The Developing Human Connectome Project: typical and disrupted perinatal functional connectivity

Abstract The Developing Human Connectome Project is an Open Science project that provides the first large sample of neonatal functional MRI data with high temporal and spatial resolution. These data enable mapping of intrinsic functional connectivity between spatially distributed brain regions under normal and adverse perinatal circumstances, offering a framework to study the ontogeny of large-scale brain organization in humans. Here, we characterize in unprecedented detail the maturation and integrity of resting state networks (RSNs) at term-equivalent age in 337 infants (including 65 born preterm). First, we applied group independent component analysis to define 11 RSNs in term-born infants scanned at 43.5–44.5 weeks postmenstrual age (PMA). Adult-like topography was observed in RSNs encompassing primary sensorimotor, visual and auditory cortices. Among six higher-order, association RSNs, analogues of the adult networks for language and ocular control were identified, but a complete default mode network precursor was not. Next, we regressed the subject-level datasets from an independent cohort of infants scanned at 37–43.5 weeks PMA against the group-level RSNs to test for the effects of age, sex and preterm birth. Brain mapping in term-born infants revealed areas of positive association with age across four of six association RSNs, indicating active maturation in functional connectivity from 37 to 43.5 weeks PMA. Female infants showed increased connectivity in inferotemporal regions of the visual association network. Preterm birth was associated with striking impairments of functional connectivity across all RSNs in a dose-dependent manner; conversely, connectivity of the superior parietal lobules within the lateral motor network was abnormally increased in preterm infants, suggesting a possible mechanism for specific difficulties such as developmental coordination disorder, which occur frequently in preterm children. Overall, we found a robust, modular, symmetrical functional brain organization at normal term age. A complete set of adult-equivalent primary RSNs is already instated, alongside emerging connectivity in immature association RSNs, consistent with a primary-to-higher order ontogenetic sequence of brain development. The early developmental disruption imposed by preterm birth is associated with extensive alterations in functional connectivity.

† These authors contributed equally to this work.
The Developing Human Connectome Project is an Open Science project that provides the first large sample of neonatal functional MRI data with high temporal and spatial resolution. These data enable mapping of intrinsic functional connectivity between spatially distributed brain regions under normal and adverse perinatal circumstances, offering a framework to study the ontogeny of large-scale brain organization in humans. Here, we characterize in unprecedented detail the maturation and integrity of resting state networks (RSNs) at term-equivalent age in 337 infants (including 65 born preterm). First, we applied group independent component analysis to define 11 RSNs in term-born infants scanned at 43.5-44.5 weeks postmenstrual age (PMA). Adult-like topography was observed in RSNs encompassing primary sensorimotor, visual and auditory cortices. Among six higher-order, association RSNs, analogues of the adult networks for language and ocular control were identified, but a complete default mode network precursor was not. Next, we regressed the subject-level datasets from an independent cohort of infants scanned at 37-43.5 weeks PMA against the group-level RSNs to test for the effects of age, sex and preterm birth. Brain mapping in term-born infants revealed areas of positive association with age across four of six association RSNs, indicating active maturation in functional connectivity from 37 to 43.5 weeks PMA. Female infants showed increased connectivity in inferotemporal regions of the visual association network. Preterm birth was associated with striking impairments of functional connectivity across all RSNs in a dose-dependent manner; conversely, connectivity of the superior parietal lobules within the lateral motor network was abnormally increased in preterm infants, suggesting a possible mechanism for specific difficulties such as developmental coordination disorder, which occur frequently in preterm children. Overall, we found a robust, modular, symmetrical functional brain organization at normal term age. A complete set of adult-equivalent primary RSNs is already instated, alongside emerging connectivity in immature association RSNs, consistent with a primary-to-higher order ontogenetic sequence of brain development. The early developmental disruption imposed by preterm birth is associated with extensive alterations in functional connectivity.

