (A) Calzavara,Mailly, & Haber (2007), (A0)Yeterian and Pandya (1993), (B)Miyata and Sasaki (1984), (B0)Yeterian and Pandya (1993), (C) Künzle (1975), and (C0) Yeterian and Pandya (1993). Blank spots without striatal images are positions that were examinedbut not shown or explicitly discussed by the original authors, implying a continuation of adjacent striatal projection patterns. Dashed outline of striatumindicates a slice at a position that was mentioned in the original text as being devoid of projections, but not illustrated.


REFERENCES

Azuar, C., Reyes, P., Slachevsky, A., Volle, E., Kinkingnehun, S.,Kouneiher, F., et al. (2014). Testing the model of caudo-rostralorganization of cognitive control in the human with frontallesions. Neuroimage, 84, 1053–1060.

Badre, D. (2008). Cognitive control, hierarchy, and therostrocaudal organization of the frontal lobes. Trends inCognitive Sciences, 12, 193–200.

Badre, D., & D’Esposito, M. (2007). Functional magneticresonance imaging evidence for a hierarchical organizationof the prefrontal cortex. Journal of Cognitive Neuroscience,19, 2082–2099.

Badre, D., & D’Esposito, M. (2009). Is the rostrocaudal axis ofthe frontal lobe hierarchical? Nature Reviews Neuroscience,10, 659–669.

Badre, D., & Frank, M. J. (2012). Mechanisms of hierarchicalreinforcement learning in cortico-striatal circuits 2: Evidencefrom fMRI. Cerebral Cortex, 22, 527–536.

Badre, D., Kayser, A. S., & D’Esposito, M. (2010). Frontal cortexand the discovery of abstract action rules. Neuron, 66,315–326.

Badre, D., & Nee, D. E. (2018). Frontal cortex and thehierarchical control of behavior. Trends in CognitiveSciences, 22, 170–188.

Bahlmann, J., Blumenfeld, R. S., & D’Esposito, M. (2015). Therostrocaudal axis of frontal cortex is sensitive to the domainof stimulus information. Cerebral Cortex, 25, 1815–1826.

Barbalat, G., Chambon, V., Domenech, P. J., Ody, C., Koechlin,E., Franck, N., et al. (2011). Impaired hierarchical controlwithin the lateral prefrontal cortex in schizophrenia.Biological Psychiatry, 70, 73–80.

Botvinick, M. M. (2008). Hierarchical models of behaviorand prefrontal function. Trends in Cognitive Sciences, 12,201–208.

Braver, T. S., & Cohen, J. D. (2000). On the control of control:The role of dopamine in regulating prefrontal function andworking memory. In S. Monsell & J. Driver (Eds.), Controlof cognitive processes: Attention and performance XVIII(pp. 713–737). Cambridge, MA: MIT Press.

Buckner, R. L. (2003). Functional-anatomic correlates ofcontrol processes in memory. Journal of Neuroscience,23, 3999–4004.

Burman, K. J., Bakola, S., Richardson, K. E., Reser, D. H., &Rosa, M. G. (2014a). Patterns of cortical input to the primarymotor area in the marmoset monkey. Journal of ComparativeNeurology, 522, 811–843.

Burman, K. J., Bakola, S., Richardson, K. E., Reser, D. H., &Rosa, M. G. (2014b). Patterns of affarent input to the caudaland rostral areas of the dorsal premotor cortex (6DC and6DR) in the marmoset monkey. Journal of ComparativeNeurology, 522, 3683–3716.

Calzavara, R., Mailly, P., & Haber, S. N. (2007). Relationshipbetween the corticostriatal terminals from areas 9 and 46, andthose from area 8A, dorsal and rostral premotor cortex andarea 24c: An anatomical substrate for cognition to action.European Journal of Neuroscience, 26, 2005–2024.

Cavada, C., & Goldman-Rakic, P. S. (1989a). Posterior parietalcortex in rhesus monkey: I. Parcellation of areas based ondistinctive limbic and sensory corticocortical connections.Journal of Comparative Neurology, 287, 393–421.

