Characterisation of Functional Brain Networks underlying Cognitive Reasoning and Intelligence Luke James Hearne BPsySci (Hons)

A thesis submitted for the degree of Doctor of Philosophy at The University of Queensland in 2017 Queensland Brain Institute

Abstract

Intelligence is a general mental capacity that encompasses the ability to plan, problem-solve, reason, think abstractly and comprehend complex relations between concepts. Measured via complex, multi- step reasoning tasks, individual differences in intelligence scores are highly predictive of a number of crucial life outcomes. Despite many decades of research, however, the neural bases of intelligence and reasoning abilities are not well understood. Early and lesion studies suggested that regions of the are exclusively responsible for intelligent behaviour and higher cognitive reasoning abilities. More recently, however, investigators have conceptualized intelligence as relying upon a widely distributed network of interconnected nodes. The aim of my thesis was to investigate human reasoning and intelligence within this newer, network-centric framework. I conducted a series of functional magnetic resonance imaging (fMRI) experiments to characterize relevant networks in the brain, and related these networks to reasoning task performance, individual differences in intelligence, and the breakdown of reasoning ability amongst individuals with developmental anomalies in brain wiring.

The thesis is organized into seven chapters. In the first chapter, I introduce the research question in the context of previous literature that has investigated the brain-basis of reasoning and intelligence by measuring brain activity using fMRI. In doing so, I discuss the origins of intelligence, why task complexity matters, as well as modern in-vivo brain imaging techniques and their biological basis. I then present the concepts of brain network organization, connectivity and graph theory as a relatively untapped avenue to explain higher cognitive reasoning. Armed with this new framework for understanding the brain, I discuss a number of studies that have used this approach to inform the key questions investigated in this thesis.

In Chapter 2 I employed an individual-differences approach to investigate the relationship between ‘resting-state’ baseline functional networks and commonly used measures of intelligence. Previous work has implicated a fronto-parietal network of brain regions that support higher cognitive reasoning, formalized in the parieto-frontal theory of intelligence. However, existing evidence for the fronto-parietal theory is mixed, at least with respect to findings from connectivity studies. To address this issue, I undertook a quantitative summary of previous literature and performed an empirical analysis of data from a large cohort of individuals (N = 317) sourced from the Human Project. I found that connectivity between the default-mode and fronto-parietal networks best explained individual differences in intelligence.

In Chapter 3 I investigated the relationship between default-mode and cognitive control networks during a complex reasoning task. Participants completed a modified version of the classic Wason Selection Task, a measure of relational reasoning, in which the number of cognitive variables is systematically varied from trial to trial. Consistent with previous work, I found that cortical regions

within fronto-parietal networks demonstrated a parametric increase in activation with increases in reasoning complexity. By contrast, default-mode network areas showed a parametric ‘deactivation’ with increases in complexity. Critically, bilateral angular gyri within the default-mode network showed increased functional connectivity with the anterior prefrontal cortex and , challenging the notion that functional segregation between default-mode and cognitive control networks supports increased cognitive reasoning demand.

In Chapter 4, I developed and tested a modified version of a relational reasoning paradigm known as the Latin Square Task (LST). In this task, participants are presented with a four-by-four matrix populated with geometric shapes, blank spaces and a single probe location. Participants are asked to identify the item belonging at the probe location by following the rule that each shape can only occur once in every row and column (analogous to the popular game known as Sudoku). I found that the LST produced large and reliable complexity effects on response accuracy and reaction time.

In Chapter 5, I combined resting-state and task-driven fMRI to examine how flexible, task- specific reconfigurations associated with increasing reasoning demands are integrated within a stable intrinsic brain topology. Participants underwent an initial resting-state scan, followed by the LST, and then a second resting-state scan. Using a suite of network analysis techniques, I found that networks that were segregated at rest merged during task states. Larger increases in network efficiency within the newly established brain modules were associated with better reasoning performance. These results shed light on the network architectures that underlie external task engagement, and highlight selective changes in brain connectivity that are brought to bear with increases in task complexity.

In Chapter 6 I used the LST to examine brain networks underlying cognitive reasoning in a small cohort of individuals with callosal dysgenesis (CD), a rare disorder in which the corpus callosum, the major tract connecting the two cerebral hemispheres, is malformed or absent. I found that despite intact interhemispheric functional connectivity, the connectivity profile of those with CD was abnormal compared with healthy controls. This aberrant connectivity was revealed only under conditions of higher reasoning demands, suggesting a ‘state’, rather than ‘trait’ network difference.

Finally in Chapter 7 I summarise the overall findings of this thesis, namely that increments in reasoning complexity are supported by flexible, moment-to-moment changes in covariance between distinct brain regions. I also discuss a novel role for default-mode and fronto-parietal connectivity in identifying individual differences in intelligence at rest, as well as during complex task execution. I suggest that recent network based reconceptualization of global workspace theory may provide an explanation for such changes; however, future work should aim to test these potential mechanisms using methods that allow causal inferences to be made, such as brain stimulation.

Declaration by author This thesis is composed of my original work, and contains no material previously published or written by another person except where due reference has been made in the text. I have clearly stated the contribution by others to jointly-authored works that I have included in my thesis.

I have clearly stated the contribution of others to my thesis as a whole, including statistical assistance, survey design, data analysis, significant technical procedures, professional editorial advice, and any other original research work used or reported in my thesis. The content of my thesis is the result of work I have carried out since the commencement of my research higher degree candidature and does not include a substantial part of work that has been submitted to qualify for the award of any other degree or diploma in any university or other tertiary institution. I have clearly stated which parts of my thesis, if any, have been submitted to qualify for another award.

I acknowledge that an electronic copy of my thesis must be lodged with the University Library and, subject to the policy and procedures of The University of Queensland, the thesis be made available for research and study in accordance with the Copyright Act 1968 unless a period of embargo has been approved by the Dean of the Graduate School.

I acknowledge that copyright of all material contained in my thesis resides with the copyright holder(s) of that material. Where appropriate I have obtained copyright permission from the copyright holder to reproduce material in this thesis.

Publications during candidature Hearne, L.J., Cocchi, L., Zalesky, A., Mattingley, J.B., 2017. Reconfiguration of brain network architectures between resting-state and complexity-dependent cognitive reasoning. J. Neurosci. 37, 485–17. doi:10.1523/JNEUROSCI.0485-17.2017 (available online: https://goo.gl/GH9tdf). Cocchi, L., Yang, Z., Zalesky, A., Stelzer, J., Hearne, L. J., Gollo, L. L., & Mattingley, J. B. (2017). Neural decoding of visual stimuli varies with fluctuations in global network efficiency. Mapping, 38(6), 3069–3080. http://doi.org/10.1002/hbm.23574. Hearne, L. J., Mattingley, J. B., & Cocchi, L. (2016). Functional brain networks related to individual differences in human intelligence at rest. Scientific Reports, 6(August), 32328. doi:10.1038/srep32328 (available online: https://goo.gl/4Wc9cb). Hearne, L. J., Cocchi, L., Zalesky, A., & Mattingley, J. B. (2015). Interactions between default mode and control networks as a function of increasing cognitive reasoning complexity. Human , 36(7), 2719–2731. doi:10.1002/hbm.22802 (available online: https://goo.gl/wmXWnS).

Publications included in this thesis Hearne, L. J., Mattingley, J. B., & Cocchi, L. (2016). Functional brain networks related to individual differences in human intelligence at rest. Scientific Reports. Incorporated as Chapter 2 (available online: https://goo.gl/4Wc9cb).

Contributor Statement of contribution Luke J. Hearne (Candidate) Designed experiment (75%) Analysed data (100%) Wrote the paper (80%) Edited paper (70%) Jason B. Mattingley Edited paper (15%) Luca Cocchi Designed experiment (25%) Wrote the paper (20%) Edited paper (15%)

Hearne, L. J., Cocchi, L., Zalesky, A., & Mattingley, J. B. (2015). Interactions between default mode and control networks as a function of increasing cognitive reasoning complexity. Human Brain Mapping. Incorporated as Chapter 3 (available online: https://goo.gl/wmXWnS).

Contributor Statement of contribution Luke J. Hearne (Candidate) Designed experiment (75%) Analysed data (100%) Wrote the paper (70%) Edited paper (75%) Luca Cocchi Designed experiment (25%) Wrote the paper (20%) Edited paper (15%) Andrew Zalesky Edited paper (5%) Jason B. Mattingley Wrote the paper (10%) Edited paper (10%)

Hearne, L.J., Cocchi, L., Zalesky, A., Mattingley, J.B., 2017. Reconfiguration of brain network architectures between resting-state and complexity-dependent cognitive reasoning. J. Neurosci. Incorporated as Chapter 5 (available online: https://goo.gl/GH9tdf).

Contributor Statement of contribution Luke J. Hearne (Candidate) Designed experiment (75%) Collected data (100%) Analysed data (100%) Wrote the paper (70%) Edited paper (75%) Luca Cocchi Experiment design (15%) Wrote the paper (20%) Edited paper (15%) Andrew Zalesky Edited paper (5%) Jason B. Mattingley Experiment design (15%) Wrote the paper (10%) Edited paper (10%)

Contributions by others to the thesis

Dr Luca Cocchi and Professor Jason Mattingley (supervisors) contributed to the conception and design of studies contained within this thesis. Dr Andrew Zalesky and Tong Wu provided analytical guidance for the work contained in Chapters 2, 3 and 5. Dr Oscar Jacoby and Zoie Nott contributed to data collection in Chapter 5. Tim Edwards, Annalisa Paolina and Dr Ryan Dean (Callosal dysgenesis research team) assisted with data collection for Chapter 6.

Statement of parts of the thesis submitted to qualify for the award of another degree

None

Acknowledgements

This thesis couldn’t have completed without the help and support of a long list of people. First, my supervisors Luca Cocchi and Jason Mattingley, you have both been fantastic mentors. I feel very lucky to have had you both as my supervisors. To Jason, thank you for your support, advice and encouragement along the way (especially during the later months). To Luca, thank you for investing so much time into my learning and growth as a young researcher. A special mention to Andrew Zalesky, I benefitted tremendously from your knowledge and feedback on my work.

On a more pragmatic note, thank you to the Australian Postgraduate Award scheme, the Queensland Brain Institute, the Science of Learning Research Centre, and the University of Queensland for funding my studies and travel.

Thank you to my friends and colleagues at the Queensland Brain Institute and Centre of Advanced Imaging. Aiman, Nicole, Oscar and Zoie – thanks for making those long hours at the MR scanner less boring. To the extended Mattingley, Dux and Cunnington lab members thank you for the support and discussions we have had over the last four years. I couldn’t ask for a better bunch of colleagues and friends to enjoy conferences, retreats and QBI coffee with (everyday at 10am sharp).

To my friends and family, thank you for supporting me throughout my PhD. To Mum and Dad, I know you still have no idea what I do, but thank you for supporting me. I love you all very much. James; thank you for all the chats we had, despite being halfway around the world. Last, but not least, Meg. You were always there when I needed you. Thank you for everything.

Keywords Reasoning, intelligence, higher cognition, networks, functional connectivity, resting-state, fMRI

Australian and New Zealand Standard Research Classifications (ANZSRC) ANZSRC code: 170205, Neurocognitive Patterns and Neural Networks, 80% ANZSRC code: 170109, Personality, Abilities and Assessment, 20%

Fields of Research (FoR) Classification FoR code: 1109, , 80% FoR code: 1701, Psychology, 20%

Table of Contents

Chapter 1. General Introduction ...... 1 1.1 Intelligence and reasoning ...... 2 1.1.1 General intelligence ...... 2 1.1.2 Complexity ...... 4 1.1.3 Investigating intelligence in the brain ...... 7 1.2 Assessing changes in human brain activity in vivo ...... 8 1.2.1 Physiological basis of fMRI ...... 8 1.2.2 Univariate analysis of the BOLD signal ...... 9 1.3 Brain activity related to intelligence and reasoning behaviour ...... 11 1.3.1 From local to global models of brain activation associated with reasoning complexity ..... 11 1.4 From functional specialisation to brain networks ...... 14 1.4.1 Analysing functional brain networks ...... 16 1.4.2 BOLD-based functional connectivity ...... 18 1.4.3 Large-scale functional networks of the brain ...... 19 1.4.4 The relationship between large-scale functional and structural networks ...... 20 1.5 Functional networks related to reasoning and intelligence ...... 21 1.5.1 Task-evoked changes in functional networks related to the processing of relational complexity ...... 22 1.5.2. Functional networks related to individual differences in intelligence and reasoning ability ...... 24 1.5.3 ‘Dysconnectivity’ and reasoning ...... 25 1.6 Thesis aims ...... 26 Chapter 2: Functional brain networks related to individual differences in human intelligence at rest ...... 28 2.1 Preamble ...... 29 2.2 Abstract ...... 29 2.3 Introduction ...... 29 2.4 Methods...... 31 2.4.1 Explorative mapping of relationship between intelligence and resting-state networks ...... 31 2.4.2 Analysis of intrinsic functional networks supporting human intelligence using HCP data 32 2.4.3 HCP Data Preprocessing and Analysis ...... 33 2.5 Results ...... 34 2.5.1 Exploratory mapping: resting-state functional connectivity related to intelligence ...... 34 2.5.2 Empirical analysis of resting-state functional networks supporting human intelligence using HCP data ...... 35 2.6 Discussion ...... 37 Chapter 3: Interactions between default mode and control networks as a function of increasing cognitive reasoning complexity ...... 40 3.1 Preamble ...... 41 3.2 Abstract ...... 41 3.3 Introduction ...... 41 3.4 Materials and Methods ...... 42 3.4.1 Participants ...... 42 3.4.2 Paradigm ...... 42

3.4.3 Imaging ...... 45 3.5 Results ...... 46 3.5.1 Complexity-induced change in neural activity ...... 47 3.5.2 Complexity-induced change in PPI connectivity ...... 48 3.6 Discussion ...... 52 Chapter 4: The Latin Square Task as a measure of relational reasoning: a replication and assessment of reliability ...... 56 4.1 Preamble ...... 57 4.2 Abstract ...... 57 4.3 Introduction ...... 57 4.4 Method ...... 59 4.4.1 Participants ...... 59 4.4.2 LST items and presentation ...... 60 4.4.3 Procedure ...... 60 4.4.4 Statistical analyses ...... 60 4.5 Results ...... 61 4.5.1. Behaviour associated with increasing relational complexity ...... 61 4.5.2. Reliability of the Latin Square Task ...... 63 4.6 Discussion ...... 64 Chapter 5: Reconfiguration of brain network architectures between resting-state and complexity-dependent cognitive reasoning ...... 66 5.1 Preamble ...... 67 5.2 Abstract ...... 67 5.3 Introduction ...... 67 5.4 Materials and Methods ...... 69 5.4.1 Participants ...... 69 5.4.2 Experimental paradigm ...... 69 5.4.3 Neuroimaging acquisition and pre-processing ...... 71 5.4.4 Functional connectivity network construction ...... 73 5.4.5 Analysis overview ...... 73 5.4.6 Figures and Visualization ...... 76 5.5 Results ...... 77 5.5.1 Behavioural results ...... 77 5.5.2 Functional brain module reconfiguration ...... 78 5.5.3 Network Based Statistic analysis ...... 81 5.5.4 Within-module global efficiency and behaviour ...... 82 5.6 Discussion ...... 84 Chapter 6: Anomalous functional network integration in response to cognitive control demands in human callosal dysgenesis ...... 88 6.1 Preamble ...... 89 6.2 Abstract ...... 89 6.3 Introduction ...... 89 6.4 Methods...... 92 6.4.1 Participants ...... 92 6.4.2 Task ...... 93 6.4.3 Imaging data analysis ...... 94

6.5 Results ...... 96 6.5.1 Behaviour ...... 96 6.5.2 Brain activity ...... 96 6.5.3 Brain connectivity ...... 99 6.6 Discussion ...... 101 Chapter 7: General Discussion ...... 104 7.1 Relational reasoning is supported by distributed and flexible functional networks ...... 106 7.2 The interplay between default-mode and fronto-parietal functional connectivity in support of reasoning and intelligence ...... 107 7.3 Global network properties that support increased cognitive demand ...... 108 7.4 Future Directions ...... 110 7.5 Conclusions ...... 112 References ...... 113 Appendices ...... 139

List of Figures and Tables Chapter 1 Figure 1-1 Example items from the Raven’s Progressive Matrices Test...... 4 Figure 1-2 Example items from common relational reasoning tasks and their link to intelligence...... 6 Figure 1-3 Components of the blood-oxygen level dependent (BOLD) response...... 9 Figure 1-4 ‘Cognitive Control’ and ‘Default-mode’ networks associated with increased task demands...... 13 Figure 1-5 Network segregation and integration...... 15 Figure 1-6 General pipeline for fMRI functional connectivity analyses...... 17 Figure 1-7 Canonical resting-state networks of the human cortex...... 20 Figure 1-8 Schematic of results and review by Cocchi and colleagues (2014a, 2013)...... 23

Chapter 2 Table 2-1 Characteristics of studies included in the resting-state explorative mapping...... 31 Figure 2-1 Pairwise functional connections associated with intelligence at rest from previous literature...... 35 Figure 2-2 Network-intelligence analysis on 317 independent HCP participants...... 36 Table 2-2 Regions implicated in the analysis of the data...... 37

Chapter 3 Figure 3-1 Schematic of the Wason Selection Task (WST) and modified version used in the current study...... 43 Table 3-1 Experimental conditions in the modified WST, and examples of rules and cards ...... 44 Figure 3-2 Complexity-based changes in regional activity...... 47 Table 3-2 General linear model results used to investigate changes in task-evoked (PPI) connectivity ...... 48 Figure 3-3 Complexity-evoked changes in PPI connectivity...... 49 Table 3-3 Pairwise changes in PPI functional connectivity between source and target regions as a functional of relational complexity ...... 50 Figure 3-4 Complexity-evoked changes in PPI connectivity within the default-mode network ...... 51 Table 3-4 Pairwise change in PPI connectivity between source and target regions encompassing the default-mode network as a function of relational complexity ...... 52

Chapter 4 Figure 4-1 Example items from the modified version of the Latin Square Task (LST)...... 58 Figure 4-2 Example LST trial sequence...... 61 Figure 4-3 Behavioural performance on the Latin Square Task (LST), and relationship with fluid intelligence as measured with the Advanced Progressive Matrices (APM)...... 62 Table 4-2 Split-half reliability of the Latin Square Task ...... 63 Figure 4-4 Comparison of mean accuracy on the LST in the current experiment, and the accuracy values reported in Birney and Bowman (2009)...... 65

Chapter 5 Figure 5-1 Experimental design and sequence of displays in a typical trial of the Latin Square Task. 71 Figure 5-2 Behavioural results for the Latin Square Task visualized as box and whisker plots...... 77 Figure 5-3 Modular structure as a function of reasoning complexity in the Latin Square Task...... 79 Figure 5-4 Changes in Variation of information as a function of reasoning complexity...... 80

Figure 5-5 Change in pairwise functional connectivity associated with reasoning complexity...... 82

Figure 5-6 Changes in global network efficiency (Eglob) across the identified reasoning task modules...... 83 Figure 5-7 Conceptual model of functional networks supporting reasoning and rest states...... 86

Chapter 6 Figure 6-1 Anatomy of the corpus callosum...... 90 Table 6-1 Demographics of the callosal dysgenesis participants and their scores on intellectual function measures...... 92 Figure 6-2 MR images of structural anomalies in the seven callosal dysgenesis (CD) participants. ... 93 Figure 6-3 Example problems and sequence of displays in a typical trial of the Latin Square Task. .. 94 Figure 6-4 Performance accuracy of the two groups (callosal dysgenesis and neurotypical controls) in the Latin Square Task...... 96 Figure 6-5 Brain activity associated with increased reasoning complexity...... 97 Table 6-2 Regions of Interest derived from GLM analysis...... 98 Figure 6-6 Follow-up analyses of beta values averaged across ROIs within the FPN and CON...... 99 Figure 6-7 Network-level functional connectivity trends associated with increasing reasoning complexity...... 100 Figure 6-8 Analysis of edges within the fronto-parietal network...... 101

Chapter 7 Figure 7-1 Global workspace theory and network reconfigurations...... 108 Figure 7-2 Combined brain stimulation and imaging as a potential future direction to establish the relations between brain networks and nodes...... 111

List of Abbreviations used in this thesis

AAL Automated anatomical labelling ACC Anterior cingulate cortex AG AI Anterior insula ANOVA Analysis of variance APM Raven's Advanced Progressive Matrices BCT Brain Connectivity Toolbox BOLD Blood-oxygen level dependent CCN Cognitive control network CD Callosal dysgenesis CON Cingulo-opercular network CSD Constrained spherical deconvolution CSF Cerebrospinal fluid DARTEL Diffeomorphic anatomical registration through exponentiated lie algebra DICOM Digital Images and Communications in Medicine DLPFC Dorsolateral prefrontal cotex DMN Default-mode network DWI Diffusion weighted images EEG EPI Echo Planar Imaging FC Functional connectivity FDR False discovery rate fMRI Functional magnetic resonance imaging FOV Field of view FPN Fronto-parietal network FPV Fronto-parietal-visual module FSIQ Full Scale Intelligence Quotient FWE Familywise error corrected g General intelligence GLM General linear modelling HCP Human Connectome Project IFJ Inferior frontal gyrus IPS Intraparietal sulcus IQ Intelligence Quotient LPFC Lateral prefrontal cortex LST Latin Square Task MEG MNI Montreal Neurological Institute MPFC Medial prefrontal cortex MR Magnetic resonance MRI Magnetic resonance imaging NART National Adult Reading Test NBS Network based statistic P-FIT Parieto-frontal integration theory

PC PCC Posterior cingulate cortex PET Positron emission tomography PFC Prefrontal cortex PIQ Performance Intelligence Quotient PMAT Penn's Progressive Matrices PPI Psychophysiological interaction Pre-SMA Pre-supplementary motor area RC Relational complexity RLPFC Rostrolateral prefrontal cortex ROI Region of interest RPM Raven's Progressive Matrices SPM Statistical parametric mapping TC Temporal cortex tDCS Transcranial direct current stimulation TE Time echo TMS Transcranial magnetic stimulation TR Time repetition VHMC Voxel-mirrored homotopic connectivity VIn Variation of Information VIQ Verbal Intelligence Quotient WAIS Wechsler Adult Intelligence Scale WASI Wechsler Abbreviated scale of Intelligence WST Wason Selection task

Chapter 1. General Introduction

1 1.1 Intelligence and reasoning

1.1.1 General intelligence

Scientists have been trying to answer the question of why some individuals are more cognitively capable than others for a long time (Spearman, 1904). The practical utility of measuring and quantifying intelligence initially drove progress in the field. Alfred Binet created an early intelligence test for the French government to predict which children might need special attention (Binet and Simon, 1916). Around the same time, the US army wanted to ‘classify soldiers according to their mental ability’ in World War I, which led to the development of the ‘Army Alpha’ test (Yerkes and Yoakum, 1920). The outcomes of these research programs, and others like them, influenced government policy. In 1944, the Butler Education Act ordered that children with a cognitive ability score sufficiently high at age 11 should complete a more advanced curriculum (Jeffereys, 1984).

Tests from the early 20th century, as well as measures used today such as Cattel’s Culture Fair Test (Cattell, 1949) and the Wechsler Adult Intelligence Scales (Weschler, 2008), rely on the concept of the ‘positive manifold’. The notion of the positive manifold is simple: the outcomes of different tests tapping a wide variety of cognitive abilities (e.g., memory, perception, reasoning, vocabulary) tend to correlate within individuals. In other words, when an individual scores highly on one test of cognitive ability, he or she is likely to score highly on other, different tests of cognitive ability. Therefore, simply measuring a wide variety of abilities within individuals allows an estimate of their general intelligence (g) to be extrapolated (Spearman, 1904). It is suggested that Spearman’s g represents the general ability to understand complex ideas, ‘adapt effectively to the environment, engage in complex reasoning and overcome obstacles by taking thought’ (Neisser et al., 1996).

General intelligence has been shown repeatedly to predict a number of important life outcomes (Plomin and Deary, 2014). For example, it has been demonstrated that g predicts school achievement from a young age (Ceci, 1991; Deary et al., 2007), as well as markers of socioeconomic success such as income and job performance (Hunter, 1986; Strenze, 2007). Moreover, several large studies (with sample sizes of over one million participants) have demonstrated that g-scores higher than one standard deviation above the mean are associated with significant reductions in mortality (Batty et al., 2009 - 32% reduction; Calvin et al., 2011 - 24% reduction). Such effects appear to propagate to entire populations as well; states and countries with higher g have shown better economic outcomes, such as higher per-capita gross domestic product, faster economic growth and more equal income distributions (Lynn, 2010; Meisenberg, 2012).

Intelligence is highly heritable. Approximately 40 – 80% of the variability in g can be attributed to genetic factors (Nisbett et al., 2012; Petrill et al., 2004). As with other traits, this relationship can be modulated by certain variables. For example, higher socioeconomic status is associated with higher levels of heritability (Turkheimer et al., 2003). Early studies were unsuccessful in consistently linking

2 specific genes to general intelligence (Sternberg, 2012). However, this was likely due to a lack of statistical power, as more recent work using larger sample sizes (e.g., 78,000 individuals; Sniekers et al., 2017) has implicated a number of genes, the majority of which are expressed in brain tissue and neural cell development.

An example of a commonly used intelligence test is the Raven’s Progressive Matrices (RPM), arguably the one test that is most predictive of cognitive ability (Marshalek et al., 1983; Snow et al., 1984; Tucker-Drob and Salthouse, 2009; see Figure 1-1). The RPM presents a series of visual puzzles in which the individual is required to select from a range of options a single ‘tile’ that is missing from a larger pattern shown above. To solve these problems, one has to consider each element of the pattern (e.g., the dot configurations in Figure 1-1b) and decide how the elements relate to each other across the rows and columns to predict which tile fits in the blank space. A core component of such tests is relational reasoning: the ability to actively link and manipulate multiple mental representations (Duncan et al., 2000; Halford et al., 1998). Relational reasoning is thought to support many higher cognitive functions, including abstract thought, problem solving and decision-making (Crone et al., 2009; Halford et al., 2010).

Since the discovery of g (Spearman, 1904), several psychometric models of sub-types of intelligence have been proposed (Cattell, 1963; Johnson and Bouchard, 2005). One of the most influential of these models is Cattel’s (1963, 1949) distinction between fluid and crystallised intelligence. In this model, fluid intelligence refers to the capacity to solve problems without prior knowledge, experience, or skills. Much like the RPM (Figure 1-1), fluid intelligence is best measured with tests devoid of cultural or scholastic influence (Horn, 1998). By contrast, crystallised intelligence represents knowledge gained by education, culture and experience in general. There is good evidence for such a distinction between intelligence subtypes. For example, fluid and crystallised intelligence measures tend to show independent trajectories across the lifespan (e.g., Tucker-Drob and Salthouse, 2009). However, there are several other popular theories with supporting evidence that refute this two-factor distinction (e.g., the verbal, perceptual and image rotation model; Johnson and Bouchard, 2005). Thus, while there is no clear consensus regarding one particular substructure of intelligence, such models do not diminish the importance, or predictive power of g.

3

Figure 1-1 Example items from the Raven’s Progressive Matrices Test. In this test participants are asked to select the missing piece of the puzzle in the top panel, from the available options in the bottom panel. While the task is the same in panels a and b, the complexity of the two items is different. In the problem in panel a, because the pattern is the same across all elements (space) there is no parallel processing; thus the complexity is low. In the problem in panel b, however, the relationship between each different element is less straightforward, requiring multiple parallel relationships to be considered at once. Source: https://goo.gl/tCY3zK.

1.1.2 Complexity

A key component of intelligence testing is the differentiation of overall ability scores by incremental task complexity. Successful performance on more complex aspects of a task is associated with higher intelligence scores (Jenson, 1987). Individuals with low fluid intelligence tend to make errors once a task contains too many operations, or rules that need to be considered in a problem (e.g., Figure 1-1b, Bhandari and Duncan, 2014; Duncan et al., 2008). This finding has been formally described as goal neglect, where participants can report task requirements before or after a testing session, but simply cannot implement them during the task. This inability to cope with multi-part, complex problems has led to the argument that intelligence represents an individual’s ability to chunk complex problems into smaller, simpler operations, while keeping in mind the overarching goal (Duncan, 2010).

Operationalising what makes a task complex is not trivial, despite its apparent simplicity. Relational complexity theory (RC theory) is a domain-independent conceptual framework that can be used to quantify the level of reasoning complexity needed to complete a given task (Halford et al., 1998). RC theory posits that the number of relations between variables predicts the complexity of a problem, regardless of the domain of the original stimulus (e.g., semantic, spatial, temporal, etc.,). For example, the statement ‘Bill is taller than Ted’ contains two elements (‘Bill’ & ‘Ted’) and a single relation (‘is taller than’). On its own, a second similar statement embodies the same level of complexity (e.g.,

4 ‘Charles is taller than Bill’), but when used in conjunction with the first statement (e.g., ‘Who is taller, Charles or Ted?’) it requires parallel processing of both single statements and is therefore of a higher complexity.

Evidence for RC theory is well established. Increasing the number of relations in a reasoning problem imposes a quantifiable cognitive load (measured via reaction time or accuracy), and will eventually result in a breakdown of the reasoning process (Halford et al., 2005). For example, in the n-term task participants are given a list of relational premises to integrate (e.g., ‘T < P’, ‘P > V’ and ‘T < V’), and they are then asked to order each element (i.e., ‘T’,‘P’,’V’) from highest to lowest (see Figure 1-2, Andrews et al., 2006). To manipulate the degree of relational complexity the number of initial premises can be altered; two initial premises would indicate a Binary level problem, three would indicate a Ternary level problem, and so on. In this paradigm greater complexity problems are associated with increased reaction time and decreased accuracy (Andrews and Halford, 1998; Birney and Bowman, 2009; Maybery et al., 1986). This ‘complexity effect’ has been shown across multiple cognitive domains, including language, spatial reasoning and (Andrews et al., 2013, 2006; Andrews and Halford, 1998; Birney et al., 2006; Hansell et al., 2015; Maybery et al., 1986).

5

Figure 1-2 Example items from common relational reasoning tasks and their link to intelligence. a. Two examples from the n-term task. Participants are given a number of premises containing letters and relations (e.g., ‘T < P’). They then attempt to sequence cards representing each element (e.g., ‘T’) from highest to lowest, keeping in mind the initial premises. b. Two examples from the sentence comprehension task. Here participants are read aloud a sentence with n nouns to be assigned to thematic roles of the verbs. They are then asked a comprehension question pertaining to the sentence (e.g., ‘Who liked?’). The number of nouns contained within a sentence determines the complexity of a given problem. Even though here the premises are presented in a different domain (i.e., language comprehension) the same complexity effects can be observed such that problems with more relations are associated with lower accuracy and longer reaction times. Adapted from Andrews, Birney and Halford (2006). c. Participants completed a relational reasoning task at three different levels of complexity (Latin Square Task, accuracy for each condition on x-axis, N=99), as well as the Raven’s Advanced Progressive Matrices (APM, y-axis). The link between performance on the task and intelligence only emerges once the problems are complex enough to elicit individual differences. Data presented in panel c were collected for studies reported in Chapters 4 & 5 of this thesis.

6 Critically, performance decrements imposed by complexity demands are related to measures of intelligence, further establishing the link between relational complexity and general intelligence (Andrews and Todd, 2008; Birney and Bowman, 2009). This association is only apparent once a task’s complexity level is high enough to elicit individual differences. Thus, for example, Binary problems are typically too easy for healthy adults (Birney and Bowman, 2009, see Figure 1-2c). Instead, the critical complexity level that differentiates between healthy adults seems to involve four relations (i.e., Quaternary; Halford et al., 2005).

In sum, intelligence is both measurable and highly predictive of crucial life outcomes. The types of tasks that are most predictive of intelligence tend to rely heavily on the processing of relations in cognitive reasoning problems. Increasing the number of relations that must be processed in parallel increases the complexity of the task. Moreover, higher task complexity is associated with a decrement in performance, which is correlated with intelligence measures.

1.1.3 Investigating intelligence in the brain

Which brain regions subserve intelligence? Historically, conclusions about brain function were made through case studies of individuals who had suffered brain injury, or who had undergone invasive surgery that damaged specific brain structures. Behavioural measures were collected post-injury, and conclusions were drawn regarding how damage to the implicated brain region might have been responsible for a particular behavioural impairment. For example, Waltz and colleagues (1999) compared reasoning performance in individuals with prefrontal cortex (PFC) or temporal cortex damage, and healthy controls. They found that individuals who had PFC damage demonstrated a reduction in reasoning ability, compared with individuals with temporal cortex damage and healthy controls. Crucially, this deficit was selective for reasoning, as separate measures of episodic memory and semantic knowledge were not impaired in the PFC group when compared with healthy controls. Such findings, in conjunction with other research, have led to the conclusion that intelligence and reasoning ability are subserved largely by the PFC (Duncan et al., 1995; Luria, 1966; Waltz et al., 1999).

When attempting to understand brain-behaviour relationships, however, results from lesion studies need to be interpreted with caution. Lesion studies typically depend on the locality assumption, i.e., the idea that discrete brain regions subserve distinct cognitive functions, and that local lesions yield only local effects on cognition and behaviour. This approach largely ignores the fact that activity in brain regions that are anatomically distant from, but functionally connected with, the damaged site may also be altered. Moreover, the study of post-injury behaviour is further complicated by the complex nature of injuries themselves, making straightforward inferences difficult. For example, different types of injuries may have different effects on the brain (e.g., traumatic brain injury versus stroke), or surrounding brain tissue might be functionally affected by the injury but appear structurally

7 intact (Rorden and Karnath, 2004). Fortunately, many of these issues can be circumvented by the use of high spatial resolution, in vivo brain imaging methods in healthy individuals.

One in vivo imaging approach is brain morphometry – the measurement of brain structure. Using techniques such as magnetic resonance imaging (MRI), the thickness of an individual’s cerebral cortex can be estimated accurately (~ resolution of 0.5mm3 to 1mm3). Recent meta-analyses of studies that have employed this technique have demonstrated a small but significant correlation between overall brain volume and intelligence measures (McDaniel, 2005). In addition, the thickness of specific brain regions within frontal and temporal cortex, as well as subcortical structures, has been shown to correlate with intelligence scores across a number of investigations (Basten et al., 2015). However, the relationship between intelligence and cortical thickness is likely to be non-linear. For example, in early childhood the opposite relationship between cortical thickness and intelligence scores has been reported (Shaw et al., 2006). While informative, the above work tells us little about how neural activity, rather than structure, is organised to support intelligent behaviour. To investigate this question, different methodological techniques are required.

1.2 Assessing changes in human brain activity in vivo

Informative, non-invasive measurement of neural activity in humans is an ongoing challenge in cognitive neuroscience. Popular techniques used to investigate neural dynamics in humans include functional magnetic resonance imaging (fMRI), positron emission tomography (PET), magnetoencephalography (MEG) and electroencephalography (EEG). Since its inception in 1990 (Ogawa et al., 1990) fMRI has arguably become the workhorse of investigations in cognitive neuroscience, largely due to its high spatial resolution (Poldrack and Farah, 2015). A typical, modern fMRI study has a spatial resolution of approximately 2-3 mm3 and the potential to collect data samples every one to two seconds. Critically, fMRI relies on the blood-oxygen level dependent (BOLD) signal, a correlate of neural activity, rather than measuring neural events themselves (Logothetis et al., 2001; Ogawa et al., 1990). In the section below I outline the known physiological basis of the BOLD signal, as well as the most common analyses employed to investigate brain activity in response to cognitive tasks.

1.2.1 Physiological basis of fMRI

Measurement of the BOLD signal exploits the magnetic properties of iron contained within deoxyhaemoglobin. Subtle changes in the ratio between oxygenated and deoxygenated blood can be detected and measured within a large magnetic field (i.e., an MR scanner). As neural activity increases in response to a stimulus, oxygen extraction from blood increases, resulting in an increase in deoxyhaemoglobin. Following this, within seconds, neurovascular signals increase blood flow causing an over-oxygenation of the tissue (thus decreasing deoxyhaemoglobin). This effect is slow, arising two to five seconds post-event, depending on the nature of the stimulus (Hillman, 2014). It is this

8 ‘overshoot’ of oxygenated blood that is measured with the BOLD signal (see Figure 1-3).

Figure 1-3 Components of the blood-oxygen level dependent (BOLD) response. After stimulus-related (blue section) neural activity there is a marked increase in blood flow which causes an increase in the ratio of oxygenated (black line) to deoxygenated blood (grey line). Such changes are relatively slow compared with the initial neural activity. Adapted from Malonek and Grinvald (1996).

The relationship between actual neuronal events and changes in cerebral blood flow is referred to as neurovascular coupling. This coupling, though initially hypothesized over 100 years ago (Roy and Sherrington, 1890), is complicated, and the exact biological mechanisms remain poorly understood (Hillman, 2014; Howarth, 2014). Broadly speaking, the BOLD signal has been linked to both synaptic activity (e.g., local field potentials) and action potentials (‘spikes’). However, these two measures are often correlated with one another. When they are separable, it is thought that cerebral blood flow likely corresponds better to local field potentials than to spikes (Thomsen et al., 2004). Thus, the BOLD signal is largely interpreted as a measure of local neuronal inputs averaged across the activity contained within a voxel (i.e., each three-dimensional “pixel” of data, Logothetis et al., 2001; Logothetis and Wandell, 2004).

An important caveat is that the BOLD response is explicitly tied to the vascularity of the brain. As such, observed differences in BOLD, as well as the spatial location of such differences, can arise from differences in the vascular anatomy across brain regions (Ances et al., 2009; Turner, 2002). For example, it is known that various disease states, aging and medication can all alter the relationship between the BOLD signal and actual neural activity (D’Esposito et al., 2003).

1.2.2 Univariate analysis of the BOLD signal

The aim of general linear modelling (GLM) in fMRI experiments is to analyse each voxel to determine whether the BOLD signal has demonstrated a statistically significant change due to an

9 experimental manipulation. With the widespread availability of fMRI, GLM became the principal tool for investigating brain function (Poldrack and Farah, 2015). There are various reasons for this, including free and widely available statistical analysis packages (e.g., statistical parametric mapping [SPM], Friston et al., 1994) and the relative simplicity and flexibility of analysis, as well as the intuitive inferences that can be made from a significant result (e.g., whether a given brain region has a higher BOLD signal, or ‘level of activity’, in one experimental condition compared with another).

In a typical fMRI experiment, several important pre-processing steps need to be undertaken prior to statistical analysis. Broadly speaking, these steps are taken to ensure meaningful comparisons across individual differences in brain anatomy, and to remove ‘nuisance’ variables from the data. Although a detailed review of these steps is beyond the scope of this thesis, it is important to emphasise that the BOLD signal is heavily processed before it can be used to infer brain function (Friston et al., 1994; Smith et al., 2004). An example pre-processing pipeline might include head motion correction, cerebrospinal fluid signal regression, spatial normalisation, high- and low-pass temporal filtering, and spatial smoothing (Chao-Gan and Yu-Feng, 2010). Critically, different statistical analyses may require altered pre-processing pipelines to emphasise various aspects of the BOLD signal (e.g., resting-state BOLD analysis usually involves aggressive head movement regression and a low frequency bandpass filter).

The GLM approach attempts to explain variance associated with the BOLD signal over time in terms of a linear combination of predictor variables. Predictors correspond to a time course of the expected BOLD response given the conditions of the experiment. Modelling a predicted BOLD response in this way is not trivial, and there are various methods that are used to achieve the desired result (Penny et al., 2011). Additional predictors can also be entered into the model, depending on the experiment. These predictors could represent variables of interest, like reaction times, or perhaps ‘noise’ variables like head motion. A statistical test is then performed to ‘find’ voxels that fit the model containing all the predictors, measured by a test statistic. This same general approach can be extrapolated to second level analyses, where the statistical test will now identify voxels that satisfy the tested hypothesis, for example separating two groups of participants (e.g., a between-subjects t-test) or differences across several experimental conditions (e.g., ANOVA), and so on.

As a concrete example, consider an fMRI experiment in which participants are instructed to tap their index finger for 20 seconds, followed by rest for 20 seconds, repeated several times. We can expect that the BOLD response in the contralateral primary motor cortex (M1) will be higher in the ‘tapping’ epochs versus the ‘rest’ epochs. To test this prediction, the expected BOLD response is modelled, and a t-test is performed to examine how well the model fits the obtained BOLD signal in each voxel. A statistical map is then generated and carried to a second-level analysis, where the hypothesis of greater activity in contralateral M1 during tapping than rest is tested statistically.