Introduction
The Developing Human Connectome Project (dHCP) is an Open Science project funded by the European Research Council to provide a large dataset of functional and structural brain images from 20 to 44 weeks of gestational age. This enables the characterization of 4D (three spatial dimensions and time) connectivity maps, which map the trajectories of human brain development to improve understanding of normal development and allow earlier detection and intervention for neurological and psychological disorders.
This paper analyses functional connectivity at the time of normal birth in infants born at term and preterm. Temporal coherences in the blood oxygen level-dependent (BOLD) contrast measured with resting state functional MRI (rs-fMRI) can be spatiotemporally decomposed into resting state networks (RSNs), 1,2 predominantly at low frequency (50.1 Hz), 3 distinct from cardiovascular signal. 4 Whilst RSNs have been extensively and robustly characterized in the mature brain, previous studies of RSN development in newborn infants have been limited by smaller sample sizes. The dHCP provides the first high quality, large-scale, 4D dataset of functional connectivity at this critical period of development, enabling us to address two key questions. First, are higher order RSNs such as the default mode network (DMN) 5 instated with adult topology in the neonatal period? Some find analogues of these at term-equivalent age (TEA) [6][7][8][9][10] while others locate their origin in later infancy or early childhood, contemporaneous with the emergence of the higher cognitive abilities these networks are believed to support. 11,12 Second, what is the effect of preterm birth on RSN development? Preterm birth is associated with significant risk for enduring neurodevelopmental and psychiatric problems in later life, [13][14][15][16] even in the absence of overt structural brain injury during the preterm period. Various alterations in the complexity, scope, strength and efficiency of functional connectivity in preterm-at-term infants have been reported 9,10,[17][18][19] ; however, the majority of studies lack the large numbers of control subjects required to characterize these effects with precision.
The mature adult RSNs are well characterized, with high intrasubject reproducibility, 20,21 largely consistent topology across healthy subjects, and anatomical mapping that reinforces both structural and task-functional MRI-derived parcellations of the cortex. 22 The identification of functional MRI-RSN signatures associated with disease offers considerable translational potential due to rs-fMRI's relatively straightforward and widely used acquisition, whole-brain coverage, and high spatial resolution compared to other functional imaging methods. The achievement of this in the immature brain requires a complete account of RSN ontogeny. The developing CNS shows spontaneous, patterned, correlated intrinsic activity from early prenatal life (reviewed in Blankenship and Feller, 23 Keunen et al. 24 and Vasung et al. 25 ); immature RSNs can be identified from as early as 26 weeks postmenstrual age (PMA) in the preterm infant. 6,9 By TEA, RSNs encompassing brain regions serving primary functions (sensorimotor, auditory, visual) are clearly apparent in mature configuration in both term and preterm-born infants [6][7][8][10][11][12]26 and show minimal change in infancy, while higher order association networks appear to be largely fragmented at TEA, with studies suggesting that they synchronize into complete adult-like core structure in the first year (DMN, dorsal attention network) or second year (salience network, bilateral frontoparietal networks) of life. 11,12 Importantly, TEA is not a single time point but rather a window encompassing a critical period of brain development in which there is intense myelination of white matter (reviewed in Dubois et al. 27 ) and rapid expansion in both the size and gyrification of the cerebral cortex. 28,29 Dense sampling across the age range is therefore required to map the associated changes in functional connectivity.
Here, we apply a data-driven approach to 337 rs-fMRI datasets acquired in term and preterm infants between 37 and 44.5 weeks PMA. We first defined a normative set of RSNs in a subsample of term-born infants scanned at 43.5-44.5 weeks PMA using probabilistic independent component analysis (ICA). 30 ICA is a dimensionality reduction technique that decomposes data into a set of components with maximal statistical independence; applied to rs-fMRI, ICA can reveal large-scale brain networks without requirement for a predefined model of network structure. We then regressed subject-level data from term and preterm infants scanned at 37-43.5 weeks PMA against these networks. The resulting whole-brain correlation maps enabled us to both characterize the ontogeny of individual RSNs, and investigate the influence of prematurity on cortical functional connectivity. We hypothesized that primary sensorimotor RSNs would be well established at TEA, while higher order RSNs would be immature and emerging; and that preterm birth would be associated with reduced functional connectivity at TEA.