Cavada, C., & Goldman-Rakic, P. S. (1989b). Posterior parietalcortex in rhesus monkey: II. Evidence for segregatedcorticocortical networks linking sensory and limbic areas withthe frontal lobe. Journal of Comparative Neurology, 287,422–445.

Chatham, C. H., & Badre, D. (2015). Multiple gates on workingmemory. Current Opinion in Behavioral Sciences, 1, 23–31.

Chatham, C. H., Frank, M. J., & Badre, D. (2014). Corticostriataloutput gating during selection from working memory.Neuron, 81, 930–942.

Chavis, D. A., & Pandya, D. N. (1976). Further observations oncorticofrontal connections in the rhesus monkey. BrainResearch, 117, 369–386.

Choi, E. Y., Tanimura, Y., Vage, P. R., Yates, E. H., & Haber, S. N.(2017). Convergence of prefrontal and parietal anatomicalprojections in a connectional hub in the striatum.Neuroimage,146, 821–832.

Choi, E. Y., Yeo, B. T., & Buckner, R. L. (2012). Theorganization of the human striatum estimated by intrinsicfunctional connectivity. Journal of Neurophysiology, 108,2242–2263.

Christoff, K., Keramatian, K., Gordon, A. M., Smith, R., &Mädler, B. (2009). Prefrontal organization of cognitive controlaccording to levels of abstraction. Brain Research, 1286,94–105.

Cole, M. W., Bassett, D. S., Power, J. D., Braver, T. S., &Petersen, S. E. (2014). Intrinsic and task-evoked networkarchitectures of the human brain. Neuron, 83, 238–251.

Collins, A. G. E., & Frank, M. J. (2013). Cognitive control overlearning: Creating, clustering, and generalizing task-setstructure. Psychological Review, 120, 190–229.

Crittenden, B. M., & Duncan, J. (2014). Task difficultymanipulation reveals multiple demand activity but nofrontal lobe hierarchy. Cerebral Cortex, 24, 532–540.

Curtis, C. E., & D’Esposito, M. (2004). The effects ofprefrontal lesions on working memory performance andtheory. Cognitive, Affective, & Behavioral Neuroscience, 4,528–539.

Dixon, M. L., Fox, K. C., & Christoff, K. (2014). Evidence forrostrocaudal functional organization in multiple brain areasrelated to goal-directed behavior. Brain Research, 1572,26–39.

Draganski, B., Kherif, F., Klöppel, S., Cook, P. A., Alexander,D. C., Parker, G. J., et al. (2008). Evidence for segregated andintegrative connectivity patterns in the human basal ganglia.Journal of Neuroscience, 28, 7143–7152.

Farooqui, A. A., Mitchell, D., Thompson, R., & Duncan, J.(2012). Hierarchical organization of cognition reflected indistributed frontoparietal activity. Journal of Neuroscience,32, 17373–17381.

Fedorenko, E., Duncan, J., & Kanwisher, N. (2013). Broaddomain generality in focal regions of frontal and parietalcortex. Proceedings of the National Academy of Sciences,U.S.A., 110, 16616–16621.

Frank, M. J., & Badre, D. (2012). Mechanisms of hierarchicalreinforcement learning in corticostriatal circuits 1:Computational analysis. Cerebral Cortex, 22, 509–526.

Frank, M. J., Loughry, B., & O’Reilly, R. C. (2001). Interactionsbetween frontal cortex and basal ganglia in working memory:A computational model. Cognitive, Affective, & BehavioralNeuroscience, 1, 137–160.

Fuster, J. (2009). The prefrontal cortex: Fourth edition. London:Academic Press.

Goldman, P. S., & Nauta, W. J. (1977). An intricately patternedprefronto-caudate projection in the rhesus monkey. Journalof Comparative Neurology, 72, 369–386.

Goldman-Rakic, P. S. (1988). Topography of cognition: Paralleldistributed networks in primate association cortex. AnnualReview of Neuroscience, 11, 137–156.

Goulas, A., Uylings, H. B., & Stiers, P. (2014). Mapping thehierarchical layout of the structural network of the macaqueprefrontal cortex. Cerebral Cortex, 24, 1178–1194.