10 A critical problem that arises with such analyses is how to correct for the large number of statistical comparisons across brain voxels. A brain volume, depending on spatial resolution and other factors, might contain anywhere from 40,000 – 900,000 individual voxels. Consequently, using a standard Bonferroni approach in which a correction is performed for every possible comparison is likely too conservative. Instead, cluster-level inferences are considered acceptable for standard fMRI analyses. This approach takes advantage of the assumption that ‘true’ significant voxels will be likely to be spatially localised in clusters. Such clustering occurs because brain regions associated with a meaningful BOLD response are likely to encompass more than a single voxel. By taking advantage of these elements, appropriately thresholded, cluster-based correction has been shown to be a suitable way to isolate areas of significant brain activity in statistical maps (Eklund et al., 2016).

1.3 Brain activity related to intelligence and reasoning behaviour

Using standard univariate fMRI, an extensive effort has been made to ‘map’ which brain regions are associated with reasoning behaviour and intelligence. In a typical experiment, participants undertake cognitive tasks thought to involve the processing of complex relations (e.g., the Raven’s Progressive Matrices task) while in the MR scanner. Expected BOLD patterns are modelled to find regions of the brain that show reliable activity in response to ‘more complex’ conditions versus ‘less complex’ conditions.

1.3.1 From local to global models of brain activation associated with reasoning complexity

The processing of increasingly complex relations in cognitive reasoning tasks has previously been associated with discrete prefrontal cortical regions. Evidence for this notion comes from brain lesion experiments, in which damage to the PFC is associated with decrements in reasoning performance (e.g., Duncan et al., 1995; Waltz et al., 1999). Early fMRI studies largely confirmed this view by demonstrating that increased lateral and rostral PFC activation was associated with increased relational complexity demand across both visuospatial (Christoff et al., 2001; Kroger, 2002; Prabhakaran et al., 1997) and semantic tasks (Wendelken et al., 2008).

More recently, however, the existence of a localised brain region critical for reasoning behaviour has been disputed. From several meta-analytical and data-driven analyses in large datasets a general consensus has emerged that a fronto-parietal set of brain regions activates in response to complex goal-oriented tasks, regardless of the cognitive domain (Duncan, 2010; Duncan and Owen, 2000; Fedorenko et al., 2013). Such regions include the dorsolateral prefrontal cortex (DLPFC), inferior frontal junction (IFJ), intraparietal sulcus (IPS), anterior cingulate cortex (ACC) and anterior insula (AI). This collective of brain regions has been termed the ‘multiple-demand system’ (Duncan, 2010), the ‘cognitive control network’ (Cole and Schneider, 2007), and the ‘task-positive network’ (Fox et al., 2005) depending on the research group penning the paper (see Figure 1-4). Hereafter, I will refer to this collection of fronto-parietal regions as the ‘cognitive control network’ (CCN).

11 Several animal studies have also cited the importance of fronto-parietal cortical regions for complex task performance (Asaad et al., 2000; Duncan, 2001; Mitchell et al., 2016; Procyk et al., 2000; Rao et al., 1997). For example, single cell recordings in the macaque during the execution of complex tasks have revealed that a wide range of task features, including task rules, cues and stimuli features can be coded by the same set of cells (Duncan, 2001; Ibos et al., 2013). Moreover, this coding is dynamic, such that neurons are reallocated to the most important task at hand, at the timescale of single trials (Kadohisa et al., 2013; Stokes et al., 2013). While it is not possible to make the same claims based on fMRI experiments in humans, due to the limited spatial and temporal resolution of fMRI, recent work using multivariate pattern analysis techniques has demonstrated that fronto-parietal cortex can indeed discriminate between a wide variety of stimuli, rules and contexts, mirroring the findings from single- cell recordings (e.g., Cole et al., 2011; Li et al., 2007; Woolgar et al., 2015).

A key property of the CCN is that the extent of activity within this network scales with complexity, such that more complex tasks tend to evoke larger BOLD responses (e.g., Gould et al., 2003). Moreover, this is mirrored by a similar ‘deactivation’ effect in brain regions commonly referred to collectively as the default-mode network (DMN), or task-negative system (Gilbert et al., 2012; Shulman et al., 1997). This set of regions includes the medial prefrontal cortex (MPFC), posterior cingulate cortex (PCC), angular gyri, precuneus and temporal cortex (Figure 1.4c, Buckner et al., 2009). It is thought that the functional antagonism between these two cortical systems (c.f. Figures 1- 4b & 1-4d) is important for healthy brain function and successful task performance (Anticevic et al., 2012).

Critically, activity within the CCN has been linked to intelligence, formalised in the parieto-frontal integration theory of intelligence (P-FIT; Jung and Haier, 2007). This theory postulates that interactions between frontal and parietal cortices underpin individual differences in reasoning ability in humans. The broad nature of this hypothesis has left many unanswered questions (Vakhtin et al., 2014). Moreover, it was developed prior to modern network definitions and thus doesn’t explicitly map onto known functional networks like the CCN.

Nevertheless, functional evidence supporting modern interpretations of this model comes from studies that have shown that variability in activation within the CCN correlates with (e.g., Gray et al., 2003), or is mediated by (e.g., Tschentscher et al., 2017), intelligence scores, such that higher intelligence is associated with increased CCN activity. In addition, DMN activity during complex tasks is also associated with intelligence, such that smaller reductions are associated with higher intelligence (Basten et al., 2013). While relatively unexplored, this result potentially highlights the delicate balance between CCN and DMN activity profiles in the service of higher cognitive reasoning performance.

12

Figure 1-4 ‘Cognitive Control’ and ‘Default-mode’ networks associated with increased task demands. Components of the cognitive control and default-mode networks are visualised in panels a & c, respectively. Panels b & d show a schematic of the typical pattern of activity within the networks during cognitively challenging, goal-directed tasks. Activity in the cognitive control network tends to increase with increasing demands, whereas activity in the default-mode network tends to decrease. Pre-SMA= pre-supplementary motor area, IPS = intraparietal sulcus, IFG = inferior frontal gyrus, ACC = anterior cingulate cortex, AI = anterior insula, MPFC = medial prefrontal cortex, PCC = posterior cingulate cortex, PC = precuneus, AG = angular gyrus, TC = temporal cortex. Data visualised from thirty participant datasets collected for Chapter 5 of this thesis.

To summarise, a fronto-parietal ‘cognitive control’ network of cortical areas is thought to respond to task complexity, regardless of the cognitive domain (reasoning, attention, working memory and so on). Physiological evidence suggests that cells within these regions have adaptive coding properties; that is, the same neurons (or populations of neurons) respond flexibly to many different types of stimuli and task rules. BOLD activity within the CCN increases as task complexity increases. The CCN is coupled with a separate, ‘default-mode’ network whose activity decreases as complexity increases. Both networks are thought to be associated with measures of reasoning.

While informative, the aforementioned literature (and ‘brain-mapping’ in general) largely focuses on

13 isolating regions specialised for a particular function, by way of examining discrete regional activation amplitudes (Friston et al., 1994; Poldrack and Farah, 2015). This approach ignores the fact that brain regions will likely need to work together in concert to achieve complex behavioural outcomes. In fact, a key prediction of the P-FIT model – that neural interactions between frontal and parietal cortices predicts intelligence – cannot be tested by analysing univariate changes in the BOLD signal. Thus, to investigate the interaction between distinct brain regions we turn to brain network analysis and graph theory.

1.4 From functional specialisation to brain networks

As described above, with the advent and availability of high-resolution brain mapping tools such as fMRI, a great deal of research has focused on ascribing specific cognitive functions to particular brain regions. More recently, with exponential increases in computing power for data analysis, the development of new analytical tools (and their widespread availability), and conceptual influences from cellular, molecular and computational neuroscience, the discipline of cognitive neuroscience has undergone something of a paradigm shift in the way it deals with brain-behaviour relationships. Specifically, rather than simply endeavouring to ascribe cognitive processes to individual brain regions, there has been a shift to thinking about cognitive processes as being subserved by complex, interconnected and highly dynamic functional networks. This change in outlook is analogous to that which has occurred in animal physiology research, which has undergone a shift from investigating activity patterns of individual spiking neurons to analysing complex spatiotemporal patterns of synchronisation and desynchronization across larger neuronal populations (Kumar et al., 2010; Mill et al., 2017). Thanks to these technical and conceptual advances, it is clear that the computational power of the brain is more than the sum of its segregated parts, and instead reflects the dynamic integration of specialised processes by widespread neural networks (Sporns and Zwi, 2004).

The fundamental building blocks of brain network organisation involve the concepts of segregation and integration (Friston, 1994; Tononi et al., 1994). Segregation refers to the specificity and modularity of brain organisation. Evidence for segregation of brain function can be found at every scale of investigation (Sporns, 2016), from single neurons that respond selectively to motion (e.g., Albright, 1984), to groups of neurons that form local processing units within sensory modalities (e.g., vision, audition and so on, Chiang et al., 2011), to macroscopic brain regions measured with fMRI that respond selectively to human faces (e.g., Kanwisher et al., 1997). By contrast, integration refers to how such specialised processing modules exchange information to produce perception and behaviour. In the brain, segregation and integration can be thought of as an intricate balance of connections within and between distinct processing units (see Figure 1-5). Network neuroscience aims to understand how these fundamental organisational principles can delineate brain networks responsible for perception, cognition and motor behaviour.

14

Figure 1-5 Network segregation and integration. a. Two network graphs are visualised. Coloured circles represent nodes and grey lines represent connections. Less distance between nodes represents stronger connections. The colour of the nodes denotes network affiliation. One graph is more segregated (left), while the other is more integrated (right). b. Simulation of networks with increasing probability of between-module (x- axis) and within-module (y-axis) connections (Guimera et al., 2005). Colours denote the modularity index of each graph, which represents the density of within-module connections relative to an appropriate null network. The modularity index was used to quantify levels of segregation and integration (exemplars visualized in panel a). Figure reproduced from Shine and Poldrack (2017).

Transfer of information between brain regions is supported by both anatomy and neural dynamics. Structural connectivity refers to the measurement of physical connections between neurons and neuronal populations (Friston, 1994). In-vivo whole brain structural networks are typically derived using diffusion , an MRI technique that leverages the properties of water molecules within axons to estimate white matter connections in the brain (Melhem et al., 2002). By contrast, functional connectivity is defined as a statistical dependency among remote neurophysiological events

15 (Friston, 2011). In fMRI data, functional connectivity is usually represented by a correlation in the temporal domain, but there are also complex bio-statistical methods for inferring the influence of one region on another. Put simply, when two regions of the brain are highly correlated in their activity, mutual information is implied (Friston, 1994). Critically, the exact information conveyed by functional connectivity measures will depend on the imaging modality used. Below I outline BOLD- based functional connectivity in detail (Section 1.4.2).

1.4.1 Analysing functional brain networks

Considering the brain as a set of distinct regions (or nodes in a network) with functional connections between them (edges in the network) allows the construction of macroscale wiring diagrams, or ‘’. Brain regions of interest are defined depending on the experimental question, ranging from voxel-wise analyses (i.e., each “pixel” of fMRI data is considered a brain region, Biswal et al., 1995) to predefined brain templates (Power et al., 2011) or a priori brain regions identified by the particular research question (Cocchi et al., 2014a). Most commonly, neural signals from such brain regions are then correlated to form a correlation matrix containing information about the extent of statistical dependency (i.e., a Pearson’s r value) of each brain region relative to every other brain region (Zalesky et al., 2012b). Other mathematical variations on this basic premise exist, but essentially all functional connectivity methods provide information about the extent of statistical dependence between distant brain regions (Smith et al., 2011).

Correlation matrices, or connectomes, can be further interrogated in a number of ways. Much like other brain imaging data, connectomes can be compared across cognitive states, between participant groups (e.g., patients versus healthy controls), or with respect to a given behavioural measure (e.g., intelligence scores). The insights that can be gained from connectomic analysis can be considered within three broad categories: spatial topology, directionality and temporal dynamics (Mill et al., 2017). Topology is perhaps the richest domain, and refers to the spatial configuration of networks. Functional connectivity – reflected by increased correlations – is taken as a measure of integration. Decreased functional connectivity is taken as a measure of segregation. Likewise, graph theory, a branch of mathematics dedicated to investigating wiring diagrams, can be used to extract topological properties of networks (see Figure 1-6). Examples of such properties might include: hubs – nodes that take on a higher burden of communication than would be expected by chance; modules – the grouping of nodes within a network into discrete communities balanced by their connections; and global efficiency – the ease with which a network can communicate across all regions (Rubinov and Sporns, 2010).

Two other broad insights can be garnered from functional connectomes. Directionality aims to characterise how one brain region influences another (so called ‘effective connectivity’). To make claims about directionality of influence is not trivial, and complex data modelling is needed (e.g.,

16 Dynamic Causal Modelling, Friston et al., 2003). This approach provides more information than a simple correlation; consider a result that suggests greater functional connectivity between the motor cortex and thalamus during finger tapping versus a resting condition. Directionality can tell us that connectivity from the thalamus to the motor cortex increases, but not vice versa. Though more informative than functional connectivity, the drawback of effective connectivity is that computation time confines the size of the resulting directed graphs (Razi et al., 2017). Second, temporal dynamics refers to changes in network activity over prolonged timescales, from several minutes to hours, days or even years (Poldrack et al., 2015; Zalesky et al., 2014). These analyses aim to characterize how functional networks develop over time, which is most commonly achieved by comparing multiple datasets across several time points (e.g., Bassett et al., 2011). Recently, this approach has been undertaken ‘within-scan’ by splitting a typical scanning session into smaller epochs to identify fluctuations in connectivity at the timescale of seconds or minutes (Shine et al., 2016; Zalesky et al., 2014), but there is ongoing debate about the neural underpinnings of such dynamics (Laumann et al., 2017).

Figure 1-6 General pipeline for fMRI functional connectivity analyses. a. The BOLD signal is extracted from brain regions of interest (orange = I; blue = J; green = K). b. Correlation matrices represent the statistical dependence (i.e., Pearson’s r) between each brain region in the network. c. The correlation matrix transposed into graph form. d. The graph can be investigated for various topological graph-theoretical metrics, such as modules and hubs. Adapted from Rubinov and Sporns (2011).

Much like GLM analyses, statistical analysis of correlation matrices poses a large multiple

17 comparisons problem. Consider that a network of 300 nodes would include 44,850 unique !"" × %&& connections (i.e., ). Recently developed tools like the Network Based Statistic (NBS, Zalesky % et al., 2010) and Spatial Pairwise Clustering (Zalesky et al., 2012) have become widely available, allowing for appropriate control of type I error while limiting the likelihood of type II error.

The NBS is used repeatedly throughout this thesis. Briefly, to assess statistical significance an independent univariate test statistic is computed for each connection in the connectivity matrix according to the hypothesis being tested. Then, an initial threshold is applied to this matrix of statistical values. This threshold is based on knowledge regarding the expected effect size (for example t = 3.5) and removes possible weak links from the matrix. Then, a breadth-first search is conducted to identify the number and size of subnetworks in which at least two nodes are connected by an edge (i.e., connected components). The size of the resulting connected component(s) can be described in terms of the number of edges (the topological extent), or the sum of the t-statistic of each element in the component (the effect intensity).

Permutation testing is then used to assess the significance of the observed component(s). To do so, labels assigned to each of the original matrices are shuffled randomly. The initial NBS analysis (as above) is repeated on this newly generated data and the size of the largest component is stored. This procedure is repeated many times (e.g., 5000) generating a null distribution of maximum component sizes. The ‘real’ observed connected component size can then be compared with this null distribution and its significance can be determined depending on where it lies within the distribution. While the NBS provides a sizeable gain in statistical power relative to available statistical methods, it is critical to note that the null hypothesis is rejected at the level of components, and not individual connections. Thus, it is incorrect to make inferences about specific connections within a component relating to the hypotheses that were tested (Fornito et al., 2015; Zalesky et al., 2010).

1.4.2 BOLD-based functional connectivity

Work on fMRI-based functional connectivity networks in humans has been largely pursued in task- free, so-called ‘resting-state’ experiments. In one seminal study, Biswal and colleagues (1995) collected fMRI data from participants while they lay in the MR scanner and simply rested rather than performing a task. Participants were instructed to ‘refrain from any cognitive, language or motor tasks as much as possible’. By simply comparing slow fluctuations (< 0.1 Hz) in activity within a region of the sensorimotor cortex with activity elsewhere in the brain, Biswal et al. observed correlations with distant but functionally related regions such as the contralateral motor cortex. Furthermore, this functional connectivity map derived from resting-state data was strikingly similar to activations produced by a simple finger-tapping task. In other words, correlations between BOLD signals in different regions at rest seemed to arise in a network-specific manner that was very similar to those observed during task activation, at least within the motor system.

18 Biswal’s discovery was initially met with indifference in the literature (Jonathan D. Power et al., 2014), with some investigators suggesting that spontaneous BOLD signal correlations at rest merely reflected noise (e.g., Fransson et al., 1999), but over the next decade the work of Biswal et al. was replicated and extended across a large number of resting-state network mapping studies. For example, resting-state recapitulations of well-known attention (Fox et al., 2006), executive control (Dosenbach et al., 2006; Seeley et al., 2007) and default-mode networks have been reported (Greicius et al., 2003). This line of research eventually led to whole-brain functional network ‘connectome’ mapping studies (e.g., Hagmann et al., 2008), and descriptions of the relationship between these networks and behaviour (e.g., van den Heuvel et al., 2009). Moreover, it has now been established that resting functional connectivity correlates with patterns of anatomical connectivity (Honey et al., 2009), is stable within and between individuals (Damoiseaux et al., 2006; Deuker et al., 2009; Zuo et al., 2010), is consistent across different scanners and imaging parameters (Shehzad et al., 2009; Van Dijk et al., 2010), holds across different brain states (e.g., during light anaesthesia; Vincent et al., 2007) and can be found using different imaging modalities (e.g., MEG, de Pasquale et al., 2010; EEG, Mantini et al., 2007). Thus, despite initial scepticism, resting-state networks have become accepted as a likely ‘intrinsic’ functional wiring map of the brain.

It is important to consider how statistical dependencies in the BOLD signal between brain regions should be interpreted in terms of underlying neural communication. Recall from Section 1.2.1 that the BOLD signal likely represents the sum of local field potentials, although there are undoubtedly other neural and non-neural contributions to the signal (Logothetis et al., 2001). A number of studies have now demonstrated, through simultaneous intracortical recordings and fMRI, that a reliable relationship exists between fMRI-based functional connectivity metrics and underlying dependencies in locally measured neural activity (Keller et al., 2013; Scholvinck et al., 2013; Shmuel and Leopold, 2008). This relationship has also been observed at the level of macroscale organisation: slow cortical potentials across entire networks demonstrate similar correlations to those observed in the BOLD signal (He et al., 2008).

1.4.3 Large-scale functional networks of the brain

A consensus has emerged that the cortex can be subdivided into robust and reproducible macro-scale functional networks (Figure 1-7, Dosenbach et al., 2010; Gordon et al., 2016; Power et al., 2011; Yeo et al., 2011). Canonical networks include the default-mode, fronto-parietal, cingulo-opercular, attention, visual, somatosensory, auditory and subcortical networks. Confusingly, some networks are labelled anatomically whereas others are based on proposed functional roles. The mapping of functional networks has now become a field of research in its own right, and it seems likely that future network parcellations will become more precise and of higher resolution. [Consider, for example, that the Dosenbach et al. (2010) parcellation contains 160 regions, the Power et al. (2011) parcellation contains 264 regions, and the Craddock et al. (2016) parcellation contains 333 regions.]

19 Importantly, as mentioned in the previous section, functional networks derived at rest are spatially similar to activation patterns observed during task-based imaging experiments. As such, both the CCN and the DMN have analogous, resting-state networks (as shown in Figure 1-7), and seem to reflect an important and intrinsic functional organisation architecture (Power et al., 2011). Recall that in complex task activation studies, the CCN and DMN tend to show an opposing activation profile (see Figure 1-4). Interestingly, they also show a similar opposing relationship at rest, such that the two networks tend to be highly negatively correlated (or anticorrelated, Fox et al., 2005).

Figure 1-7 Canonical resting-state networks of the human cortex. Resting-state networks derived from over 300 participants using graph analyses. Several canonical networks arise during the resting-state, including (but not limited to) default-mode (red), fronto-parietal (yellow), cingulo-opercular (purple & black), attention (green and teal), subcortical (brown), somatosensory (light blue), and visual (dark blue). Figure adapted from Power et al. (2011).

Furthermore, evidence from resting-state network mapping suggests that task-evoked CCN activations (shown in Figure 1-4) can be subdivided into meaningful cingulo-opercular (CON) and fronto- parietal network (FPN) subcomponents (Figure 1-7, shown in purple/black and yellow colours, respectively; Dosenbach et al., 2007, 2006). These subnetworks are anatomically distinct and are postulated to be functionally distinct as well; the CON is thought to support sustained attention during task performance, whereas the FPN is thought to support trial-by-trial cognitive control (Dosenbach et al., 2006). In the following section I will describe some of the research that has tied these functional networks to reasoning complexity and intelligence.

1.4.4 The relationship between large-scale functional and structural networks

The organization of functional networks is thought to reflect, in part, the underlying structural

20 connectivity of the brain (Hagmann et al., 2008; Passingham et al., 2002). This notion is intuitive; functional interactions in the brain must be constrained by the anatomical scaffolding to some degree (Honey et al., 2009). However, recall that functional connectivity is a measure of statistical dependence. Thus, functional connectivity could also be supported by many indirect anatomical projections. Moreover, while reliable (Damoiseaux et al., 2006; Deuker et al., 2009), functional connectivity is known to exhibit temporal dynamics (Zalesky et al., 2014), as well as reconfigurations in response to task demands (Cocchi et al., 2014a). Such properties of functional networks render simple ‘one-to-one’ structure-function mapping impractical (Mišić et al., 2015).

The complex relationship between structure and function is highlighted by disorders characterized by disruptions of white matter fibre tracts (Cocchi et al., 2014b; Lynall et al., 2010; Zalesky et al., 2011). For example, callosal dysgensis is a disorder in which individuals are born without, or with a partially formed, corpus callosum (the major interhemispheric white matter tract in the brain). Despite this anatomical abnormality, several studies have observed intact bilateral resting-state networks (Owen et al., 2013; Tyszka et al., 2011), suggesting that functional networks can exist with severely altered structural underpinnings.

1.5 Functional networks related to reasoning and intelligence

Previous investigations of the brain basis of complex reasoning ability have focused on the contributions of discrete brain regions (e.g., Waltz et al., 1999). With recent advances in imaging methodology, however, it has been proposed that reasoning ability and intelligence, as complex higher cognitive abilities, are likely an emergent property of widespread neural interactions (Cole et al., 2012; Duncan, 2010). In contrast to earlier univariate GLM analyses (Bunge et al., 2009), network analysis represents a novel, potentially fruitful approach to investigate how communication between separate brain regions might give rise to complex behaviour (Sporns, 2013).

To date there have been two distinct lines of research relating functional networks to reasoning abilities. The first line of research has endeavoured to isolate task-evoked connectivity changes associated with processing of increasingly complex relations (Cocchi et al., 2014a; Parkin et al., 2015). This approach is similar to that taken in classic activation studies in which the BOLD signal associated with a low complexity reasoning condition is ‘subtracted’ from that evoked by a more complex reasoning condition. In network-based approaches, however, the experimenter is interested in widespread network changes, rather than local differences in BOLD signal amplitude. A second line of research has taken an individual-differences approach, in which network properties (most commonly derived from resting-state data), such as functional connectivity strength, or higher-order graph theoretical descriptions, have been related to individual differences in reasoning ability (Song et al., 2008; van den Heuvel et al., 2009). This approach has benefitted from the relative ease of acquiring resting-state MR data and behavioural measures outside the scanner, resulting in large

21 (statistically well-powered) sample sizes.

1.5.1 Task-evoked changes in functional networks related to the processing of relational complexity

To capture functional brain network properties that support the processing of increasingly complex relations, it is necessary to study brain dynamics during systematically designed cognitive tasks. This line of research builds upon previous literature which utilised mass univariate statistical methods to identify the ‘cognitive control’ network involved in complex cognitive tasks (see Section 1.3.1).

In a key study by Cocchi and colleagues (2013), participants performed a modified version of the classic Wason Selection Task (WST; Wason, 1968) while undergoing MR scanning. In the traditional WST, participants are shown four cards placed on a table. Each card has a letter on one side and a number on the opposite side. For example, the four cards might show ‘A’,’7’,’C’ and ‘5’. Participants are then given a logical proposition, such as ‘if there is an A on one side, the opposite side is a 7’. They are then asked ‘which card (or cards) must be turned over to test the truth of this proposition?’ Despite the apparent simplicity of the task, most people tend to make consistent errors of logic by choosing to verify the rule, rather than falsify it (Wason and Shapiro, 1971). Moreover, errors on the WST correlate with measures of general cognitive ability and other reasoning measures (Stanovich and West, 1998). To adapt the WST paradigm for the MR scanner, each trial in the experiment of Cocchi et al. (2013) involved the presentation of a logical rule followed by a single argument. The task was to ascertain whether the argument could disconfirm the previously established rule (for further details, see Chapter 3 of this thesis). By varying the rules and subsequent arguments, five levels of relational complexity were tested. As expected, participants showed a clear complexity effect, such that performance was reduced in terms of accuracy and reaction time as the complexity of rule implementation increased.

In the imaging data, functional connectivity was shown to increase in line with reasoning complexity within the FPN, as well as between the FPN and CON (Cocchi et al., 2014a). In addition, several nodes in subcortical areas and showed increased functional connectivity with the FPN and CON. These transient changes in connectivity were interpreted as forming a task-relevant ‘meta- system’ (see Figure 1-8). Finally, the left dorsolateral and right rostrolateral PFC were highlighted as key network nodes within the identified ‘reasoning’ network. Both brain regions demonstrated a relatively high ‘degree’ (i.e., number of significant connections) compared with other nodes in the network.

Following from the work of Cocchi et al. (Cocchi et al., 2014a, 2013) and others (Dosenbach et al., 2006), further research has sought to clarify the roles of the FPN and CON during complex tasks. For example, using multivariate pattern analysis Crittenden et al. (2016) found that the FPN was superior to the CON, in terms of decoding accuracy, at discriminating between multiple, complex task-rule conditions. This result was interpreted as evidence that the FPN is involved in moment-to-moment

22 ‘task-implementation’, whereas the CON is responsible for ‘task-set’ and maintenance of overall goals (Seeley et al., 2007).

In a similar vein, recent work by Parkin and colleagues (2015) incorporated a spatial reasoning task with high and low relational complexity. Using independent component analysis, they validated the observation of separable FPN and CON systems during task performance. In addition, they used dynamic causal modelling to investigate directional network models that best fitted the obtained data. They found that global changes in functional connectivity during task performance were best described by a change in top-down connectivity from the rostrolateral PFC, via the inferior frontal gyrus, to the anterior insula.

Figure 1-8 Schematic of results and review by Cocchi and colleagues (Cocchi et al., 2014a, 2013). a. Nodes affiliated with fronto-parietal (yellow), cingulo-opercular (purple), and visual networks (blue) rendered on a glass brain. Coordinates from Power et al., (2011). b. Graph representing the networks during low demand. Here circles (nodes) represent brain regions affiliated with fronto-parietal, cingulo-opercular and visual networks, and lines (edges) represent functional connectivity. At rest and during low complexity reasoning conditions the networks of the brain are distinct and separable. c. Graph representing networks during increased reasoning complexity. More complex conditions were associated with transient connectivity changes between networks (red edges) that formed a ‘meta-system’ of networks relevant to the task (dashed box).

In summary, there is preliminary evidence that increased relational reasoning complexity is supported by interactions within and between fronto-parietal and cingulo-opercular networks. These observations have led to the idea of a ‘meta-system’ that transiently forms to support complex cognition (see Figure 1-8, Cocchi et al., 2013). In addition, regions such as the lateral prefrontal

23 cortex, which have been implicated in previous brain lesion and univariate imaging analyses (e.g., Bunge et al., 2009; Waltz et al., 1999), play a key role in these network dynamics (Cocchi et al., 2014a), possibly influencing other brain regions in a top-down manner (Parkin et al., 2015). While this research area is still in its infancy, by taking into account both global dynamics (e.g., the meta- system) and local properties (e.g., the LPFC), this work has started to provide a way of reconciling global and local perspectives on the brain networks underlying reasoning complexity. Important questions for the future include how other ‘higher order’ networks, such as the default-mode network, might contribute to the processing of complex relations, as well as how exploring the network effects of meta-system transitions influence or predict behavioural outcomes.

1.5.2. Functional networks related to individual differences in intelligence and reasoning ability

A parallel line of research has investigated functional network properties associated with individual differences in reasoning ability and intelligence. Considering previous empirical brain activation research (e.g., Basten et al., 2013; Gray et al., 2003), as well as theoretical accounts (recall the parieto-frontal integration theory of intelligence; Jung and Haier, 2007), the clear next question has been to ask whether fronto-parietal connections predict individual differences in intelligence. Indeed, an early study by Song and colleagues (2008) found that functional connectivity values within the frontal lobe, and between frontal and parietal cortex, predicted intelligence scores. More recent work has highlighted that regions within the FPN tend show increased between-network connectivity compared with other regions of the brain. That is to say, despite ‘belonging’ to the FPN, the lateral PFC also shows high connectivity with other large-scale networks of the brain (Cole et al., 2012; Michael W Cole et al., 2015). Moreover, this global connectivity property correlates with measures of fluid intelligence (Michael W Cole et al., 2015). A similar pattern of results has been demonstrated across the whole FPN, such that the variability in connectivity within the FPN is predictive of intelligence, as well as a host of other behavioural measures (Finn et al., 2015; Smith et al., 2015).

Other attempts to isolate network-intelligence relationships, however, have provided less conclusive evidence for the importance of the FPN. As a striking example, in a follow-up study using the same data reported above (Song et al., 2008), Song and colleagues (2009) found that several connections within the default-mode network also correlated with intelligence, suggesting that the existing ‘FPN- centric’ literature may reflect biased analysis decisions, such as region of interest choice. In line with this, a number of studies have related individual differences in intelligence to resting-state connectivity of the default-mode network and cingulo-opercular network (Pamplona et al., 2015; Santarnecchi et al., 2015; Wang et al., 2011; Yuan et al., 2012). While the CON is often seen as a critical counterpart to the FPN (Seeley et al., 2007), the DMN is considered antagonistic to it (Anticevic et al., 2012).

Outside of specific functional networks, global network properties have also been related to

24 intelligence and reasoning measures, the foremost being global efficiency. As mentioned previously, the global efficiency metric is the inverse of the shortest path length in a graph, and is thus thought to be a marker of neural efficiency (Achard and Bullmore, 2007). Speaking to this directly, van den Heuvel and colleagues (2009) found a strong positive relationship between functional network global efficiency and intelligence. This finding has been replicated for different behavioural measures of intelligence (e.g., Pamplona et al., 2015), as well as across imaging techniques (Langer et al., 2011). Moreover, a recent study by Schultz and colleagues (2016) demonstrated that higher general intelligence scores are associated with a smaller change in network efficiency between rest, reasoning, working memory and language task states. The authors interpreted this finding as suggesting that more intelligent individuals need fewer ‘network updates’ to perform complex tasks than less intelligent individuals. Taken together, these findings suggest that more intelligent individuals possess functional networks that have an increased potential for information transfer.

In conclusion, the investigation of individual differences is broadly consistent with the P-FIT model of intelligence. Integration within the FPN, as well as whole brain communication, is related to individual differences in intelligence measured via complex reasoning tasks. The resting-state literature also highlights the role of the DMN and CON, however, as well as whole brain efficiency metrics. To clarify this discrepancy in the literature, in Chapter 2 I undertook a literature review and empirical analysis of data from a large cohort of individual resting-state network studies.

1.5.3 ‘Dysconnectivity’ and reasoning

A relatively recent line of research has implemented network analysis tools in cohorts of individuals with various brain abnormalities or injuries. This line of research has investigated the effects of network disruption on behaviour and disease states (Aerts et al., 2016; Fornito et al., 2015a). For example, consider the case of an individual with a localised brain injury. Using recent brain network approaches, it should be possible to map any effects of the injury on spatially distant but functionally connected structures of the brain (Thiebaut De Schotten et al., 2014). Such maps can then be compared across cohorts, or related to behavioural measures, in much the same way as classic lesion approaches, but without many of the pitfalls of traditional lesion analyses (Rorden and Karnath, 2004).

While the network approach is relatively new, one recent study by Urbanski and colleagues (2016) examined 27 patients with focal damage to the frontal lobes. In their analysis they inferred the anatomical connections that were likely to have been affected by each participant's lesion, and related these connectivity changes to behavioural performance on an analogical reasoning task. They found that damage to the left RLPFC and changes to anatomical connectivity with the RLPFC in the contralateral hemisphere, as well as changes in long range connections with the left parietal cortex, predicted worse performance on the reasoning task.

25 More broadly, a key hypothesis to emerge from the study of healthy functional networks predicts that damage to network hubs that connect multiple systems will be more likely to have widespread effects on cognition than damage to non-hub regions (Fornito et al., 2015a). In addition to computational modelling (Honey and Sporns, 2008), recent empirical work has also provided initial support for this hypothesis (Warren et al., 2014). Lesions to brain regions with between-network hub properties are associated with worse behavioural outcomes on a variety of measures, including concept formation and reasoning, when compared to control lesions of similar connection density (Warren et al., 2014).

The investigation of ‘dysconnection’ is an exciting and underexplored avenue for researching the mechanisms that underpin reasoning and intelligence. While the majority of this thesis examines functional networks and reasoning ability in healthy adults, in Chapter 6 I utilize this approach in a group of individuals with callosal dysgenesis, a disorder in which individuals are born without the corpus callosum, or with this structure only partially formed. Participants with callosal dysgenesis are particularly intriguing as they present with intact bilateral functional networks, despite their lack of interhemispheric connectivity (Owen et al., 2013; Tyszka et al., 2011).

1.6 Thesis aims

The overarching aim of my PhD was to characterise changes in functional brain network dynamics associated with cognitive reasoning, and to investigate the link between such dynamics and individual differences in ability. In Chapter 2, I summarise previous literature and report on an investigation of individual differences in intelligence associated with resting-state network connectivity. This work reveals that intelligence is best predicted by connectivity between the DMN and the FPN. In Chapter 3, I investigated connectivity within and between the DMN and cognitive control networks during a modified version of the classic Wason Selection Task. I found that the angular gyri of the parietal cortex (regions of the DMN) increased their functional connectivity with prefrontal and subcortical regions, despite showing an overall decrease in BOLD magnitude in response to the task. In Chapter 4, I report the results of a behavioural experiment in which I tested the reliability of a modified relational reasoning paradigm called the Latin Square Task (LST), originally introduced by Birney et al. (2006). The Latin Square Task involves the presentation of a grid populated with different shapes and a blank ‘probe’ location. Participants are asked to solve for the probe, keeping in mind that each shape can only appear once in every row or column, similar to the popular game Sudoku. This paradigm elicited similar behavioural effects to those reported previously (Birney and Bowman, 2009), and showed both internal consistency and test-retest reliability. In Chapter 5, healthy adult participants completed an initial resting-state scan, followed by the modified LST, developed in Chapter 4, and then a second resting-state scan. Using network analysis I found that increased relational reasoning complexity was associated with decreased network modularity and increased network efficiency, and that this latter effect was correlated with reasoning performance. In Chapter 6, a small group of individuals with callosal dysgenesis performed the LST while undergoing fMRI. I

26 investigated common task-related networks (the FPN and CON; Duncan, 2010; Seeley et al., 2007). Under low reasoning demands, brain activity and network topology were similar to those observed in normal adults (as described in Chapter 5), whereas under higher reasoning demands there was a notably diminished response in the fronto-parietal network.

Finally in Chapter 7 I present a general discussion of the findings from the empirical chapters. I discuss how changes in covariance between spatially distributed brain regions support reasoning behaviour and, specifically, I focus on the critical role of DMN-FPN interactions. I suggest that global workspace theory (Baars, 1988; Dehaene et al., 1998) might provide an explanatory framework for these observations, and conclude by suggesting that future work should aim to test this framework using methods that allow causal inferences to be made, such as non-invasive brain stimulation.

27

Chapter 2: Functional brain networks related to individual differences in human intelligence at rest

28 2.1 Preamble

The work contained in this chapter has been published in Scientific Reports (Hearne et al., 2016). Figures, tables and formatting have been adapted to fit the formatting of the thesis.

2.2 Abstract

Intelligence is a fundamental ability that sets humans apart from other animal species. Despite its importance in defining human behaviour, the neural networks responsible for intelligence are not well understood. The dominant view from neuroimaging work suggests that intelligent performance on a range of tasks is underpinned by segregated interactions in a fronto-parietal network of brain regions. Here we asked whether fronto-parietal interactions associated with intelligence are ubiquitous, or emerge from more widespread associations in a task-free context. First we undertook an exploratory mapping of the existing literature on functional connectivity associated with intelligence. Next, to empirically test hypotheses derived from the exploratory mapping, we performed network analyses in a cohort of 317 unrelated participants from the Human Connectome Project. Our results revealed a novel contribution of across-network interactions between default-mode and fronto-parietal networks to individual differences in intelligence at rest. Specifically, we found that greater connectivity in the resting-state was associated with higher intelligence scores. Our findings highlight the need to broaden the dominant fronto-parietal conceptualisation of intelligence to encompass more complex and context-specific network dynamics.

2.3 Introduction

Human intelligence can be broadly defined as the capacity to understand complex ideas, adapt effectively to the environment and engage in complex reasoning (Neisser et al., 1996). Measures of intelligence can be related to performance on virtually any cognitive task, from sensory discrimination (Melnick et al., 2013) to challenging cognitive tasks such as the identification of patterns in the Raven’s Progressive Matrices test (Raven, 2000). Importantly, scores on intelligence tests can accurately predict various life outcomes, including academic success(Jensen, 1998), job performance (Hunter, 1986), and adult morbidity and mortality (Gottfredson and Deary, 2004). Brain imaging studies have suggested that neural activity in frontal and parietal cortices during the execution of cognitive tasks is related to individual differences in intelligence (Basten et al., 2015; Duncan, 2010). These findings have been formalised in the influential Parieto-Frontal Integration Theory of intelligence (P-FIT, Jung and Haier, 2007) and have been proposed to extend to intrinsic networks of the brain (Langeslag et al., 2013; Song et al., 2008). By contrast, recent work investigating brain activity at rest (i.e., in the absence of any specific cognitive task) has suggested intelligence is underpinned by communication between widespread brain regions including, but not limited to, parieto-frontal areas (Cole et al., 2012; Song et al., 2009; van den Heuvel et al., 2009). Here we asked whether the P-FIT extends to intrinsic brain networks by undertaking an explorative summary of

29 recent literature and conducting an empirical analysis using a dataset of 317 unrelated participants from the Human Connectome Project (Van Essen et al., 2013).

Functional magnetic resonance imaging (fMRI) has been used to examine the relationship between individual differences in intelligence and brain activity during the engagement of cognitive abilities such as working memory (Gray et al., 2003) and reasoning (Duncan et al., 2000; Lee et al., 2006). Typically, in these studies regions associated with intelligence are isolated by subtracting fMRI signals between two conditions with different ‘intelligence-loadings’ (e.g., easy versus difficult reasoning problems, Duncan et al., 2000; Lee et al., 2006). Local changes in brain activity are then correlated with standard intelligence scores to identify regions that are related to individual differences in intelligence. Although this approach has been useful for identifying functionally segregated neural correlates of intelligence, it is insensitive to the integration of information processes across spatially and functionally segregated brain regions (i.e., functional connectivity).

Recent studies have investigated the relationship between individual intelligence scores and patterns of functional connectivity during resting-state scans (Finn et al., 2015; Song et al., 2008; van den Heuvel et al., 2009). One hypothesis to emerge from this work is that resting-state functional connectivity associated with intelligence should recapitulate the functional topology of frontal-parietal networks (Song et al., 2008). Attempts to test this prediction, however, have so far produced inconclusive findings. For example, individual differences in intelligence have been related to changes in resting-state connectivity in neural networks broadly involved in -referential mental activity (default-mode network), attentional control processes () and task-set maintenance (cingulo-opercular network; Pamplona et al., 2015; Santarnecchi et al., 2015; Smith et al., 2015; Song et al., 2009; Wang et al., 2011; Yuan et al., 2012).

Here we used convergent approaches to assess whether the P-FIT can be extended to task-free (resting-state) contexts. We started by conducting an exploratory mapping of previous findings from studies that had investigated the relationship between resting-state functional connectivity and measures of intelligence. Specifically, we mapped significant pairwise connections from four previous studies (see Table 2-1) into a validated topological characterisation of resting-state brain networks (Gordon et al., 2016). We found that the previously reported functional connections associated with intelligence were not restricted to the fronto-parietal system (Figure 2-1). We next tested this qualitative observation by mapping brain-intelligence relationships using a large and independent set of neuroimaging and behavioural data from the Human Connectome Project (HCP, (Van Essen et al., 2013). Within the HCP data, general intelligence is defined as individual scores on a shortened version of the Raven’s Progressive Matrices and the Picture Vocabulary test. According to a context- invariant interpretation of the P-FIT (Song et al., 2008), intelligence should be related to connectivity within a fronto-parietal network as assessed during task performance and in the resting-state.