Subjects
Research participants were prospectively recruited as part of the dHCP, an observational, cross-sectional Open Science programme approved by the UK National Research Ethics Authority (14/LO/ 1169). Written consent was obtained from all participating families prior to imaging. Term-born infants were recruited from the postnatal wards and approached on the basis of being clinically well. Preterm-born infants were recruited from the neonatal unit and postnatal wards. Infants were not approached for study inclusion if there was a history of severe compromise at birth requiring prolonged resuscitation, a diagnosed chromosomal abnormality or any contraindication to MRI scanning (e.g. due to incompatible implants). No infants included in the final study group required treatment for clinically significant brain injury. We selected 416 structural-functional datasets acquired at TEA from the 2019 (second) dHCP data release. Only infants scanned at 37-44.5 weeks PMA in term-born infants, or 37-43.5 weeks PMA in preterm-born infants, were considered for inclusion. One infant was included twice because of the two datasets being acquired at different ages; only the second dataset was used. Thirty-five infants were excluded because of a history of neurodevelopmental disorder in a first-degree relative. Forty-three were excluded because of motion (see the 'Functional data preprocessing' section). The final study population therefore consisted of 337 infants, divided into three groups: (i) term-born infants scanned at 43.5-44.5 weeks PMA, who were used to define the normative set of RSNs and excluded from all subsequent subject-level analyses; and the remaining infants scanned at 37-43.5 weeks PMA, including both (ii) term-born; and (iii) preterm-born infants (Table 1). No infants in the preterm study group had major brain injury such as periventricular leukomalacia, grade 3 or 4 intraventricular haemorrhage, major intracerebral haemorrhage, or ischaemic focal brain lesions. As is commonly seen in preterm infants, there was evidence of mild pathologies including 15 (23%) with a history of grade 1 or 2 intraventricular haemorrhage; 10 (15%) with small cerebellar haemorrhage with no involvement of the vermis; 22 (34%) with punctate white matter lesions; and 22 (34%) with diffuse excessively high signal intensity (DEHSI) of the white matter. These mild pathologies are known to be poorly predictive of later neurodevelopmental outcome at the individual level, [31][32][33][34] and there were no significant differences in early developmental outcome between the term and preterm-born groups ( Table 1).

MRI data acquisition
Neuroimaging was acquired in a single scan session for each infant at the Evelina Newborn Imaging Centre, Evelina London Children's Hospital, using a 3 T Philips Achieva system (Philips Medical Systems). All infants were scanned without sedation in a scanner environment optimized for safe and comfortable neonatal imaging, including a dedicated transport system, positioning device and a customized 32-channel receive coil, with a custommade acoustic hood. 35 MRI-compatible ear putty and earmuffs were used to provide additional acoustic noise attenuation. Infants were fed, swaddled and comfortably positioned in a vacuum jacket prior to scanning to promote natural sleep. All scans were supervised by a neonatal nurse and/or paediatrician who monitored heart rate, oxygen saturation and temperature throughout the scan. High temporal resolution BOLD functional MRI optimized for neonates was acquired over 15 min 3 s (2300 volumes) using a multislice gradient-echo echo planar imaging (EPI) sequence with multiband excitation (multiband factor 9). Repetition time was 392 ms, echo time was 38 ms, flip angle was 34 , and the acquired spatial resolution was 2.15 mm isotropic. 36 For registration of the functional MRI data, high-resolution T 1 -and T 2 -weighted anatomical imaging was also acquired in the same scan session, with a spatial resolution of 0.8 mm isotropic (T 1 -weighted: field of view 145 Â 122 Â 100 mm, repetition time 4795 ms; T 2 -weighted: field of view 145 Â 145 Â 108 mm, repetition time 12 000 ms, echo time 156 ms).