Hazy, T. E., Frank, M. J., & O’Reilly, R. C. (2007). Towards anexecutive without a homunculus: Computational models ofthe prefrontal cortex/basal ganglia system. Philosophical


Transactions of the Royal Society of London, Series B:Biological Sciences, 362, 1601–1613.

Jarbo, K., & Verstynen, T. D. (2015). Converging structural andfunctional connectivity of orbitofrontal, dorsolateral prefrontal,and posterior parietal cortex in the human striatum. Journalof Neuroscience, 35, 3865–3878.

Jones, E. G., & Powell, T. P. (1970). An anatomical study ofconverging sensory pathways within the cerebral cortex ofthe monkey. Brain, 93, 793–820.

Jubault, T., Ody, C., & Koechlin, E. (2007). Serial organization ofhuman behavior in the inferior parietal cortex. Journal ofNeuroscience, 27, 11028–11036.

Kemp, J. M., & Powell, T. P. (1970). The cortico-striate projectionin the monkey. Brain, 93, 525–546.

Koechlin, E., & Jubault, T. (2006). Broca’s area and thehierarchical organization of human behavior. Neuron, 50,963–974.

Koechlin, E., Ody, C., & Kouneiher, F. (2003). The architectureof cognitive control in the human prefrontal cortex. Science,302, 1181–1185.

Kouneiher, F., Charron, S., & Koechlin, E. (2009). Motivationand cognitive control in the human prefrontal cortex. NatureNeuroscience, 12, 939–945.

Künzle, H. (1975). Bilateral projections from precentral motorcortex to the putamen and other parts of the basal ganglia.An autoradiographic study in Macaca fascicularis. BrainResearch, 88, 195–209.

Lee, M. H., Hacker, C. D., Snyder, A. Z., Corbetta, M., Zhang, D.,Leuthardt, E. C., et al. (2012). Clustering of resting statenetworks. PLoS One, 7, e40370.

Margulies, D. S., & Petrides, M. (2013). Distinct parietal andtemporal connectivity profiles of ventrolateral frontal areasinvolved in language production. Journal of Neuroscience,33, 16846–16852.

Mesulam, M. M. (1981). A cortical network for directedattention and unilateral neglect. Annals of Neurology, 10,309–325.

Mesulam, M. M. (1990). Large-scale neurocognitive networksand distributed processing for attention, language, andmemory. Annals of Neurology, 28, 597–613.

Mesulam, M. M. (1998). From sensation to cognition. Brain,121, 1013–1052.

Miyata, M., & Sasaki, K. (1984). Horseradish peroxidase studieson thalamic and striatal connections of the mesial part ofarea 6 in the monkey. Neuroscience Letters, 49, 127–133.

Morecraft, R. J., Geula, C., & Mesulam, M. M. (1992).Cytoarchitecture and neural afferents of orbitofronal cortexin the brain of the monkey. Journal of ComparativeNeurology, 323, 341–358.

Morecraft, R. J., Stilwell-Morecraft, K. S., Cipolloni, P. B., Ge, J.,McNeal, D. W., & Pandya, D. N. (2012). Cytoarchitecture andcortical connections of the anterior cingulate and adjacentsomatomotor fields in the rhesus monkey. Brain ResearchBulletin, 87, 457–497.

Nee, D. E., & Brown, J. W. (2012). Rostral–caudal gradientsof abstraction revealed by multi-variate pattern analysis ofworking memory. Neuroimage, 63, 1285–1294.

Nee, D. E., & Brown, J. W. (2013). Dissociable frontal-striataland frontal-parietal networks involved in updating hierarchicalcontexts in working memory. Cerebral Cortex, 23, 2146–2158.

Nee, D. E., & D’Esposito, M. (2016). The hierarchical organizationof the lateral prefrontal cortex. eLife, 5, e12112.

Nee, D. E., Jahn, A., & Brown, J. W. (2014). Prefrontal cortexorganization: Dissociating effects of temporal abstraction,relational abstraction, and integration with fMRI. CerebralCortex, 24, 2377–2387.