30 Conversely, absence of overlap between task and resting-state networks would be more consistent with a context-specific neurophysiological model of intelligence.

2.4 Methods

2.4.1 Explorative mapping of relationship between intelligence and resting-state networks

For the explorative mapping of data on the relationship between intelligence scores and intrinsic neural activity we performed a manual literature search of English-language peer-reviewed fMRI studies linking measures of pairwise resting-state functional connectivity with behavioural measures of intelligence in healthy human adults (Table 2-1). The literature review was conducted using PubMed, Web of Science® (Thomson Reuter) and Scopus® (Elsevier), and was last updated the 9th of December 2015. Corresponding authors were contacted and asked to provide additional details or whole-brain results if these were not included in the published papers. Our final sample included data from 207 healthy adult participants. In one case (Song et al., 2009, 2008), the same participant cohort was used across two studies. However, in these studies orthogonal region-of-interest analyses were conducted. Conversely, a recent study was not included (Malpas et al., 2015) because it involved the same cohort and similar analyses as an already included study (Santarnecchi et al., 2015). Note that we included studies that utilised both individual differences and group differences in intelligence (details in Table 2-1). We also conducted a summary of several studies that investigated global and local changes in functional connectivity (see Appendix A). Due to the lack of overlap in analysis methods, the outcome was not included in the final analysis.

Table 2-1 Characteristics of studies included in the resting-state explorative mapping. Sample Author N Males Age Measure Brain-behaviour Analysis ROI (M±SD) relationship Song et al. 59a 49% 24.6±3.5 WAIS Correlation Seed to voxel Bilateral DLPFC Song et al. 59a 49% 24.6±3.5 WAIS Correlation Multi-region 13 default- pairwise mode regions analysis Pamplona 29 52% 26.8±5.8 WAIS Correlation Multi-region AAL atlasb et al. pairwise analysis. Santarnecc 119 50% 33±13 WASI Between-group Seed to voxel Six seed hi et al. median-split regions Note: a The samples used in the indicated studies were not independent. b MNI centroids were used as regions of interest. ROI = Regions of interest, WAIS = Wechsler Adult Intelligence Scale, WASI = Wechsler Abbreviated Scale of Intelligence, DLPFC = dorsolateral prefrontal cortex, PCC = posterior cingulate cortex, AAL = Automated Anatomical Labelling, VHMC = voxel-mirrored homotopic connectivity

31 Cortical regions resulting from the above mapping were transposed into a common functional brain parcellation comprising 333 cortical regions (Gordon et al., 2016). This brain parcellation was selected because it has been shown to be a more refined, homogeneous extension of widely used, functionally defined resting-state parcellations (Power et al., 2011; Yeo et al., 2011). It is important to note that some of the networks isolated in the adopted parcellation represent sub-networks of the fronto-parietal network defined by the P-FIT. Specifically, the fronto-parietal, cingulo- opercular/parietal and dorsal/ventral attention networks defined by the current brain parcellation are considered to be part of the same fronto-parietal system in the P-FIT. Each region was then mapped into the adopted parcellation by generating 5mm radius spheres from the reported MNI coordinates and quantitatively assessing the spatial overlap with regions of the adopted parcellation. A parcel was defined as overlapping with the region(s) reported in a previous study when it covered at least 20 contiguous voxels (1mm3) of the MNI sphere. In instances where several parcels were implicated from a single coordinate, only the parcel that overlapped the most was included. Edges were drawn between implicated parcels using their associated r-values from the original studies. Note that changing our criteria for an overlapping region, either by increasing or decreasing the voxel limit, or by increasing the sphere size to 10mm, yielded very similar results to those reported below.

2.4.2 Analysis of intrinsic functional networks supporting human intelligence using HCP data

We next conducted an analysis of data from a large, independent sample of healthy adult participants, to examine associations between intelligence and functional connectivity across the whole brain in a task-free context. The relationship between measures of intelligence and neural activity was assessed using high quality resting-state fMRI data from the HCP (Van Essen et al., 2013). Specifically, we used data from 317 genetically unrelated participants included in the S900 data release (173 female,

Mage = 28.43 years, SDage = 3.77, rangeage = 22 – 36 years). Two relevant behavioural tasks were used as measures of general intelligence. The Penn’s Progressive Matrices (PMAT), a shortened version of Raven’s Progressive Matrices (Bilker et al., 2012), was used as a measure of fluid intelligence (M = 17.06, SD = 4.77, range = 4-24). In the Raven’s Progressive Matrices participants are presented with puzzles containing visual patterns with a piece missing. They are instructed to ‘fill in the blank piece’ from a given selection of possible answers. The Picture Vocabulary Test, a component of the National Institutes of Health toolbox, was adopted as a measure of crystallized intelligence (Akshoomo et al., 2013) (M = 116.59, SD = 9.57, range = 92.84 – 153.09). In this task participants are presented with an audio recording of a word, and are shown four pictures. They are asked to select the picture that most closely matches the meaning of the spoken word. Individual scores on the PMAT and Picture Vocabulary Test were significantly correlated (r = 0.38, p < .0001).

32 2.4.3 HCP Data Preprocessing and Analysis

Data consisted of whole brain echo-planar images (EPIs) with sub-second temporal resolution (time repetition of 720 ms) and high spatial resolution (2mm3 voxels) (Moeller et al., 2010; Ugurbil et al., 2013). The data used for this study were downloaded as per the Human Connectome Project minimally preprocessed pipeline with denoising procedures (for details see (Glasser et al., 2013; S. M. Smith et al., 2013) and included both left-to-right and right-to-left acquisitions from the first resting- state dataset (i.e., resting-state fMRI 1 FIX-denoised package). The average time series from the voxels comprising each of the 333 regions in the adopted parcellation (Gordon et al., 2016) were extracted using the Matlab toolbox DPARSF V.3 (Chao-Gan and Yu-Feng, 2010). As per the studies included in the resting-state explorative mapping (Pamplona et al., 2015; Santarnecchi et al., 2015; Song et al., 2009, 2008) , we calculated functional connectivity per participant in each acquisition as a Pearson correlation between each pair of regions, which were subsequently Fisher-Z transformed. Each pair of Z-matrices (left-to right, and right-to-left) was then averaged resulting in a 333 x 333 functional connectivity matrix for each of the 317 participants. Note that no global signal regression was performed for consistency with previously published studies included in the resting-state explorative mapping.

To assess the relationship between resting-state functional connectivity and individual intelligence scores we used the network based statistic [NBS(Zalesky et al., 2012a, 2010), https://sites.google.com/site/bctnet/comparison/nbs]. The NBS is a powerful and sensitive statistical tool that controls for Type I error at the network-level. The use of NBS represents a distinct advantage in term of sensitivity over previous studies that corrected for multiple comparisons at the edge level (Pamplona et al., 2015; Song et al., 2008). Unthresholded functional connectivity matrices were first used as input into the NBS (Zalesky et al., 2012a). All possible pairs of connections (333 x 332/2 = 55,278) were examined for putative associations with intelligence. To this end, Z-normalised fluid intelligence scores (PMAT) and crystalized intelligence scores (Picture Vocabulary Test) were used as separate variables of interest in the NBS. Age and gender were considered as covariates of no interest. Following this procedure, a matrix of brain-behaviour associations was obtained. The matrix was thresholded using an exploratory t-statistic of 3.5. A slightly higher final threshold (t = 3.7) was adopted because it allowed the detection of medium sized effects while discarding small or spurious effects (Zalesky et al., 2012a). Note that additional exploratory analyses showed that networks arising using higher or lower t-thresholds resembled the original results. Familywise error corrected (FWE) p-values were ascribed to the resulting networks using a null distribution obtained by 5000 permutations. Only components that survived a network-level threshold of p < 0.05 FWE were declared significant. Analyses were performed using both the extent criterion (number of connections in a network) and intensity criterion (sum of test statistic values in a network) in NBS, for both positive and negative associations with intelligence. It is important to note that the NBS is a network-

33 sensitive method, and does not test for significance at the level of individual edges. Therefore, our analysis provides a network-level characterization of resting-state functional connectivity correlates of intelligence that can guide further, and more local, investigations.

Figures were generated using BrainNet Viewer (Xia et al., 2013) , NeuroMArVL, (http://immersive.erc.monash.edu.au/neuromarvl/) and in-house Matlab scripts.

2.5 Results

2.5.1 Exploratory mapping: resting-state functional connectivity related to intelligence

Results from our mapping of studies (Table 2-1) assessing the relationship between resting-state functional connectivity and intelligence scores are presented in Figure 2-1. Significant patterns of pairwise functional connectivity positively associated with intelligence (Figure 2-1a) suggest a key role for connections between prefrontal and frontal cortices comprising the dorsal attention network (dark green). Significant brain-behaviour associations were also observed for the posterior cingulate/precuneus (red, default-mode network), the superior parietal cortex (yellow, fronto-parietal network) and the occipital cortices (dark blue, visual network). Resting-state functional connectivity between bilateral prefrontal cortices encompassing the dorsal attention network and the right insula (, black) was also associated with intelligence scores.

Correlations between lower resting-state functional connectivity (i.e., reduced positive correlations and/or increased anticorrelations) and higher intelligence scores (Figure 2-1b) involved connections within cortical areas comprising the default-mode network (red) as well as functional interactions between these areas and regions within the dorsal attention network, including the visual, cingulo- parietal and somatosensory (both hand and mouth) regions.

34

Figure 2-1 Pairwise functional connections associated with intelligence at rest from previous literature. a. Connections in which higher functional connectivity was associated with higher intelligence. b. Connections in which lower functional connectivity (i.e., reduced positive correlations and/or increased anticorrelations) was associated with higher intelligence. Edges are weighted by level of correlation reported in the original studies. In the case where no r-value was provided (i.e., in between-group contrasts) edges were weighted at the minimum value for visualization purposes (±0.25).

2.5.2 Empirical analysis of resting-state functional networks supporting human intelligence using HCP data

Though suggestive, the above results represent merely the overlap of findings from a small number of studies with varying sample sizes and regions of interest. Thus, to empirically test the hypothesis that resting-state connectivity correlates of intelligence extend beyond the fronto-parietal network, we utilized resting-state data from the Human Connectome Project. Specifically, we assessed positive and negative linear relationships between whole-brain resting-state functional connectivity and intelligence scores of 317 unrelated participants (i.e., participants did not have the same mother or father).

35

Figure 2-2 Network-intelligence analysis on 317 independent HCP participants. a. Pairwise functional connections associated with intelligence scores [p = 0.045 (extent), p = 0.032 (intensity), both FWE corrected at the network level]. Cortical colours reflect their network allegiance, and edge weights reflect the uncorrected edge t-statistics. Note that the light blue regions in (a) were not linked to a specific network by Gordon and colleagues25. Panel b shows the same results as those depicted in panel a, but outside of anatomical space. Here the edge t-statistics are represented by colour. Circles represent network nodes comprising the default-mode and fronto-parietal and non-affiliated networks. Pie charts show the percentage of significant connections that were within (white) or across (coloured) different networks. c. Scatterplot of the average functional connectivity (FC) values in the whole implicated network (panel a) as a function of general intelligence scores (r = 0.38), DMN = default-mode network, FPN = fronto-parietal network.

Functional connectivity in a concentrated resting-state network comprising regions of the fronto- parietal (yellow), default-mode (red), and cortex not associated with any specific network (aqua; note that these regions have been labeled as default-mode in other parcellations, (Yeo et al., 2011) showed a significant positive relationship with intelligence scores (p = 0.045/0.032 familywise error corrected at network level for extent and intensity effects, respectively; see Methods for details, Figure 2-2). Specifically, the resulting regions included the bilateral superior medial frontal cortex, superior orbital

36 gyrus and temporal cortex, as well as the left middle cingulate cortex and right middle frontal and supramarginal gyrus (details in Table 2-2). No other networks were implicated in the analysis. Connections within the default-mode and fronto-parietal networks accounted for the majority of edges detected (Figure 2-2b). A follow-up correlation between the mean connectivity value of all implicated edges and general intelligence (as measured by the average of z-scored fluid and crystallized intelligence measures) showed that higher positive connectivity values were associated with higher intelligence scores (r = 0.38).

Table 2-2 Regions implicated in the analysis of the Human Connectome Project data.

Gordon MNI Coordinates Resting-state network Anatomy region x y z 25 -5.6 42.2 35.1 Default-mode Superior medial frontal gyrus 26 -1.7 -17.7 39.1 Default-mode Middle cingulate cortex 114 -27.5 53.6 0 Default-mode Superior frontal gyrus 115 -23.4 61 -6.8 None Superior orbital gyrus 128 -53.2 -13 -29.2 None Inferior temporal gyrus 150 -6.5 54.7 18.1 Default-mode Superior medial frontal gyrus 151 -15.7 64.7 13.7 Default-mode Superior frontal gyrus 165 11.9 21.9 59.9 Default-mode Posterior medial frontal gyrus 167 47.9 -42.5 41.5 Fronto-parietal Supramarginal gyrus 277 28.4 57 -5.1 Fronto-parietal Superior orbital gyrus 291 54.7 -7.8 -26.9 None Inferior temporal gyrus 321 16 61 19.8 Default-mode Superior medial frontal gyrus 322 8.2 53.8 14 Default-mode Superior medial frontal gyrus 327 42.4 19.5 48.2 Fronto-parietal Middle frontal gyrus 328 38.9 9.6 42.7 Fronto-parietal Middle frontal gyrus Note: In some cases the implicated parcels cross multiple anatomical boundaries; here we have simply tried to provide the most accurate anatomical description.

No associations were found between increased intelligence and decreased resting-state functional connectivity. Considering the large sample size this is unlikely to be related to a lack of statistical power. Nevertheless, we performed the NBS again with two lower exploratory statistical thresholds (t = 3.0 and t = 2.5). No networks showed significant negative associations between intelligence and functional connectivity using these lower statistical thresholds.

2.6 Discussion

We assessed whether the dominant parieto-frontal integration theory of intelligence (P-FIT; Jung and Haier, 2007) can be extended to networks supporting intelligence in task-free contexts (i.e., at rest) by

37 conducting a functional connectivity analysis of Human Connectome Project data. Specifically, we tested whether resting-state functional connectivity within frontal and parietal brain regions, and between these regions and the rest of the brain, can account for individual variability in intelligence scores. While our findings confirm a key role for fronto-parietal networks in supporting intelligence, they also highlight the importance of connectivity between regions associated with the fronto-parietal, default-mode and regions not strongly associated with homogeneous networks (although these regions have been identified as comprising the default-mode network before; Yeo et al., 2011), particularly in the prefrontal cortex. More broadly, our results suggest that interactions between fronto-parietal and default-mode networks are important for explaining individual differences in intelligence in a state of rest.

Recent evidence suggests that the default-mode and frontal-parietal networks represent overarching systems of the brain, composed of several sub-networks that dynamically interact (Eldaief et al., 2011; Hugdahl et al., 2015). Engaging in demanding external tasks has traditionally been associated with increased activity and functional connectivity in fronto-parietal networks, on the one hand, and reduced activity and connectivity in default-mode areas on the other (Anticevic et al., 2012). The opposite functional relationship between these two systems has also been observed during the resting- state (Fox et al., 2005; Sonuga-Barke and Castellanos, 2007). Likewise, during cognitive tasks, it has been shown that individuals with higher and lower intelligence tend to activate these networks differentially (Basten et al., 2013; Lipp et al., 2012). Specifically, individuals with higher intelligence deactivate the default-mode network less (i.e., a smaller task-induced decrease in the BOLD signal, (Basten et al., 2013) and activate fronto-parietal and cingulo-opercular network regions more than individuals with lower intelligence (Basten et al., 2013; Gray et al., 2003; Lee et al., 2006). Evidence for the second claim, however, is mixed (Perfetti et al., 2009; Tang et al., 2010).

Somewhat at odds with this functional dichotomy between fronto-parietal and default-mode network activity, our analyses suggest that greater cooperation between distinct brain regions comprising default-mode and fronto-parietal networks in the resting-state are associated with higher intelligence scores. This pattern of connectivity-intelligence associations is consistent with other findings suggesting that higher global network efficiency is related to higher general intelligence measures (van den Heuvel et al., 2009) and that the across-network connectivity of the fronto-parietal network is critical for fluid intelligence (Michael W. Cole et al., 2015). More broadly, our findings are compatible with recent conceptualisations of default-mode network function as critical in maintaining a large “dynamic repertoire” of possible neural states at rest (Deco et al., 2011; Deco and Jirsa, 2012), facilitating the flexible emergence of task-specific dynamics (Gu et al., 2015; Hellyer et al., 2014; Vatansever et al., 2015).

How the brain self-reorganizes to achieve optimal configurations of functional networks across individuals with varying levels of intelligence is an open question. Recent neuroimaging work has

38 suggested that transient cooperation between different neural systems, including fronto-parietal, cingulo-opercular and default-mode networks, is integral to complex cognitive tasks such as reasoning (Cocchi et al., 2014a; Hearne et al., 2015), memory recollection (Fornito et al., 2012) and working memory performance (Leech et al., 2011; Vatansever et al., 2015). Future studies should test the notion that individual differences in intelligence rely on dynamic, context-specific, reconfigurations of local activity and connectivity within a diffuse system comprising fronto-parietal, cingulo-opercular and default-mode regions (Cocchi et al., 2013).

A strength of the current work is the use of a statistically robust network-based method to isolate brain-intelligence associations at rest. While sensitive to network-level associations between functional connectivity and intelligence, our approach may have overlooked edge-specific associations detected in previous work (see Figure 2-1). For example, we found no significant negative association between differences in individual functional connectivity strength and intelligence, despite two previous studies included in our explorative mapping reporting such a relationship (Santarnecchi et al., 2015; Song et al., 2008). To enhance comparability, we attempted to keep the current analysis as similar to previous work as possible. In fact, both our study and previous work used similar data for calculating functional connectivity (i.e., Z-normalised pairwise Pearson correlations on data without global signal regression). One possible explanation for the discrepancy between our work and earlier results may relate to the use of different statistical methods to infer connectivity-intelligence associations. Previous work used edge-specific correlations with intelligence scores. In contrast, our analysis focused upon significant relationships at the level of whole brain networks. It is possible our approach was less sensitive to circumscribed negative associations between functional connectivity and intelligence. Finally, while our study assessed the functional relationship between pairwise changes in connectivity and intelligence, other studies assessed the link between intelligence and more complex measures of functional connectivity patterns (see Appendix A). However, we note that these analyses broadly validated the current results by implicating key default-mode regions and other brain areas (see Appendix A).

In summary, our study provides a novel characterization of large-scale networks that explain individual differences in intelligence in a state of rest. Our results suggest that intelligence is supported by activity within a diffuse neural system comprised of brain regions encompassing fronto- parietal and default-mode networks. Consistent with these findings, we propose the influential parieto-frontal intelligence theory (P-FIT) may need to be extended to address context-specific network interactions. The functional links between transitions from diffuse resting-state dynamics and more segregated task dynamics and intelligence will be an important topic for ongoing research.

39

Chapter 3: Interactions between default mode and control networks as a function of increasing cognitive reasoning complexity

40 3.1 Preamble

In Chapter 2 I related resting-state fronto-parietal and default-mode functional network properties to measures of intelligence. As a direct follow-up, I explored the same networks while participants completed a complex reasoning task. The work contained in this chapter has been published in Human Brain Mapping (Hearne et al., 2015).

3.2 Abstract

Successful performance of challenging cognitive tasks depends upon a consistent functional segregation of activity within the default-mode network, on the one hand, and control networks encompassing fronto-parietal and cingulo-opercular areas on the other. Recent work, however, has suggested that in some cognitive control contexts nodes within the default-mode and control networks may actually cooperate to achieve optimal task performance. Here we used functional magnetic resonance imaging to examine whether the ability to relate variables while solving a cognitive reasoning problem involves transient increases in connectivity between default-mode and control regions. Participants performed a modified version of the classic Wason Selection Task, in which the number of variables to be related is systematically varied across trials. As expected, areas within the default-mode network showed a parametric deactivation with increases in relational complexity, compared with neural activity in null trials. Critically, some of these areas also showed enhanced connectivity with task-positive control regions. Specifically, task-based connectivity between the striatum and the angular gyri, and between the thalamus and right temporal pole, increased as a function of relational complexity. These findings challenge the notion that functional segregation between regions within default-mode and control networks invariably support cognitive task performance, and reveal previously unknown roles for the striatum and thalamus in managing network dynamics during cognitive reasoning.

3.3 Introduction

The execution of increasingly demanding cognitive tasks, such as logical problem solving, has been associated with proportional increases in neural activity and functional connectivity in fronto-parietal and cingulo-opercular control networks (Cocchi et al., 2014; Dosenbach et al., 2006). Increased task difficulty is also associated with decreases in neural activity within regions encompassing the default- mode network (Lawrence et al., 2003; McKiernan et al., 2003). An increase in functional antagonism (i.e., anticorrelation) between activity in control and default-mode regions as a function of increased task demands is thought to be critical for optimal cognitive performance (Anticevic et al., 2012; Fox et al., 2005; Kelly et al., 2008; Sonuga-Barke and Castellanos, 2007; Weissman et al., 2006).

Recent evidence from studies suggests that cognitive processes such as reasoning, attention and memory recall are supported by complex reconfigurations in the patterns of cooperation and competition between widespread resting-state networks (Cocchi et al., 2013; Dwyer

41 et al., 2014; Leech et al., 2012; Leech et al., 2011). Such changes in neural network dynamics preserve the general functional network architecture of the brain (Cole et al., 2014), appear task- specific (Cocchi et al., 2014; Cole et al., 2013; Sridharan et al., 2008), and may include cooperation between regions and networks that are otherwise functionally segregated in the resting-state (Bluhm et al., 2011; Dwyer et al., 2014; Fornito et al., 2012; Leech et al., 2011; Liang et al., 2015; Popa et al., 2009). This research suggests that cognitive task performance does not necessarily require antagonism between default-mode and control networks (Anticevic et al., 2012; Sonuga-Barke and Castellanos, 2007).

We recently demonstrated that the process of actively relating variables to solve logical reasoning problems involves enhancement of neural activity and task-based connectivity between discrete regions within the fronto-parietal and cingulo-opercular control networks (Cocchi et al., 2014). The ability to relate multiple variables to achieve internal goals is a critical component of human intelligence (Johnson-Laird, 2010), and has been formally quantified with the relational complexity metric (Halford et al., 1998; Halford et al., 2010). The task used in our study was an adaptation of the Wason Selection Task, a classic deductive reasoning paradigm (Wason, 1966) (Figure 3-1a). In this paradigm participants are asked to test if a given set of propositions could potentially disconfirm a logical rule (Figure 3-1b). Our original analysis focused exclusively on positive, complexity-evoked changes in local activity and connectivity, compared with task unrelated activity. Here we undertook new analyses of these data to directly assess complexity-induced changes in dynamics between control and default-mode regions. Our aim was to determine whether complexity in relational reasoning is supported by a consistent increase in competition (i.e. anticorrelation), or selective integration, between regions encompassing control and default-mode networks. By testing these competing predictions, our study provides critical information for understanding task-evoked neural dynamics supporting complex reasoning processes.

3.4 Materials and Methods

3.4.1 Participants

Twenty-one participants aged from 21 to 39 years (mean ± SD = 28.6 ± 5.0 years, 12 females) were included in the analysis. Participants provided informed written consent to participate in the study. The study was approved by The University of Queensland Human Research Ethics Committee and conducted in accordance with the Declaration of Helsinki.

3.4.2 Paradigm

In the classic Wason Selection Task (WST) participants are shown four cards and provided with a conditional rule (Figure 3-1a). Participants are then asked which card(s) must be turned over to disconfirm the rule. To solve the WST, the participant must consider each card’s relationship with the rule, and determine whether the card can disconfirm it (Wason, 1968).

42

Figure 3-1 Schematic of the Wason Selection Task (WST) and modified version used in the current study. a. In the classic WST participants are shown four cards (in this example, ‘3’, ‘8’, green and orange) and are provided with a conditional rule such as ‘If a card shows an even number on one face, then its opposite face will be green. Participants are then asked which card(s) must be turned over to test the conditional rule provided. The only cards that can disconfirm the rule are the orange (reverse side even) and 8 (reverse side orange) cards. If the 3 card is orange on its opposite face it does not invalidate the rule (the rule does not say anything about odd cards). Likewise, if the reverse of the green card is an odd number the rule cannot be disconfirmed. b. Trial structure of the modified WST used in the fMRI study. Participants were presented with a rule (e.g. ‘If A then 7’), followed by a single ‘card’, and were then asked to judge if the presented card could disconfirm the rule. By presenting cards serially the level of relational complexity needed to disconfirm the rule is manipulated on a trial-by-trial basis.

In the current study we implemented an adapted version of the WST (Figure 3-1b). The key difference between the original version of the task and the one used in our study is that the rules and ‘cards’ were presented sequentially, and that only one card was assessed on each trial (rather than all four simultaneously). The task involved presentation of a logical rule (3000 ms) that established a defined relationship between two alphanumeric variables (e.g. “If A then 7”). Rules always consisted

43 of a single letter and a single digit (from 1 to 9). Letters and digits were presented with equivalent frequency as the first and second elements of the rule. One hundred and eighty different rules were presented throughout the experiment. As such, it was not possible for participants to reduce the task to a mere matching of elements. Following a variable 3000–5000 ms period, a single letter (on 50% of the trials) or number was displayed for 3000 ms. Participants were asked to think of this character as one side of a card and to indicate, as quickly and accurately as possible, if the card could disconfirm the rule, assuming the other side of the card could contain any possible number (if the face side contained a letter) or letter (if the face side contained a number). Importantly, participants were instructed that the rule was not bidirectional. For example, the rule “If A then 7” does not imply “If 7 then A”.

Table 3-1 Experimental conditions in the modified WST, and examples of rules and cards Condition Example Example Correct Answer Logic Rule Card Binary If A then 7 A This card can The ‘A’ card matches the first element of the rule disconfirm the rule and the reverse side (e.g. 5) can be used to disconfirm the rule. This is a binary relation. Inverse 7 Not possible to The ‘7’ card matches the second element of the rule Binary disconfirm the rule and the reverse side cannot disconfirm the rule. This is a binary relation in the reverse direction defined by the rule. Ternary C (not- Not possible to This card contains neither of the two rule elements A) disconfirm the rule (‘A’ or ‘7’). The ‘C’ card must contain a number on the reverse side, and therefore is not informative about the rule ‘If A then 7’. This is a ternary relation. Quaternary 5 (not-7) This card can This card contains neither of the two rule elements disconfirm the rule (‘A’ or ‘7’). The ‘5’ card must contain a letter on the reverse side. This letter could potentially be ‘A’, therefore disconfirming the rule ‘If A then 7’. This is a quaternary relation. Active # - No reasoning required control Null No rule - - No reasoning required presented Note: It is important to remember that the rule is not bidirectional. In the above table ‘A’ implies ‘7’, but ‘7’ does not imply A. Therefore showing a ‘7’ on the reverse side of ‘C’ is not informative, but showing ‘5’ on the reverse of ‘A’ is.

Five unique card conditions were presented (see Table 3-1 for an example of each). In accordance with the relational complexity metric (Halford et al., 2010), each condition had a specific level of complexity: i) Binary. The card presents an element that corresponds to the first element of the rule. ii) Inverse binary. The card presents an element that corresponds to the second element of the rule. iii)

44 Ternary. The card element is different from the first element of the rule, but of the same category (i.e. a letter or number). iv) Quaternary. The card presents an element that is different to the second element of the rule, but of the same category. v) Active control. The active control condition involved the presentation of a non-alphanumeric element completely unrelated to the rule (e.g. #). Lastly, the paradigm included null trials in which a white fixation cross was presented for the whole duration of the trial. Participants were instructed to rest and do nothing when presented with a null trial. These six trial types were presented with equal frequency in pseudo-random order across five runs lasting approximately 10 minutes, each consisting of 36 task trials. Before performing the task in the MR scanner participants performed a training session. In this practice session the relations to be established to solve the different card conditions were explained (further details in Cocchi et al., 2014).

3.4.3 Imaging

The study was conducted using a 3T Siemens Trio scanner (32 channel head coil, TR=2550 ms; TE= 32 ms; flip angle= 90°, FOV= 210 × 210 mm, 36 axial slices). Images were preprocessed using a standard pipeline implemented in the software SPM8 [for details see (Cocchi et al., 2014)]. This pipeline comprised a correction for acquisition time, realignment, normalization, high-pass filtering, and a correction for first-order autocorrelations. Analyses comprised the assessment of neural activity and task-based functional connectivity related to the processing of increasingly complex card stimuli. Regional activity was isolated using a general linear model framework, as implemented in SPM8. At the first level, conditions of interest were modelled as boxcar functions convolved with a canonical hemodynamic response function and its temporal derivative. The model included the rule epochs and the card-processing epoch (3000 ms) for all six conditions (the four card types, plus the control and null trials). Errors due to increased relational complexity were one of our predictors for changes in brain activity and connectivity; as such, the card regressors included both correct and incorrect trials. We also performed the analysis on correct trials alone, however, to investigate whether the exclusion of errors had any effect on changes in connectivity. First-level contrasts were used for second-level random effects analyses isolating regions showing changes in neural activity as a function of relational complexity. Specifically, we isolated the positive average card effect using a t-contrast (1 1 1 1 1 -5, with -5 being the null trials). The opposite contrast (-1 -1 -1 -1 -1 5) was used to isolate regions showing a negative task effect (i.e. deactivation). A p-value lower than 0.05, family-wise error corrected (FWE) at cluster level, was used as the threshold to declare significance. Beta regressor values were extracted from a 5 mm radius sphere located on the local maxima of each region showing a significant average effect of card complexity to explore changes in the pattern of activity as a function of complexity.

Functional interactions supporting card processing were investigated with a multiregional psychophysiological interaction (PPI) modelling approach (Cocchi et al., 2014; Gerchen et al., 2014;

45 McLaren et al., 2012). Instead of assessing connectivity (i.e., functional integration) between a single- seed region and all brain voxels, connectivity was assessed between pairs of regions isolated using the GLM (positive and negative average card effects). Twenty-four regions were included; 18 showed an average positive card effect [the same regions used in (Cocchi et al., 2014)] and 6 showed an average negative card effect [p < 0.05 FWE corrected at the cluster level, see Table 3-2]. For each participant, brain activity (first eigenvariate) was extracted from a 5 mm spherical seed region located around the peak activation voxel. As in a standard PPI analysis, the PPI signal was estimated for each region by multiplying the region's activity with the card versus the baseline regressor. Thus, the estimated PPI signal is equal to the region's activity during the processing of a specific card, but zero for all other card conditions and -1 for the baseline. Next, a GLM was used to model card-dependent influences of any given region upon another. Activity within the target regions was the dependent variable. The explanatory variable was the PPI term corresponding to the source region. The card versus baseline regressor and the main effects of the psychological and physiological factors related to the activity of the region used to determine the PPI term were included as nuisance covariates. This procedure was repeated for every possible pair of regions (24 x 23= 552) in each individual card type versus rest. The result was a 24 × 24 connectivity matrix for each individual and card type, where each element (x,y) of the matrix stored the parameter estimate (β) for the equivalent PPI term. Unlike a regular functional connectivity analysis, this resulted in an asymmetric matrix: half the matrix contained connectivity estimates in one direction (e.g. A to B) while the second half contained connectivity estimates for the opposite direction (e.g. B to A). The β estimates quantified the card-dependent influence region x exerts on region y. Finally, the extent of the differences in the card-evoked influence of one region on another across the five active card conditions (A, 7, C, 5, # in the example depicted in Table 3-1) was tested using a within-subjects analysis of variance (ANOVA). Pairs of regions with a t-statistic exceeding an uncorrected threshold of three (equivalent to p< 0.01 uncorrected) were searched for complexity-modulated networks. Using the network based statistic (NBS) a family-wise error corrected p-value (p < 0.05) was then attributed to each surviving network using permutation testing [10,000 permutations (Zalesky et al., 2012; Zalesky et al., 2010)]. This procedure was repeated three times: once as described above, a second time using only correct trials, and a third using a different set of regions encompassing the default mode brain network (DMN; 32 regions defined by Dosenbach et al., 2010). For this analysis the best sensitivity was obtained using an exploratory t-threshold of 3.5. All brain images presented were visualized using BrainNet viewer (Xia et al., 2013).

3.5 Results

As reported in Cocchi et al. (2014), the mean response accuracy was above 80% in all card conditions, but declined significantly as a function of card complexity (X2 = 25.85, df = 4, p < 0.01). As expected, reaction time increased significantly with card complexity (X2 = 21.96, df = 4, p < 0.01).

46 3.5.1 Complexity-induced change in neural activity

In line with our original analysis (Cocchi et al., 2014), increments in card complexity increased activity within regions encompassing the fronto-parietal and cingulo-opercular networks (Figure 3- 2a, Table 3-2). The analyses of complexity-induced deactivations (null minus average card effect) identified a set of brain regions that are part of the default-mode brain network [(Dosenbach et al., 2010; Greicius et al., 2003; Harrison et al., 2008; Power et al., 2010; Raichle et al., 2001), Figure 3- 2a, details in Table 3-2]. Post hoc analysis of the condition-specific (e.g. A card vs null card period) regional beta regressor values showed a significant parametric decrease in activity as a function of increasing relational complexity (Figure 3-2b).

Figure 3-2 Complexity-based changes in regional activity. a. Average brain activity during card processing. The task-positive contrast (average positive card minus null trials) showed significant neural activity in regions encompassing both fronto-parietal and cingulo-opercular networks (Dosenbach et al., 2010, see Table 3-2 for details). The task-negative contrast (null trials minus average positive card effect) showed brain regions that form part of the default-mode network (Fox et al., 2005, Table 3-2). b. Mean brain activity change with increases in card complexity in task-positive and task-negative contrasts. Error bars represent the standard error of the mean (SEM). Overall, task-positive regions increased their activity as a function of card complexity. By contrast, task-negative regions decreased their activity as a function of complexity. Note that these averaged results for the group were representative for all individual regions.

47 Table 3-2 General linear model results used to investigate changes in task-evoked (PPI) connectivity Anatomy a Stats b Resting-state network c x y z KE Z Pcorr Task Positive ROIs (average task effect minus null trials) Parietal cortex -45 -34 46 12636 6.84 < 0.001 Fronto-parietal 42 -55 46 5.25 Anterior insula -30 17 7 5.79 Cingulo- 36 14 7 5.87 opercular Cingulate cortex 3 8 52 5.58 Striatum 21 5 16 6.19 Thalamus -15 -19 10 5.68 Cuneus -9 -73 10 5.51 Visual 12 -70 13 4.95 Lateral frontal cortex -57 8 22 6.46 Fronto-parietal Lateral prefrontal cortex -36 47 28 5.88 36 47 34 5.87 Dorsolateral prefrontal cortex -39 23 31 4.32 42 26 31 4.82 Motor cortex -30 -16 58 5.96 Sensorimotor Cerebellum 3 -52 -14 6.16 Rostrolateral prefrontal cortex 36 50 1 4.22 Fronto-parietal -28 50 -8 62 4.41 0.07 (<0.05 FDR) Task Negative ROIs (null trials minus average task effect) Medial frontal cortex -3 47 -8 1916 6.10 < 0.001 Default-mode Angular gyrus -39 -78 30 277 6.36 < 0.001 48 -66 22 127 5.82 0.005 Posterior cingulate cortex 0 -46 43 1138 5.13 < 0.001 Right temporal lobe 54 2 -20 145 5.10 0.002 Left superior temporal gyrus 54 -9 -12 154 4.51 0.002 Note. a Coordinates (x, y, z) are given in Montreal Neurological Institute (MNI) atlas space. b If not otherwise indicated, p values are Family Wise Error (FWE) corrected for multiple comparisons at cluster level. FDR= False Discovery Rate. c ROI resting-state network allegiances are based on Dosenbach et al., (2010, supplementary table S6.).

3.5.2 Complexity-induced change in PPI connectivity

PPI connectivity analysis revealed a significant increase in integration between cingulo-opercular, fronto-parietal, and default-mode regions as a function of increased card complexity (Figure 3-3a and Table 3-3). Although the PPI analysis was performed on a larger set of regions, changes in

48 connectivity between fronto-parietal and cingulo-opercular regions replicated our previously reported results (Cocchi et al., 2014, red edges in Figure 3-3a, Table 3-3). There was also a significant modulation in complexity-evoked connectivity between discrete regions encompassing task-positive and default-mode networks (green edges in Figure 3-3b, Table 3-3). Specifically, integration between the angular gyri and the striatum progressively increased as a function of relational complexity. The two angular gyri also showed increased integration with the rostrolateral prefrontal cortex. A similar pattern of results was found between the right temporal pole (one of the task- negative regions) and the thalamus and striatum. It is important to note that NBS only allows inferences at the level of the whole network (p < 0.05 FWE). As such, while the detected edge- specific effects were consistent and relatively robust (partial Eta-squared > 0.13, p < 0.02 uncorrected), inferences on the functional significance of such specific connectivity changes need to be drawn with caution.

Figure 3-3 Complexity-evoked changes in PPI connectivity. a. Red nodes correspond to task-positive regions isolated in the general linear model analysis (Figure 3.2; see Table 3.2 for details). Blue nodes correspond to task-negative (‘deactivated’) regions. The red edges indicate increases in PPI connectivity between task-positive regions, whereas the green edges represent increases in connectivity between task-positive and task-negative regions (details in Table 3.3). b. Changes in average PPI connectivity values between task-positive and task- negative regions as a function of increased card complexity. Average changes were representative of edge-

49 specific connectivity changes due to changes in complexity.

Within the network of interest, there was no change in integration as a function of complexity between regions showing significant deactivations with complexity (blue spheres in Figure 3-3a). A follow-up analysis was performed to further investigate the nature of this unexpected result (see below). When only correct trials were included in the PPI analysis, a similar network and pattern of connectivity changes were observed. The only exception was the absence of a complexity-based increase in PPI connectivity between the left rostrolateral prefrontal cortex and the angular gyri.

Table 3-3 Pairwise changes in PPI functional connectivity between source and target regions as a functional of relational complexity Source region Target region Direction Task positive – Task positive connections Left anterior insula Right anterior insula Increasing Left dorsolateral prefrontal cortex

Right anterior insula Left dorsolateral prefrontal cortex Cingulate cortex Left dorsolateral prefrontal cortex Left cuneus Striatum Left dorsolateral prefrontal cortex Right cuneus Striatum Thalamus Left dorsolateral prefrontal cortex Right dorsolateral prefrontal cortex Left rostrolateral prefrontal cortex Right dorsolateral prefrontal cortex Left dorsolateral prefrontal cortex Lateral frontal cortex Left dorsolateral prefrontal cortex Left lateral prefrontal cortex Striatum Right parietal cortex Left dorsolateral prefrontal cortex Right rostrolateral prefrontal cortex Cingulate cortex Left cuneus

Right cuneus Left dorsolateral prefrontal cortex Task negative – Task positive connections Angular gyrus left Striatum Increasing Angular gyrus right Striatum Temporal pole right Striatum Thalamus Right rostrolateral prefrontal cortex Angular gyrus left Angular gyrus right

50 Note: Connections were corrected at the network level using NBS (P < 0.05 FWE, Zalesky et al., 2010, 2012).

Follow-up analysis of default-mode network dynamics as a function of complexity We found no change in connectivity within the default-mode network as a function of increasing task complexity. This result was unexpected, but might be due to our approach of selecting the regions of interest based on complexity-evoked activation (Gerchen et al., 2014). Likewise, the default-mode network is thought to be comprised of functionally heterogeneous sub-networks (Andrews-Hanna et al., 2010; Eldaief et al., 2011). The fact that regions were selected based on complexity-induced neural activations may therefore have biased our selection toward one or more specific sub- network(s). As such, it is possible that complexity-induced deactivations did not capture complexity- induced network dynamics in full. To test this, we performed the same PPI connectivity analysis using 32 regions of interest encompassing the current task-negative network and the wider as defined by Dosenbach and colleagues (2010, see Appendix B).

Figure 3-4 Complexity-evoked changes in PPI connectivity within the default-mode network. a. Purple edges indicate increased PPI connectivity as a function of complexity, whereas blue edges represent decreased connectivity (details in Table 3-4). b. Average PPI connectivity patterns for pairwise connections showing an increase in connectivity (in purple) and decreased connectivity (in blue) as a function of increased relational complexity.

51 Table 3-4 Pairwise change in PPI connectivity between source and target regions encompassing the default-mode network as a function of relational complexity

Source region Target region Direction Default-mode regions Anterior prefrontal cortex Ventral medial prefrontal cortex Increasing Ventral medial prefrontal cortex Left angular gyrus Anterior cingulate cortex Left inferior temporal Left occipital cortex Left occipital cortex Right Angular gyrus Left Angular gyrus Left inferior temporal cortex Precuneus Left angular gyrus Intra parietal sulcus Right angular gyrus Medial occipital cortex Left angular gyrus Right occipital cortex Precuneus Left occipital cortex Right angular gyrus Left angular gyrus Ventral medial prefrontal cortex Right occipital cortex Decreasing Superior frontal cortex Note: Connections were corrected at the network level using NBS (P < 0.05 FWE, Zalesky et al., 2010, 2012).