Functional data preprocessing
Data were preprocessed using an in-house pipeline optimized for neonatal imaging and specifically developed for the dHCP, detailed in Fitzgibbon et al. 37 In brief, susceptibility dynamic distortion together with intra-and intervolume motion effects were corrected in each subject using a bespoke pipeline including slice-to-volume and rigid-body registration. [38][39][40][41] To regress out signal artefacts related to head motion, cardiorespiratory fluctuations and multiband acquisition, 42 24 extended rigid-body motion parameters were regressed together with single-subject ICA noise components identified with the FSL FIX tool (Oxford Centre for Functional Magnetic Resonance Imaging of the Brain's Software Library, version 5.0). Denoised data were registered into T 2 -weighted native space using boundary-based registration 43 and non-linearly registered to a standard space based on a weekly template from the dHCP volumetric atlas 44 using a diffeomorphic multimodal (T 1 /T 2 ) registration. 45 While the functional MRI preprocessing pipeline for the dHCP 37 addresses the potential problem of head motion in rs-fMRI data, 46,47 motion is also a surrogate marker of the arousal state of the infant, which interacts with the underlying neural activity. 48,49 To address this issue, we opted for a conservative approach consisting of the selection of a continuous subsample of the data ($70%) with lowest motion for each subject, and excluding those subjects with a high level of motion from further analyses. Specifically, volumes with DVARS (the root mean square intensity difference between successive volumes) 41.5 interquartile range (IQR) above the 75th centile, after motion and distortion correction, were considered as motion outliers. 37 Mean DVARS was 90.5 [standard deviation (SD) 18.5] in the term-born group and 95.5 (SD 20.3) in the preterm-born group (P = 0.056, unpaired t-test). As DVARS is a relative measure, the absolute DVARS cut-off varied between subjects. Within each acquired dataset (2300 volumes), the continuous set of 1600 volumes with the minimum number of motion-outlier volumes was identified, and the dataset cropped accordingly for all subsequent analyses. Subjects with more than 160 motion-outlier volumes (10% of the cropped dataset) were excluded entirely. This allowed us to minimize the potential effect of different states of arousal even after appropriately denoising the data. The number of motion-outlier volumes remaining in the cropped dataset was recorded for each subject and included as a covariate in all subsequent regression analyses. The median number of motion-outlier volumes in the term-born group was 49.5 (IQR 27-86.5) and in the preterm-born group was 34 (IQR 12-83) [group difference not significant under assumption of normality (P = 0.185, unpaired t-test) or non-normality (P = 0.052, Mann-Whitney U-test)].

Group-level network definition
We first defined the normative set of RSNs by group ICA in 24 healthy term-born infants scanned at 43.5-44.5 weeks PMA. These subjects were excluded from all subsequent regression analyses. Probabilistic group ICA by temporal concatenation across subjects was carried out using FSL MELODIC. 30 The ICA dimensionality was set at 30, representing a pragmatic balance between robustness and interpretability (as in Toulmin et al. 19 ). The output comprised 30 group-average spatial maps representing 30 independent components. The maps were visually inspected and each component manually labelled as RSN or noise, following guidelines in Fitzgibbon et al. 37