O’Reilly, R. C., & Frank, M. J. (2006). Making workingmemory work: A computational model of learning in theprefrontal cortex and basal ganglia. Neural Computation,18, 283–328.

Pandya, D. N., & Yeterian, E. H. (1990). Prefrontal cortex inrelation to other cortical areas in rhesus monkey: Architectureand connections. Progress in Brain Research, 85, 63–94.

Parkin, B. L., Hellyer, P. J., Leech, R., & Hampshire, A. (2015).Dynamic network mechanisms of relational integration.Journal of Neuroscience, 35, 7660–7673.

Petrides, M., & Pandya, D. N. (1994). Comparative architectonicanalysis of the human and macaque frontal cortex. In F.Boller & J. Grafman (Eds.), Handbook of neuropsychology(pp. 17–58). Amsterdam: Elsevier.

Petrides, M., & Pandya, D. N. (1999). Dorsolateral prefrontalcortex: Comparative cytoarchitectonic analysis in thehuman and the macaque brain and corticocorticalconnection patterns. European Journal of Neuroscience, 11,1011–1036.

Petrides, M., & Pandya, D. N. (2006). Efferent associationpathways originating in the caudal prefrontal cortex in themacaque monkey. Journal of Comparative Neurology, 498,227–251.

Power, J. D., Cohen, A. L., Nelson, S. M., Wig, G. S., Barnes,K. A., Church, J. A., et al. (2011). Functional networkorganization of the human brain. Neuron, 72, 665–678.

Reynolds, J. R., O’Reilly, R. C., Cohen, J. D., & Braver, T. S.(2012). The function and organization of lateral prefrontalcortex: A test of competing hypotheses. PLoS One, 7,e30284.

Saleem, K. S., Miller, B., & Price, J. L. (2014). Subdivisions andconnectional networks of the lateral prefrontal cortex inthe macaque monkey. Journal of Comparative Neurology,522, 1641–1690.

Selemon, L. D., & Goldman-Rakic, P. S. (1985). Longitudinaltopography and interdigitation of corticostriatal projectionsin the rhesus monkey. Journal of Neuroscience, 5,776–794.

Van Dijk, K. R., Sabuncu, M. R., & Buckner, R. L. (2012). Theinfluence of head motion on intrinsic functional connectivityMRI. Neuroimage, 59, 431–438.

Van Essen, D. C. (2005). A Population-Average, Landmark- andSurface-based (PALS) atlas of human cerebral cortex.Neuroimage, 28, 635–662.

Verstynen, T. D., Badre, D., Jarbo, K., & Schneider, W. (2012).Microstructural organizational patterns in the humancorticostriatal system. Journal of Neurophysiology, 107,2984–2995.

Vogt, B. A., & Pandya, D. N. (1987). Cingulate cortex of therhesus monkey: II. Cortical afferents. Journal ofComparative Neurology, 262, 271–289.

Voytek, B., Kayser, A. S., Badre, D., Fegen, D., Chang, E. F.,Crone, N. E., et al. (2015). Oscillatory dynamics coordinatinghuman frontal networks in support of goal maintenance.Nature Neuroscience, 18, 1318–1324.

Yeo, B. T., Krienen, F. M., Sepulcre, J., Sabuncu, M. R.,Lashkari, D., Hollinshead, M., et al. (2011). The organizationof the human cerebral cortex estimated by intrinsicfunctional connectivity. Journal of Neurophysiology, 106,1125–1165.

Yeterian, E. H., & Pandya, D. N. (1993). Striatal connections ofthe parietal association cortices in rhesus monkeys. Journalof Comparative Neurology, 332, 175–197.

Yeterian, E. H., & Van Hoesen, G. W. (1978). Cortico-striateprojections in the rhesus monkey: The organization of certaincortico-caudate connections. Brain Research, 139, 43–63.


Documents

2018-Evidence for a Functional Hierarchy of Association ... · Evidence for a Functional Hierarchy of Association Networks Eun Young Choi1, Garrett K. Drayna1, and David Badre2 Abstract