Result showed that increased task complexity induced two distinct sets of changes within default- mode network dynamics (Figure 3-4, Table 3-4). A first set of regions, primarily consisting of the bilateral angular gyri, precuneus, occipital cortex and medial prefrontal cortex, showed an increase in functional connectivity as a function of card complexity (purple connections in Figure 3-4). Conversely, a small subset of regions showed a decrease in integration as card complexity increased (blue connections in Figure 3-4). Our findings also suggest that the angular gyri were involved in both types of dynamics, as well as in cross-network integration (i.e., DMN-control networks, Figure 3-3). Note that there was also a small set of connections within the network that did not show a consistent increase or decrease in PPI connectivity across increasing reasoning complexity (see Appendix B).

3.6 Discussion

The aim of this study was to assess changes in the functional interplay between brain areas showing increased or decreased neural activity as a function of increasing complexity in a well-known cognitive reasoning task. Specifically, we tested whether performance is underpinned by a consistent

52 anticorrelation between regions within the default-mode and control networks, or, alternatively, whether emergent integration between these networks arises in response to increasing relational complexity.

Task-induced reductions (‘deactivations’) in neural activity as a function of cognitive load have been consistently observed in brain areas within the default-mode network (Lawrence et al., 2003; McKiernan et al., 2003; Sonuga-Barke and Castellanos, 2007). Such context-driven deactivations are thought to accompany shifts in attentional focus between self-directed mental activity and external stimuli and tasks. Consistent with this notion, it has been shown that deactivation within default-mode regions becomes more pronounced when interoceptive thoughts are reduced and task difficulty increases (Lawrence et al., 2003; McKiernan et al., 2006; McKiernan et al., 2003). Likewise, individuals with attentional problems show reduced segregation (anticorrelation) between regions of the control and default-mode networks (Chabernaud et al., 2012; Cocchi et al., 2012). Together, these observations have led to the view that segregation or antagonism between default-mode and task- positive regions is a predictor of optimal cognitive performance (Anticevic et al., 2012; Fox et al., 2005; Sonuga-Barke and Castellanos, 2007). Recent findings have challenged this view, however, suggesting instead that in some task contexts performance may be supported by transient cooperation - or reduced antagonism - between regions that are otherwise segregated in a state of rest (Fornito et al., 2012; Leech et al., 2012; Leech et al., 2011; Liang et al., 2015). The current study further adds to this literature by showing that increasing cognitive demands in a reasoning task is accompanied by a loss of segregation and a progressive enhancement of connectivity between regions within control and default-mode networks. These changes co-occurred alongside an increased functional interplay between fronto-parietal and cingulo-opercular regions (Cocchi et al., 2014).

The current results show that parametric increases in task complexity proportionally enhance integration between regions demonstrating either a local increase (striatum and thalamus) or decrease (angular gyri and right temporal pole) in neural activity. The pattern of network-level connectivity revealed here is consistent with previous results suggesting that the striatum and the thalamus are functionally and anatomically related to cortical structures supporting cognitive control functions (Haber, 2003; Haber and Knutson, 2010). Increased striatal activity is critical to the performance of cognitive control tasks (Bunge and Wright, 2007; Mestres-Misse et al., 2012). Likewise, striatal activity appears important for rapidly linking acquired representations with specific actions (Bunge et al., 2005; Pasupathy and Miller, 2005). Our findings go beyond these observations by suggesting that the striatum and the thalamus are key relays for facilitating enhancement of task-related processes, as well as suppression of task-unrelated processes. This hypothesis is consistent with recent evidence that the caudate and thalamus are transitional nodes that change their coupling with control and default-mode networks as a function of task demands (Dwyer et al., 2014). Although the PPI method does not allow an unequivocal assessment of causal interactions between distinct neural populations,

53 our results suggest that deactivation of specific default-mode regions may drive the initiation of fronto-parietal and cingulo-opercular activity via the basal ganglia (Table 3-2). Likewise, increased functional integration between default-mode areas and the basal ganglia may be essential to manage the integration of self-referential processes during task performance.

Greater task complexity was also related to increased connectivity between the right rostrolateral prefrontal cortex and the angular gyrus, bilaterally. The rostrolateral prefrontal cortex is involved in delayed, context-specific implementation of acquired rules (Gilbert, 2011; Koechlin et al., 1999; Sakai and Passingham, 2003). We have suggested that, alongside neurons within the anterior insular cortex (Menon and Uddin, 2010), neurons within this anterior prefrontal region play an important role in managing the task-based interplay between fronto-parietal and cingulo-opercular networks (Cocchi et al., 2014). While inferences on pairwise changes in connectivity need to be interpreted cautiously, the current findings suggest that the right rostrolateral prefrontal cortex might be involved in managing the engagement and disengagement of diffuse patterns of activity within the default-mode network. In line with this proposition, when only correct trials were included in our analyses, connectivity changes between the rostrolateral prefrontal cortex and the angular gyri were absent. Other than this discrete change in connectivity, the topology of the resulting network was identical to the network isolated when including both correct and error trials. This result is unlikely to be due to the small reduction in statistical power, as participants maintained mean accuracy above 80% in all card conditions. Rather, it suggests that increased integration between the rostrolateral prefrontal cortex and default-mode regions as a function of increased task complexity is related to contamination of task-unrelated activity with task-relevant processes (Christoff et al., 2009; Weissman et al., 2006). As such, default-mode interference with rostrolateral prefrontal cortex activity may cause, or be the consequence of, increased difficulties in efficiently managing the cingulo-opercular (task-set) and fronto-parietal (trial-by-trial control) dynamics. These results should encourage further investigation of the task-based interplay between default-mode and rostrolateral prefrontal cortex as a function of cognitive load.

Previous studies have suggested that cognitive task performance is accompanied by an increase in integration between default-mode regions (Hampson et al., 2006). The generalizability of these findings has been challenged by evidence suggesting that the default-mode network may be comprised of sub-networks that have distinct functional roles in different task contexts (Andrews- Hanna et al., 2010). Our analyses support the latter hypothesis by showing that functional integration within regions of the default mode network may either increase or decrease as a function of task complexity. Our findings also highlight a role for the angular gyrus in managing complexity-based integration and segregation within the default-mode network, and between the default-mode and control networks.

54 In summary, the current study highlights the dynamic functional interplay between default-mode and control networks as a function of increased reasoning complexity. Our findings challenge the notion that functional segregation between these networks invariably supports cognitive task performance. By contrast, the results provide further evidence in favor of recent accounts which suggest that transient cooperation between regions encompassing default-mode and control networks is critical to performing challenging cognitive tasks (Cocchi et al., 2013). We have also shown that the striatum and thalamus play an important part in managing interactions between control and default-mode processes during cognitive reasoning.

55

Chapter 4: The Latin Square Task as a measure of relational reasoning: a replication and assessment of reliability

56 4.1 Preamble

This chapter of my thesis represents the beginning of a series of experiments that were run concurrently with those presented in Chapters 2 and 3. In this series of experiments I tested the reliability and suitability of a relational reasoning task for a planned fMRI experiment in healthy adult participants (Chapter 5). I also used the task to examine brain networks associated with complex reasoning in individuals with dysgenesis of the corpus callosum (Chapter 6).

4.2 Abstract

The Latin Square Task (LST) is a relational reasoning paradigm developed by Birney and colleagues (Birney et al., 2006). The LST is informed by relational complexity theory, which posits that the number of relations between variables predicts the complexity of a problem, regardless of the domain of the original stimulus (e.g., semantic, spatial, temporal, etc.). Previous work has shown that the LST elicits typical reasoning complexity effects, such that increases in complexity are associated with decrements in task accuracy and increases in response times. Moreover, accuracy on the LST correlates with scores on measures of fluid intelligence such as Raven’s Progressive Matrices (Birney and Bowman, 2009). Here we modified the LST for use in functional brain imaging experiments, in which presentation durations must be strictly controlled, and assessed its validity and reliability. Modifications included presenting the components within each trial serially, such that the reasoning and motor response periods were separated. In addition, the inspection time for each LST problem was constrained to five seconds. We replicated previous findings of higher error rates and slower

2 response times with increasing relational complexity, and observed relatively large effect sizes (n p > 0.70, r > 0.50). Moreover, two reliability analyses confirmed that individual performance on the LST is stable across separate testing sessions. Interestingly, we found that limiting the inspection time for individual problems in the LST had little effect on accuracy relative to the unconstrained times used in previous work, a finding that is important for future brain imaging experiments aimed at investigating the neural correlates of relational reasoning.

4.3 Introduction

Relational reasoning refers to the ability to actively link and manipulate multiple mental representations (Halford et al., 1998). Increased complexity in relational reasoning is associated with reductions in task accuracy and increases in response times (Birney et al., 2006; Birney and Bowman, 2009). Critically, individual differences in reasoning ability are strongly linked with measures of fluid intelligence (Bhandari and Duncan, 2014; Birney and Bowman, 2009; Duncan et al., 2008). Designing tasks that measure relational complexity that are based on theory, have sound psychometric properties, and that are straightforward to construct and administer, is challenging (Birney et al., 2012). Here we took a recently developed relational reasoning task (the Latin Square Task, Birney et

57 al., 2006) and modified it to increase its utility for use in a typical ‘event-related’ functional brain- imaging experiment.

Birney and colleagues (2006, 2009) developed the Latin Square Task (LST) as a non-verbal behavioural task in which reasoning complexity (as defined by relational complexity theory) and working memory load are varied (see Figure 4-1). Each LST puzzle involves the presentation of a four-by-four matrix that is populated with geometric shapes, blank spaces, and a single probe location indicated by a question mark. The participant is asked to solve for the target probe according to the rule that each shape can only occur once in each row and column (analogous to the game ‘Sudoku’).

Figure 4-1 Example items from the modified version of the Latin Square Task (LST). Each LST item is presented as a 4 x 4 grid. Within the grid there is a target cell denoted by a red question mark. The remaining cells either contain geometric shapes or are left blank. The shapes are organized within each grid to require two, three or four relations to be processed. a. In the above examples the correct response is ‘square’ for the binary problem, ‘cross’ for the ternary problem, and ‘cross’ for the quaternary problem. All these problems only require one processing step (see text for details). b. Step 2 problems require the participant to solve an initial relation (as in panel a), before solving a second relation. In this example both the initial and second relation are binary-level difficulty. The correct response is ‘cross’.

The items correspond to three incremental levels of complexity: binary, ternary and quaternary. Binary problems require integration of information across a single row or column. Ternary problems involve integration across a single row and column. Quaternary problems, which are the most complex, require integration of information across multiple rows and columns. A further way in which item difficulty is manipulated is through the magnitude of working memory load in each problem (referred to as processing steps). Each puzzle can be categorized as either a ‘step 1’ problem or a ‘step 2’ problem. Step 2 problems require solving one relation, followed by a second relation to find the target Thus, the participant must keep the answer to the initial relation in mind in order to solve the second, increasing the working memory load (see Figure 4-1b). Previous work has shown that accuracy decreases with increasing relational complexity, and also with step 2 relative to step 1 items (Birney et al., 2006; Birney and Bowman, 2009).

58 The LST has been used in several different participant groups, including children (Birney et al., 2006; Perret et al., 2011), stroke patients (Andrews et al., 2013) and healthy adults (e.g., Birney et al., 2012). Importantly, previous research has confirmed a link between LST performance and fluid intelligence scores. For example, Birney et al., (2012) reported a high correlation between overall LST performance and scores on Raven’s Advanced Progressive Matrices, a commonly used fluid intelligence test (APM, r = 0.63). Birney and Bowman (2009) also reported moderate correlations between each level of relational complexity in the LST (binary, ternary, quaternary) and measures of fluid intelligence.

In the original studies reported by Birney and colleagues, participants viewed each problem and produced an answer at their leisure. For example, Birney and Bowman (2009) found that a group of university students of average fluid intelligence responded to binary, ternary and quaternary problems after 11, 21, and 35 seconds, respectively. Here we adapted the LST so that items were presented for a relatively short duration (just 5 seconds each), and responses were indicated via a button press after a brief delay. We undertook this important modification of the original task so that it would be better suited to delivery in the context of a functional brain imaging experiment. Based on the findings of Birney and Bowman (2009), we expected that constraining the time spent on each problem would reduce accuracy in the task, particularly in the higher complexity (ternary and quaternary) conditions. We also hypothesized that our modified LST would produce typical complexity effects, such that higher relational complexity and processing steps would be associated with lower accuracy on the task. Moreover, we predicted that there would be a significant positive correlation between performance on the LST and scores on a test of fluid intelligence. Forty-one participants completed the modified version of the LST paradigm and the Advanced Raven’s Progressive Matrices (APM). We also invited half the participants back to the lab after eight months to undertake the LST a second time, with the goal of assessing the task’s test-retest reliability.

4.4 Method

4.4.1 Participants

Forty-one participants were included in the experiment (Mage = 23.02, SD = 3.15, range = 19-34, 26 female). Data from two participants were excluded as their performance on the APM test was more than two standard deviations below the group average. Participants included staff and students of The University of Queensland (average years of education = 16.4) and were reimbursed $20AUD for volunteering. The University of Queensland Human Research Ethics Committee approved the study.

Eight months after the testing session (mean = 8.4 months), 20 participants returned to the lab to undertake the LST a second time. One participant was excluded due to a testing error. These participants were reimbursed a further $20AUD for their time.

59 4.4.2 LST items and presentation

The LST items used in the experiment were modified versions of those created by Birney and colleagues (2006, 2009; used with permission). In addition to the original items, each item was ‘copied’ by rotating the positions of the elements in the matrix by 90 and 180 degrees, generating three versions of each 36 items (108 items in total). Within the 108 trials there were three levels of complexity (binary, ternary, quaternary: 36 trials each), and two levels of processing steps (step 1, step 2: 54 trials each), yielding a three by two factorial design (18 trials per cell).

A further 36 ‘null’ trials were also included, to yield a total of 144 trials in the experiment. The null trials involved presentation of the standard 4 x 4 matrix, but instead of a target question mark (‘?’) an asterisk was presented (‘*’) to signal that no reasoning was required in this problem. The identity of the shapes that appeared in null trials was random, but the number of shapes and their spatial locations were matched to those included in the ‘active’ LST trials. Null trials matched all the visual characteristics of the standard problems, but with no relational reasoning demands, and were included for a later fMRI experiment (reported in Chapter 5). Administration of the items was pseudo- randomized such that no two items of the same complexity occurred in succession.

4.4.3 Procedure

Participants completed the APM in an initial session, with a 40-minute time limit. In a second session they completed a short, 12-problem practice of the LST, followed by the main LST task. Feedback on accuracy was given after each block. The LST was presented in 18 blocks of eight problems using MATLAB software (The Mathworks, Inc., Natick, MA) running the Psychophysics Toolbox (Brainard, 1997; Kleiner et al., 2007; Pelli, 1997). Each trial contained a variable fixation period, followed by an LST item. This was followed by a second, variable fixation period, a response screen, and a confidence-rating screen (see Figure 4-2 for trial structure). The three-point confidence scale probed whether participants felt certain the problem had been answered correctly (3), felt unsure of their accuracy (2), or felt certain the problem had been answered incorrectly (1). Analysis of confidence ratings was not considered in the current work.

4.4.4 Statistical analyses

Two-way repeated-measures ANOVAs were conducted with factors of complexity (binary, ternary, quaternary) and processing steps (step 1, step 2) for both accuracy and reaction time data. Bonferroni corrected follow up t-tests were performed for accuracy scores. Two reliability analyses were performed. First, a split-half reliability analysis was conducted. This involved randomly splitting data from each condition into equal partitions for each individual, and then correlating the mean score across individuals for each condition. This operation was undertaken 1000 times, permuting the random partitions for both accuracy and reaction times. Second, test-retest reliability was performed using a Spearman ranked correlation on overall LST accuracy scores from the initial test and the

60 eight-month follow-up.

Figure 4-2 Example LST trial sequence. Each trial contained a jittered fixation period, followed by an LST item, a second jittered fixation period, a response screen, and a confidence rating scale. In Null trials the motor response screen had one geometric shape replaced with an asterisk, which indicated to participants the appropriate button to press.

4.5 Results

4.5.1. Behaviour associated with increasing relational complexity

The range of APM scores captured in our sample (M = 24.90, SD = 4.60, range = 14-33) was comparable to that reported in previous work (Birney and Bowman, 2009; Bors and Stokes, 1998). On average, participants completed the test in 30.14 minutes (SD = 7.65). The use of parametric statistics was justified as the normality assumption of the LST and APM data was not violated (Kolmogorov- Smirnov test, p = 0.16 and 0.20, respectively).

Consistent with previous work (Birney and Bowman, 2009), a significant positive correlation was found between the modified LST and APM scores, such that higher fluid intelligence scores were correlated with higher scores on the LST, r = 0.55, p < 0.001 (see Figure 4-3). A repeated measures ANOVA with factors of complexity and processing steps revealed a significant main effect of complexity, confirming that as relational complexity increased, accuracy decreased (F2, 80 = 113.25, 2 n p = 0.74, p < 0.001, see Figure 4-3b). In addition, there was a significant main effect of the number of processing steps, such that participants were less accurate for step 2 problems than for step 1

2 problems (F1, 40 = 248.42, n p = 0.86, p < .001). There was no significant interaction between the

61 2 factors of complexity and processing steps (F2, 39 = 0.84, n p = 0.02, p = 0.44).

Figure 4-3 Behavioural performance on the Latin Square Task (LST), and relationship with fluid intelligence as measured with the Advanced Progressive Matrices (APM). a. Correlation between performance on the LST and APM. b. Accuracy on the LST as a function of reasoning complexity and processing steps. c. Reaction time on the LST as a function of reasoning complexity and processing steps. Error bars represent 95% confidence intervals.

Bonferroni-corrected follow-up pairwise t-tests revealed that accuracy was higher for binary items than for both quaternary items (p < 0.001, see Table 4.1 for descriptive statistics) and ternary items (p = 0.017, marginally significant; the corrected significance threshold is 0.017 for this comparison). Accuracy was also higher for ternary items than for quaternary items (p < 0.001).

A significant main effect was also found for reaction time, such that participants took longer to

2 respond to problems with higher relational complexity (F2, 80 = 110.34, n p = 0.73, p < 0.001, see Figure 4-3c). Reaction times for step 2 problems were also slower when compared with those for step

2 1 problems (F1, 40 = 79.20, n p = 0.66, p < 0.001). Finally, there was a significant interaction between 2 relational complexity and processing steps (F2, 80 = 17.31, n p = 0.30, p < 0.001). Follow up analysis revealed this effect was driven by a larger difference in step 1 and step 2 Binary items, compared with Ternary (p < 0.001) and Quaternary (p = 0.001).

62 Table 4-1 Mean accuracy and response times for the Latin Square Task Accuracy (%) Response Time (s) Condition Mean SD Mean SD LST 71.34 9.89 0.98 0.23 Binary 83.20 12.15 0.88 0.18 S1 94.85 6.84 0.76 0.19 S2 71.49 17.44 1.00 0.17 Ternary 77.71 14.39 0.94 0.19 S1 86.99 10.80 0.91 0.19 S2 68.43 17.98 0.97 0.19 Quaternary 53.12 17.38 1.13 0.20 S1 66.12 23.33 1.03 0.21 S2 40.11 11.42 1.22 0.18 Note: S1 = one processing step, S2 = two processing steps.

4.5.2. Reliability of the Latin Square Task

To investigate the reliability of the LST, a split-half reliability analysis was performed. Median Pearson r and p values across the 1000 permutations are shown in Table 4-2 (see Method for details). Broadly speaking, there was high reliability across all conditions with the exception of step 2- quaternary items. Reliability scores for accuracy were generally lower than those for reaction times. Finally, to examine the test-rest reliability of the LST, 19 participants completed the task again eight months after the initial session. Overall, LST accuracy scores between the two sessions showed a high correlation (Spearman’s rank r = 0.75, p < 0.001), suggesting that performance on the LST is relatively stable over time.

Table 4-2 Split-half reliability of the Latin Square Task

Accuracy Reaction Time Condition r p r p Binary S1 0.31 0.046 0.81 < 0.001 Binary S2 0.51 < 0.001 0.53 < 0.001 Ternary S1 0.34 0.031 0.67 < 0.001 Ternary S2 0.50 < 0.001 0.63 < 0.001 Quaternary S1 0.60 < 0.001 0.55 < 0.001 Quaternary S2 0.31 0.65 0.34 0.029 Note: S1 = one processing step, S2 = two processing steps.

63 4.6 Discussion

Relational reasoning is the ability to actively link and manipulate multiple mental representations (Halford et al., 1998). This foundational ability supports many higher cognitive functions, including abstract thought, problem solving and decision-making (Crone et al., 2009; Halford et al., 2010). Critically, relational reasoning task performance has been linked to intelligence (Birney et al., 2006; Birney and Bowman, 2009; Duncan et al., 2008; Halford et al., 1998). Designing valid and reliable behavioural tasks that systematically manipulate the number of relations in a given problem is a difficult endeavour (Birney et al., 2012). In the current work we aimed to investigate the replicability and reliability of a modified version of the LST (Birney et al., 2006; Birney and Bowman, 2009). To do so, we tested a group of young adult participants on a novel version of the LST and a separate measure of fluid intelligence (Advanced Raven's Progressive Matrices, APM), and examined split- half and test-retest reliability.

In contrast to previous studies of the LST, we constrained the problem-solving time to five seconds. We expected that this might reduce overall accuracy on the task relative to the open-ended administration of previous studies (e.g., Birney et al., 2006; Birney and Bowman, 2009), particularly for the highest complexity problems. In addition, we hypothesized that typical relational complexity and processing step (working memory) effects would be evident, such that accuracy would decline and reaction times would increase with increments in complexity and working memory load.

2 We replicated both complexity and processing-step effects on accuracy with large effect sizes (n p > 0.80). A similar pattern was evident for participants’ reaction times, such that response times were more prolonged with increasing complexity. Moreover, we replicated previous findings of a significant positive correlation between LST accuracy and fluid intelligence (Birney et al., 2012; Birney and Bowman, 2009). Participants performed somewhat more poorly on step 2 problems relative to previous studies in which time constraints were not placed on participants (Birney and Bowman, 2009). However, this difference was surprisingly small considering the discrepancy in participants’ response times. In Figure 4-4 we present a comparison of mean accuracy of participants in the current experiment and those reported in the study of Birney and Bowman (2009), who tested the LST in a similar cohort of young adult University of Queensland students. For the quaternary problems, there was approximately a one-percent difference in accuracy between the current study, in which participants had five seconds to consider each problem, and the study of Birney & Bowman (2009), in which participants required, on average, 35 seconds to register their answers for each problem.

64

Figure 4-4 Comparison of mean accuracy on the LST in the current experiment, and the accuracy values reported in Birney and Bowman (2009). Birney and Bowman collected LST data with no time restriction in a group of adult university students. Performance on step 1 problems was virtually identical for the two studies, and followed a similar overall pattern for step 2 problems.

We used two complementary analyses to assess reliability. First, for both accuracy and reaction times, we found that the internal consistency of each condition was high, with the exception of the most difficult condition (quaternary, step 2). Second, we observed that the LST was highly reliable across time. Although the test-retest analysis was likely somewhat statistically underpowered (N = 19), the actual correlation value was relatively large (rs = 0.75). Further work in larger samples may be beneficial in the future.

To conclude, using a modified version of the LST, we replicated key relational complexity effects and their association with fluid intelligence. Moreover, the LST was highly reliable across two complementary tests of statistical reliability. A new and interesting finding was that performance accuracy was very similar here as in previous work, even though we limited the time available to solve each problem to just 5 seconds (see Figure 4-4). This observation begs further investigation, particularly regarding the role of processing speed in fluid intelligence (Conway et al., 2002). Our development of a time-restricted, reliable version of the LST may be of interest to clinicians who wish to examine cognitive reasoning efficiently in patient groups (e.g., Andrews et al., 2013), and to those undertaking brain imaging investigations in which short trial durations permit the completion of more repetitions in the scanner.

65

Chapter 5: Reconfiguration of brain network architectures between resting-state and complexity- dependent cognitive reasoning

66 5.1 Preamble

After developing the modified Latin Square Task in Chapter 4, we conducted a large fMRI study using the same paradigm. The work contained in this chapter has been published in the Journal of Neuroscience (Hearne et al., 2017).

5.2 Abstract

Our capacity for higher cognitive reasoning has a measureable limit. This limit is thought to arise from the brain’s capacity to flexibly reconfigure interactions between spatially distributed networks. Recent work, however, has suggested that reconfigurations of task-related networks are modest when compared with intrinsic ‘resting-state’ network architecture. Here we combined resting-state and task- driven functional magnetic resonance imaging to examine how flexible, task-specific reconfigurations associated with increasing reasoning demands are integrated within a stable intrinsic brain topology. Human participants (21 males and 28 females) underwent an initial resting-state scan, followed by a cognitive reasoning task involving different levels of complexity, followed by a second resting-state scan. The reasoning task required participants to deduce the identity of a missing element in a 4 x 4 matrix, and item difficulty was scaled parametrically as determined by relational complexity theory. Analyses revealed that external task engagement was characterized by a significant change in functional brain modules. Specifically, resting-state and null-task demand conditions were associated with more segregated brain network topology, whereas increases in reasoning complexity resulted in merging of resting-state modules. Further increments in task complexity did not change the established modular architecture, but impacted selective patterns of connectivity between fronto- parietal, subcortical, cingulo-opercular and default-mode networks. Larger increases in network efficiency within the newly established task modules were associated with higher reasoning accuracy. Our results shed light on the network architectures that underlie external task engagement, and highlight selective changes in brain connectivity supporting increases in task complexity.

5.3 Introduction

Humans are unparalleled in their ability to reason and solve complex problems in the service of goal- directed behaviour (Johnson-Laird, 2010; Penn et al., 2008). Nevertheless, our ability to reason successfully is limited by the complexity of the task at hand (Halford et al., 1998, 2005). Increasing reasoning demands are supported by the flexible reconfiguration of large-scale functional brain networks (Cocchi et al., 2014a, 2013), but recent work has demonstrated that such reconfigurations are relatively modest and occur within a preserved global network architecture (Cole et al., 2014; Krienen et al., 2014). Here we assessed changes in functional brain architecture induced by engagement in a complex reasoning task, as well as changes in communication across regions with parametric increases in reasoning complexity. To do so, we used high-field functional magnetic

67 resonance imaging (fMRI) to measure brain activity at rest, and during performance of a behavioural task in which task complexity was manipulated parametrically.

Higher cognitive functions are supported by the adaptive reconfiguration of large-scale functional networks (Bassett et al., 2011; Braun et al., 2015; Cohen and D’Esposito, 2016; Cole et al., 2013; Yue et al., 2017). Previous empirical and theoretical work suggests that a multitude of complex tasks are related to activity and communication within and between select fronto-parietal, cingulo-opercular, and default-mode networks (Bolt et al., 2017; Cocchi et al., 2014a; Crittenden et al., 2016; Hearne et al., 2015; Knowlton et al., 2012). Such networks are flexible and tend to increase their functional relationship in line with task demands across a wide range of domains, including reasoning (Cocchi et al., 2014a), working memory (Vatansever et al., 2017) and decision making (Cole et al., 2013).

Recent empirical work has shown that task-induced network reconfigurations are modest when compared with intrinsic, ‘resting-state’ networks (Cole et al., 2014; Krienen et al., 2014). For example, Cole and colleagues reported a matrix-level correlation between rest and task states of r = 0.90 (on average 38% of connections demonstrated change, with an average change of r = 0.04). Likewise, it is now apparent that task-induced activity can be well predicted and modelled from resting-state data alone (Cole et al., 2016; Tavor et al., 2016). These results suggest that while behaviourally meaningful, selective task-induced reconfigurations occur against a backdrop of stable, large-scale networks that support diverse cognitive functions (Crossley et al., 2013; Power et al., 2011). An important unresolved question is how selective, ‘flexible’ task driven reconfigurations emerge amongst ‘stable’ intrinsic brain topology. Moreover, it is critical to understand how such global and selective changes are related to behaviour(Bolt et al., 2017; Mill et al., 2017).

To investigate this question we measured functional brain networks at rest, as well as during several discrete levels of reasoning complexity. To systematically manipulate task complexity, we exploited relational complexity theory (Halford et al., 1998), which posits that the number of relations between variables quantifies the complexity of a problem, regardless of the domain of the original stimulus (e.g., semantic, spatial, etc.). Using this theoretical framework it has been shown that increasing the number of relations imposes a quantifiable cognitive load (measured via reaction time and accuracy), and eventually results in a breakdown of the reasoning process (Halford et al., 2005). We collected 7T fMRI data from 65 individuals while they undertook a non-verbal reasoning task known as the Latin Square Task (Birney et al., 2006). During the task, participants solved problems with three discrete levels of difficulty, defined formally in terms of their relational complexity (Binary, Ternary, Quaternary). In addition, just prior to the task, and again immediately afterwards, participants underwent a resting-state scan. To examine network reconfigurations across rest and reasoning states we utilized modularity to assess segregation and integration and global efficiency to assess changes in network communication. Further to examine selective changes, we employed the network-based

68 statistic to identify circumscribed changes in connectivity patterns (Zalesky et al., 2010), and related such network metrics to behaviour.

5.4 Materials and Methods

5.4.1 Participants

Sixty-five healthy, right-handed participants undertook the current study, of whom 49 were included in the final analysis (M = 23.35 years, SD = 3.6 years, range = 18 – 33 years, 28 females). Four participants were excluded due to MR scanning issues, one participant was excluded due to an unforeseen brain structure abnormality, a further participant was excluded due to low accuracy in the behavioural task (total score more than 3 standard deviations below the mean) and ten participants were excluded due to excessive head movement (see pre-processing section for head movement exclusions). Participants provided informed written consent to participate in the study. The research was approved by The University of Queensland Human Research Ethics Committee.

5.4.2 Experimental paradigm

Each participant completed two behavioural sessions and one imaging session. In the imaging session participants underwent a resting-state scan, followed by three, 12-minute runs of the Latin Square Task (LST; described below), a structural scan and finally a second resting-state scan (see Figure 5- 1a).

In the two behavioural sessions, participants completed the Raven’s Advanced Progressive Matrices (40 minute time limit), which is a standard and widely used measure of fluid intelligence (Raven, 2000). Of the 49 participants included in the analysis, 43 also completed a conjunction visual search task in which they were instructed to report the orientation of a target letter ‘L’ (rotated 90 degrees leftward or rightward) amongst ‘T’ distractors in set sizes of 8, 16 or 24 items. The search cost was defined as the increase in reaction time between the smallest and largest set sizes. This task was chosen as a ‘low reasoning’ counterpart to the Raven’s Progressive Matrices in order to demonstrate the specificity of brain-behaviour correlations, as described in detail in the Results.

Participants also completed a modified version of the LST (Birney et al., 2006; LST, Birney and Bowman, 2009). The LST is a non-verbal relational reasoning paradigm in which reasoning complexity is parametrically varied with minimal working memory demands (Halford, 1998, Birney et al., 2006). Each LST ‘puzzle’ involves the presentation of a four-by-four matrix populated with a small number of geometric shapes (square, circle, triangle or cross), blank spaces and a single target, denoted by a yellow question mark (‘?’; see Figure 5-1b). Participants were asked to solve for the target according to the rule that each shape can only occur once in every row and once in every column (similar to the game of Sudoku). Binary problems require integration of information across a single row or column. Ternary problems involve integration across a single row and column.

69 Quaternary problems, the most complex, require integration of information across multiple rows and columns (see Figure 5-1b for examples of each of these problems). Null trials involved presentation of an LST grid, but instead of a target question mark (‘?’) an asterisk was presented (‘*’) to cue the participant that no reasoning was required in this puzzle. The identity of the shapes that appeared in Null trials was random, but the number of shapes and their spatial locations were matched to those in the active LST trials. In total, 144 LST items were presented in the MR session across 16 blocks, with 36 items in each relational complexity condition; Null, Binary, Ternary and Quaternary. Prior to the MR session participants completed 20 practice trials of the LST (12 with corrective feedback). The visual angle subtended by the LST matrices was ~7.7 degrees, so that the entire stimulus fell within the parafoveal region of the visual field. Stimuli were projected onto a screen located at the head end of the MR scanner, and participants viewed the projected stimuli via a mirror mounted on the head coil.

Administration of all items was pseudo-randomized such that no two items of the same complexity occurred sequentially, and each block had two problems from each level of complexity (see Figure 5- 1c for trial structure). Motor responses were counterbalanced across individuals, such that equal numbers of participants had the same shape-response mapping. Confidence ratings were used to determine participants’ subjective feeling of success, and to identify any trials in which participants inadvertently disengaged from the task altogether (e.g., due to a momentary lapse of attention). A three-point confidence scale indicated whether participants felt certain the problem had been answered correctly (4), felt unsure of their accuracy (3), or felt certain the problem had been answered incorrectly (2). On the far left (demarcated by a vertical line, see Figure 5-1c) was an additional ‘inattention’ rating point (1) that participants were instructed to select if they felt they had not attempted to solve the problem due to a momentary lapse of attention, fatigue or other factors. This response was used to separate incorrect choices arising from failures in reasoning, from those due to non-specific “off-task” mind wandering (Smallwood and Schooler, 2015).

70

Figure 5-1 Experimental design and sequence of displays in a typical trial of the Latin Square Task. a. Functional magnetic resonance imaging session outline. Participants completed resting-state scans before and after three runs of task imaging. b. Examples of each reasoning complexity condition. The correct answers are square, cross and cross, respectively, for the Binary, Ternary and Quaternary problems illustrated. c. Example trial sequence. Each trial contained a jittered fixation period, followed by an LST item, a second, jittered fixation period, a response screen, and a confidence rating scale. In Null trials the motor response screen had one geometric shape replaced with an asterisk, representing the correct button to press.

5.4.3 Neuroimaging acquisition and pre-processing

Imaging data were collected using a 7 Tesla Siemens MR scanner fitted with a 32-channel head coil, at the Centre for Advanced Imaging, The University of Queensland. For both resting-state and task fMRI, whole brain echo-planar images were acquired using a multi-band sequence (acceleration factor of five; Moeller et al., 2010). In each of the two resting scans, 1050 volumes were collected

71 (~10 minutes each). In the each of the three runs of the task, 1250 volumes were collected (~12 minutes each) with the following parameters: voxel size = 2 mm3, TR = 586 ms, TE = 23 ms, flip angle = 40°, FOV = 208 mm, 55 slices. Structural images were also collected to assist functional data pre-processing. These images were acquired using the following parameters: MP2RAGE sequence, voxel size = 0.75 mm3, TR = 4300 ms, TE = 3.44 ms, 256 slices (see Figure 5-1a for session structure).

Imaging data were pre-processed using an adapted version of the MATLAB (MathWorks, USA) toolbox Data Processing Assistant for Resting-State fMRI (DPARSF V 3.0, Chao-Gan and Yu-Feng, 2010). Both resting-state and task data were pre-processed with the same pipeline (except where noted). DICOM images were first converted to Nifti format and realigned. T1 images were re- oriented, skull-stripped (FSL BET), and co-registered to the Nifti functional images using statistical parametric mapping (SPM8) functions. Segmentation and the DARTEL algorithm were used to improve the estimation of non-neural signal in subject space and the spatial normalization (Ashburner, 2007). From each gray matter voxel the following signals were regressed: undesired linear trends, signals from the six head motion parameters (three translation, three rotation), white matter and cerebrospinal fluid (estimated from single-subject masks of white matter and cerebrospinal fluid). The CompCor method (Behzadi et al., 2007) was used to regress out residual signal unrelated to neural activity (i.e., five principal components derived from noise regions-of-interest in which the time series data were unlikely to be modulated by neural activity). Global signal regression was not performed due to the ongoing controversy associated with this step (Caballero-Gaudes and Reynolds, 2017; Saad et al., 2012). This choice may increase motion artifacts in the data (Ciric et al., 2017). For this reason we employed a strict head motion censoring approach (see below). Single-subject functional images were subsequently normalized and smoothed using DARTEL (4mm3). Data processing steps also involved filtering (0.01–0.15 Hz) at a low frequency component of the BOLD signal known to be sensitive to both resting-state and task-based functional connectivity (Sun et al., 2004), therefore allowing comparison of both resting-state and task data.

Head movement

Participants with head displacement exceeding 3mm in more than 5% of volumes in any one scan were excluded. In addition to gross head movement, it has also been shown that functional connectivity can be influenced by small volume-to-volume ‘micro’ head movements (Jonathan D Power et al., 2014; Van Dijk et al., 2012). To ensure micro-head movement artifacts did not contaminate our findings, both resting-state and task-based data with frame-to-frame displacements greater than 0.40 mm were censored (Power et al., 2014). Participants with less than 85% of data remaining in any condition were excluded.

72 5.4.4 Functional connectivity network construction

For each subject, regionally averaged time series were extracted for 264 spheres of 5mm radius sampled across cortical and subcortical gray matter. Spheres were positioned according to an existing brain parcellation, based on task activations induced by a wide range of behavioural tasks (Power et al., 2011). This parcellation and associated network definitions were generated from a large cohort of participants (N > 300), and has the advantage of being independent of the imaging data obtained in the current study.

For both sets of resting-state data (pre- and post-task), functional connectivity was estimated using a temporal Pearson correlation between each pair of time series (Zalesky et al., 2012b). This resulted in a 264 × 264 connectivity matrix for each subject. For the task-based functional connectivity analyses we used a regression approach (i.e., Cole et al., 2014) rather than psycho-physiological interactions (PPI) as others have used previously (Cocchi et al., 2014a; Gerchen et al., 2014; McLaren et al., 2012). We opted for this approach rather than PPI due to our interest in assessing connectivity across both rest and task states. For each brain region of interest, a task regressor composed of the condition onsets modelled as boxcar functions convolved with a canonical hemodynamic response function was regressed from the time series. This step was taken to remove variance associated with task-related coactivation (Cole et al., 2014). Then, after accounting for the hemodynamic lag, the residual time series from each five-second reasoning period was concatenated to form a condition-specific time series of interest, in each brain region. A Pearson correlation was performed on the resulting regional time series for each condition separately resulting in a 4 (condition) x 264 x 264 connectivity matrix for each subject. Finally, both resting-state and task-based matrices were converted to z-scores. Analysis decisions such as z-normalization and thresholding were employed so as to be consistent with previous, related work aimed at assessing dynamic reconfiguration of connectivity patterns as a function of task demands (e.g., Cole et al., 2014, Power et al., 2011). Such choices do, however, affect the resulting graph metrics (Rubinov and Sporns, 2011). Thus, unless otherwise noted (see network based statistic analysis, below) weighted graphs of proportional densities from the top 5% to the top 30% of connections were considered for analysis. Such network densities have been shown to provide robust functional brain network characterizations (Garrison et al., 2015) and are similar to those used in previous, related work (e.g., Power et al., 2011).

5.4.5 Analysis overview

We undertook three complementary analyses to identify functional network reorganization due to increasing relational complexity. First, we calculated and compared community partitions that arose in each of the resting-state and task conditions. Following this, we performed an analysis to identify changes in connectivity associated with performance of the Latin Square Task using the network based statistic (NBS, Zalesky et al., 2010), a sensitive statistical tool that controls for Type I error at

73 the network level. To assess the functional and behavioural impact of the connectivity changes identified in the previous two analyses, we calculated changes in global efficiency (Achard and Bullmore, 2007) for each functional module detected. Moreover, to assess the behavioural implications of the observed network changes, we correlated metrics of changes in module efficiency with performance accuracy on the LST. When appropriate, nonparametric statistics were used for repeated-measures comparisons (Friedman test), follow-up tests (Wilcoxon signed rank) and measures of effect size (Kendall’s coefficient of concordance, W).

Community detection

A module is a group of nodes in a graph that contains stronger connections within-module than expected in an appropriate random network null model. A modularity partition represents the subdivision of a graph into non-overlapping modules (Fortunato, 2010). The degree of modularity in a network can be characterized by the Q index (Newman and Girvan, 2004), which represents the density of within-module connections relative to an appropriate random network null model. The aim of community detection is to isolate a module partition that maximizes Q.

1 '()) = / − )3 4(5 5 ) 2. 01 01 0 1 01

Above is the modularity equation, where a78 represents the weight of the edge between 9 and :, 301 = ; ; < = represents the expected number of links according to the so-called configuration null model %>

(Newman et al., 2001), where ?0 is the degree of node 9; in this case the null preserves the node degree while forming connections at random. 2. represents the total number of connections in the network, 50 denotes the community to which node 9 is assigned, the Kronecker delta function,

4(5051), is 1 if 50 = 51 and 0 if otherwise. Finally, ) is the resolution parameter; when ) < 1, larger communities are resolved, if ) > 1, smaller communities are resolved.