Subject-level analyses
We next regressed the group-level spatial maps into the subjectlevel 4D space-time datasets of the subjects scanned at 37-43.5 weeks PMA (248 term-born, 65 preterm-born). Specifically, the group-level spatial maps (including both RSN signal and artefact components) were used to generate subject-specific versions of the spatial maps and associated time series using dual regression. 50 Artefact components were included to better account for confound variance (noise) in the regression model. 51 First, for each subject, the set of group-level RSN spatial maps was regressed (as spatial regressors in a multiple regression) into the subject's 4D space-time dataset. This resulted in a set of subject-specific time series, one per group-level spatial map. Next, those time series were regressed (as temporal regressors, again in a multiple regression) into the same 4D dataset, resulting in a set of subject-specific spatial maps, one per group-level spatial map.
We then performed cross-subject analysis using general linear models (GLM) to test for the effects of group (term versus preterm birth, sex) and continuous variables (gesational age at birth, PMA at scan) on the subject-level RSN spatial maps, including the number of motion-compromised volumes as a nuisance covariate. Specifically, in the model evaluating the effect of PMA at scan (term-born infants only) the covariates were sex and motion; in the model evaluating the effect of sex the covariates were PMA at scan, gesational age at birth and motion; in the model evaluating the effect of term versus preterm birth the covariates were PMA at scan, sex and motion; and in the model evaluating the effect of gesational age at birth the covariates were PMA at scan, sex and motion. A further group-level analysis was conducted in which term-born infants were separated into weekly bins according to their PMA at scan, enabling group-average maps of functional connectivity at each week of brain development to be generated for qualitative comparison. For this we entered data from the 20 subjects in each bin (37.5-38.5 weeks, 38.5-39.5 weeks, 39.5-40.5 weeks, 40.5-41.5 weeks, and 41.5-42.5 weeks) with the lowest number of postnatal days of life at time of scan, to maximize similarity between groups for meaningful visual comparison. The covariates in this model were sex and motion. Voxel-wise statistical tests were implemented in FSL randomise 52 using threshold-free cluster enhancement 53 with 5000 permutations. As all contrasts were twotailed, family-wise error-rate (FWE) corrected (for multiple comparisons across voxels) P-values 5 0.025 were accepted as significant. Because of the exploratory nature of this study, the main results are presented without correction for the effect of multiple RSNs assessed; however, we also provide the Bonferroni-corrected results (i.e. P 5 0.025/n RSNs) in the Supplementary material.
To quantify longitudinal changes in within-network functional connectivity, we analysed the relationship between PMA at scan and a derived parameter we term 'core network strength'. This measure was determined for each RSN for each subject by masking the RSN-specific spatial map (the output of stage two of dual regression) by the corresponding group-ICA network template thresholded at Z 4 3, then calculating the mean b-parameter value (regression coefficient) within the masked image. The partial Spearman's correlation between core network strength and PMA at scan was calculated in term-born infants while controlling for sex and motion (number of motion-compromised volumes), and a GLM was used to test for group differences in core network strength between term and preterm infants while controlling for PMA at scan, sex and motion. Correlation and GLM analyses of core network strength were implemented in Python 3.7 with Pingouin 0.2.9 and statsmodels 0.10.1.

Anatomical localization and data visualization
Results were localized in the standard space using an in-house adaptation of the neonatal version 54 of the AAL atlas, 55 projected to the 40-week high-resolution neonatal dHCP template. 44 Data were displayed using FSLeyes for planar visualization and Connectome Workbench for cortical surface visualization.

Data availability
The dHCP is an open-access project. The imaging and collateral data used in this study were included in the 2019 (second) dHCP data release, which can be downloaded by registering at https:// data.developingconnectome.org/.

Resting state networks
Eleven RSNs were identified by group ICA in a subsample of termborn infants scanned between 43.5 and 44.5 weeks PMA (n = 24), who were excluded from any further analyses. Five RSNs included primary motor or sensory cortical areas and were categorized as primary networks (Fig. 1A): medial motor, lateral motor, somatosensory, auditory and visual. The remaining six were categorized as association networks (Fig. 1C): motor association (including the premotor and supplementary motor areas), temporoparietal (including Broca's area and the extended Wernicke's area), posterior parietal (including the precuneus and posterior cingulate cortices), frontoparietal (including the frontal, supplementary and parietal eye fields), prefrontal and visual association. The full cortical surface parcellation is provided in Supplementary Fig. 1 and Supplementary Video 1.