In the present study, modules were identified using the Louvain greedy algorithm (Blondel et al., 2008) implemented in the Brain Connectivity Toolbox (BCT, Rubinov and Sporns, 2010). The resolution parameter was set to unity (γ = 1). Testing across several levels of γ showed consistent results. For clarity, we highlight the BCT scripts used throughout the Method section. There are multiple possible module partitions that maximize Q for each graph, resulting in community assignments that vary across each run of the algorithm (Good et al., 2010; Sporns and Betzel, 2016). To resolve this variability we used a consensus approach (Lancichinetti and Fortunato, 2012), whereby module partitions are calculated a number of times (103 iterations, for each participant and condition) and used to calculate an agreement matrix (agreement.m). The agreement matrix represents the tendency for each pair of nodes to be assigned to the same module across iterations. Finally, the agreement matrix was subjected to an independent module partitioning (consensus_und.m), resulting

74 in an individual-level module partition for each participant in each condition. In this step, the resolution parameter was also set to unity (τ = 1), representing the level at which the agreement matrix was thresholded before being subjected to the consensus procedure. For example, τ = 1 thresholds the matrix such that only nodes consistently partitioned into the same community across all permutations are included. Testing across several levels of τ showed consistent results. A general community structure including motor-sensory, auditory, visual, default-mode and fronto- parietal/cingulo-opercular modules was entered into the algorithm as the initial community partition. In our data, this choice decreased computation time, presumably because the initial community structure was associated with a Q-value that was close to the true maximum. Module reconfiguration results were replicated using different community affiliation priors, including variations of the original community partitions (Cole et al., 2013; Power et al., 2011) and purely data-driven methods (i.e., no community affiliation input).

The procedure for group-level modular decomposition was implemented in a similar fashion to the individual-level decompositions described above. The critical difference was that instead of creating an individual-level agreement matrix, the agreement matrix represented the tendency for each pair of nodes to be assigned to the same module across participants. The same consensus procedure followed, resulting in a single module partition for each condition for the group of 49 participants. Resolution parameters were kept identical to the previous individual-level modularity analysis.

Significance testing for within-participant differences in modular structure

To investigate differences in the nodal composition of modules across conditions, we used the Variation of Information metric (VIn, Meilă, 2007), an information-theoretic measure of partition distance (partition_distance.m). To ascribe statistical significance to differences in partition structure we used a repeated-measures permutation procedure to compare real VIn values to appropriate null distributions (Dwyer et al., 2014). Specifically, half of the participants’ condition labels were randomly switched in the contrast of interest (e.g., Binary versus Ternary). This resulted in two new sets of individual-level module structures for the contrast (albeit with shuffled data). The shuffled module structures were then subjected to the previously used pipeline to generate group-level module partitions. Finally, VIn was used to quantify the difference between these partitions. This procedure was repeated 104 times to build a null distribution for each contrast of interest, with which the real data were compared.

Pairwise functional connectivity analysis

The Network Based Statistic (NBS, Zalesky et al., 2010) was used to identify changes in pairwise functional connectivity at rest and during the task. For the first contrast, a paired t-test was performed between the Pre- and Post-task resting-state data. For the second contrast, a one-way repeated measures ANOVA was used to compare all four task states (Null, Binary, Ternary, Quaternary). For

75 the analysis, unthresholded functional connectivity matrices were used as input into the NBS. Briefly, all possible pairs of connections (264 x 263/2 = 34,716) were tested against the null hypothesis, endowing each connection with a test statistic, which was subsequently thresholded. Here an exploratory F-statistic of 20 (equivalent to a t-statistic of 4.47) was used as the threshold, though additional exploratory analyses showed that networks arising using higher or lower t-thresholds resembled the original results. This threshold was adopted because it allowed the detection of effects of medium size while discarding small or spurious effects. Family-wise error corrected (FWE) p- values were ascribed to the resulting networks using a null distribution obtained by 5,000 permutations. Only components that survived a network-level threshold of p < 0.001 FWE were declared significant. This analysis allowed us to identify sub-networks that significantly increased or decreased their functional connectivity across relational reasoning task conditions, providing complementary results to the graph analyses. Results were also replicated using a summed t-statistic rather than component size (i.e., the intensity NBS).

Network efficiency analysis

Global efficiency is defined as the inverse of the average characteristic path length between all nodes in a network (Latora and Marchiori, 2001). Assuming that information follows the most direct path, global efficiency provides an index for parallel information transfer in a network (Rubinov and Sporns, 2010). In the context of functional brain networks, global efficiency is thought to be an index of increased capacity for information exchange (Achard and Bullmore, 2007). The link between indices of global efficiency and global neural information transfer is, however, not yet clear. Nevertheless, a number of studies have shown that high global brain network efficiency can enhance neurophysiological (Cocchi et al., 2017; de Pasquale et al., 2016) and cognitive processes (Bassett et al., 2009; Shine et al., 2016; van den Heuvel et al., 2009).

Here we wanted to investigate differences in network communication within module, and determine how such difference might relate to behaviour. To do so, we computed global efficiency for each participant, in each condition, for the three major modules identified in the initial modularity analysis (using efficiency_wei.m from the BCT). Importantly, matrix thresholding was performed after dividing the modules to ensure any efficiency effects were not due to differences in degree across modules. Finally, we computed the difference in efficiency between the most and least difficult conditions (i.e., Quaternary versus Null) and correlated this change in efficiency with overall accuracy scores on the Latin Square Task.

5.4.6 Figures and Visualization

Figures were generated with a combination of MATLAB, and online network visualization tools (alluvial diagram; (http://www.mapequation.org/apps/MapGenerator.html, and the connectogram; http://immersive.erc.monash.edu.au/neuromarvl/).

76 5.5 Results

5.5.1 Behavioural results

A nonparametric Friedman test revealed a significant effect of reasoning complexity on both LST accuracy (E% = 86.20, Kendall’s W = 0.88, p < 0.001) and reaction time (E% = 63.71, W = 0.65, p < 0.001, see Figure 5-2). Bonferroni corrected follow-up Wilcoxon signed-rank test comparisons revealed that accuracy was significantly higher for the Binary condition (M = 34.96, SD = 1.04) than for both the Ternary condition (M = 31.63, SD = 3.52, z = 5.59, p < 0.001) and the Quaternary condition (M = 22.78, SD = 6.71, z = 6.10, p < 0.001). Accuracy was also higher for Ternary items than for Quaternary items (z = 5.74, p < 0.001). The reaction time results followed a similar pattern, such that responses were faster in the Binary condition (M = 771.60 ms, SD = 201.60 ms) than in the Ternary (M = 844.20 ms, SD = 217.30 ms, z = -4.62, p < 0.001) and Quaternary (M = 933.70 ms, SD = 205.00 ms, z = -5.83, p < 0.001) conditions. Likewise, reaction times in the Ternary condition were significantly faster than those in the Quaternary condition (z = -4.79, p < 0.001).

Figure 5-2 Behavioural results for the Latin Square Task visualized as box and whisker plots. Here the boxes represent the median and interquartile ranges, and the whiskers show the minimum and maximum values. a. Accuracy as a function of reasoning complexity. b. Reaction time as a function of reasoning complexity. Significance markers indicate p < 0.001.

As expected, there was a significant positive correlation between scores on the Raven’s Advanced Progressive Matrices, a measure of fluid intelligence, and overall accuracy on the LST (r = 0.44, p = 0.002). By contrast, for the visual search task, there was no correlation between reaction time cost and LST score (N = 43, r = -0.09, p = 0.58). A Steiger z-test (Lee and Preacher, 2013) demonstrated that these two correlations were significantly different from one another (N = 43, z = -2.90, p = 0.003), confirming that LST performance is linearly related to an established measure of fluid intelligence, but not to a widely used test of visual attention (Triesman & Gelade, 1980).

77 Participants’ confidence was assessed on each trial. Importantly, we included an explicit rating for when participants had not attempted the reasoning problem due to an attention lapse, mind wandering, or fatigue. Averaging across all conditions, the mean number of such lapses was fewer than one out of 36 trials (M across conditions = 0.47 trials, SD across conditions = 0.98 trials). We can thus conclude that overall, participants were able to engage as instructed in cognitive reasoning across all three levels of relational complexity in the LST.

5.5.2 Functional brain module reconfiguration

Modularity analysis revealed four major modules in the baseline (Pre-task) resting-state. For clarity, these modules are represented in reference to Power and colleagues’ (2011) initial network affiliations. The modules broadly correspond to the sensory, default-mode, visual and fronto-parietal networks.

Variation of Information analysis (Meilă, 2007) revealed a significant difference in the community structure of Binary, Ternary and Quaternary conditions compared with the Pre-task resting-state (mean statistics across thresholds are reported in text; see Appendix C for extensive results, VIn = 0.20, p = 0.006, VIn = 0.21, p = 0.001, VIn = 0.20, p = 0.003, respectively) and Post-task resting-state (VIn = 0.18, p = 0.039, VIn = 0.20, p = 0.005, VIn = 0.19, p = 0.016). The difference between rest and reasoning states was associated with the emergence of a single, conjoined fronto-parietal-visual module that was composed of several large-scale networks identified by Power et al. (2011). The transitory nature of the fronto-parietal-visual module was confirmed by its switch back to its original configuration in the Post-task resting-state (i.e., after completion of the LST). Figure 5-3b shows a representation of the reconfiguration of modules across experimental conditions. There was no significant difference between the Pre-task and Post-task resting-state community structure (VIn = 0.09, p = 0.788). There was also no consistent difference between the Pre- and Post-task resting-state communities and the Null task condition (VIn = 0.16, p = 0.08, VIn = 0.15 p = 0.196, respectively). These main effects were broadly replicated across all thresholds tested (Figure 5-4a).

78

Figure 5-3 Modular structure as a function of reasoning complexity in the Latin Square Task. a. Alluvial ‘flow’ demonstrating the network affiliations (as per Power et al., 2011) compared with the Pre-task resting- state (Rosvall et al., 2009). Each individual streamline represents a node in the network, colored by its original resting-state affiliation as shown on the left (Power et al., 2011). b. Changes in modular structure across the experimental conditions. Visual and fronto-parietal modules merged to form a ‘Task-related’ module during Binary, Ternary and Quaternary conditions of the LST. Results for 15% network density are shown, but statistics were performed across several thresholds. c. Anatomical rendering of the task-related modules in the Quaternary condition. Each sphere is color-coded by its initial resting-state module allegiance.

Having established a difference in community structure we sought to test the relative contribution of each module to the observed reconfiguration. To do so we implemented a similar strategy to Braun and colleagues (2015), whereby VIn was calculated at the individual-community level for our nodes

79 of interest, and compared using repeated-measures statistics. Thus we compared visual, sensory, fronto-parietal and default-mode modules across all network densities for the Binary-Rest, Ternary- Rest, and Quaternary-Rest contrasts (conceptually similar to follow-up parametric statistics). Results revealed that the fronto-parietal module had higher VIn values (i.e., larger differences in community structure) than visual (mean p value across thresholds, p < 0.001), sensory (p < 0.001) and default- mode modules (p < 0.001), see Figure 5-4b) across all contrasts.

Figure 5-4 Changes in Variation of information as a function of reasoning complexity. a. Variation of Information (VIn) values (black markers) compared with a null distribution (gray markers, 5th/ 95th percentile in bold line, 1st/ 99th percentile shown in tails) for the three main contrasts across all network densities. Only the right-most contrast (Binary versus Pre-task rest) showed a consistent difference between partitions. b. Comparison of VIn values across visual, sensory, fronto-parietal and default-mode modules in each task condition compared with rest across all network densities. The fronto-parietal module was consistently more variable in relation to other modules. Error bars represent 95% confidence intervals.

Finally, the index of modularity, Q, was compared across conditions. This index of modularity increases as more intramodular connections are found than expected by chance (Newman & Girvan, 2004). Non-parametric Friedman tests revealed a significant difference in Q across conditions (mean statistic across thresholds, E% = 51.48, p < 0.001). Bonferonni corrected follow-up tests were performed to compare each task state (Null, Binary, Ternary, Quaternary) with each resting-state (Pre- and Post-task). Results revealed that Q was significantly lower in the Ternary (Mean Q = 0.39) and Quaternary conditions (Mean Q = 0.39) when compared with both Pre- (Mean Q = 0.44, z = 3.74, z = 4.07, p < 0.001) and Post-task resting-states (Mean Q = 0.45, z = 4.49, z = 4.52, p < 0.001). No effect

80 was found when comparing Null or Binary conditions with rest. Complementing the observed changes in community structure, analysis of Q scores highlight a significant reduction in modularity compared with the resting-state, but only in the task conditions that imposed higher demands on cognitive reasoning.

5.5.3 Network Based Statistic analysis

To further refine our account of module reconfiguration, we assessed changes in whole brain connectivity using the network based statistic (NBS, Zalesky et al., 2010). In line with the result from the first analysis, a paired t-test between Pre- and Post-task resting-states revealed no significant differences. Our second contrast, a one-way repeated measures ANOVA was performed comparing all four reasoning complexity conditions (Null, Binary, Ternary, Quaternary).

A sub-network comprising 63 nodes and 85 edges changed in response to reasoning complexity demands (p < 0.001, FWE corrected at the network level, Figure 5-5). The majority of edges within the subnetwork demonstrated increased functional connectivity (86% of edges, shown as warm colors in Figure 5-5), but a number of edges also demonstrated a decrease in positive correlations with increasing reasoning complexity (see Figure 5-5b for trend across conditions). Consistent with our previous work on changes in functional connectivity during complex reasoning (Cocchi et al., 2014a; Hearne et al., 2015), the network was largely composed of nodes encompassing fronto-parietal (17%), subcortical (19%), cingulo-opercular (12%) and default-mode networks (24%, as per Power et al., 2011 network affiliations, see Appendix C for a list of regions implicated by this analysis). Moreover, nearly all edges (95%) were across-network. Two further visual-parietal subnetworks were identified by the NBS, consisting of two and three nodes respectively (not visualized).

81

Figure 5-5 Change in pairwise functional connectivity associated with reasoning complexity. a. Connectogram representation of significant changes in pairwise functional connectivity that scaled with relational complexity. Edges are colored by the direction of the change in correlation across relational complexity. Warm colors represent increases in connectivity and cool colors represent decreases in connectivity. Lighter colors represent higher F-statistics. Network nodes, plotted as circles, are colored by their initial resting- state networks (Power et al., 2011). Outside the connectogram the colored bars represent the modules identified in the previous analysis of data from the Quaternary condition: Sensory (orange), Default-mode (red) and Fronto-parietal-visual modules (green-blue). b. Each individual connection in the subnetwork (averaged across subjects) plotted as a function of reasoning complexity. Average values for positive and negative connections are shown as bold lines.

5.5.4 Within-module global efficiency and behaviour

Our final analysis sought to investigate changes in global efficiency within each major module evident during the task. Global efficiency has previously been taken to be an index of increased capacity for information exchange (Achard and Bullmore, 2007). Specifically, we were interested in whether each module showed changes in efficiency, and whether any such changes were related to reasoning performance.

Nonparametric Friedman tests revealed that both the Sensory (orange in Figure 5-6a) and Fronto- parietal-visual modules (FPV, green in Figure 5-6a) demonstrated significant differences in efficiency across conditions (Sensory E% = 44.17, p < 0.001, FPV: E% = 77.78, p < 0.001, see Figure

82 5-6a). No such effect was found for the Default-mode module (p = 0.23). Bonferroni corrected follow-up tests confirmed that the effect was driven by increased efficiency within all task states compared with Pre- and Post-task resting-states (Sensory: z-range = 2.4 – 4.69, p < 0.02, FPV: z- range = 3.78 – 5.22, p < 0.001).

Figure 5-6 Changes in global network efficiency (Eglob) across the identified reasoning task modules. a. Global network efficiency levels within each module across experiment conditions. Error bars represent 95% confidence intervals. R1= Pre-task rest, B = Binary, T = Ternary, Q = Quaternary, R2 = Post-task rest. b. Correlation between accuracy in the Latin Square Task (LST) and changes in fronto-parietal-visual (FPV) module efficiency during the task. Changes in network efficiency were correlated with overall reasoning performance, such that increased efficiency correlated with better task performance (r = 0.33, p < 0.01). Results are visualized at 15% network density.

We also investigated the relationship between individual differences in module efficiency and behavioural performance. To do so, we correlated reasoning accuracy scores with changes in module efficiency between the Pre-task resting-state and the most complex reasoning condition (Quaternary). Only module efficiency within the FPV module was significantly correlated with behaviour(see Figure 5-6b, mean statistics across thresholds: r = 0.33, p = 0.026; Spearman’s rs = 0.27, p = 0.084), such that larger increases in efficiency within the FPV module were associated with better reasoning

83 performance. Neither the Default-mode or Sensory modules demonstrated such a relationship (p = 0.19, p = 0.29 respectively). Further, to probe the reliability of the above finding, we compared change in efficiency from Quaternary to the Null-task state, which yielded a similar result (r = 0.35, p = 0.021, Spearman’s rs = 0.30, p = 0.048). The correlation was also robust to partialing out fluid intelligence scores based on the Raven’s Matrices test (r = 0.35, p = 0.025). By contrast, there was no correlation between performance in the visual search task and module efficiency (N = 43, p = 0.65), suggesting the module efficiency-behaviourrelationships were specific to the LST. Finally, we also replicated these results, as well as the follow up Variation of Information results (Figure 5-4b), using the original visual, sensory, default-mode and fronto-parietal networks defined by Power and colleagues (2011), instead of our own data-driven modules.

5.6 Discussion

Human reasoning has a quantifiable capacity limit (Halford et al., 1998). This limit is thought to arise from the brain’s ability to reconfigure interactions between spatially distributed networks (Cocchi et al., 2014a; Parkin et al., 2015; Schultz and Cole, 2016), but recent work has highlighted the circumscribed nature of such interactions when compared with whole brain ‘resting’ architecture (Cole et al., 2014). In light of these recent findings, we examined how global and selective network properties change from resting to reasoning states, and how such changes relate to reasoning behaviour. We found that complexity-based limits in reasoning ability rely on selective patterns of connectivity that emerge in addition to a more general task-induced functional architecture.

We used a non-verbal reasoning task, originally designed to test predictions from relational complexity theory, to systematically manipulate reasoning complexity (Birney et al., 2006; Birney and Bowman, 2009; Halford et al., 1998). In doing so we replicated previous behavioural results by demonstrating a reliable reduction in accuracy and an increase in reaction time as a function of increased complexity (Birney et al., 2006; Zeuch et al., 2011; Zhang et al., 2009). Importantly, an analysis of participants’ trial-by-trial ratings indicated that task errors were related to complexity demands and not factors such as transitory lapses in attention or disengagement from the task. The behavioural results also confirmed previous reports that individual reasoning capacity limits are correlated with scores on standard measures of fluid intelligence (Bhandari and Duncan, 2014; Birney et al., 2006) such as Raven’s Matrices.

Parametric increases in relational complexity have previously been tied to neural activity of segregated regions of the prefrontal cortices (Bunge et al., 2009; Christoff et al., 2001; Golde et al., 2010; Kroger, 2002) as well as to functional connectivity within fronto-parietal and cingulo-opercular “multiple-demand” networks (Cocchi et al., 2014a; Crittenden et al., 2016; Parkin et al., 2015). Cingulo-opercular connectivity has been associated with initiating and maintaining task sets (Dosenbach et al., 2006) whereas the fronto-parietal network has been associated with moment-to-

84 moment cognitive control (Cole and Schneider, 2007). Here we found that reasoning performance was best explained by a module composed of brain regions within the fronto-parietal, salience, subcortical and visual networks. First, functional connectivity and global efficiency of this subnetwork increased in line with increased reasoning demands (Figures 5-5 & 5-6). Second, larger increases in global efficiency within this module were associated with higher accuracy in the reasoning task. Finally, edge-wise connectivity of the fronto-parietal network was shown to increase in line with relational complexity, largely between default-mode and subcortical networks. These findings are broadly consistent with previous work showing that enhanced global network efficiency, and connectivity within the default-mode and fronto-parietal networks at rest, can predict intelligence and reasoning performance (Finn et al., 2015; Hearne et al., 2016; Song et al., 2009, 2008; van den Heuvel et al., 2009). Taken together the results confirm the central role of flexible fronto-parietal connectivity in implementing external goal-directed cognitive control (Cole et al., 2013, Cocchi et al., 2014).

It has been proposed that the cingulo-opercular network can be further divided to include a separate ‘salience’ system associated with bottom-up attention (Power et al., 2011; Seeley et al., 2007). Here we found that the salience network was implicated in the fronto-parietal-visual module but did not show complexity-induced edge-wise connectivity changes. On the other hand, the cingulo-opercular network did show edge-wise connectivity changes in line with reasoning complexity, but was not implicated in the fronto-parietal-visual module. This set of results is consistent with the notion that the cingulo-opercular network is a control-related counterpart of the fronto-parietal network, and suggests that the cingulo-opercular network might have a distinct role from that of the salience aspect of the system (Power et al., 2011).

It remains unclear precisely how ‘resting-state’ networks coordinate flexible patterns of integration and segregation as a function of task complexity. Nevertheless, our findings support a key role for subcortical structures such as the thalamus in mediating such relationships (Bell et al., 2016; Sherman, 2016). Specifically, bilateral putamen and thalamus were implicated in both subnetworks that increased and decreased functional connectivity as task complexity increased (general trend shown in Figure 5-5b). This finding is in line with recent descriptions of the thalamus as a ‘global kinless’ hub, with evidence of activation in multiple cognitive contexts and strong connectivity across multiple large-scale functional networks (Guimera et al., 2007, Hwang et al., 2017, van den Heuval & Sporns, 2011). Such subcortical regions might be interpreted as managing relationships between task-related networks to form a coherent modular structure. Further work will be needed to elucidate the particular role of subcortical regions in this relatively unexplored area (Bell & Shine, 2016).

Functional brain module reconfigurations have previously been related to performance on a range of higher cognitive tasks, including learning (Bassett et al., 2011), working memory (Braun et al., 2015; Vatansever et al., 2017, 2015) and cognitive control (Dwyer et al., 2014). Here we found that the community architecture of the brain is flexible, but only in response to large cognitive shifts. For

85 example, resting-state visual and fronto-parietal modules, each of which is composed of several known sub-networks (Power et al., 2011), merged together during the reasoning task (Figure 5-3). Importantly, this reorganization was relatively isolated; follow up analyses indicated that the rest of the brain remained stable across changes in task complexity. In line with this observation, recent network-based re-conceptualizations of global workspace theory (Dehaene et al., 1998; Kitzbichler et al., 2011) have suggested that large, task-based module reconfigurations arise to better serve network communication underpinning behaviour. Our work refines this idea by showing that once resting-state modules reconfigure in response to external task demands, the majority of connectivity changes occur without interrupting the newly established modular architecture, as illustrated schematically in Figure 5-7. Moreover, it is these connectivity changes within the newly reconfigured modules that seem to be most related to behaviour.

Figure 5-7 Conceptual model of functional networks supporting reasoning and rest states. a. At rest, functional modules are relatively independent. b. External, goal-directed task states are accompanied by broad module-level changes; a fronto-parietal-visual module forms (green), amongst stable default-mode (red) and sensory-motor modules (orange). c. Increased task demands are accompanied by specific increases (solid lines) and decreases (dashed-lines) in functional connectivity, rather than further modular reconfiguration. Ultimately, in the most complex conditions, the entire network reaches a similar level of correlation through both integrated and segregated dynamics (see Figure 5-5b).

Our finding that increased demands on cognitive reasoning are paralleled by a reduction in network modularity and increased efficiency has now been reported in several different task contexts (Bola and Sabel, 2015; Cohen and D’Esposito, 2016; Godwin et al., 2015; Kitzbichler et al., 2011; Shine et al., 2016; Vatansever et al., 2015; Westphal et al., 2017; Yue et al., 2017). Using the network-based statistic we found that these global changes were supported by increases and decreases in functional connectivity across multiple large-scale networks. Interestingly, the default-mode network has been suggested to act as a ‘global integrator’, facilitating fronto-parietal cognitive control networks during conscious processing of information (Dehaene et al., 1998; Guldenmund et al., 2012; Leech et al., 2012; Vatansever et al., 2015). In line with this notion, we found that default-mode regions such as

86 medial frontal cortex, angular gyri and posterior cingulate cortex demonstrated increased functional connectivity with fronto-parietal, cingulo-opercular and visual networks as task demands increased.

Modulations of visual network connectivity might be related to the visual nature of the LST, and specifically the requirement that participants search the 4 x 4 matrix to identify a shape at the probed location. A previous behavioural study found that participants’ eye fixation patterns differed for one- object and two-object relational problems (Gordon & Moser, 2007), raising the possibility that changes in relational complexity might be associated with changes in search patterns (and by extension, associated network connectivity). We cannot unequivocally rule out potentially small differences in eye movement patterns between complexity conditions in our study, but there are at least two reasons why such findings are unlikely to be directly relevant here. First, Gordon and Moser (2007) actively encouraged visual search by having their participants compare two different picture stimuli arranged one above the other on a page. By contrast, our LST paradigm involved the relatively brief presentation of a single 4 x 4 matrix at fixation. Second, the stimuli used by Gordon and Moser (2007) were visually complex line drawings that included several different object types that varied in size, shape and semantic content across trials. By contrast, our LST stimuli involved a single matrix containing identical shapes across complexity conditions, and did not explicitly require active search to solve for the target. Finally, we found no correlation between performance on a standard visual search task and reasoning performance or brain-based network efficiency metrics (subset of participants, N=43). If increasing visual search demands across task conditions was responsible for the observed network differences, performance on this visual search task should have correlated with the brain-derived metrics.

In conclusion, our findings suggest that reasoning demands rely on selective patterns of connectivity within fronto-parietal, salience, cingulo-opercular, subcortical and default-mode networks, which emerge in addition to a more general, task-induced modular architecture. Further work will be needed to elucidate the network processes that bring about the intricate and coordinated changes in connectivity patterns at the level of edges, modules and the whole brain, in the service higher cognition. Meanwhile, the current results provide novel insights into the roles of both specific and global network changes in reasoning.

87

Chapter 6: Anomalous functional network integration in response to cognitive control demands in human callosal dysgenesis

88 6.1 Preamble

The work described in this chapter follows directly from that presented in Chapter 5, but here the focus is on a small group of individuals with a rare condition called callosal dysgensis. I have attempted to limit methodological redundancies by referring where appropriate to information already provided in Chapter 5. I would like to acknowledge the important contribution of Dr. Ryan Dean, from the Queensland Brain Institute, who conducted the diffusion MRI pre-processing steps reported here. At the time of writing the work presented in this chapter was being prepared for submission to the journal Brain.

6.2 Abstract

Higher cognitive reasoning is thought to rely on dynamic functional interactions between networks of cortical areas distributed across the two cerebral hemispheres. Here we used high-field magnetic resonance imaging and brain network analyses to investigate the functional brain networks underlying cognitive reasoning in a group of individuals with callosal dysgenesis (CD), a structural abnormality that primarily affects white matter connections between the two hemispheres. Participants were asked to solve Sudoku-like problems while their brain activity was measured. The complexity of these problems was parametrically varied by changing the number of relations that needed to be established between shapes in a matrix. Behaviourally, participants showed a reduction in response accuracy as task complexity increased. Brain activity evoked during the task was observed in cortical regions known to constitute two key cognitive control systems: the fronto-parietal and cingulo-opercular networks. Under low reasoning demands, patterns of local neural activity and network topology in the affected group closely resembled those observed in healthy controls. By contrast, under higher reasoning demands the CD cohort demonstrated diminished activity and functional connectivity within the fronto-parietal network. These ‘state’ rather than ‘trait’ differences in functional network integration help explain the link between previously observed neurotypical resting-state networks and the heterogeneous cognitive deficits in CD.

6.3 Introduction

The corpus callosum is the major white matter commissure of the brain. It connects the left and right cerebral hemispheres, and consists of more than 190 million axonal projections (Tomasch, 1954). These connections are thought to support the coupling of bilateral functional networks (Honey et al., 2009; Putnam et al., 2008; Shen et al., 2015; van der Knaap and van der Ham, 2011) that underpin both simple and complex behaviour (Cocchi et al., 2014a; Duncan, 2010; Schulte and Müller- Oehring, 2010). A range of florid and distinctive behavioural deficits typically arises in adults who have had their corpus callosum surgically sectioned for the relief on intractable seizures (Concha et al., 2006). Most notably, these individuals tend to display a ‘disconnection syndrome’ by failing to complete tasks that require shared information between the two cerebral hemispheres (Gazzaniga et

89 al., 1962; Seymour et al., 1994). By contrast, in individuals with callosal dysgenesis (CD), a developmental condition in which the corpus callosum is only partially formed or fails to develop altogether, interhemispheric communication deficits are generally more subtle (Lassonde et al., 1991; Paul et al., 2007).

Evidence for interhemispheric transfer of information across a range of behavioural tasks has been observed in individuals with CD (Lassonde et al., 1991; Sauerwein et al., 1981), suggesting that this capacity is preserved by alternative interhemispheric routes, such as the anterior and posterior commissures (Tovar-Moll et al., 2014). In line with this, recent investigations of individuals with CD using functional magnetic resonance imaging (fMRI) have detected intact bilateral functional networks during resting-state MR scans (e.g., the default-mode network; Tyszka et al., 2011). It remains unclear, however, how these developmentally ‘rewired’ functional networks respond during cognitive reasoning tasks. To investigate this question, we employed high field fMRI to measure brain activity while individuals with CD performed a behavioural task in which the complexity of cognitive reasoning demands was systematically varied.

Figure 6-1 Anatomy of the corpus callosum. A diagram of a healthy corpus callosum with basic anatomical landmarks superimposed over a MR scan of the author (sagittal plane).

CD is a heterogeneous condition, both anatomically and behaviourally. Such heterogeneity is underpinned by a number of potential environmental and genetic contributions that are also variable and complex (Bedeschi et al., 2006; Paul et al., 2007). Anatomically, the extent of corpus callosum malformation in CD ranges from complete absence, to partially formed segments (e.g., absence of the genu, see Figure 6-1 for a diagram of anatomical landmarks), and hypoplasia (homogeneous reduction in callosal size: Paul et al., 2007; Tovar-Moll et al., 2007). Furthermore, CD often occurs alongside other brain abnormalities (Hetts et al., 2006). Aberrant fibres are known to form during development, such as the so-called Probst bundles, which consist of misrouted callosal tracts that run parallel to the interhemispheric fissure (Tovar-Moll et al., 2007). The behavioural and

90 neuropsychological profiles of people with CD are equally heterogeneous, ranging from subtle cognitive deficits to severe impairments across a wide variety of cognitive domains (Paul et al., 2004; Siffredi et al., 2013). Only mild performance deficits have been observed on tests of interhemispheric communication, however, suggesting that despite the structural abnormalities, the primary functions of the corpus callosum in individuals with CD are at least partly preserved (Lassonde et al.,

1991; Sauerwein et al., 1981).

The question of how interhemispheric communication is maintained in CD is puzzling. Using fMRI, recent work has observed intact canonical resting-state functional brain networks in individuals with CD (Tyszka et al., 2011). These networks seem to be directly supported by anomalous interhemispheric white matter pathways, such as those travelling through the anterior and posterior commissures, suggesting that alternative routes are available to support functional networks (Lassonde et al., 1991; Tovar-Moll et al., 2014). However, such routes are likely to be indirect, polysynaptic and functionally costly (Marco et al., 2012; Owen et al., 2013).

The roles and limitations of such developmentally rewired functional networks are not well understood. While there is evidence in CD of preserved network activity during tactile stimulation within sensorimotor networks (Duquette et al., 2008), less is known about networks associated with higher cognitive abilities (Hinkley et al., 2016). In neurotypical individuals, complex cognitive control tasks are associated with increased activity within, and coupling between, fronto-parietal (FPN) and cingulo-opercular networks (CON) (Cocchi et al., 2014a; Crittenden et al., 2016; Duncan, 2010). It is not known, however, whether individuals with CD produce similar task-evoked brain patterns, and to what degree indirect pathways can support functional network coupling as task demands increase. One hypothesis is that while bilateral functional networks may be detectable in states of low demand (e.g., the resting-state, Tovar-Moll et al., 2014; Tyszka et al., 2011), these networks become dysfunctional as cognitive demands increase, ultimately resulting in reduced behavioural performance.

To investigate activity patterns associated with cognitive reasoning across brain networks in participants with CD, we collected fMRI data at 7T from seven affected individuals while they undertook the Latin Square Task (LST; Birney et al., 2006), which is a relational reasoning task akin to the popular game known as ‘Sudoku’. The LST is a non-verbal relational reasoning paradigm in which reasoning complexity is parametrically varied with minimal working memory demands (Birney et al., 2006; Halford et al., 1998). Critically, this task allowed us to incrementally load task complexity and track the resultant brain activity patterns and network responses. Considering previous work, we expected to find relatively normal bilateral FPN and CON networks in the low-demand reasoning conditions (Duquette et al., 2008; Tovar-Moll et al., 2014; Tyszka et al., 2011). We also hypothesized that high reasoning complexity would reveal deficits in functional integration within and

91 between these networks. To test these hypotheses, we compared brain activation and network responses obtained in the seven CD individuals with those from a normative sample of 30 healthy participants.

6.4 Methods

6.4.1 Participants

The current study included seven participants with CD aged 24–64 years (see Table 6-1). CD1, CD2 & CD4 presented with complete agenesis of the corpus callosum. CD3 presented with a genu and posterior body remnant. CD5 presented with a body remnant. CD 6 presented with an intact genu and a rostrum remnant. CD7 presented with intact genu, rostrum and splenium, as well as the anterior portion of the body (see Figure 6-2a for depiction of anatomy). In line with previous work (Tovar- Moll et al., 2014), the CD sample was heterogeneous, with different degrees of inter- and intra- hemispheric structural connectivity (Figure 6-2b). To benchmark participants’ performance and patterns of task-induced brain activity and functional connectivity, we included a sample of 30 healthy controls (Age: M = 23.60, SD = 3.95 years, 17 females) who had completed the same task and imaging protocols as part of a previously published study (Hearne et al., 2017), and whose data were selected to best match the behavioural performance of the CD group. This research was approved by The University of Queensland Human Research Ethics Committee.

Table 6-1 Demographics of the callosal dysgenesis participants and their scores on intellectual function measures. Age Sex VIQ PIQ FSIQ NART APM Callosal abnormality (Years) (100±15) (100±15) (100±15) (100±15) (10.2±1.4) CD1 24 M 94 99 97 96 8 Complete CD2 40 M 88 103 96 96 5 Complete CD3 24 M 128 115 124 110 10 Partial: intact genu, posterior body remnant CD4 47 F 92 105 99 112 9 Complete CD5 46 M 91 88 88 113 7 Partial: body remnant CD6 64 M 108 127 118 112 11 Partial: intact genu and rostrum remnant CD7 24 F 91 80 94 100 5 Partial: intact genu, rostrum, splenium anterior body Note: FSIQ, Full Scale Intelligence Quotient; VIQ, Verbal Intelligence Quotient; PIQ, Performance Intelligence Quotient; NART, National Adult Reading Test derived premorbid Intelligence Quotient; APM, Raven’s Advanced Progressive Matrices Set One. Typical mean and standard deviations are displayed underneath each test. For the APM these values were calculated from the 30 healthy individuals included in the study.

92

Figure 6-2 MR images of structural anomalies in the seven callosal dysgenesis (CD) participants. a. Sagittal view of midline brain structure in individuals with CD. B. Fibre tract index of inter- and intra- hemispheric connections for CD2 – CD7 (details in Methods). As expected, participants with complete agenesis of the corpus callosum had fewer interhemispheric white matter connections than those with a partially formed callosum. This trend was similar for intrahemispheric connections.

6.4.2 Task

Participants underwent 3 x 12-minute runs of the Latin Square Task (LST, Birney et al., 2006), high- angular resolution diffusion MRI, and an MP2RAGE anatomical scan. We used the same cognitive task as a previous study (Hearne et al., 2017). Each LST trial involved the presentation of a four-by- four matrix populated with a small number of geometric shapes (square, circle, triangle or cross), blank spaces and a single target probe, denoted by a yellow question mark (‘?’; see Figure 6-3a). The matrix was presented at fixation and participants were asked to solve for the target according to the rule: “each shape can only occur once in every row and once in every column” (similar to the game of Sudoku). Binary problems require integration of information across a single row or column. Ternary problems involve integration across a single row and column. Quaternary problems, the most complex, require integration of information across multiple rows and columns. Null trials involved presentation of an LST grid, but instead of a target probe (‘?’) an asterisk was presented (‘*’) to inform the participant that no reasoning was required for that trial. Puzzles were presented for 5 seconds, followed by a motor response period and a confidence rating (see Figure 6-3b). Administration of all items was pseudo-randomized such that no two items of the same complexity occurred sequentially, and each block had two problems from each level of complexity. In total, 144

93 LST items were presented in the MR session across 16 blocks, with 36 items in each relational complexity condition; null, binary, ternary and quaternary. Prior to the MRI session participants completed 20 practice trials of the LST (12 with corrective feedback).

Figure 6-3 Example problems and sequence of displays in a typical trial of the Latin Square Task. a. Examples of each reasoning complexity condition. The correct answers are square, cross and cross, respectively, for the binary, ternary and quaternary problems illustrated. b. Example trial sequence. Each trial contained a jittered fixation period, followed by an LST item, a second, jittered fixation period, a response screen, and a confidence rating scale. In null trials the motor response screen had one geometric shape replaced with an asterisk, which indicated which button to press for that trial. The task was exactly the same as that described in Chapter 5.

6.4.3 Imaging data analysis

Data acquisition

Imaging data were acquired using a 7 Tesla Siemens MRI fitted with a 32-channel head coil. The functional MRI protocol was identical to that reported for the same task in a previous study (see Chapter 5; Hearne et al., 2017). All diffusion weighted images (DWI) were acquired using Siemens’ Advanced EPI Diffusion Imaging sequence (ver. 511E). Diffusion space was evenly sampled over two half-shells (32 sample points with b = 700s mm-2 and 64 sample points with b = 2000s mm-2, respectively). The voxel dimensions were 1.5 × 1.5 × 1.5 mm, with an axial slice gap of 0.3 mm. The

94 repetition time of the diffusion pulse sequence was 12,000ms, with an echo time of 54.4ms. Four B0 images were also acquired, one of which was reverse phase-encoded for correcting EPI distortions. Unfortunately an error occurred with the recording of data for CD1 during the diffusion scan excluding them from the structural analysis (i.e. Figure 6-2b), however they were included in all the functional analyses.

Preprocessing

Functional MRI data were pre-processed as outlined in Chapter 5 (Hearne et al., 2017). However, due to recent evidence that CompCor (Behzadi et al., 2007) noise signal regression can detrimentally affect between-group imaging comparisons (Parkes et al., 2017), CSF and white matter signal regression were computed by manually placed seed regions. DWI were corrected for EPI distortion, motion and bias field inhomogeneity; the diffusion gradient direction matrix was rotated to account for these corrections.

Brain activity analysis

Regional activity changes associated with increments in relational reasoning complexity were isolated using a general linear model (GLM; SPM12). The CD individuals performed at chance in the most difficult Quaternary problems (see Results, below), and so we excluded this condition for the remaining analysis. First-level t-contrasts were used to isolate the average positive effect of increased complexity [Null (-2) versus Binary (1) and Ternary (1)]. Following this, using the SPM toolbox MarsBar (Brett et al., 2002), beta values for each individual for each contrast were extracted across 12 regions of interest (ROI, 5mm spheres) defined in the GLM analysis (see Table 6-2).

Task-evoked functional connectivity

Task-evoked functional connectivity was estimated using the same method as in Chapter 5 (Hearne et al., 2017), and employing the 12 ROIs defined above. Regionally averaged time-series were extracted for the ROIs, and a task nuisance covariate was regressed from each signal (Cole et al., 2014). A Pearson correlation was performed on the residual time-series and converted to z-scores.

Structural connectivity

Structural connectivity analyses were undertaken using MRtrix3 (www.mrtrix.org). Constrained spherical deconvolution (CSD) and probabilistic tractography were jointly used to generate 20 million streamlines across the whole-brain white matter. These were pruned to the 2 million streamlines that represented the most probable propagations of fibre tracts between brain regions via CSD-informed filtering of tractograms (R. E. Smith et al., 2013). A combined Harvard-Oxford cortical and subcortical structural atlas and the filtered tractogram were then used to generate binarised structural connectivity matrices. We created an index to estimate the degree of inter- and intra-hemispheric

95 structural connectivity by summing the number of connections (i) within, and (ii) across hemispheres (Figure 6-2b).