Effect of postmenstrual age at scan
To characterize normal maturation in functional connectivity from 37-43.5 weeks in term-born infants, we analysed the association between previously calculated RSNs independently regressed to each subject and PMA at scan, while controlling for sex and motion. Brain regions showing increasing connectivity with older PMA at scan were identified in four RSNs, all association networks ( Fig. 2 and Supplementary Fig. 2, Bonferroni-corrected). Localization of significant voxels to the AAL atlas is provided in Supplementary Table 1. There were no brain tissue regions showing negative association with older age at scan.
To illustrate maturational changes in functional connectivity, we produced spatial maps of average network structure in termborn infants categorized into weekly groups according to their PMA at scan, while controlling for sex and motion (Fig. 3).
To quantify longitudinal changes in within-network functional connectivity, we analysed the relationship between PMA at scan and a derived parameter we term 'core network strength', defined as the mean b-parameter value in each subject's RSN-specific spatial map (the outputs of stage two of dual regression) after masking by the corresponding group-ICA network template thresholded at Z 4 3. Three RSNs showed a positive partial correlation between PMA at scan and core network strength (Fig. 4). There were no RSNs with negative correlation.

Effect of sex
To determine differences in functional connectivity between male and female infants, we analysed this as a group effect, while controlling for gesational age at birth, PMA at scan and motion. Female infants showed increased connectivity of inferior occipitotemporal regions (including the posterior fusiform gyrus) within the visual association network (Fig. 5 and Supplementary Fig. 3, Bonferroni-corrected).

Effect of preterm birth
To determine differences in functional connectivity between termand preterm-born infants we first analysed this as a group effect, while controlling for PMA at scan, sex and motion. There was extensive impairment of functional connectivity across all RSNs in preterm-born infants; uncorrected core network strength was 23-41% reduced relative to term-born infants across the 11 networks (all P 5 0.001, independent samples t-tests). Conversely, pretermborn infants showed increased connectivity of the bilateral superior parietal lobule within the lateral motor network ( Fig. 6 and Supplementary Fig. 4, Bonferroni-corrected). The association of younger gesational age at birth with impaired functional connectivity was replicated across all networks in a separate analysis in which gesational age at birth was entered as a continuous variable, indicating a dose-dependent effect of prematurity on functional connectivity ( Supplementary Figs 5 and 6, Bonferroni-corrected).

Discussion
In this large cohort of newborn infants we provide detailed characterization of the maturational trajectories of normal functional network development at TEA, and show that the early developmental disruption imposed by preterm birth is associated with significant and widespread alterations in functional connectivity.

Network architecture and maturation in term-born infants
Overall, we found a robust, modular, symmetrical functional organization of the brain at TEA. Our results confirm and further elucidate the primary-to-higher order maturational sequence of RSN development. Using a whole-brain, voxel-level approach, we depicted changes in the posterior parietal, frontoparietal, prefrontal and visual association networks, showing expansions in network shape with increasing age at scan (Fig. 2). We also investigated the effect of age at scan across a predefined area (the core region, as defined by the template network), demonstrating global changes in network strength or integrity in the somatosensory, posterior parietal and visual association networks (Fig. 4).