6.5 Results

6.5.1 Behaviour

A typical complexity effect was observed in individuals with CD, as previously found in neurotypical adult participants (Birney and Bowman, 2009; Cocchi et al., 2014a), such that reasoning problems of greater complexity resulted in poorer response accuracy, F(2,12) = 60.45, p < 0.001, F% = 0.91. As can be seen in Figure 6-4, follow-up tests confirmed that participants were more accurate in the Binary than in the Ternary condition (t = 4.15, p = 0.018), and in the Ternary than in the Quaternary condition (t = 6.89, p = 0.001).

Figure 6-4 Performance accuracy of the two groups (callosal dysgenesis and neurotypical controls) in the Latin Square Task. Average values for controls are plotted in grey. Grey shading indicates 95% repeated measures confidence intervals. Individual results for the callosal dysgenesis (CD) individuals are plotted in different colours (see legend at right). Statistical tests were performed on the CD cohort only; asterisks indicate p < 0.05.

6.5.2 Brain activity

Our initial brain imaging analysis aimed to test whether individuals with CD showed similar brain activity patterns to healthy controls in response to LST demands. Previous work on cognitive reasoning has isolated co-activations within fronto-parietal (FPN) and cingulo-opercular networks (CON) that respond to task complexity, such that increased complexity is coupled with an increased magnitude of BOLD response (Cocchi et al., 2014a; Cole and Schneider, 2007; Dosenbach et al., 2006; Duncan, 2010). Due to the small sample size of the CD cohort, we generated binarized activity

96 maps containing the highest t-statistics for each individual (top 5% values). These maps were then overlaid to create a ‘consistency map’ (shown in Figure 6-5a). This consistency map was qualitatively compared with the results of a random effects analysis in the neurotypical control group.

Figure 6-5 Brain activity associated with increased reasoning complexity. a. Brain activity evoked in the CD group, comparing null versus binary and ternary conditions contrasted with brain activity evoked in the neurotypical controls. The blue-green heat map (left) represents the overlap of brain activity across CD individuals (top 5% t-values), such that green represents complete overlap and blue represents less overlap (three CD individuals). The red-yellow heat map (right) shows the results of a standard random effects analysis in the controls (initial uncorrected search threshold of p < 0.0001, p < 0.001 family-wise error corrected at cluster level) b. Regions of interest derived from the analysis in panel a, highlighting the fronto-parietal (FPN) and cingulo-opercular networks. Only cortical regions with the highest degree of overlap are included (see Table 6-2). Brain renderings were created using SurfIce software (www.nitrc.org/projects/surfice).

97

Table 6-2 Regions of Interest derived from GLM analysis.

Anatomy Network MNI Coordinate Overlap x y z Right Sup. frontal FPN 26 -4 59 6 Left Sup. frontal FPN -26 -3 64 5 Right RLPFC CON 38 34 25 6 Left RLPFC CON -43 31 25 4 Right Inf. frontal FPN 49 8 29 5 Left Inf. frontal FPN -48 4 29 6 Right ACC CON 9 4 45 6 Left ACC CON -6 4 45 4 Right inf. parietal FPN 54 -32 47 4 Left inf. parietal FPN -48 -32 41 6 Right Ant. insula CON 39 23 -7 3 Left Ant. Insula CON -37 24 -7 4 Note: Overlap represents the number of CD individuals (out of 7) who demonstrated high t-statistics (top 5%) at a given MNI coordinate.

Qualitatively, both groups demonstrated an overlapping pattern of co-activations in response to the reasoning task, involving cortical regions of the FPN and CON (Dosenbach et al., 2006). These findings were supported by the analyses of average change in the BOLD signal as a function of task complexity within regions of the FPN and CON (see Figure 6-5b and Table 6-2 for regions of interest). A mixed ANOVA (complexity x group) on average beta values across each network revealed significant main effects, such that the Ternary condition yielded a larger increase in brain activity than the Binary condition when compared with Null, F(1,35) = 21.45, p < 0.001, F% = 0.34, F(1,35) = 6.24, p = 0.017, F% = 0.15, for FPN and CON respectively (see Figure 6-6). For the FPN, a significant interaction was also observed, F(1,35) = 6.30, p < 0.017, F% = 0.10, suggesting that individuals with CD showed a diminished BOLD response in the Ternary condition compared with controls.

98

Figure 6-6 Follow-up analyses of beta values averaged across ROIs within the FPN and CON. Average values for neurotypical controls are plotted in grey. Grey shaded areas indicate 95% repeated measures confidence intervals. Individual results for the CD sample are plotted in colour. Asterisks indicate p < 0.05.

6.5.3 Brain connectivity

Next, we tested whether individuals with CD demonstrated a similar network correlation (i.e., functional connectivity) to healthy controls in response to increasing relational complexity. Previous work has suggested that increased correlations can be found both within and between the FPN and CON in response to increased task demands (Cocchi et al., 2014a; Crittenden et al., 2016). Task-based functional connectivity was calculated between regions that showed consistent BOLD activation across CD participants (see Figure 6-5). The CD cohort demonstrated similar levels of mean functional connectivity compared with controls in the Null condition, suggesting intact bilateral networks in low cognitive demand conditions (Owen et al., 2013; Tyszka et al., 2011). A mixed

ANOVA (complexity x group) demonstrated a significant interaction between complexity and group in the FPN, F(2,70)=3.79, p=0.03, F%=0.09. No effect was apparent for the CON or across-network connections. These statistics, while exploratory, support the qualitative trends observed in Figure 6-7.

99

Figure 6-7 Network-level functional connectivity trends associated with increasing reasoning complexity. Mean functional connectivity values (y-axis, z-score) plotted as a function of increasing reasoning complexity demands (x-axis, Null, Binary, Ternary) in FPN, CON, and across-network connections for neurotypical controls (grey) and CD participants (black). Grey shaded areas indicate 95% repeated measures confidence intervals. The bottom row shows the results for each individual with CD (coloured lines). Asterisks indicate p < 0.05.

The effects shown in Figure 6-7 have multiple edge-related explanations. Thus, a follow-up analysis was conducted to investigate the role of individual connections in the FPN. To do so, each individual functional connection within the FPN was considered independently within the contrast of Ternary G × H minus Null. To identify edges of interest, for each possible connection ( = 15) a distribution of % control edge weights was built, and the results for each CD individual were compared against this distribution. An arbitrary cut-off of the top or bottom 10th percentile was taken to indicate a ‘significant’ connection (see Figure 6-8a).

Visual inspection of averaged connectivity matrices (Figure 6-8b) highlights the difference between the neurotypical response to reasoning complexity in relation to that of the CD participants. The neurotypical response was characterized by larger correlation values to/from the superior frontal gyrus (top/left rows/columns), whereas the CD response demonstrated negative edge weights, particularly

100 to/from the inferior parietal cortex (bottom/right rows/columns). In line with this, only edges in the bottom of the distribution were identified (Figure 6-8c), suggesting CD is largely characterized by decreased functional connectivity (as in Figure 6-7). Across the CD cohort, 23 edges were identified, the majority of which were interhemispheric (N=17). In general, the results of this analysis support the idea of a network-level abnormality in CD, rather than a specific edge impairment.

Figure 6-8 Analysis of edges within the fronto-parietal network. a. ‘Significant’ edges were selected by building a distribution of control edge weights and displaying only those that were in the top or bottom 10%. b. Visualization of the mean correlation matrix (z-scores) for the contrast ‘Ternary minus Null’. Each ‘pixel’ represents the mean edge weight across participants for a single connection. Warmer colours represent an increase in connectivity from Null to Ternary, whereas cooler colours indicate a decrease in connectivity. c. Brain network representation of significant edges summed across CD individuals. Blue-green edge colours indicate the number of connections that were found across participants, with green indicating more consistency. All connections were in the bottom 10th percentile, and no connections were within the top 10th percentile.

6.6 Discussion

Despite the complete absence or malformation of the corpus callosum, intact bilateral functional networks have been observed in individuals with CD (Owen et al., 2013; Tovar-Moll et al., 2014;

101 Tyszka et al., 2011). These networks are considered to be supported by atypical interhemispheric white matter pathways that form during development (Lassonde et al., 1991; Tovar-Moll et al., 2014). Less is known about the role and properties of such ‘rewired’ functional networks. Here we asked individuals with CD to undergo fMRI while solving non-verbal reasoning problems known to elicit responses within the FPN and CON (Hearne et al., 2017). We found that activity patterns within the FPN and CON in CD participants were largely preserved in low-demand reasoning conditions, but were abnormal compared with those of neurotypical controls under high-demand conditions.

Overall, we found compelling evidence that functional networks supporting higher cognitive processes can be preserved in a variety of structural configurations (Mišić et al., 2016; Owen et al., 2013; Tyszka et al., 2011). In low demand conditions (i.e. null and binary) both activation and functional connectivity was preserved in the FPN and CON. Strikingly, even those with complete agenesis of the corpus callosum demonstrated this effect, suggesting that the corpus callosum is not necessary for bilateral functional networks. It is currently unclear how such networks are preserved in people with CD. Recent evidence suggests that functional networks are supported by developmentally rewired connections through the anterior and posterior commissures (Tovar-Moll et al., 2014). Another potential explanation is that subcortical structures provide common input to both cerebral hemispheres leading to observable bilateral functional networks. One candidate structure, the thalamus, is endowed with whole brain structural and functional connectivity necessary for such a role (Behrens et al., 2003; Hwang et al., 2017; Schmitt et al., 2017), as well as the potential to communicate interhemispherically through the interthalamic adhesion, which is an understudied bridge of fibres connecting the bilateral thalami (Damle et al., 2017). However, both theories of structural rewiring and common subcortical inputs are not mutually exclusive.

Critically, we found that the FPN demonstrated an inability to adapt to increased reasoning demand. With regard to brain activity, we observed a decreased (or more variable) task-evoked response to ternary level problems within the FPN. In terms of functional connectivity, we observed that the FPN showed a marked decline in correlation as reasoning complexity increased. Previously, the FPN has been linked to moment-to-moment task demands (Cocchi et al., 2014a); thus, it is possible that functional network deficits in CD will manifest in whichever network is currently active. Alternatively, this effect could be specific to the FPN regardless of the cognitive domain. There is now considerable evidence that more complex cognitive tasks rely on more spatially distributed processing than simple tasks (Cohen and D’Esposito, 2016; Hearne et al., 2017; Yue et al., 2017). Therefore, as task complexity increases, individuals with CD may have to rely more heavily on suboptimal long-range anatomical wiring. There is some evidence to this claim, as interhemispheric functional connectivity in association cortex is reduced in CD compared to sensory cortex (Owen et al., 2013).

102 Understandably, considering the rarity of CD in the population (Glass et al., 2008), the current study consisted of a small sample (N=7), and we were unable to adhere to best practice regarding, for example, cluster-based inference, or the identification of statistically significant network edges (Eklund et al., 2016; Zalesky et al., 2010). However, when averaging values within regions of interest (notwithstanding the loss of information involved in the gross aggregation of complex data) traditional statistics largely confirmed the qualitative trends we observed. It is entirely possible that, with appropriate statistical power, individuals with CD would demonstrate significantly altered activation topology (e.g., Hinkley et al., 2016). A further limitation of the current work is the lack of age matching between the CD cohort and neurotypical controls. The latter were derived as a convenience sample from those who had volunteered for the study described in Chapter 5. While the BOLD response is known to change with age (D’Esposito et al., 2003), the main findings from the current work (Figure 6-7) were consistent despite the different ages of the individual participants. It is also possible that other, unmeasured group differences between the controls and the CD participants, such as IQ and socioeconomic status, could have influenced the observed results. However, we think this is unlikely given the similarity between measures of intellectual functioning performed in the CD group and those derived from neurotypical norms (e.g., FSIQ score mean = 102.3, standard deviation = 13.4).

Our results suggest that the general functional architecture supporting cognitive control is preserved in individuals with CD (Tyszka et al., 2011), implying that functional networks can diverge from the available underlying anatomical scaffolding, at least under some conditions (Mišić et al., 2016). Critically, we found that in CD this architecture struggles to accommodate increasingly complex demands, a finding that would not be detected in conventional resting-state paradigms. This effect may arise from the developmentally rewired structural connections that support these functional networks. When considered in conjunction with individual differences in structural connectivity, task- driven deficits in FPN integration may provide a possible mechanism to explain the broad, heterogeneous cognitive deficits in CD. However, future work with larger sample sizes will be needed to confirm this conjecture.

103

Chapter 7: General Discussion

104 The ability to reason and solve complex problems is an important cognitive function that underpins human intelligence. Individual differences in intelligence are measureable and highly predictive of a number of crucial life outcomes (e.g., Deary et al., 2007). However, the neural bases of intelligence and reasoning ability are not well understood. Early neuropsychological studies suggested that injury to the prefrontal cortex caused a decrement in the ability to reason successfully (e.g., Waltz et al., 1999), a premise that was later supported by functional brain imaging studies using fMRI (e.g., Gray et al., 2003). More recently it has been suggested that higher cognitive abilities, such as reasoning and intelligence, likely rely upon interactions between widespread brain regions (Cole et al., 2012; Jung and Haier, 2007). In this context, the broad aim of my thesis was to investigate the neural correlates of human reasoning, with a particular emphasis on the contributions of functional brain networks. I conducted a series of fMRI experiments to characterize relevant networks in the brain, and related these networks to reasoning performance, individual differences in intelligence, and the breakdown of reasoning ability in people with callosal dysgenesis, a developmental abnormality of brain wiring.

In Chapter 2 I examined functional brain network correlates of intelligence in 317 individuals from the Human Connectome Project dataset (Van Essen et al., 2013). I found that intelligence was best predicted by a set of functional connections between default-mode and fronto-parietal networks. In Chapter 3 I investigated these same networks in participants who underwent fMRI while performing the Wason Selection Task (Wason, 1968), a well-known test of deductive reasoning. This experiment revealed that the angular gyri of the parietal cortex increased their functional connectivity with prefrontal and subcortical regions, despite showing an overall decrease in BOLD magnitude in response to the task. In Chapter 4 I developed a new version of an existing relational reasoning paradigm known as the Latin Square Task (LST; Birney et al., 2006), and tested its reliability and its correlation with a standard measure of fluid intelligence (the Advanced Progressive Matrices; Raven, 2000). A central aim of this experiment was to develop a reasoning task in which relational complexity could be varied systematically, and in which trials could be administered repeatedly over a relatively short time frame, in anticipation of conducting a subsequent fMRI study. In Chapter 5 young adult participants underwent an initial resting-state scan, followed by the LST task validated in Chapter 4, and then a second resting-state scan. Using network analysis I found that increased reasoning complexity was associated with decreased network modularity and increased network efficiency, and that this latter effect was correlated with reasoning performance. Finally, in Chapter 6 a small group of individuals with callosal dysgenesis performed the LST while undergoing fMRI. I investigated common task-related fronto-parietal (FPN) and cingulo-opercular networks (CON; Duncan, 2010; Seeley et al., 2007). Under low reasoning demands, brain activity and network topology were similar to those observed in normal adults (as described in Chapter 5), whereas under higher reasoning demands there was a notably diminished response in the fronto-parietal network of individuals with callosal dysgenesis.

105 Below I discuss the broader implications of the empirical findings from Chapters 2 – 6, and offer some thoughts on the implications of these results for recent models of higher cognition. Toward the end of the chapter I outline some limitations of the current work and offer some suggestions for future research.

7.1 Relational reasoning is supported by distributed and flexible functional networks

Throughout my thesis I used relational complexity theory (Halford et al., 1998) to manipulate and interpret brain responses to changes in reasoning complexity. I observed consistent behavioural effects across two different task paradigms, and multiple independent datasets. I also found that reasoning performance was correlated with established measures of fluid intelligence. The work presented in Chapters 4, 5 and 6 confirmed that the Latin Square Task is a robust and reliable paradigm for examining brain network responses to complex reasoning problems.

In the experiments involving measures of brain activity, the cognitive control network (CCN) was associated with changes in reasoning complexity. Recall that the CCN is composed of a number of brain regions, including the dorsolateral prefrontal cortex (DLPFC), inferior frontal junction (IFJ), intraparietal sulcus (IPS), dorsal anterior cingulate cortex (DACC) and anterior insula (AI). Together, these regions constitute the FPN and CON. These observations were made across three independent datasets: two obtained in healthy adult cohorts, and another from a small group of individuals with callosal dysgenesis. Critically, CCN activity scaled with increases in the level of reasoning complexity. This finding is in line with a ‘multiple-demand’ account of cognitive control (Duncan, 2010; Duncan and Owen, 2000), which postulates that a variety of goal-directed behaviours are associated with widespread, network-level co-activation, rather than with discrete, focal activity in a specific brain region. It is notable that the same general pattern of network activity was observed in individuals with callosal dysgenesis (Chapter 6), implying that co-ordinated activity within the CCN is robust to a major developmental anomaly in white matter structure.

A central finding of the thesis was that increased network integration within the FPN was associated with increments in reasoning complexity. This pattern of connectivity reached a plateau in the most complex conditions, coinciding with a breakdown in successful task performance (Cocchi et al., 2014a). This result suggests that processing complex relations relies upon strengthened communication between spatially distributed brain regions. Indeed, a novel finding from this work was the involvement of regions outside of the CON and FPN, especially the default-mode network (DMN) and subcortical networks (see Section 7.2 for an in-depth discussion of the DMN). More broadly, this body of work suggests that the functional network architecture of the brain is not static but can be reorganized to match trial-by-trial task demands (Cole et al., 2013).

106 7.2 The interplay between default-mode and fronto-parietal functional connectivity in support of reasoning and intelligence

Current theories of the brain basis of intelligence suggest that activity within, and connectivity between, frontal and parietal cortices is related to individual differences in intelligence (Basten et al., 2013; Jung and Haier, 2007). Across experiments in this thesis, I found that the DMN was associated with reasoning behaviour. Specifically, individual differences in intelligence measures were associated with increased resting-state functional connectivity within and between anterior and temporal portions of the DMN (medial prefrontal cortex and temporal cortex), as well as regions of the FPN (right frontal gyrus and superior parietal cortex). I also observed that the angular gyri, key regions of the DMN, increased functional connectivity as relational reasoning complexity increased during the Wason Selection Task. A whole brain analysis of neural activity during the LST showed a similar effect: 24% of the connections that changed in response to increased reasoning demands involved the DMN.

The DMN has been proposed to activate in response to conditions which might broadly be described as internally generated cognition independent of the ‘here and now’ (Andrews-Hanna, 2012; Buckner et al., 2008). Such cognitive processes include recollection (Greicius and Menon, 2004), autobiographical thought (Svoboda et al., 2006), (Amodio and Frith, 2006), envisioning the future (Schacter et al., 2007), and other self-reflective thoughts (Gusnard et al., 2001). In line with this, the DMN is traditionally thought to be functionally antagonistic to control networks (Anticevic et al., 2012; Shulman et al., 1997). Evidence for this comes from work which has shown that at rest, activity within the DMN is negatively correlated with activity in attention networks (e.g., Fox et al., 2005), the extent of which can predict attention task performance (Kelly et al., 2008). Moreover, during external goal-directed cognitive control tasks, the DMN has been shown to decrease in activity (as shown in Chapter 3; see also Shulman et al., 1997). There is evidence that attenuated DMN suppression (i.e., a reduced negative change in the BOLD signal) is associated with worse performance during working memory tasks (Anticevic et al., 2010; Daselaar et al., 2004), suggesting that suppression of the DMN is a marker of attentional focus and disengagement from distracting thoughts (Anticevic et al., 2010; Binder, 2012).

The results from Chapters 2, 3 and 5 suggest it is likely that in certain task contexts increased functional dependence between the FPN and DMN is desirable. It might be important for communication between the FPN and DMN to increase with task complexity to successfully suppress activity associated with distracting or non-relevant thoughts (e.g., daydreaming; Buckner et al., 2008). Alternatively, such interactions might support memory (e.g., ‘What is the task rule?’) or meta- cognitive processes (Fornito et al., 2012). Regardless, the results validate an emerging view that FPN- DMN interactions are more variable across time and context than previously thought (Dixon et al., 2017). Critically, this developing literature encourages an update to the P-FIT model of intelligence. It

107 is now clear that properties of networks outside of the FPN can be tied to both individual differences in measures of intelligence, as well as the performance of tasks that are thought to ‘tap into’ intelligence (e.g., the WST and LST). An important future direction for research in this area will be to investigate how these specific network interactions coexist within a flexible global topology.

7.3 Global network properties that support increased cognitive demand

When investigating global network metrics it was observed that greater reasoning complexity was coupled with a reduction in network modularity and an increase in network efficiency. These effects essentially represent two ways of expressing a single network configuration shift (Rubinov, 2016) that are likely driven by changes in between-module connectivity. Moreover, the large-scale community structure of the brain was found to be flexible, but only in response to large cognitive shifts (e.g., from resting-state to complex problems, see Figure 7-1). This effect was underpinned by changes in between-network functional connections across relational complexity conditions (e.g., 95% of significant edges observed in Chapter 5 were between-network edges). Individuals with callosal dysgensis demonstrated intact networks but did not show the prototypical increase in functional connectivity in line with reasoning complexity. These results hint that disturbed connectomes may limit global reorganization needed for successful task performance.

Figure 7-1 Global workspace theory and network reconfigurations. a. Conceptual model of results from Chapter 5. At rest (left) modules are independent, but external goal-directed task states are accompanied by broad module-level changes (green). Increased task demands (far right) are accompanied by specific increases and decreases in connectivity, rather than further modular reconfiguration. b. Schematic of the global workspace

108 model (Dehaene et al., 1998). The global workspace (green) forms in order to share information across specialized processing modules (coloured nodes in the outer layer).

The parallel effect of decreased modularity and increased network efficiency is not specific to relational reasoning, and has now been reported in several different cognitive domains, including selective attention, working memory and visual discrimination (Bola and Sabel, 2015; Cohen and D’Esposito, 2016; Godwin et al., 2015; Kitzbichler et al., 2011; Shine et al., 2016; Vatansever et al., 2015; Westphal et al., 2017; Yue et al., 2017). It has been theorized that such reconfigurations arise to service increased information transfer (Achard and Bullmore, 2007; Shine et al., 2016; Shine and Poldrack, 2017). However, this increased capacity for information processing comes at a cost; segregated brain networks will prioritise communication within modules that are anatomically close, thus reducing the cost associated with long distance neural communication (Bullmore and Sporns, 2012).

The pattern of global connectivity reported in Chapter 5 is reminiscent of the global workspace theory of conscious processing (Baars, 1993; Dehaene et al., 1998). Global Workspace Theory proposes two main computational spaces. One is a set of specialized modules associated with cognitive processes such as attention, perception and memory. The other is a ‘global workspace’ that coalesces during effortful cognition in order to share information between the specialized modules. This idea has parallels in graph theoretical notions of modules (communities of brain regions that likely represent similar functions) and integration (increased functional connectivity). Thus, it has been argued that decreased network modularity and increased efficiency are markers of the emergence of a global workspace (Finc et al., 2017; Vatansever et al., 2015). This is in line with recent ideas of multi-network ‘meta-systems’ that form dynamically to underpin complex behaviour (Cocchi et al., 2013).

In addition to the work contained within this thesis, published research suggests that both the FPN and DMN express important characteristics of the global workspace during conscious information processing (Dehaene et al., 1998; Guldenmund et al., 2012; Leech et al., 2012; Cocchi et al., 2013; Vatansever et al., 2015). For instance, both networks show increased functional connectivity in response to more complex task demands. Also, both networks are comprised of domain-general cortex associated with behaviours that involve integrating information across more specialized brain regions (Duncan, 2010, Federenko et al., 2013, Buckner et al., 2008), and both networks possess anatomical hubs and long-range connections (van den Heuval, 2011, Collin et al., 2013). Moreover, it has been shown that increased efficiency and decreases in modularity correlate with better performance across a variety of domains (as described in detail in Chapter 5; see also Bassett et al., 2009; Bola and Sabel, 2015; Kitzbichler et al., 2011; van den Heuvel et al., 2009; Deniz Vatansever et al., 2015).

109 7.4 Future Directions

Understanding the neural mechanisms that give rise to intelligence and the processing of complex relations is not trivial. Global Workspace Theory provides a broad explanation for several key findings in my thesis, including the existence and function of reconfigurations in large-scale network topology (Section 7.1), the involvement of both the FPN and the DMN in response to increased cognitive demands (Section 7.2), and trends in network efficiency and modularity. Ultimately, however, this account is descriptive; a key remaining question is what neural mechanisms produce this phenomenon. The current thesis (and much of the literature) is limited by the correlative nature of brain imaging studies. Typical experiments assess whether the manipulation of behaviour (A) causes neural activity (B); it does not necessarily follow that B causes A. Therefore, in order to understand the neural principles that drive large-scale network reconfiguration supporting behaviour, future studies should utilize computational modelling or causal approaches, such as brain stimulation or lesion studies, in combination with brain imaging (Silvanto and Pascual-Leone, 2012).

While there are many possible future directions for the work outlined in my thesis (some of which are mentioned in each individual chapter), here I will briefly outline two non-invasive brain stimulation approaches that might profitably be used in conjunction with functional network analysis (see Figure 7-2). The advantage of these techniques over other methods (e.g., brain lesions or computational modelling) is that they allow the researcher to reversibly alter neural activity within a local or diffuse set of brain regions in vivo, and to measure the effects of such manipulations on behaviour. One of these approaches is transcranial magnetic stimulation (TMS), a technique that can alter activity of a focal area of the cortex (10-30 mm) and can produce immediate, or longer lasting (~20 minutes) excitation and inhibition effects within the cortex (Huang et al., 2005). Despite the focal nature of the direct influence of TMS on the cortex (Pascual-Leone et al., 1997), the indirect effects can propagate to distant areas, and are likely to depend on the connectivity profile of the stimulated region (Eldaief et al., 2011; Fox et al., 2012). For instance, stimulation of regions with hub properties will likely reveal more diffuse effects than stimulation of regions without hub properties (Cocchi et al., 2016).

The second stimulation technique is transcranial direct current stimulation (tDCS). Unlike TMS, tDCS uses an electrical current that is passed between two electrodes placed on the scalp (Nitsche et al., 2000). Contrary to TMS, which can cause neurons to fire action potentials, tDCS is thought to work by altering cells’ resting membrane potentials, thus rendering them more or less likely to fire under the influence of an external stimulus or task (Filmer et al., 2014). The electrode pads employed with tDCS can be placed to target a specific system (e.g., the motor network; Stagg et al., 2014). Like TMS, tDCS can be used to induce long-lasting increased (anodal stimulation) or decreased (cathodal stimulation) cortical excitability (Nitsche and Paulus, 2000).

110

Figure 7-2 Combined brain stimulation and imaging as a potential future direction to establish the relations between brain networks and nodes. In these hypothetical experiments participants would undertake a cognitive control task while brain activity is measured using fMRI (or EEG/MEG). a. TMS can be used to stimulate a focal patch of cortex (e.g., a single node in a brain network at the scale of fMRI) to suppress or enhance its activity. This method could be used to investigate the causal role of the targeted cortex (red node), as well as the functional connections (red edges) and nodes (orange nodes) that depend upon it. b. The influence of tDCS on cortical tissue is less focal than the effect of TMS, but the former could be used to modulate network- level or whole-brain network properties, for example by increasing fronto-parietal network connectivity (shown in green). Such diffuse stimulation might in fact be critical for investigating network-level interactions across widespread brain areas.

Armed with these brain stimulation tools, several hypotheses regarding the nature of reasoning complexity and global workspace theory could be empirically tested. One hypothesis is that disrupting key brain regions that form the global workspace should cause a detriment to behaviour (Baars et al., 2013). It is plausible that nodes with high degree during complex tasks, such as the lateral prefrontal cortex, could have a distinct role in managing such transitions. Critically, the causal effect of disrupting this node on the global workspace could be measured using network metrics, such as modularity and efficiency, in addition to traditional behavioural measures.

In addition, up to this point the association between DMN and higher cognitive ability has been largely correlational. Using TMS, this relationship could be teased apart to test the extent to which the DMN is critical for successful task performance, and what underlying cognitive processes are taking place. As mentioned previously, the DMN could be implicated in a host of different behaviours, from supressing non-relevant thoughts (Buckner et al., 2008) to memory (Fornito et al., 2012). Future behavioural paradigms, in conjunction with neuroimaging and TMS, could shed light on these unanswered questions.

Across my thesis, better performance on complex tasks was associated with increased FPN correlation and efficiency. This result has now been observed across a variety of task contexts (e.g., Shine et al., 2016; Yue et al., 2017). A question arising from this work is the directionality of the relationship.

111 Does increased network efficiency cause better behavioural performance? Or does better behavioural performance simply manifest in increased efficiency? One approach would be to use tDCS to experimentally manipulate the brain into a state of high or low efficiency (see, for example, Polanía et al., 2012) and measure behavioural outcomes. As demonstrated throughout this thesis (and through other work, e.g., Cole et al., 2013) the FPN would be the clear target for stimulation. Importantly, the causal effect of experimentally manipulating this network could be measured using fMRI-derived measures of network efficiency, as well as behavioural outcomes.

7.5 Conclusions

Reasoning is a complex, multifaceted behaviour that underpins intelligence. My aim in this thesis was to characterise changes in functional brain networks associated with cognitive reasoning, and to investigate the link between such dynamics and individual differences in ability. Overall, the findings revealed the dynamic nature of functional brain networks and the importance of default-mode and fronto-parietal network connectivity in the service of increasingly complex reasoning behaviour. It is suggested that recent network-based reconceptualizations of global workspace theory may provide a possible framework for understanding such changes. Future work should aim to test these potential accounts using methods that lend themselves to causal inferences, such as non-invasive brain stimulation.

112

References

113

Achard, S., Bullmore, E., 2007. Efficiency and Cost of Economical Brain Functional Networks. PLoS Comput. Biol. 3, 174–183. Aerts, H., Fias, W., Caeyenberghs, K., Marinazzo, D., 2016. Brain networks under attack: Robustness properties and the impact of lesions. Brain 139, 3063–3083. doi:10.1093/brain/aww194 Akshoomo, N., Beaumont, J.L., Bauer, P.J., Dikmen, S.S., Gershon, R.C., Mungas, D., Slotkin, J., Tulsky, D., Weintraub, S., Zelazo, P.D., Heaton, R.K., 2013. NIH Toolbox Cognition Battery (Cb): Composite Scores of Crystallized, Fluid, and Overall Cognition. Monogr. Soc. Res. Child Dev. 78, 119–132. Albright, T.D., 1984. Direction and orientation selectivity of neurons in visual area MT of the macaque Direction and Orientation Selectivity of Neurons in Visual Area MT of the Macaque. J. Neurophysiol. 52, 1106–1130. Amodio, D.M., Frith, C.D., 2006. Meeting of minds: the medial frontal cortex and social cognition. Nat. Rev. Neurosci. 7, 268. Ances, B.M., Liang, C.L., Leontiev, O., Perthen, J.E., Fleisher, A.S., Lansing, A.E., Buxton, R.B., 2009. Effects of aging on cerebral blood flow, oxygen metabolism, and blood oxygenation level dependent responses to visual stimulation. Hum. Brain Mapp. 30, 1120–1132. doi:10.1002/hbm.20574 Andrews-Hanna, J.R., 2012. The brain’s default network and its adaptive role in internal mentation. Neurosci. 18, 251–270. Andrews, G., Birney, D., Halford, G.S., 2006. Relational processing and working memory capacity in comprehension of relative clause sentences. Mem. Cognit. 34, 1325–1340. doi:10.3758/BF03193275 Andrews, G., Halford, G., Shum, D., Maujean, A., Chappell, M., Birney, D., 2013. Relational processing following stroke. Brain Cogn. 81, 44–51. doi:10.1016/j.bandc.2012.09.003 Andrews, G., Halford, G.S., 1998. Children’s ability to make transitive inferences: The importance of premise integration and structural complexity. Cogn. Dev. 13, 479–513. doi:10.1016/S0885- 2014(98)90004-1 Andrews, G., Todd, J.M., 2008. Two sources of age-related decline in comprehension of complex relative clause sentences. New Res. Short-Term Mem. 93–123. Anticevic, A., Cole, M.W., Murray, J.D., Corlett, P.R., Wang, X.-J., Krystal, J.H., 2012. The role of default network deactivation in cognition and disease. Trends Cogn. Sci. 16, 584–92. doi:10.1016/j.tics.2012.10.008 Anticevic, A., Repovs, G., Shulman, G.L., Barch, D.M., 2010. When less is more: TPJ and default network deactivation during predicts working memory performance. Neuroimage 49, 2638–2648.

114 Asaad, W.F., Rainer, G., Miller, E.K., 2000. Task-Specific Neural Activity in the Primate Prefrontal Cortex. J. Neurophysiol. 21, 451–459. doi:10.1016/S0896-6273(00)80658-3 Ashburner, J., 2007. A fast diffeomorphic image registration algorithm. Neuroimage 38, 95–113. doi:http://dx.doi.org/10.1016/j.neuroimage.2007.07.007 Baars, B.J., Franklin, S., Ramsoy, T.Z., 2013. Global workspace dynamics: Cortical “binding and propagation” enables conscious contents. Front. Psychol. 4, 1–22. doi:10.3389/fpsyg.2013.00200 Baars, B.J.B., 1988. A cognitive theory of consciousness, Cambridge University Press. Cambridge University Press. Bassett, D.S., Bullmore, E.T., Meyer-Lindenberg, A., Apud, J.A., Weinberger, D.R., Coppola, R., 2009. Cognitive fitness of cost-efficient brain functional networks. Proc. Natl. Acad. Sci. 106, 11747–11752. doi:10.1073/pnas.0903641106 Bassett, D.S., Wymbs, N.F., Porter, M. a, Mucha, P.J., Carlson, J.M., Grafton, S.T., 2011. Dynamic reconfiguration of human brain networks during learning. Proc. Natl. Acad. Sci. U. S. A. 108, 7641–6. doi:10.1073/pnas.1018985108 Basten, U., Hilger, K., Fiebach, C.J., 2015. Where smart brains are different: A quantitative meta- analysis of functional and structural brain imaging studies on intelligence. Intelligence 51, 10– 27. doi:10.1016/j.intell.2015.04.009 Basten, U., Stelzel, C., Fiebach, C.J., 2013. Intelligence is differentially related to neural effort in the task-positive and the task-negative brain network. Intelligence 41, 517–528. doi:10.1016/j.intell.2013.07.006 Batty, G.D., Wennerstad, K.M., Smith, G.D., Gunnell, D., Deary, I.J., Tynelius, P., Rasmussen, F., 2009. IQ in Early Adulthood and Mortality By Middle Age. Epidemiology 20, 100–109. doi:10.1097/EDE.0b013e31818ba076 Bedeschi, M.F., Bonaglia, M.C., Grasso, R., Pellegri, A., Garghentino, R.R., Battaglia, M.A., Panarisi, A.M., Di Rocco, M., Balottin, U., Bresolin, N., Bassi, M.T., Borgatti, R., 2006. Agenesis of the corpus callosum: Clinical and genetic study in 63 young patients. Pediatr. Neurol. 34, 186–193. doi:10.1016/j.pediatrneurol.2005.08.008 Behrens, T.E.J., Johansen-Berg, H., Woolrich, M.W., Smith, S.M., Wheeler-Kingshott, C. a M., Boulby, P. a, Barker, G.J., Sillery, E.L., Sheehan, K., Ciccarelli, O., Thompson, a J., Brady, J.M., Matthews, P.M., 2003. Non-invasive mapping of connections between human thalamus and cortex using diffusion imaging. Nat. Neurosci. 6, 750–7. doi:10.1038/nn1075 Behzadi, Y., Restom, K., Liau, J., Liu, T.T., 2007. A component based noise correction method (CompCor) for BOLD and perfusion based fMRI. Neuroimage 37, 90–101. doi:http://dx.doi.org/10.1016/j.neuroimage.2007.04.042 Bell, P.T., Shine, J.M., Shine, M., 2016. Subcortical contributions to large-scale network

115 communication. Neurosci. Biobehav. Rev. 71, 313–322. doi:http://dx.doi.org/10.1016/j.neubiorev.2016.08.036 Bhandari, A., Duncan, J., 2014. Goal neglect and knowledge chunking in the construction of novel behaviour. Cognition 130, 11–30. doi:10.1016/j.cognition.2013.08.013 Bilker, W.B., Hansen, J.A., Brensinger, C.M., Richard, J., Gur, R.E., Gur, R.C., 2012. Development of abbreviated nine-item forms of the Raven’s standard progressive matrices test. Assessment 19, 354–69. doi:10.1177/1073191112446655 Binder, J.R., 2012. Task-induced deactivation and the “resting” state. Neuroimage. doi:10.1016/j.neuroimage.2011.09.026 Binet, A., Simon, T., 1916. The development of intelligence in children: The Binet-Simon Scale. Williams & Wilkins Company. Birney, D.P., Halford, G.S., Andrews, G., 2006. Measuring the Influence of Complexity on Relational Reasoning: The Development of the Latin Square Task. Educ. Psychol. Meas. 66, 146–171. doi:10.1177/0013164405278570 Birney, Bowman, 2009. An experimental-differential investigation of cognitive complexity. Psychol. Sci. Q. 51, 449–469. Birney, Bowman, Beckmann, Seah, Z., 2012. Assessment of Processing Capacity. Eur. J. Psychol. Assess. 28, 216–226. doi:10.1027/1015-5759/a000146 Biswal, B., Zerrin Yetkin, F., Haughton, V.M., Hyde, J.S., 1995. Functional connectivity in the motor cortex of resting human brain using echo-planar mri. Magn. Reson. Med. 34, 537–541. Blondel, V.D., Guillaume, J.-L., Lambiotte, R., Lefebvre., E., 2008. Fast unfolding of communities in large networks. J. Stat. Mech. Theory Exp. 2008, P10008. Bola, M., Sabel, B. a., 2015. Dynamic reorganization of brain functional networks during cognition. Neuroimage 114, 398–413. doi:10.1016/j.neuroimage.2015.03.057 Bolt, T., Nomi, J.S., Rubinov, M., Uddin, L.Q., 2017. Correspondence between evoked and intrinsic functional brain network configurations. Hum. Brain Mapp. 38, 1992–2007. doi:10.1002/hbm.23500 Bors, D.A., Stokes, T.L., 1998. Raven’s Advanced Progressive Matrices: Norms for first-year university students and the development of a short form. Educ. Psychol. Meas. 58, 382–398. Brainard, D.H., 1997. The Psychophysics Toolbox. Spat. Vis. 10, 433–436. doi:10.1163/156856897X00357 Braun, U., Schäfer, A., Walter, H., Erk, S., Romanczuk-Seiferth, N., Haddad, L., Schweiger, J.I., Grimm, O., Heinz, A., Tost, H., Meyer-Lindenberg, A., Bassett, D.S., 2015. Dynamic reconfiguration of frontal brain networks during executive cognition in humans. Proc. Natl. Acad. Sci. 112, 11678–11683. doi:10.1073/pnas.1422487112 Brett, M., Anton, J.-L., Valabregue, R., Poline, J.-B., 2002. Region of interest analysis using an SPM

116 toolbox. Buckner, R.L., Andrews-Hanna, J.R., Schacter, D.L., 2008. The brain’s default network: Anatomy, function, and relevance to disease. Ann. N. Y. Acad. Sci. 1124, 1–38. doi:10.1196/annals.1440.011 Buckner, R.L., Sepulcre, J., Talukdar, T., Krienen, F.M., Liu, H., Hedden, T., Andrews-hanna, J.R., Sperling, R. a., Johnson, K. a., 2009. Cortical Hubs Revealed by Intrinsic Functional Connectivity: Mapping, Assessment of Stability, and Relation to Alzheimer’s Disease. J. Neurosci. 29, 1860–1873. doi:10.1523/JNEUROSCI.5062-08.2009.Cortical Bullmore, E., Sporns, O., 2012. The economy of brain network organization. Nat. Rev. Neurosci. 13, 336–49. doi:10.1038/nrn3214 Bunge, S.A., Helskog, E.H., Wendelken, C., 2009. Left, but not right, rostrolateral prefrontal cortex meets a stringent test of the relational integration hypothesis. Neuroimage 46, 338–342. doi:10.1016/j.neuroimage.2009.01.064 Caballero-Gaudes, C., Reynolds, R.C., 2017. Methods for cleaning the BOLD fMRI signal. Neuroimage 154, 128–149. doi:10.1016/j.neuroimage.2016.12.018 Calvin, C.M., Deary, I.J., Fenton, C., Roberts, B.A., Der, G., Leckenby, N., Batty, G.D., 2011. Intelligence in youth and all-cause-mortality: Systematic review with meta-analysis. Int. J. Epidemiol. 40, 626–644. doi:10.1093/ije/dyq190 Cattell, R., 1963. Theory of fluid and crystallized intelligence: A critical experiment. J. Educ. Psychol. 1–22. Cattell, R., 1949. The dimensions of culture patterns by factorization of national characters. J. Abnorm. Soc. Psychol. 44, 443–469. doi:10.1037/h0054760 Ceci, S.J., 1991. How much does schooling influence general intelligence and its cognitive components? A reassessment of the evidence. Dev. Psychol. 27, 703–722. doi:10.1037/0012- 1649.27.5.703 Chao-Gan, Y., Yu-Feng, Z., 2010. DPARSF: A MATLAB Toolbox for “Pipeline” Data Analysis of Resting-State fMRI. Front. Syst. Neurosci. 4, 1–7. doi:10.3389/fnsys.2010.00013 Chiang, A.S., Lin, C.Y., Chuang, C.C., Chang, H.M., Hsieh, C.H., Yeh, C.W., Shih, C.T., Wu, J.J., Wang, G.T., Chen, Y.C., Wu, C.C., Chen, G.Y., Ching, Y.T., Lee, P.C., Lin, C.Y., Lin, H.H., Wu, C.C., Hsu, H.W., Huang, Y.A., Chen, J.Y., Chiang, H.J., Lu, C.F., Ni, R.F., Yeh, C.Y., Hwang, J.K., 2011. Three-dimensional reconstruction of brain-wide wiring networks in drosophila at single-cell resolution. Curr. Biol. 21, 1–11. doi:10.1016/j.cub.2010.11.056 Christoff, K., Prabhakaran, V., Dorfman, J., Zhao, Z., Kroger, J.K., Holyoak, K.J., Gabrieli, J.D., 2001. Rostrolateral prefrontal cortex involvement in relational integration during reasoning. Neuroimage 14, 1136–1149. doi:10.1006/nimg.2001.0922 Ciric, R., Wolf, D.H., Power, J.D., Roalf, D.R., Baum, G., Ruparel, K., Shinohara, R.T., Elliott, M.A.,

117 Eickhoff, S.B., Davatzikos, C., Gur, R.C., Gur, R.E., Bassett, D.S., Satterthwaite, T.D., 2017. Benchmarking of participant-level confound regression strategies for the control of motion artifact in studies of functional connectivity. Neuroimage 154, 174–187. doi:10.1016/j.neuroimage.2017.03.020 Cocchi, L., Halford, G.S., Zalesky, A., Harding, I.H., Ramm, B.J., Cutmore, T., Shum, D.H.K., Mattingley, J.B., 2014a. Complexity in relational processing predicts changes in functional brain network dynamics. Cereb. Cortex 24, 2283–2296. doi:10.1093/cercor/bht075 Cocchi, L., Harding, I.H., Lord, A., Pantelis, C., Yucel, M., Zalesky, A., 2014b. Disruption of structure-function coupling in the connectome. NeuroImage. Clin. 4, 779–87. doi:10.1016/j.nicl.2014.05.004 Cocchi, L., Sale, M. V., Gollo, L.L., Bell, P.T., Nguyen, V.T., Zalesky, A., Breakspear, M., Mattingley, J.B., 2016. A hierarchy of timescales explains distinct effects of local inhibition of primary visual cortex and frontal eye fields. Elife 5, 1–17. doi:10.7554/eLife.15252 Cocchi, L., Yang, Z., Zalesky, A., Stelzer, J., Hearne, L.J., Gollo, L.L., Mattingley, J.B., 2017. Neural decoding of visual stimuli varies with fluctuations in global network efficiency. Hum. Brain Mapp. 38, 3069–3080. doi:10.1002/hbm.23574 Cocchi, L., Zalesky, A., Fornito, A., Mattingley, J.B., 2013. Dynamic cooperation and competition between brain systems during cognitive control. Trends Cogn. Sci. 17, 493–501. doi:10.1016/j.tics.2013.08.006 Cohen, J.R., D’Esposito, M., 2016. The Segregation and Integration of Distinct Brain Networks and Their Relationship to Cognition. J. Neurosci. 36, 12083–12094. doi:10.1523/JNEUROSCI.2965- 15.2016 Cole, M., Etzel, J., Zacks, J., Schneider, W., Braver, T., 2011. Rapid Transfer of Abstract Rules to Novel Contexts in Human Lateral Prefrontal Cortex. Front. Hum. Neurosci. 5, 1–13. doi:10.3389/fnhum.2011.00142 Cole, M.W., Bassett, D.S., Power, J.D., Braver, T.S., Petersen, S.E., 2014. Intrinsic and task-evoked network architectures of the human brain. Neuron 83, 238–251. doi:10.1016/j.neuron.2014.05.014 Cole, M.W., Ito, T., Bassett, D.S., Schultz, D.H., 2016. Activity flow over resting-state networks shapes cognitive task activations. Nat. Neurosci. 19, 1718–1726. doi:10.1038/nn.4406 Cole, M.W., Ito, T., Braver, T.S., 2015. Lateral Prefrontal Cortex Contributes to Fluid Intelligence Through Multinetwork Connectivity. Brain Connect. 5, 497–504. doi:10.1089/brain.2015.0357 Cole, M.W., Reynolds, J.R., Power, J.D., Repovs, G., Anticevic, A., Braver, T.S., 2013. Multi-task connectivity reveals flexible hubs for adaptive task control. Nat. Neurosci. 16, 1348–1355. doi:10.1038/nn.3470 Cole, M.W., Schneider, W., 2007. The cognitive control network: Integrated cortical regions with

118 dissociable functions. Neuroimage 37, 343–60. doi:10.1016/j.neuroimage.2007.03.071 Cole, M.W., Takuya, I., Braver, T.S., Ito, T., Braver, T.S., 2015. Lateral prefrontal cortex contributes to fluid intelligence via multi-network connectivity. Brain Connect. 5, 497–504. doi:10.1089/brain.2015.0357 Cole, M.W., Yarkoni, T., Repovs, G., Anticevic, A., Braver, T.S., 2012. Global connectivity of prefrontal cortex predicts cognitive control and intelligence. J. Neurosci. 32, 8988–99. doi:10.1523/JNEUROSCI.0536-12.2012 Concha, L., Gross, D.W., Wheatley, B.M., Beaulieu, C., 2006. Diffusion tensor imaging of time- dependent axonal and myelin degradation after corpus callosotomy in epilepsy patients. Neuroimage 32, 1090–1099. Crittenden, B.M., Mitchell, D.J., Duncan, J., 2016. Task Encoding across the Multiple Demand Cortex Is Consistent with a Frontoparietal and Cingulo-Opercular Dual Networks Distinction. J. Neurosci. 36, 6147–6155. doi:10.1523/JNEUROSCI.4590-15.2016 Crone, E. a, Wendelken, C., van Leijenhorst, L., Honomichl, R.D., Christoff, K., Bunge, S. a, 2009. Neurocognitive development of relational reasoning. Dev. Sci. 12, 55–66. doi:10.1111/j.1467- 7687.2008.00743.x Crossley, N.A., Mechelli, A., Vertes, P.E., Winton-Brown, T.T., Patel, A.X., Ginestet, C.E., McGuire, P., Bullmore, E.T., 2013. Cognitive relevance of the community structure of the human brain functional coactivation network. Proc. Natl. Acad. Sci. 110, 11583–11588. doi:10.1073/pnas.1220826110 D’Esposito, M., Deouell, L.Y., Gazzaley, A., 2003. Alterations in the BOLD fMRI signal with ageing and disease: a challenge for neuroimaging. Nat. Rev. Neurosci. 4, 863–872. doi:10.1038/nrn1246 Damle, N.R., Ikuta, T., John, M., Peters, B.D., DeRosse, P., Malhotra, A.K., Szeszko, P.R., 2017. Relationship among interthalamic adhesion size, thalamic anatomy and neuropsychological functions in healthy volunteers. Brain Struct. Funct. 222, 2183–2192. doi:10.1007/s00429-016- 1334-6 Damoiseaux, J.S., Rombouts, S.A.R.B., Barkhof, F., Scheltens, P., Stam, C.J., Smith, S.M., Beckmann, C.F., 2006. Consistent resting-state networks across healthy subjects. Proc. Natl. Acad. Sci. U. S. A. 103, 13848–53. doi:10.1073/pnas.0601417103 Daselaar, S.M., Prince, S.E., Cabeza, R., 2004. When less means more: Deactivations during encoding that predict subsequent memory. Neuroimage 23, 921–927. doi:10.1016/j.neuroimage.2004.07.031 de Pasquale, F., Della Penna, S., Snyder, A.Z., Lewis, C., Mantini, D., Marzetti, L., Belardinelli, P., Ciancetta, L., Pizzella, V., Romani, G.L., Corbetta, M., 2010. Temporal dynamics of spontaneous MEG activity in brain networks. Proc. Natl. Acad. Sci. U. S. A. 107, 6040–6045.

119 doi:10.1073/pnas.0913863107 de Pasquale, F., Della Penna, S., Sporns, O., Romani, G.L., Corbetta, M., 2016. A Dynamic Core Network and Global Efficiency in the Resting Human Brain. Cereb. Cortex 26, 4015–4033. doi:10.1093/cercor/bhv185 Deary, I.J., Strand, S., Smith, P., Fernandes, C., 2007. Intelligence and educational achievement. Intelligence 35, 13–21. doi:10.1016/j.intell.2006.02.001 Deco, G., Jirsa, V.K., 2012. Ongoing cortical activity at rest: Criticality, multistability, and ghost attractors. J. Neurosci. 32, 3366–3375. doi:10.1523/JNEUROSCI.2523-11.2012 Deco, G., Jirsa, V.K., McIntosh, A.R., 2011. Emerging concepts for the dynamical organization of resting-state activity in the brain. Nat. Rev. Neurosci. 12, 43–56. doi:10.1038/nrn2961 Dehaene, S., Kerszberg, M., Changeux, J.-P., 1998. A neuronal model of a global workspace in effortful cognitive tasks. Proc. Natl. Acad. Sci. 95, 14529–14534. doi:10.1073/pnas.95.24.14529 Deuker, L., Bullmore, E.T., Smith, M., Christensen, S., Nathan, P.J., Rockstroh, B., Bassett, D.S., 2009. Reproducibility of graph metrics of human brain functional networks. Neuroimage 47, 1460–1468. Dixon, M.L., Andrews-hanna, J.R., Spreng, R.N., Irving, Z.C., Mills, C., Girn, M., Christo, K., 2017. Interactions between the default network and dorsal attention network vary across default subsystems, time and cognitive states. Neuroimage 147, 632–649. doi:10.1016/j.neuroimage.2016.12.073 Dosenbach, N.U.F., Fair, D.A., Miezin, F.M., Cohen, A.L., Wenger, K.K., Dosenbach, R.A.T., Fox, M.D., Snyder, A.Z., Vincent, J.L., Raichle, M.E., Schlaggar, B.L., Petersen, S.E., 2007. Distinct brain networks for adaptive and stable task control in humans 104. Dosenbach, N.U.F., Nardos, B., Cohen, A.L., Fair, D. a, Power, J.D., Church, J. a, Nelson, S.M., Wig, G.S., Vogel, A.C., Lessov-Schlaggar, C.N., Barnes, K.A., Dubis, J.W., Feczko, E., Coalson, R.S., Pruett, J.R., Barch, D.M., Petersen, S.E., Schlaggar, B.L., 2010. Prediction of individual brain maturity using fMRI. Science 329, 1358–61. doi:10.1126/science.1194144 Dosenbach, N.U.F., Visscher, K.M., Palmer, E.D., Miezin, F.M., Wenger, K.K., Kang, H.C., Burgund, E.D., Grimes, A.L., Schlaggar, B.L., Petersen, S.E., 2006. A core system for the implementation of task sets. Neuron 50, 799–812. doi:10.1016/j.neuron.2006.04.031 Duncan, J., 2010. The multiple-demand (MD) system of the primate brain: mental programs for intelligent behaviour. Trends Cogn. Sci. 14, 172–9. doi:10.1016/j.tics.2010.01.004 Duncan, J., 2001. An adaptive coding model of neural function in prefrontal cortex. Nat. Rev. Neurosci. 2, 820–829. doi:10.1038/35097575 Duncan, J., Burgess, P., Emslie, H., 1995. Fluid intelligence after frontal lobe lesions. Neuropsychologia 33, 261–268. doi:0028-3932(94)00124-3 Duncan, J., Owen, a M., 2000. Common regions of the human frontal lobe recruited by diverse

120 cognitive demands. Trends Neurosci. 23, 475–83. Duncan, J., Parr, A., Woolgar, A., Thompson, R., Bright, P., Cox, S., Bishop, S., Nimmo-Smith, I., 2008. Goal neglect and Spearman’s g: competing parts of a complex task. J. Exp. Psychol. Gen. 137, 131–48. doi:10.1037/0096-3445.137.1.131 Duncan, J., Seltz, R., Kolodny, J., Bor, D., Herzog, H., Ahmed, A., Newell, F., Emslie, H., 2000. A neural basis for general intelligence. Science (80-. ). 289, 457–460. Duquette, M., Rainville, P., Alary, F., Lassonde, M., Lepore, F., 2008. Ipsilateral cortical representation of tactile and painful information in acallosal and callosotomized subjects. Neuropsychologia 46, 2274–2279. doi:10.1016/j.neuropsychologia.2008.02.017 Dwyer, D.B., Harrison, B.J., Yucel, M., Whittle, S., Zalesky, a., Pantelis, C., Allen, N.B., Fornito, a., 2014. Large-Scale Brain Network Dynamics Supporting Adolescent Cognitive Control. J. Neurosci. 34, 14096–14107. doi:10.1523/JNEUROSCI.1634-14.2014 Eklund, A., Nichols, T.E., Knutsson, H., 2016. Cluster failure: Why fMRI inferences for spatial extent have inflated false-positive rates. Proc. Natl. Acad. Sci. 201602413. doi:10.1073/pnas.1602413113 Eldaief, M.C., Halko, M. a., Buckner, R.L., Pascual-Leone, a., 2011. Transcranial magnetic stimulation modulates the brain’s intrinsic activity in a frequency-dependent manner. Proc. Natl. Acad. Sci. 108, 21229–21234. doi:10.1073/pnas.1113103109 Fedorenko, E., Duncan, J., Kanwisher, N., 2013. Broad domain generality in focal regions of frontal and parietal cortex. Proc. Natl. Acad. Sci. 110, 16616–16621. doi:10.1073/pnas.1315235110 Filmer, H.L., Dux, P.E., Mattingley, J.B., 2014. Applications of transcranial direct current stimulation for understanding brain function. Trends Neurosci. 37, 742–753. doi:http://dx.doi.org/10.1016/j.tins.2014.08.003 Finc, K., Bonna, K., Lewandowska, M., Wolak, T., Nikadon, J., Dreszer, J., Duch, W., Kühn, S., 2017. Transition of the functional brain network related to increasing cognitive demands. Hum. Brain Mapp. 0. doi:10.1002/hbm.23621 Finn, E.S., Shen, X., Scheinost, D., Rosenberg, M.D., Huang, J., Chun, M.M., Papademetris, X., Constable, R.T., 2015. Functional connectome fingerprinting: identifying individuals using patterns of brain connectivity. Nat. Neurosci. 1–11. doi:10.1038/nn.4135 Fornito, A., Harrison, B.J., Zalesky, A., Simons, J.S., 2012. Competitive and cooperative dynamics of large-scale brain functional networks supporting recollection. Proc. Natl. Acad. Sci. 109, 12788– 12793. doi:10.1073/pnas.1204185109/- /DCSupplemental.www.pnas.org/cgi/doi/10.1073/pnas.1204185109 Fornito, A., Zalesky, A., Breakspear, M., 2015a. The connectomics of brain disorders. Nat. Rev. Neurosci. 16, 159–172. doi:10.1038/nrn3901 Fornito, A., Zalesky, A., Bullmore, E.T., 2015b. Fundamentals of Brain Network Analysis, 1st ed.

121 Elsevier Inc. Fortunato, S., 2010. Community detection in graphs. Phys. Rep. 486, 75–174. doi:10.1016/j.physrep.2009.11.002 Fox, M.D., Corbetta, M., Snyder, A.Z., Vincent, J.L., Raichle, M.E., 2006. Spontaneous neuronal activity distinguishes human dorsal and ventral attention systems. Proc. Natl. Acad. Sci. U. S. A. 103, 10046–51. doi:10.1073/pnas.0604187103 Fox, M.D., Halko, M.A., Eldaief, M.C., Pascual-Leone, A., 2012. Measuring and manipulating brain connectivity with resting state functional connectivity magnetic resonance imaging (fcMRI) and transcranial magnetic stimulation (TMS). Neuroimage 62, 2232–2243. doi:10.1016/j.neuroimage.2012.03.035 Fox, M.D., Snyder, A.Z., Vincent, J.L., Corbetta, M., Van Essen, D.C., Raichle, M.E., 2005. The human brain is intrinsically organized into dynamic, anticorrelated functional networks. Proc. Natl. Acad. Sci. U. S. A. 102, 9673–8. doi:10.1073/pnas.0504136102 Fransson, P., Krüger, G., Merboldt, K.D., Frahm, J., 1999. Temporal and spatial MRI responses to subsecond visual activation. Magn. Reson. Imaging 17, 1–7. doi:10.1016/S0730- 725X(98)00163-5 Friston, K., 2011. Functional and effective connectivity: a review. Brain Connect. 1, 13–36. doi:10.1089/brain.2011.0008 Friston, K.J., 1994. Functional and effective connectivity in neuroimaging: a synthesis. Hum. Brain Mapp. 2, 56–78. Friston, K.J., Harrison, L., Penny, W., 2003. Dynamic causal modelling. Neuroimage 19, 1273–1302. Friston, K.J., Holmes, A.P., Worsley, K.J., Poline, J., Frith, C.D., Frackowiak, R.S.J., 1994. Statistical parametric maps in functional imaging: a general linear approach. Hum. Brain Mapp. 2, 189– 210. Garrison, K.A., Scheinost, D., Finn, E.S., Shen, X., Constable, R.T., 2015. The (in)stability of functional brain network measures across thresholds. Neuroimage 118, 651–661. doi:10.1016/j.neuroimage.2015.05.046 Gazzaniga, M.S., Bogen, J.E., Sperry, R.W., 1962. Some functional effects of sectioning the cerebral commissures in man. Proc. Natl. Acad. Sci. U. S. A. 48, 1765–1769. doi:10.1073/pnas.48.10.1765 Gerchen, M.F., Bernal-Casas, D., Kirsch, P., 2014. Analyzing task-dependent brain network changes by whole-brain psychophysiological interactions: A comparison to conventional analysis. Hum. Brain Mapp. 35, 5071–5082. doi:10.1002/hbm.22532 Gilbert, S.J., Bird, G., Frith, C.D., Burgess, P.W., 2012. Does “task difficulty” explain “task-induced deactivation?” Front. Psychol. 3, 1–12. doi:10.3389/fpsyg.2012.00125 Glass, H.C., Shaw, G.M., Ma, C., Sherr, E.H., 2008. Agenesis of the corpus callosum in California

122 1983-2003: A population-based study. Am. J. Med. Genet. Part A 146, 2495–2500. doi:10.1002/ajmg.a.32418 Glasser, M.F., Sotiropoulos, S.N., Wilson, J.A., Coalson, T.S., Fischl, B., Andersson, J.L., Xu, J., Jbabdi, S., Webster, M., Polimeni, J.R., Van Essen, D.C., Jenkinson, M., 2013. The minimal preprocessing pipelines for the Human Connectome Project. Neuroimage 80, 105–124. doi:10.1016/j.neuroimage.2013.04.127 Godwin, D., Barry, R.L., Marois, R., 2015. Breakdown of the brain’s functional network modularity with awareness. Proc Natl Acad Sci U S A 112, 3799–3804. doi:10.1073/pnas.1414466112 Golde, M., von Cramon, D.Y., Schubotz, R.I., 2010. Differential role of anterior prefrontal and premotor cortex in the processing of relational information. Neuroimage 49, 2890–900. doi:10.1016/j.neuroimage.2009.09.009 Good, B.H., de Montjoye, Y.-A.A., Clauset, A., 2010. Performance of modularity maximization in practical contexts. Phys. Rev. E 81, 46106. doi:10.1103/PhysRevE.81.046106 Gordon, E.M., Laumann, T.O., Adeyemo, B., Huckins, J.F., Kelley, W.M., Petersen, S.E., 2016. Generation and Evaluation of a Cortical Area Parcellation from Resting-State Correlations. Cereb. Cortex 26, 288–303. doi:10.1093/cercor/bhu239 Gottfredson, L.S., Deary, I.J., 2004. Intelligence Predicts Health and Longevity, but Why? Curr. Dir. Psychol. Sci. 13, 1–4. doi:10.1111/j.0963-7214.2004.01301001.x Gould, R.L., Brown, R.G., Owen, A.M., H, D., Howard, R.J., 2003. fMRI BOLD response to increasing task difficulty during successful paired associates learning 20, 1006–1019. doi:10.1016/S1053-8119(03)00365-3 Gray, J.R., Chabris, C.F., Braver, T.S., 2003. Neural mechanisms of general fluid intelligence. Nat. Neurosci. 6, 316–22. doi:10.1038/nn1014 Greicius, M.D., Krasnow, B., Reiss, A.L., Menon, V., 2003. Functional connectivity in the resting brain: a network analysis of the default mode hypothesis. Proc. Natl. Acad. Sci. U. S. A. 100, 253–8. doi:10.1073/pnas.0135058100 Greicius, M.D., Menon, V., 2004. Default-Mode Activity during a Passive Sensory Task: Uncoupled from Deactivation but Impacting Activation. J. Cogn. Neurosci. 16, 1484–1492. doi:10.1162/0898929042568532 Gu, S., Satterthwaite, T.D., Medaglia, J.D., Yang, M., Gur, R.E., Gur, R.C., Bassett, D.S., 2015. Emergence of system roles in normative neurodevelopment. Proc. Natl. Acad. Sci. 112, 201502829. doi:10.1073/pnas.1502829112 Guimera, R., Guimerà, R., Nunes Amaral, L. a, 2005. Functional cartography of complex metabolic networks. Nature 433, 895–900. doi:10.1038/nature03288 Guldenmund, P., Vanhaudenhuyse, A., Boly, M., Laureys, S., Soddu, A., 2012. A default mode of brain function in altered states of consciousness. Arch. Ital. Biol. 150, 107–121.

123 doi:10.4449/aib.v150i2.1373 Gusnard, D.A., Akbudak, E., Shulman, G.L., Raichle, M.E., 2001. Medial prefrontal cortex and self- referential mental activity: relation to a default mode of brain function. Proc. Natl. Acad. Sci. U. S. A. 98, 4259–64. doi:10.1073/pnas.071043098 Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C.J., Van Wedeen, J., Sporns, O., 2008. Mapping the structural core of human cerebral cortex. PLoS Biol. 6, 1479–1493. doi:10.1371/journal.pbio.0060159 Halford, G., Wilson, W., Phillips, S., 2010. Relational knowledge: the foundation of higher cognition. Trends Cogn. Sci. 14, 497–505. doi:10.1016/j.tics.2010.08.005 Halford, G.S., Baker, R., McCredden, J.E., Bain, J.D., 2005. How many variables can humans process? Psychol. Sci. 16, 70–76. doi:10.1111/j.0956-7976.2005.00782.x Halford, Wilson, Phillips, S., 1998. Processing capacity defined by relational complexity: implications for comparative, developmental, and cognitive psychology. Behav. Brain Sci. 21, 803-31-64. Hansell, N.K., Halford, G.S., Andrews, G., Shum, D.H.K., Harris, S.E., Davies, G., Franic, S., Christoforou, A., Zietsch, B., Painter, J., Medland, S.E., Ehli, E. a., Davies, G.E., Steen, V.M., Lundervold, A.J., Reinvang, I., Montgomery, G.W., Espeseth, T., Hulshoff Pol, H.E., Starr, J.M., Martin, N.G., Le Hellard, S., Boomsma, D.I., Deary, I.J., Wright, M.J., 2015. Genetic Basis of a Cognitive Complexity Metric. PLoS One 10, e0123886. doi:10.1371/journal.pone.0123886 He, B.J., Snyder, A.Z., Zempel, J.M., Smyth, M.D., Raichle, M.E., 2008. Electrophysiological correlates of the brain’s intrinsic large-scale functional architecture. Proc Natl Acad Sci U S A 105, 16039–16044. doi:10.1073/pnas.0807010105 Hearne, L.J., Cocchi, L., Zalesky, A., Mattingley, J.B., 2017. Reconfiguration of brain network architectures between resting-state and complexity-dependent cognitive reasoning. J. Neurosci. 37, 485–17. doi:10.1523/JNEUROSCI.0485-17.2017 Hearne, L.J., Cocchi, L., Zalesky, A., Mattingley, J.B., 2015. Interactions between default mode and control networks as a function of increasing cognitive reasoning complexity. Hum. Brain Mapp. 36, 2719–2731. doi:10.1002/hbm.22802 Hearne, L.J., Mattingley, J.B., Cocchi, L., 2016. Functional brain networks related to individual differences in human intelligence at rest. Sci. Rep. 6, 32328. doi:10.1038/srep32328 Hellyer, P.J., Shanahan, M., Scott, G., Wise, R.J.S., Sharp, D.J., Leech, R., 2014. The control of global brain dynamics: opposing actions of frontoparietal control and default mode networks on attention. J. Neurosci. 34, 451–61. doi:10.1523/JNEUROSCI.1853-13.2014 Hetts, S.W., Sherr, E.H., Chao, S., Gobuty, S., Barkovich, a J., 2006. Anomalies of the corpus callosum: an MR analysis of the phenotypic spectrum of associated malformations. AJR. Am. J. Roentgenol. 187, 1343–8. doi:10.2214/AJR.05.0146

124 Hillman, E., 2014. Coupling Mechanism and Significance of the BOLD Signal: A Status Report. Annu. Rev. Neurosci. 37, 161–181. doi:10.1146/annurev-neuro-071013-014111 Hinkley, L.B.N., Marco, E.J., Brown, E.G., Bukshpun, P., Gold, J., Hill, S., Findlay, A.M., Jeremy, R.J., Wakahiro, M.L., Barkovich, A.J., Mukherjee, P., Sherr, E.H., Nagarajan, S.S., 2016. The Contribution of the Corpus Callosum to Language Lateralization. J. Neurosci. 36, 4522–4533. doi:10.1523/JNEUROSCI.3850-14.2016 Honey, C.J., Sporns, O., 2008. Dynamical consequences of lesions in cortical networks. Hum. Brain Mapp. 29, 802–809. doi:10.1002/hbm.20579 Honey, C.J., Sporns, O., Cammoun, L., Gigandet, X., Thiran, J.P., Meuli, R., Hagmann, P., 2009. Predicting human resting-state functional connectivity from structural connectivity. Proc. Natl. Acad. Sci. U. S. A. 106, 2035–40. doi:10.1073/pnas.0811168106 Horn, J.L., 1998. A basis for research on age differences in cognitive capabilities. Hum. Cogn. Abil. theory Pract. 57–91. Howarth, C., 2014. The contribution of astrocytes to the regulation of cerebral blood flow. Front. Neurosci. 8, 1–9. doi:10.3389/fnins.2014.00103 Huang, Y.Z., Edwards, M.J., Rounis, E., Bhatia, K.P., Rothwell, J.C., 2005. Theta burst stimulation of the human motor cortex. Neuron 45, 201–206. doi:10.1016/j.neuron.2004.12.033 Hugdahl, K., Raichle, M.E., Mitra, A., Specht, K., 2015. On the existence of a generalized non- specific task-dependent network. Front. Hum. Neurosci. 9, 1–15. doi:10.3389/fnhum.2015.00430 Hunter, J.E., 1986. Cognitive ability, cognitive aptitudes, job knowledge, and job performance. J. Vocat. Behav. 29, 340–362. doi:10.1016/0001-8791(86)90013-8 Hwang, K., Bertolero, M.A., Liu, W.B., D’Esposito, M., 2017. The human thalamus is an integrative hub for functional brain networks. J. Neurosci. 37, 5594–5607. Ibos, G., Hamed, S. Ben, Cognitive, C.D.N., Umr, C., Lyon, C.B., 2013. A Functional Hierarchy within the Parietofrontal Network in Stimulus Selection and Attention Control 33, 8359–8369. doi:10.1523/JNEUROSCI.4058-12.2013 Jeffereys, K., 1984. R. a. Butler, the Board of Education and the 1944 Education Act, University of London. doi:10.1111/j.1468-229X.1984.tb01429.x Jensen, A.R., 1998. The Science of Mental Ability. Praeger Publishers. Jenson, A., 1987. The Influence of Cognitive Psychology on Testing. Johnson-Laird, P.N., 2010. Mental models and human reasoning. Proc. Natl. Acad. Sci. 107, 18243– 18250. doi:10.1073/pnas.1012933107 Johnson, W., Bouchard, T.J., 2005. The structure of human intelligence: It is verbal, perceptual, and image rotation (VPR), not fluid and crystallized. Intelligence 33, 393–416. doi:10.1016/j.intell.2004.12.002

125 Jung, R.E., Haier, R.J., 2007. The Parieto-Frontal Integration Theory (P-FIT) of intelligence: converging neuroimaging evidence. Behav. brain Sci. 30, 135–154; doi:10.1017/S0140525X07001185 Kadohisa, M., Petrov, P., Stokes, M., Sigala, N., Buckley, M., Gaffan, D., Kusunoki, M., Duncan, J., 2013. Article Dynamic Construction of a Coherent Attentional State in a Prefrontal Cell Population. Neuron 80, 235–246. doi:10.1016/j.neuron.2013.07.041 Kanwisher, N., McDermott, J., Chun, M.M., 1997. The fusiform face area: a module in human extrastriate cortex specialized for face perception. J. Neurosci. 17, 4302–4311. Keller, C.J., Bickel, S., Honey, C.J., Groppe, D.M., Entz, L., Craddock, R.C., Lado, F. a, Kelly, C., Milham, M., Mehta, A.D., 2013. Neurophysiological investigation of spontaneous correlated and anticorrelated fluctuations of the BOLD signal. J. Neurosci. 33, 6333–42. doi:10.1523/JNEUROSCI.4837-12.2013 Kelly, A.M.C., Uddin, L.Q., Biswal, B.B., Castellanos, F.X., Milham, M.P., 2008. Competition between functional brain networks mediates behavioral variability 39, 527–537. doi:10.1016/j.neuroimage.2007.08.008 Kitzbichler, M.G., Henson, R.N.A., Smith, M.L., Nathan, P.J., Bullmore, E.T., 2011. Cognitive Effort Drives Workspace Configuration of Human Brain Functional Networks. J. Neurosci. 31, 8259– 8270. doi:10.1523/JNEUROSCI.0440-11.2011 Kleiner, M., Brainard, D., Pelli, D., Ingling, A., Murray, R., 2007. What’s new in Psychtoolbox-3. Perception 36, 1. doi:10.1068/v070821 Knowlton, B.J., Morrison, R.G., Hummel, J.E., Holyoak, K.J., 2012. A neurocomputational system for relational reasoning. Trends Cogn. Sci. 16, 373–381. doi:10.1016/j.tics.2012.06.002 Krienen, F.M., Yeo, B.T.T., Buckner, R.L., 2014. Reconfigurable task-dependent functional coupling modes cluster around a core functional architecture. Philos. Trans. R. Soc. Lond. B. Biol. Sci. 369. doi:10.1098/rstb.2013.0526 Kroger, J.K., 2002. Recruitment of Anterior Dorsolateral Prefrontal Cortex in Human Reasoning: a Parametric Study of Relational Complexity. Cereb. Cortex 12, 477–485. doi:10.1093/cercor/12.5.477 Kumar, A., Rotter, S., Aertsen, A., 2010. Spiking activity propagation in neuronal networks: reconciling different perspectives on neural coding. Nat. Rev. Neurosci. 11, 615–627. doi:10.1038/nrn2886 Lancichinetti, A., Fortunato, S., 2012. Consensus clustering in complex networks. Sci. Rep. 2, 336. doi:10.1038/srep00336 Langer, N., Pedroni, A., Gianotti, L.R.R.R., Hänggi, J., Knoch, D., Jäncke, L., 2011. Functional brain network efficiency predicts intelligence. Hum. Brain Mapp. 33, 1393–1406. doi:10.1002/hbm.21297

126 Langeslag, S.J.E., Schmidt, M., Ghassabian, A., Jaddoe, V.W., Hofman, A., van der Lugt, A., Verhulst, F.C., Tiemeier, H., White, T.J.H., 2013. Functional connectivity between parietal and frontal brain regions and intelligence in young children: the Generation R study. Hum. Brain Mapp. 34, 3299–307. doi:10.1002/hbm.22143 Lassonde, M., Sauerwein, H., Chicoine, A.J., Geoffroy, G., 1991. Absence of disconnexion syndrome in callosal agenesis and early callosotomy: Brain reorganization or lack of structural specificity during ontogeny? Neuropsychologia 29, 481–495. doi:10.1016/0028-3932(91)90006-T Latora, V., Marchiori, M., 2001. Efficient Behavior of Small-World Networks. Phys. Rev. Lett. 87, 198701. Laumann, T.O., Snyder, A.Z., Zhou, H.J., Yeo, B.T.T., 2017. Interpreting Temporal Fluctuations in Resting-State Functional Connectivity MRI. doi:10.1101/135681 Lee, I.A., Preacher, K.J., 2013. Calculation for the test of the difference between two dependent correlations with one variable in common. Comput. Softw. doi:http://quantpsy.org/corrtest/corrtest2.htm Lee, K.H., Choi, Y.Y., Gray, J.R., Cho, S.H., Chae, J.H., Lee, S., Kim, K., 2006. Neural correlates of superior intelligence: Stronger recruitment of posterior parietal cortex. Neuroimage 29, 578–586. doi:10.1016/j.neuroimage.2005.07.036 Leech, R., Braga, R., Sharp, D.J., 2012. Echoes of the Brain within the Posterior Cingulate Cortex. J. Neurosci. 32, 215–222. doi:10.1523/JNEUROSCI.3689-11.2012 Leech, R., Kamourieh, S., Beckmann, C.F., Sharp, D.J., 2011. Fractionating the default mode network: distinct contributions of the ventral and dorsal posterior cingulate cortex to cognitive control. J. Neurosci. 31, 3217–24. doi:10.1523/JNEUROSCI.5626-10.2011 Li, S., Ostwald, D., Giese, M., Kourtzi, Z., 2007. Flexible coding for categorical decisions in the human brain. J. Neurosci. 27, 12321–30. doi:10.1523/JNEUROSCI.3795-07.2007 Lipp, I., Benedek, M., Fink, A., Koschutnig, K., Reishofer, G., Bergner, S., Ischebeck, A., Ebner, F., Neubauer, A., 2012. Investigating neural efficiency in the visuo-spatial domain: an FMRI study. PLoS One 7, e51316. doi:10.1371/journal.pone.0051316 Logothetis, N.K., Pauls, J., Augath, M., Trinath, T., Oeltermann, a, 2001. Neurophysiological investigation of the basis of the fMRI signal. Nature 412, 150–157. doi:10.1038/35084005 Logothetis, N.K., Wandell, B. a, 2004. Interpreting the BOLD signal. Annu. Rev. Physiol. 66, 735– 769. doi:10.1146/annurev.physiol.66.082602.092845 Luria, A., 1966. Higher cortical functions in man. Tavistock, London. Lynall, M.-E., Bassett, D.S., Kerwin, R., McKenna, P.J., Kitzbichler, M., Muller, U., Bullmore, E., 2010. Functional Connectivity and Brain Networks in Schizophrenia. J. Neurosci. 30, 9477– 9487. doi:10.1523/JNEUROSCI.0333-10.2010 Lynn, R., 2010. In Italy, north-south differences in IQ predict differences in income, education, infant

127 mortality, stature, and literacy. Intelligence 38, 93–100. doi:10.1016/j.intell.2009.07.004 Malonek, D., Grinvald, A., 1996. Interactions between Electrical Activity and Cortical Microcirculation Revealed by Imaging Spectroscopy : Implications for Functional Brian Mapping. Science (80-. ). 272, 551–554. Malpas, C.B., Genc, S., Saling, M.M., Velakoulis, D., Desmond, P.M., O’Brien, T.J., Brien, T.J.O., O’Brien, T.J., 2015. MRI correlates of general intelligence in neurotypical adults. J. clincal Neurosci. 8, 128–134. doi:10.1016/j.jocn.2015.07.012 Mantini, D., Perrucci, M.G., Del Gratta, C., Romani, G.L., Corbetta, M., 2007. Electrophysiological signatures of resting state networks in the human brain. Proc. Natl. Acad. Sci. U. S. A. 104, 13170–5. doi:10.1073/pnas.0700668104 Marco, E.J., Harrell, K.M., Brown, W.S., Hill, S.S., Jeremy, R.J., Kramer, J.H., Sherr, E.H., Paul, L.K., 2012. Processing speed delays contribute to executive function deficits in individuals with agenesis of the corpus callosum. J. Int. Neuropsychol. Soc. 18, 521–9. doi:10.1017/S1355617712000045 Marshalek, B., Lohman, D.F., Snow, R.E., 1983. The complexity continuum in the radex and hierarchical models of intelligence. Intelligence 7, 107–127. doi:10.1016/0160-2896(83)90023-5 Maybery, M.T., Bain, J.D., Halford, G.S., 1986. Information-processing demands of transitive inference. J. Exp. Psychol. Learn. Mem. Cogn. 12, 600–613. doi:10.1037//0278-7393.12.4.600 McDaniel, M.A., 2005. Big-brained people are smarter: A meta-analysis of the relationship between in vivo brain volume and intelligence. Intelligence 33, 337–346. doi:10.1016/j.intell.2004.11.005 McLaren, D.G., Ries, M.L., Xu, G., Johnson, S.C., 2012. A generalized form of context-dependent psychophysiological interactions (gPPI): a comparison to standard approaches. Neuroimage 61, 1277–86. doi:10.1016/j.neuroimage.2012.03.068 Meilă, M., 2007. Comparing clusterings—an information based distance. J. Multivar. Anal. 98, 873– 895. doi:10.1016/j.jmva.2006.11.013 Meisenberg, G., 2012. National IQ and economic outcomes. Pers. Individ. Dif. 53, 103–107. doi:10.1016/j.paid.2011.06.022 Melhem, E.R., Mori, S., Mukundan, G., Kraut, M.A., Pomper, M.G., van Zijl, P.C.M., 2002. Diffusion tensor MR imaging of the brain and white matter tractography. Am. J. Roentgenol. 178, 3–16. Melnick, M.D., Harrison, B.R., Park, S., Bennetto, L., Tadin, D., 2013. A strong interactive link between sensory discriminations and intelligence. Curr. Biol. 23, 1013–1017. doi:10.1016/j.cub.2013.04.053 Mill, R.D., Ito, T., Cole, M.W., 2017. From connectome to cognition: The search for mechanism in human functional brain networks. Neuroimage. doi:10.1016/j.neuroimage.2017.01.060 Mišić, B., Betzel, R.F., de Reus, M.A., van den Heuvel, M.P., Berman, M.G., McIntosh, A.R., Sporns,

128 O., 2016. Network-Level Structure-Function Relationships in Human Neocortex. Cereb. Cortex bhw089. doi:10.1093/cercor/bhw089 Mišić, B., Betzel, R.F., Nematzadeh, A., Goñi, J., Griffa, A., Hagmann, P., Flammini, A., Ahn, Y.-Y., Sporns, O., 2015. Cooperative and Competitive Spreading Dynamics on the Human Connectome. Neuron 86, 1518–1529. doi:10.1016/j.neuron.2015.05.035 Mitchell, D.J., Bell, A.H., Buckley, M.J., Mitchell, A.S., Sallet, J., Duncan, J., 2016. A Putative Multiple-Demand System in the Macaque Brain. J. Neurosci. 36, 8574–8585. doi:10.1523/JNEUROSCI.0810-16.2016 Moeller, S., Yacoub, E., Olman, C. a., Auerbach, E., Strupp, J., Harel, N., Uǧurbil, K., 2010. Multiband multislice GE-EPI at 7 tesla, with 16-fold acceleration using partial parallel imaging with application to high spatial and temporal whole-brain FMRI. Magn. Reson. Med. 63, 1144– 1153. doi:10.1002/mrm.22361 Neisser, U., Boodoo, G., Bouchard, T.J.. J., Boykin, a. W., Brody, N., Ceci, S.J., Halpern, D.F., Loehlin, J.C., Perloff, R., Sternberg, R.J., Urbina, S., 1996. Intelligence: Knowns and unknowns. Am. Psychol. 51, 77–101. doi:10.1037//0003-066X.51.2.77 Newman, M.E.J., Girvan, M., 2004. Finding and evaluating community structure in networks. Phys. Rev. E - Stat. Nonlinear, Soft Matter Phys. 69, 1–15. doi:10.1103/PhysRevE.69.026113 Newman, M.E.J., Strogatz, S.H., Watts, D.J., 2001. Random graphs with arbitrary degree distributions and their applications. Phys. Rev. E 64, 26118. doi:10.1103/PhysRevE.64.026118 Nisbett, R.E., Aronson, J.B., Dickens, C., Flynn, W., Halpern, D.F., Al., E., 2012. Intelligence: New findings and theoretical developments. Am. Psychol. 67, 130–159. doi:10.1037/a0026699 Nitsche, M. a, Nitsche, M. a, Paulus, W., Paulus, W., 2000. Excitability changes induced in the human motor cortex by weak transcranial direct current stimulation. J. Physiol. 527, 633–9. doi:PHY_1055 Ogawa, S., Lee, T.M., Kay, A.R., Tank, D.W., 1990. Brain magnetic resonance imaging with contrast dependent on blood oxygenation. Proc. Natl. Acad. Sci. U. S. A. 87, 9868–72. doi:10.1073/pnas.87.24.9868 Owen, J.P., Li, Y.-O., Yang, F.G., Shetty, C., Bukshpun, P., Vora, S., Wakahiro, M., Hinkley, L.B.N., Nagarajan, S.S., Sherr, E.H., Mukherjee, P., 2013. Resting-state networks and the functional connectome of the human brain in agenesis of the corpus callosum. Brain Connect. 3, 547–62. doi:10.1089/brain.2013.0175 Pamplona, G.S.P., Santos Neto, G.S.G.S., Rosset, S.R.E., Rogers, B.P., Salmon, C.E.G., 2015. Analyzing the association between functional connectivity of the brain and intellectual performance. Front. Hum. Neurosci. 9, 1–11. doi:10.3389/fnhum.2015.00061 Parkes, L., Fulcher, B.D., Yucel, M., Fornito, A., 2017. An evaluation of the efficacy, reliability, and sensitivity of motion correction strategies for resting-state functional MRI. bioRxiv.