Primary networks
We identified five primary RSNs (Fig. 1A), which showed adult-like topology from the earliest ages studied (Fig. 3A) and no significant change in architecture from 37 to 43.5 weeks PMA. Primary, unimodal RSNs mature earlier than higher-order networks in the preterm brain 6,9,56 ; our finding of an adult-like configuration of primary RSNs at TEA is in agreement with previous studies at this age. [6][7][8][9]11,12,26 The precise localization of sensorimotor networks along the central sulcus is especially striking in our data, even in the youngest infants studied (Fig. 3A). Determination of somatotopic maps in primary sensorimotor cortical areas occurs as early as mid-third trimester equivalent age, with similar stimulation response to adults observed by TEA. 57,58 We additionally observed a significant increase in core network strength within the somatosensory network from 37 to 43.5 weeks PMA (Fig. 4), possibly reflecting increasing integration of secondary somatosensory cortex at this age, 57 and/or increased influence of ex utero experience on this network. The bilateral insula (Fig. 1A) and thalamus ( Supplementary Fig. 1C) were strongly connected within the medial motor network, consistent with previous studies finding strong thalamocortical connectivity in sensorimotor networks. 10,19 Association networks We identified six RSNs representing higher-order association networks (Fig. 1C). Using quantitative (Figs 2 and 4) and qualitative (Fig. 3) methods, we found modest expansions in both the spatial extent and core temporal coherence of higher-order association networks from 37 to 43.5 weeks PMA. To our knowledge this is the first time these changes have been quantified over this brief but developmentally critical period. The heterogeneous timing of functional network development, in which primary networks mature earlier than higher-order association networks, can be related to parallel changes in brain structure (reviewed in Keunen et al. 24 ). Structural connectivity of the cortex begins with thalamic connections to frontal, auditory, visual and somatosensory cortices at 24-32 weeks gestation, while long-range cortico-cortical connections are not established until 33-35 weeks (reviewed in Dubois et al. 59 and Kostovic and Jovanov-Milosevic 60 ). The same sequence is later repeated in cortical myelination, with the 'primordial' sensorimotor and visual cortices histologically more mature at the time of birth. 61 White matter tracts connecting to these regions, such as the corticospinal tract and optic radiation, are also the first to mature later in infancy (reviewed in Dubois et al. 27 ). The structural and functional ontogeny mirrors the observed behavioural sequence of developmental 'milestones' in young children, in which sensorimotor, auditory and visual competencies are acquired before higher-order cognitive functions. 24 The two RSNs showing greatest increase in intrinsic connectivity (core network strength) from 37 to 43.5 weeks PMA were the posterior parietal network and visual association network (Fig. 4). The former encompasses the medial precuneus and posterior cingulate cortices (Supplementary Fig. 1C), an area of emerging functional connectivity at TEA. 62 In adulthood these regions are a prominent component of the DMN, leading some to label infant RSNs encompassing these as DMN precursors. [6][7][8][9][10] However, the mature DMN also incorporates distinct modules in the anterior cingulate/medial prefrontal cortex, orbitofrontal cortex, lateral Figure 1 Resting state networks identified by group independent component analysis. Spontaneous BOLD activity patterns (RSNs) derived from group ICA in 24 term-born infants scanned at 43.5-44.5 weeks PMA. Panels: Example axial, coronal, and sagittal slices for meaningful spatial patterns in primary (A) and association (C) RSNs, thresholded at Z 4 3 and overlaid on a T 1 structural template, displayed in radiological convention. Centre: Functional parcellation of the brain using a 'winner-takes-all' approach based on the RSNs from group ICA. RSNs were spatially smoothed and thresholded at Z 4 1 prior to determination of the 'winning' RSN at each voxel. The resulting volume was projected to the midthickness cortical surface using enclosed (nearest neighbour) volume-to-surface mapping, here displayed on the pial surface of an individual subject scanned at 42 weeks PMA and viewed from the dorsal (B) and left lateral (D) aspects. temporal cortex and hippocampus (reviewed in Raichle 63 ). We observed no temporal involvement in the posterior parietal network, and only sparse frontal involvement, specifically at the right anterior cingulate (Supplementary Fig. 1C) and bilateral orbitofrontal cortex (Supplementary Fig. 1E). This dominant posterior hub with limited frontoparietal connectivity bears more similarity to the adult DMN under anaesthesia. 64,65 Overall we find support for the concept of fragmented local modules prevailing over longrange integration at this period of development, preceding the emergence of a full analogue of the adult DMN at 6-12 months of age. 11,12,62 The visual association network comprises lateral occipital (Supplementary Fig. 1D) and inferotemporal (Supplementary Fig.  1B) cortices; regions that contribute to the ventral stream of visual processing, in which simple features coded by primary visual cortex are transformed into higher-level representations of objects, invariant of their size, rotation and position, enabling downstream object recognition and semantic processing. 66,67 It was therefore not surprising to find significant growth in the strength of this network from 37-43.5 weeks PMA (Figs 2 and 4), a period in which infants are increasingly exposed to, and able to resolve, objects in the visual field. 68 Furthermore, after controlling for differences in age, we found areas of increased connectivity within this network in female infants across inferotemporal regions including the posterior fusiform gyrus (Fig. 5). The fusiform is sensitive to complex visual stimuli including faces and facial expressions 69 ; in the corresponding region of the macaque brain the code determining face cell firing was recently deciphered. 70 In humans, reduced functional connectivity of the fusiform face area is associated with developmental prosopagnosia. 71 The sex difference in functional connectivity we have identified in this region is especially interesting in the context of behavioural data in which female neonates, compared to males, show increased preference for looking at faces. 72 Sex differences in visual attention to social stimuli have also been described in older infants 73 and in other newborn primates. 74 Structural brain development is sexually dimorphic, with small differences in tissue morphometry observed across the lifespan, frequently involving the temporal lobes, and under the influence of foetal testosterone in males. 75,76 Our finding of a sex difference in functional brain development in the inferior temporal lobes is interesting in this context, although the biological mechanisms for this and any correlations with behavioural data remain unknown at this time. Further investigation of functional connectivity in the ventral stream and social-cognitive development might elucidate mechanisms for sex differences in this domain.
Two RSNs comprised segregated (i.e. non-contiguous) brain regions, revealing anatomically meaningful patterns of functional connectivity. The temporoparietal network (Fig. 1D) connects a posterior module encompassing the extended Wernicke's area to a smaller anterior module corresponding to Broca's area. Integrated structural-functional analysis in adults showed this network is facilitated by the arcuate fasciculus. 77 The instatement of a putative 'language network' in early infancy is supported by stimulusfunctional MRI showing activation of these regions in response to speech. 78,79 The frontoparietal network (Fig. 1B) connects the frontal, supplementary and parietal eye fields, with close resemblance to the adult dorsal attention network. 80 Ocular control relies on widespread white matter connections between cortical and subcortical regions, the microstructural integrity of which correlates with visual fixation behaviour in the neonate. 81 Striatal projections of the frontal and supplementary eye fields converge upon the caudate nucleus 82 ; we found a positive association between older PMA at scan and functional connectivity of the caudate nucleus within this frontoparietal network (Fig. 2), consistent with active development of the oculomotor corticostriatal system at this age.