129 Parkin, B.L., Hellyer, P.J., Leech, R., Hampshire, a., 2015. Dynamic Network Mechanisms of Relational Integration. J. Neurosci. 35, 7660–7673. doi:10.1523/JNEUROSCI.4956-14.2015 Pascual-Leone, A., Grafman, J., Cohen, L.G., Roth, B.J., Hallett, M., 1997. Transcranial magnetic stimulation: a new tool for the study of higher cognitive functions in humans. Handb. Neuropsychol. 11, 267–292. Passingham, R.E., Klaas, S.E., Kötter, R., 2002. The anatomical basis of functional localization in the cortex. Nat. Rev. Neurosci. 3, 606–616. doi:10.1038/nrn893 Paul, L.K., Brown, W.S., Adolphs, R., Tyszka, J.M., Richards, L.J., Mukherjee, P., Sherr, E.H., 2007. Agenesis of the corpus callosum: genetic, developmental and functional aspects of connectivity. Nat. Rev. Neurosci. 8, 287–99. doi:10.1038/nrn2107 Paul, L.K., Schieffer, B., Brown, W.S., 2004. Social processing deficits in agenesis of the corpus callosum: narratives from the Thematic Appreciation Test. Arch. Clin. Neuropsychol. 19, 215– 25. doi:10.1016/S0887-6177(03)00024-6 Pelli, D.G., 1997. The VideoToolbox software for visual psychophysics: transforming numbers into movies. Spat. Vis. doi:10.1163/156856897X00366 Penn, D.C., Holyoak, K.J., Povinelli, D.J., 2008. Darwin’s mistake: explaining the discontinuity between human and nonhuman minds. Behav. Brain Sci. 31, 109-30-178. doi:10.1017/S0140525X08003543 Penny, D.W., Friston, J.K., Ashburner, T.J., Kiebel, J.S., Nichols, E.T., 2011. Statistical parametric mapping: the analysis of functional brain images: the analysis of functional brain images, Functional neuroimaging: Technical. Academic press. Perfetti, B., Saggino, A., Ferretti, A., Caulo, M., Romani, G.L., Onofrj, M., 2009. Differential patterns of cortical activation as a function of fluid reasoning complexity. Hum. Brain Mapp. 30, 497– 510. doi:10.1002/hbm.20519 Perret, P., Bailleux, C., Dauvier, B., 2011. The influence of relational complexity and strategy selection on children’s reasoning in the Latin Square Task. Cogn. Dev. 26, 127–141. doi:10.1016/j.cogdev.2010.12.003 Petrill, S.A., Hewitt, J.K., Plomin, R., Cherny, S.S., 2004. Genetic and environmental contributions to general cognitive ability through the first 16 years of life. Dev. Psychol. 40, 805–812. doi:10.1037/0012-1649.40.5.805.Genetic Plomin, R., Deary, I., 2014. Genetics and intelligence differences: five special findings. Mol. Psychiatry 20, 98–108. doi:10.1038/mp.2014.105 Polanía, R., Paulus, W., Nitsche, M.A., 2012. Reorganizing the intrinsic functional architecture of the human primary motor cortex during rest with non-invasive cortical stimulation. PLoS One 7. doi:10.1371/journal.pone.0030971 Poldrack, R.A., Laumann, T.O., Koyejo, O., Gregory, B., Hover, A., Chen, M.-Y., Gorgolewski, K.J.,

130 Luci, J., Joo, S.J., Boyd, R.L., Hunicke-Smith, S., Simpson, Z.B., Caven, T., Sochat, V., Shine, J.M., Gordon, E., Snyder, A.Z., Adeyemo, B., Petersen, S.E., Glahn, D.C., Reese Mckay, D., Curran, J.E., Göring, H.H.H., Carless, M.A., Blangero, J., Dougherty, R., Leemans, A., Handwerker, D.A., Frick, L., Marcotte, E.M., Mumford, J.A., 2015. Long-term neural and physiological phenotyping of a single human. Nat. Commun. 6, 8885. doi:10.1038/ncomms9885 Poldrack, R. a., Farah, M.J., 2015. Progress and challenges in probing the human brain. Nature 526, 371–379. doi:10.1038/nature15692 Power, J.D., Cohen, A.L., Nelson, S.M., Wig, G.S., Barnes, K.A., Church, J. a., Vogel, A.C., Laumann, T.O., Miezin, F.M., Schlaggar, B.L., Petersen, S.E., 2011. Functional Network Organization of the Human Brain. Neuron 72, 665–678. doi:10.1016/j.neuron.2011.09.006 Power, J.D., Mitra, A., Laumann, T.O., Snyder, A.Z., Schlaggar, B.L., Petersen, S.E., 2014. Methods to detect, characterize, and remove motion artifact in resting state fMRI. Neuroimage 84, 320– 341. doi:http://dx.doi.org/10.1016/j.neuroimage.2013.08.048 Power, J.D., Schlaggar, B.L., Petersen, S.E., 2014. Studying Brain Organization via Spontaneous fMRI Signal. Neuron 84, 681–696. doi:10.1016/j.neuron.2014.09.007 Prabhakaran, V., Smith, J. a, Desmond, J.E., Glover, G.H., Gabrieli, J.D., 1997. Neural substrates of fluid reasoning: an fMRI study of neocortical activation during performance of the Raven’s Progressive Matrices Test. Cogn. Psychol. 33, 43–63. doi:10.1006/cogp.1997.0659 Procyk, E., Tanaka, Y.L., Joseph, J.P., 2000. Anterior cingulate activity during routine and non- routine sequential behaviors in macaques. Nat. Neurosci. 3, 502–508. doi:10.1038/74880 Putnam, M.C., Wig, G.S., Grafton, S.T., Kelley, W.M., Gazzaniga, M.S., 2008. Structural organization of the corpus callosum predicts the extent and impact of cortical activity in the nondominant hemisphere. J. Neurosci. 28, 2912–8. doi:10.1523/JNEUROSCI.2295-07.2008 Rao, S.C., Rainer, G., Miller, E.K., 1997. Integration of what and where in the primate prefrontal cortex. Science (80-. ). 276, 821–824. Raven, J., 2000. The Raven’s progressive matrices: change and stability over culture and time. Cogn. Psychol. 41, 1–48. doi:10.1006/cogp.1999.0735 Razi, A., Seghier, M.L., Zhou, Y., McColgan, P., Zeidman, P., Park, H.-J., Sporns, O., Rees, G., Friston, K.J., 2017. Large-scale DCMs for resting state fMRI. Netw. Neurosci. 1–41. doi:10.1162/NETN_a_00015 Rorden, C., Karnath, H.-O., 2004. Using human brain lesions to infer function: a relic from a past era in the fMRI age? Nat. Rev. Neurosci. 5, 813–819. doi:10.1038/nrn1521 Rosvall, M., Axelsson, D., Bergstrom, C.T., 2009. The map equation. Eur. Phys. J. Spec. Top. 178, 13–23. doi:10.1140/epjst/e2010-01179-1 Roy, C., Sherrington, C., 1890. On the Regulation of the Blood-Supply of the Brain. J. Physiol. 11, 85–158. doi:10.1186/1471-244X-11-115

131 Rubinov, M., Sporns, O., 2011. Weight-conserving characterization of complex functional brain networks. Neuroimage 56, 2068–2079. doi:10.1016/j.neuroimage.2011.03.069 Rubinov, M., Sporns, O., 2010. Complex network measures of brain connectivity: uses and interpretations. Neuroimage 52, 1059–69. doi:10.1016/j.neuroimage.2009.10.003 Saad, Z.S., Gotts, S.J., Murphy, K., Chen, G., Jo, H.J., Martin, A., Cox, R.W., 2012. Trouble at Rest: How Correlation Patterns and Group Differences Become Distorted After Global Signal Regression. Brain Connect. 2, 25–32. doi:10.1089/brain.2012.0080 Santarnecchi, E., Tatti, E., Rossi, S., Serino, V., Rossi, A., 2015. Intelligence-related differences in the asymmetry of spontaneous cerebral activity. Hum. Brain Mapp. 36, 3586–3602. doi:10.1002/hbm.22864 Sauerwein, H.C., Lassonde, M.C., Cardu, B., Geoffroy, G., 1981. Interhemispheric integration of sensory and motor functions in agenesis of the corpus callosum. Neuropsychologia 19, 445–454. doi:10.1016/0028-3932(81)90074-9 Schacter, D.L., Addis, D.R., Buckner, R.L., 2007. Remembering the past to imagine the future: the prospective brain. Nat. Rev. Neurosci. 8, 657. Schmitt, L.I., Wimmer, R.D., Nakajima, M., Happ, M., Mofakham, S., Halassa, M.M., 2017. Thalamic amplification of cortical connectivity sustains attentional control. Nature 1–24. doi:10.1038/nature22073 Scholvinck, M.L., Leopold, D.A., Brookes, M.J., Khader, P.., 2013. The contribution of electrophysiology to functional connectivity mapping. Neuroimage 297–306. doi:10.1016/j.neuroimage.2013.04.010.The Schulte, T., Müller-Oehring, E.M., 2010. Contribution of callosal connections to the interhemispheric integration of visuomotor and cognitive processes. Neuropsychol. Rev. 20, 174–190. Schultz, D.H., Cole, M.W., 2016. Higher Intelligence Is Associated with Less Task-Related Brain Network Reconfiguration. J. Neurosci. 36, 8551–8561. doi:10.1523/JNEUROSCI.0358-16.2016 Seeley, W.W., Menon, V., Schatzberg, A.F., Keller, J., Glover, G.H., Kenna, H., Reiss, A.L., Greicius, M.D., 2007. Dissociable intrinsic connectivity networks for salience processing and executive control. J. Neurosci. 27, 2349–56. doi:10.1523/JNEUROSCI.5587-06.2007 Seymour, S.E., Reuter-lorenz, P.A., Gazzaniga, M.S., 1994. The disconnection syndrome: Basic findings reaffirmed. Brain 117, 105–115. doi:10.1093/brain/117.1.105 Shaw, P., Greenstein, D., Lerch, J., Clasen, L., Lenroot, R., Gogtay, N., Evans, A., Rapoport, J., Giedd, J., 2006. Intellectual ability and cortical development in children and adolescents. Nature 440, 676–679. doi:10.1038/nature04513 Shen, K., Mišić, B., Cipollini, B.N., Bezgin, G., Buschkuehl, M., Hutchison, R.M., Jaeggi, S.M., Kross, E., Peltier, S.J., Everling, S., Jonides, J., McIntosh, A.R., Berman, M.G., 2015. Stable long-range interhemispheric coordination is supported by direct anatomical projections. Proc.

132 Natl. Acad. Sci. 201503436. doi:10.1073/pnas.1503436112 Sherman, S.M., 2016. Thalamus plays a central role in ongoing cortical functioning. Nat. Neurosci. 16, 533–41. doi:10.1038/nn.4269 Shine, J.M., Bissett, P.G., Bell, P.T., Koyejo, O., Balsters, J.H., Gorgolewski, K.J., Moodie, C.A., Poldrack, R.A., 2016. The Dynamics of Functional Brain Networks: Integrated Network States during Cognitive Task Performance. Neuron 92, 544–554. doi:10.1016/j.neuron.2016.09.018 Shine, J.M., Poldrack, R.A., 2017. Principles of Dynamic Network Reconfiguration across Diverse Brain States. Neuroimage. doi:10.20944/preprints201707.0014.v1 Shmuel, A., Leopold, D.A., 2008. Neuronal correlates of spontaneous fluctuations in fMRI signals in monkey visual cortex: Implications for functional connectivity at rest. Hum. Brain Mapp. 29, 751–761. doi:10.1002/hbm.20580 Shulman, G.L., Fiez, J.A., Corbetta, M., Buckner, R.L., Miezin, F.M., Raichle, M.E., Petersen, S.E., 1997. Common Blood Flow Changes across Visual Tasks: II. Decreases in Cerebral Cortex. J. Cogn. Neurosci. 9, 648–663. doi:10.1162/jocn.1997.9.5.648 Siffredi, V., Anderson, V., Leventer, R.J., Spencer-Smith, M.M., 2013. Neuropsychological profile of agenesis of the corpus callosum: a systematic review. Dev. Neuropsychol. 38, 36–57. doi:10.1080/87565641.2012.721421 Silvanto, J., Pascual-Leone, A., 2012. Why the assessment of causality in brain–behavior relations requires brain stimulation. J. Cogn. Neurosci. 24, 775–777. Smallwood, J., Schooler, J.W., 2015. The Science of Mind Wandering: Empirically Navigating the Stream of Consciousness. Annu. Rev. Psychol. 66, 487–518. doi:10.1146/annurev-psych- 010814-015331 Smith, R.E., Tournier, J.D., Calamante, F., Connelly, A., 2013. SIFT: Spherical-deconvolution informed filtering of tractograms. Neuroimage 67, 298–312. doi:10.1016/j.neuroimage.2012.11.049 Smith, S.M., Jenkinson, M., Woolrich, M.W., Beckmann, C.F., Behrens, T.E.J., Johansen-Berg, H., Bannister, P.R., De Luca, M., Drobnjak, I., Flitney, D.E., Niazy, R.K., Saunders, J., Vickers, J., Zhang, Y., De Stefano, N., Brady, J.M., Matthews, P.M., 2004. Advances in functional and structural MR image analysis and implementation as FSL. Neuroimage 23, 208–219. doi:10.1016/j.neuroimage.2004.07.051 Smith, S.M., Miller, K.L., Salimi-Khorshidi, G., Webster, M., Beckmann, C.F., Nichols, T.E., Ramsey, J.D., Woolrich, M.W., 2011. Network modelling methods for FMRI. Neuroimage 54, 875–91. doi:10.1016/j.neuroimage.2010.08.063 Smith, S.M., Nichols, T.E., Vidaurre, D., Winkler, A.M., Behrens, T.E.J., Glasser, M.F., Ugurbil, K., Barch, D.M., Van Essen, D.C., Miller, K.L., 2015. A positive-negative mode of population covariation links brain connectivity, demographics and behavior. Nat. Neurosci. 1–7.

133 doi:10.1038/nn.4125 Smith, S.M., Vidaurre, D., Beckmann, C.F., Glasser, M.F., Jenkinson, M., Miller, K.L., Nichols, T.E., Robinson, E.C., Salimi-Khorshidi, G., Woolrich, M.W., Barch, D.M., Uǧurbil, K., Van Essen, D.C., Uğurbil, K., Van Essen, D.C., 2013. Functional connectomics from resting-state fMRI. Trends Cogn. Sci. 17, 666–82. doi:10.1016/j.tics.2013.09.016 Sniekers, S., Stringer, S., Watanabe, K., Jansen, P.R., Coleman, J.R.I., Krapohl, E., Taskesen, E., Hammerschlag, A.R., Okbay, A., Zabaneh, D., Najaf Amin, Breen, G., Cesarini, D., Chabris, C.F., Iacono, W.G., Ikram, M.A., Johannesson, M., Koellinger, P.D., Lee, J.J., Magnusson, P.K.E., McGue, M., Miller, M.B., Ollier, W.E.R., Payton, A., Pendleton, N., Plomin, R., Rietveld, C.A., Tiemeier, H., van Duijn, C.M., Posthuma, D., 2017. Genome-wide association meta-analysis of 78,308 individuals identifies new loci and genes influencing human intelligence. Nat. Genet. doi:doi:10.1038/ng.3869 Snow, Kyllonen, Marshalek, 1984. The topography of ability and learning correlations. Adv. Psychol. Hum. Intell. 2, 47–103. Song, M., Liu, Y., Zhou, Y., Wang, K., Yu, C., Jiang, T., 2009. Default network and intelligence difference., in: IEEE TRANSACTIONS ON AUTONOMOUS MENTAL DEVELOPMENT. pp. 101–109. doi:10.1109/IEMBS.2009.5334874 Song, M., Zhou, Y., Li, J., Liu, Y., Tian, L., Yu, C., Jiang, T., 2008. Brain spontaneous functional connectivity and intelligence. Neuroimage 41, 1168–76. doi:10.1016/j.neuroimage.2008.02.036 Sonuga-Barke, E.J.S., Castellanos, F.X., 2007. Spontaneous attentional fluctuations in impaired states and pathological conditions: a neurobiological hypothesis. Neurosci. Biobehav. Rev. 31, 977– 86. doi:10.1016/j.neubiorev.2007.02.005 Spearman, C., 1904. “General Intelligence,” Objectively Determined and Measured. Am. J. Psychol. 15, 201–292. Sporns, O., 2016. Connectome Networks: From Cells to Systems, Micro-, Meso- and Macro- Connectomics of the Brain. doi:10.1007/978-3-319-27777-6 Sporns, O., 2013. Making sense of brain network data. Nat. Methods 10, 491–493. Sporns, O., Betzel, R.F., 2016. Modular Brain Networks. Annu. Rev. Psychol. 67, 613–640. doi:10.1146/annurev-psych-122414-033634 Sporns, O., Zwi, J.D., 2004. The small world of the cerebral cortex. Neuroinformatics 2, 145–62. doi:10.1385/NI:2:2:145 Stagg, C.J., Bachtiar, V., Amadi, U., Gudberg, C.A., Ilie, A.S., Sampaio-Baptista, C., O’Shea, J., Woolrich, M., Smith, S.M., Filippini, N., Near, J., Johansen-Berg, H., 2014. Local GABA concentration is related to network-level resting functional connectivity. Elife 2014, 1–9. doi:10.7554/eLife.01465 Stanovich, K.E., West, R.F., 1998. Individual differences in rational thought. J. Exp. Psychol. Gen.

134 127, 161–188. doi:10.1037/0096-3445.127.2.161 Sternberg, R.J., 2012. Intelligence. Dialogues Clin. Neurosci. 14, 19–27. doi:10.1002/0470018860.s00555 Stokes, M.G., Kusunoki, M., Sigala, N., Nili, H., Gaffan, D., Duncan, J., 2013. Dynamic coding for cognitive control in prefrontal cortex. Neuron 78, 364–375. doi:10.1016/j.neuron.2013.01.039 Strenze, T., 2007. Intelligence and socioeconomic success: A meta-analytic review of longitudinal research. Intelligence 35, 401–426. doi:10.1016/j.intell.2006.09.004 Sun, F.T., Miller, L.M., D’Esposito, M., 2004. Measuring interregional functional connectivity using coherence and partial coherence analyses of fMRI data. Neuroimage 21, 647–658. doi:http://dx.doi.org/10.1016/j.neuroimage.2003.09.056 Svoboda, E., McKinnon, M.C., Levine, B., 2006. The functional neuroanatomy of : a meta-analysis. Neuropsychologia 44, 2189–2208. Tang, C.Y., Eaves, E.L., Ng, J.C., Carpenter, D.M., Mai, X., Schroeder, D.H., Condon, C.A., Colom, R., Haier, R.J., 2010. Brain networks for working memory and factors of intelligence assessed in males and females with fMRI and DTI. Intelligence 38, 293–303. doi:10.1016/j.intell.2010.03.003 Tavor, I., Jones, O.P., Mars, R.B., Smith, S.M., Behrens, T.E., Jbabdi, S., 2016. Task-free MRI predicts individual differences in brain activity during task performance. Science (80-. ). 352, 216–220. doi:10.1126/science.aad8127 Thiebaut De Schotten, M., Tomaiuolo, F., Aiello, M., Merola, S., Silvetti, M., Lecce, F., Bartolomeo, P., Doricchi, F., 2014. Damage to white matter pathways in subacute and chronic spatial neglect: A group study and 2 single-case studies with complete virtual “in vivo” tractography dissection. Cereb. Cortex 24, 691–706. doi:10.1093/cercor/bhs351 Thomsen, K., Offenhauser, N., Lauritzen, M., 2004. Principal neuron spiking: neither necessary nor sufficient for cerebral blood flow in rat cerebellum. J. Physiol. 560, 181–9. doi:10.1113/jphysiol.2004.068072 Tononi, G., Sporns, O., Edelman, G.M., 1994. A measure for brain complexity: relating functional segregation and integration in the nervous system. Proc. Natl. Acad. Sci. 91, 5033–5037. doi:10.1073/pnas.91.11.5033 Tovar-Moll, F., Moll, J., de Oliveira-Souza, R., Bramati, I., Andreiuolo, P. a, Lent, R., 2007. Neuroplasticity in human callosal dysgenesis: a diffusion tensor imaging study. Cereb. Cortex 17, 531–41. doi:10.1093/cercor/bhj178 Tovar-Moll, F., Monteiro, M., Andrade, J., Bramati, I.E., Vianna-Barbosa, R., Marins, T., Rodrigues, E., Dantas, N., Behrens, T.E.J., de Oliveira-Souza, R., Moll, J., Lent, R., 2014. Structural and functional brain rewiring clarifies preserved interhemispheric transfer in humans born without the corpus callosum. Proc. Natl. Acad. Sci. U. S. A. doi:10.1073/pnas.1400806111

135 Tschentscher, N., Mitchell, D., Duncan, J., 2017. Fluid Intelligence Predicts Novel Rule Implementation in a Distributed Frontoparietal Control Network. J. Neurosci. 37, 4841–4847. doi:10.1523/JNEUROSCI.2478-16.2017 Tucker-Drob, E.M., Salthouse, T. a., 2009. Confirmatory Factor Analysis and Multidimensional Scaling for Construct Validation of Cognitive Abilities. Int. J. Behav. Dev. 33, 277–285. doi:10.1177/0165025409104489 Turkheimer, E., Haley, A., Waldron, M., D’Onofrio, B., Gottesman, I.I., 2003. Socioeconomic status modified heritability of IQ in young children. Psychol. Sci. 14, 623–628. doi:10.1046/j.0956- 7976.2003.psci_1475.x Turner, R., 2002. How much cortex can a vein drain? Downstream dilution of activation-related cerebral blood oxygenation changes. Neuroimage 16, 1062–1067. doi:10.1006/nimg.2002.1082 Tyszka, J.M., Kennedy, D.P., Adolphs, R., Paul, L.K., 2011. Intact bilateral resting-state networks in the absence of the corpus callosum. J. Neurosci. 31, 15154–62. doi:10.1523/JNEUROSCI.1453- 11.2011 Ugurbil, K., Xu, J., Auerbach, E.J., Moeller, S., Vu, A.T., Duarte-Carvajalino, J.M., Lenglet, C., Wu, X., Schmitter, S., Van de Moortele, P.F., Strupp, J., Sapiro, G., De Martino, F., Wang, D., Harel, N., Garwood, M., Chen, L., Feinberg, D. a., Smith, S.M., Miller, K.L., Sotiropoulos, S.N., Jbabdi, S., Andersson, J.L.R., Behrens, T.E.J., Glasser, M.F., Van Essen, D.C., Yacoub, E., 2013. Pushing spatial and temporal resolution for functional and diffusion MRI in the Human Connectome Project. Neuroimage 80, 80–104. doi:10.1016/j.neuroimage.2013.05.012 van den Heuvel, M.P., Stam, C.J., Kahn, R.S., Hulshoff Pol, H.E., 2009. Efficiency of functional brain networks and intellectual performance. J. Neurosci. 29, 7619–24. doi:10.1523/JNEUROSCI.1443-09.2009 van der Knaap, L.J., van der Ham, I.J.M., 2011. How does the corpus callosum mediate interhemispheric transfer? A review. Behav. Brain Res. 223, 211–221. Van Dijk, K.R.A., Sabuncu, M.R., Buckner, R.L., 2012. The influence of head motion on intrinsic functional connectivity MRI. Neuroimage 59, 431–438. doi:http://dx.doi.org/10.1016/j.neuroimage.2011.07.044 Van Essen, D.C., Smith, S.M., Barch, D.M., Behrens, T.E.J., Yacoub, E., Ugurbil, K., 2013. The WU- Minn Human Connectome Project: An overview. Neuroimage 80, 62–79. doi:10.1016/j.neuroimage.2013.05.041 Vatansever, D., Manktelow, A.E., Sahakian, B.J., Menon, D.K., Stamatakis, E.A., 2017. Angular default mode network connectivity across working memory load. Hum. Brain Mapp. 38, 41–52. doi:10.1002/hbm.23341 Vatansever, D., Menon, X.D.K., Manktelow, A.E., Sahakian, B.J., Stamatakis, E.A., 2015. Default Mode Dynamics for Global Functional Integration. J. Neurosci. 35, 15254–15262.

136 doi:10.1523/JNEUROSCI.2135-15.2015 Waltz, J.A., Knowlton, B.J., Holyoak, K.J., Boone, K.B., Mishkin, F.S., Santos, M. de M., Thomas, C.R., Miller, B.L., 1999. A system for relational reasoning in human prefrontal cortex. Psychol. Sci. 119–125. doi:10.1111/1467-9280.00118 Wang, L., Song, M., Jiang, T., Zhang, Y., Yu, C., 2011. Regional homogeneity of the resting-state brain activity correlates with individual intelligence. Neurosci. Lett. 488, 275–8. doi:10.1016/j.neulet.2010.11.046 Warren, D.E., Power, J.D., Bruss, J., Denburg, N.L., Waldron, E.J., Sun, H., Petersen, S.E., Tranel, D., 2014. Network measures predict neuropsychological outcome after brain injury. Proc. Natl. Acad. Sci. 111, 14247–14252. doi:10.1073/pnas.1322173111 Wason, P.C., 1968. Reasoning about a rule. Q. J. Exp. Psychol. 20, 273–281. Wason, P.C., Shapiro, D., 1971. Natural and contrived experience in a reasoning problem. Q. J. Exp. Psychol. 23, 63–71. doi:10.1080/00335557143000068 Wendelken, C., Nakhabenko, D., Donohue, S.E., Carter, C.S., Bunge, S.A., 2008. “Brain Is to Thought as Stomach Is to ??”: Investigating the Role of Rostrolateral Prefrontal Cortex in Relational Reasoning. J. Cogn. Neurosci. 20, 682–693. doi:10.1162/jocn.2008.20055 Weschler, D., 2008. Wechsler Adult Intelligence Scale - Fourth Edition 1–3. Westphal, A.J., Wang, S., Rissman, J., 2017. Episodic memory retrieval benefits from a less modular brain network organization. J. Neurosci. 2509–16. doi:10.1523/JNEUROSCI.2509-16.2017 Woolgar, A., Afshar, Williams, Rich, A.N., 2015. Flexible Coding of Task Rules in Frontoparietal Cortex: An Adaptive System for Flexible Cognitive Control. J. Cogn. Neurosci. 1–10. doi:10.1162/jocn Xia, M., Wang, J., He, Y., 2013. BrainNet Viewer: A Network Visualization Tool for Human Brain Connectomics. PLoS One 8. doi:10.1371/journal.pone.0068910 Yeo, B.T., Krienen, F.M., Sepulcre, J., Sabuncu, M.R., Lashkari, D., Hollinshead, M., Roffman, J.L., Smoller, J.W., Zollei, L., Polimeni, J.R., Fischl, B., Liu, H., Buckner, R.L., 2011. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J. Neurophysiol. 106, 1125–1165. doi:10.1152/jn.00338.2011. Yerkes, R.M., Yoakum, C.S., 1920. Army Mental Tests. Yuan, Z., Qin, W., Wang, D., Jiang, T., Zhang, Y., Yu, C., 2012. The salience network contributes to an individual’s fluid reasoning capacity. Behav. Brain Res. 229, 384–390. doi:10.1016/j.bbr.2012.01.037 Yue, Q., Martin, R.C., Fischer-Baum, S., Ramos-Nuñez, A.I., Ye, F., Deem, M.W., 2017. Brain Modularity Mediates the Relation between Task Complexity and Performance. J. Cogn. Neurosci. 1–15. doi:10.1162/jocn_a_01142 Zalesky, A., Cocchi, L., Fornito, A., Murray, M.M., Bullmore, E., 2012a. Connectivity differences in

137 brain networks. Neuroimage 60, 1055–62. doi:10.1016/j.neuroimage.2012.01.068 Zalesky, A., Fornito, A., Bullmore, E., 2012b. On the use of correlation as a measure of network connectivity. Neuroimage 60, 2096–106. doi:10.1016/j.neuroimage.2012.02.001 Zalesky, A., Fornito, A., Bullmore, E.T., 2010. Network-based statistic: identifying differences in brain networks. Neuroimage 53, 1197–207. doi:10.1016/j.neuroimage.2010.06.041 Zalesky, A., Fornito, A., Cocchi, L., Gollo, L.L., Breakspear, M., 2014. Time-resolved resting-state brain networks. Proc. Natl. Acad. Sci. U. S. A. 111, 10341–6. doi:10.1073/pnas.1400181111 Zalesky, A., Fornito, A., Seal, M.L., Cocchi, L., Westin, C.-F., Bullmore, E.T., Egan, G.F., Pantelis, C., 2011. Disrupted axonal fiber connectivity in schizophrenia. Biol. Psychiatry 69, 80–9. doi:10.1016/j.biopsych.2010.08.022 Zeuch, N., Holling, H., Kuhn, J.-T., 2011. Analysis of the Latin Square Task with linear logistic test models. Learn. Individ. Differ. 21, 629–632. doi:10.1016/j.lindif.2011.03.004 Zhang, L., Xin, Z., Lin, C., Li, H., 2009. The complexity of the Latin Square task and its influence on children’s performance. Chinese Sci. Bull. 54, 766–775. doi:10.1007/s11434-009-0079-5 Zuo, X.-N., Kelly, C., Adelstein, J.S., Klein, D.F., Castellanos, F.X., Milham, M.P., 2010. Reliable intrinsic connectivity networks: test–retest evaluation using ICA and dual regression approach. Neuroimage 49, 2163–2177.

138

Appendices

139

Appendix A Table A-1 Characteristics of studies utilizing complex graph metrics included in the supplementary resting-state meta-analysis. Sample Author N Mal Age Behaviour Brain- Analysis type Regions of es (M±S al measure behaviour interest D) relationship Van den 19 74 29±7. WAIS Correlation Global efficiency Whole brain Heuval et al. % 8 Wang et al. 59 49 24.6± WAIS Correlation Regional Whole brain % 3.5 (Chinese) Homogeneity Cole et al. 94 42 22±4. RAPM Correlation Global brain LPFC % 7 and CCFT connectivity Yuan et al. 28 46 22.8± RPM Correlation Regional Whole brain 4 % 2.4 Homogeneity Santarnecchi 98 50 34±1 WASI Between- Global efficiency 90 AAL atlas et al. % 4 group: regionsa high, average and low defined by clustering analysis Pamplona et 29 52 26.8± WAIS Correlation Local efficiency 82 AAL atlas al. % 5.8 (Portugues regionsa e-Brazil) Note: aMNI centroids were used as regions of interest in the case of the AAL template. CCFT = Cattell Culture Fair Test, RPM = Ravens Standard Progressive Matrices, RAPM = Ravens Advanced Progressive Matrices, WAIS = Wechsler adult intelligence scale, WASI = Wechsler abbreviated scale of intelligence, LPFC = lateral prefrontal cortex, AAL = Automated Anatomical Labeling.

Figure A-1 Regions of the brain that demonstrated global or local graph properties at rest that were associated with intelligence. a. Efficiency metrics b. Regional homogeneity.

Appendix B Table B-1 Regions of interest used in the follow-up analysis of complexity-induced default mode network dynamics.

Default-mode regions b Anatomy a

x y z Ventral medial prefrontal cortex 6 64 3 Medial prefrontal cortex 0 51 32 Left Anterior prefrontal cortex -25 51 27 Ventral medial prefrontal cortex 9 51 16 Ventral medial prefrontal cortex -11 45 17 Ventral medial prefrontal cortex 8 42 -5 Anterior cingulate cortex 9 39 20 Ventral lateral prefrontal cortex 46 39 -15 Superior frontal cortex 23 33 47 Superior frontal cortex -16 29 54 Inferior temporal cortex 52 -15 -13 Inferior temporal cortex -59 -25 -15 Posterior cingulate cortex 1 -26 31 Fusiform 28 -37 -15 Posterior cingulate cortex -8 -41 3 Inferior temporal cortex -61 -41 -2 Occipital cortex -28 -42 -11 Posterior cingulate cortex -5 -43 25 Precuneus 9 -43 25 Precuneus 5 -50 33 Posterior cingulate cortex -5 -52 17 Posterior cingulate cortex 10 -55 17 Precuneus -6 -56 29 Posterior cingulate cortex -11 -58 17 Angular gyrus 51 -59 34 Angular gyrus -48 -63 35 Precuneus 11 -68 42 Intra-parietal sulcus -36 -69 40 Occipital cortex -9 -72 41 Occipital cortex 45 -72 29 Occipital cortex -2 -75 32 Occipital cortex -42 -76 26

Note. a Coordinates (x, y, z) are given in Montreal Neurological Institute (MNI) atlas space. b Coordinates and labels are from Dosenbach and colleagues (2010).

Figure B-1 Significant default-mode connections that did not show a consistent increase or decrease across task complexity.

Appendix C Table C-1 Variation of Information statistics. Network density 5% 10% 15% 20% 25% 30% Contrast VIn p VIn p VIn p VIn p VIn p VIn p Pre-task Post-task rest 0.056 0.878 0.065 0.863 0.109 0.656 0.098 0.836 0.124 0.552 0.092 0.944 rest Null 0.146 0.013 0.139 0.077 0.124 0.397 0.189 0.009* 0.198 0.005* 0.186 0.029 Binary 0.148 0.006* 0.179 0.001* 0.218 0.001* 0.219 < 0.001* 0.216 0.007* 0.209 0.023 Ternary 0.166 0.002* 0.194 < 0.001* 0.226 < 0.001* 0.226 0.001* 0.221 0.002* 0.215 0.002* Quaternary 0.148 0.009* 0.189 < 0.001* 0.225 < 0.001* 0.22 < 0.001* 0.221 < 0.001* 0.203 0.008* Post-task Null 0.117 0.231 0.154 0.019 0.098 0.81 0.153 0.084 0.185 0.015 0.191 0.018 rest Binary 0.128 0.099 0.185 0.001* 0.176 0.033 0.181 0.074 0.212 0.01* 0.203 0.018 Ternary 0.16 0.006* 0.204 < 0.001* 0.194 0.006* 0.203 0.003* 0.2 0.016 0.22 0.001* Quaternary 0.128 0.083 0.189 0.002* 0.195 0.005* 0.202 0.001* 0.2 0.005* 0.211 0.001* Null Binary 0.028 1 0.121 0.245 0.162 0.199 0.1 0.698 0.1 0.674 0.106 0.534 Ternary 0.094 0.711 0.133 0.237 0.193 0.023 0.13 0.645 0.119 0.567 0.132 0.246 Quaternary 0.052 0.987 0.117 0.272 0.181 0.04 0.14 0.539 0.132 0.467 0.132 0.291 Binary Ternary 0.096 0.681 0.154 0.159 0.097 0.801 0.1 0.485 0.093 0.571 0.11 0.306 Quaternary 0.038 0.998 0.121 0.375 0.086 0.91 0.111 0.507 0.106 0.484 0.108 0.387 Ternary Quaternary 0.082 0.747 0.084 0.894 0.041 1 0.051 0.986 0.061 0.943 0.051 0.969 Note: * indicates contrasts where p < 0.01, VIn = Variation of Information

Table C-2 Significant pairwise changes in functional connectivity associated with increasing relational complexity. MNI RSN Mod Anatomy Degree x y z -7 -52 61 SShand FPV Precuneus 12 4 -48 51 Mem FPV Precuneus 10 47 10 33 Fpn FPV Precentral gyrus 7 -2 -35 31 Mem Dmn Middle cingulate cortex 6 -34 3 4 Co Sens Middle insula 6 23 10 1 Sc FPV Putamen 6 49 8 -1 Co Sens Superior temporal pole 5 -11 -56 16 Dmn Dmn Calcarine gyrus 5 -2 38 36 Dmn Dmn Medial superior frontal gyrus 5 -42 -55 45 Fpn FPV Inferior parietal cortex 5 -45 0 9 Co Sens Rolandic operculum 4 52 -59 36 Dmn Dmn Angular gyrus 4 -35 20 51 Dmn Dmn Middle frontal gyrus 4 9 -4 6 Sc FPV Thalamus 4 -2 -13 12 Sc FPV Thalamus 4 -47 11 23 Fpn FPV Inferior frontal gyrus (operculum) 4 -42 45 -2 Fpn FPV Inferior frontal gyrus (orbital) 4 54 -28 34 Co Sens Supramarginal gyrus 3 8 -48 31 Dmn Dmn Posterior cingulate cortex 3 13 30 59 Dmn Dmn Superior frontal gyrus 3 11 -39 50 Sal Sens Middle cingulate cortex 3 15 -77 31 Vis Sens Cuneus 3 -41 6 33 Fpn FPV Precentral gyrus 3 -53 -49 43 Fpn Dmn Inferior parietal cortex 3 -44 2 46 Fpn FPV Precentral gyrus 3 3 -17 58 SShand Sens Supplementary motor area 2 65 -33 20 Aud Sens Superior temporal gyrus 2 -51 8 -2 Co Sens Insula 2 15 -63 26 Dmn FPV Cuneus 2 11 -54 17 Dmn Dmn Precuneus 2 36 10 1 Co Sens Insula 2 -10 -18 7 Sc FPV Thalamus 2 -22 7 -5 Sc FPV Putamen 2 43 -78 -12 Vis FPV Inferior occipital cortex 2 -42 -60 -9 DorAtt FPV Inferior temporal gyrus 2 -3 26 44 Fpn FPV Superior frontal gyrus (medial) 2 44 -53 47 Fpn FPV Inferior parietal cortex 2 32 14 56 Fpn FPV Middle frontal gyrus 2 -40 -19 54 SShand Sens Precentral gyrus 1 0 -15 47 SShand Sens Middle cingulate cortex 1 -10 -2 42 Co Sens Middle cingulate cortex 1 -50 -34 26 Aud Sens Supramarginal gyrus 1 59 -17 29 Aud Sens Postcentral gyrus 1 37 1 -4 Co Sens Insula 1 6 -59 35 Dmn Dmn Precuneus 1 -41 -75 26 Dmn FPV Middle occipital cortex 1 65 -31 -9 Dmn Dmn Middle temporal gyrus 1 43 -72 28 Dmn FPV Middle occipital cortex 1 -3 42 16 Dmn Dmn Anterior cingulate cortex 1 -16 29 53 Dmn Dmn Superior frontal gyrus 1 2 -24 30 Mem Dmn Middle cingulate cortex 1 12 36 20 Dmn Dmn Anterior cingulate cortex 1 6 -24 0 Sc FPV Thalamus 1 12 -17 8 Sc FPV Thalamus 1 -5 -28 -4 Sc FPV Lingual gyrus 1 31 -14 2 Sc Sens Putamen 1 29 1 4 Sc Sens Putamen 1 -31 -11 0 Sc Sens Putamen 1 15 5 7 Sc FPV Pallidum 1 37 -84 13 Vis FPV Middle occipital cortex 1 37 -81 1 Vis FPV Middle occipital cortex 1 -33 -79 -13 Vis FPV Inferior occipital cortex 1 49 -42 45 Fpn FPV Inferior parietal cortex 1 Automated Anatomical Labeling atlas was used to define anatomical regions. RSN, Initial Power et al., 2011, resting-state network affiliation; Mod, modules defined by modularity analysis in the Quaternary condition (i.e., Figure 3); Co, cingulo-opercular; Dmn, default-mode; Fpn, frontoparietal; Vis, visual; Sens, sensory; Sc, subcortical; Sal, salience; Aud, auditory; Mem, memory; Sshand, somatosensory hand; DorAtt, dorsal attention; FPV, fronto-parietal-visual.