Impact of preterm birth
Preterm birth confers a high risk of neurodevelopmental impairment 13,14,16 and psychiatric illness in later life. 15 Pervasive deficiencies and delays in structural brain maturation have been identified in preterm infants scanned at TEA, even in those without focal brain injury, including macrostructural differences in tissue volume and gyrification 29,83-85 and microstructural alterations in both grey and white matter. [86][87][88] The overall structural network architecture appears unchanged, with preservation or even abnormal strengthening of the rich-club organization of highly connected cortical hubs, at the expense of diminished peripheral Figure 2 Changes in network shape with increasing age at scan. Brain regions showing increasing functional connectivity with older PMA at scan in term-born infants scanned at 37-43.5 weeks PMA. Example sagittal, coronal, and axial slices for meaningful spatial patterns in four RSNs are shown, overlaid on a T 1 structural template and displayed in radiological convention. T-statistic maps were thresholded at P 5 0.025 (FWE corrected). White lines represent the outlines of the group ICA RSNs, thresholded at Z 4 3. connectivity and specific disruptions to thalamocortical, corticalsubcortical and short-distance corticocortical connectivity. [89][90][91][92][93] Widespread impairment of functional connectivity Now we show that, similar to structural connectivity, functional connectivity is profoundly affected by preterm birth. We found striking deficiencies in within-network connectivity across the full range of RSNs studied (Fig. 6), also replicated as a dose-dependent relationship, such that increased exposure to prematurity (younger gesational age at birth) was associated with decreased functional connectivity ( Supplementary Fig. 5); effects we were not powered to detect in a prior study due to a limited number of term-born infants. 6 Our results suggest that although functional connectivity increases across the preterm period, 6,9,10,94,95 it does not reach a normal configuration at TEA. Instead, there appears to be an aberrant developmental trajectory, in which connections between brain regions are reconfigured by premature exposure to the extra-uterine environment. Graph theoretical approaches have shown global network measures of clustering, integration and modularity at TEA are all reduced in preterm infants compared to full-term controls. 18,96 Hypothesis-driven seed-based approaches have identified disrupted thalamocortical connectivity, 9,19 consistent with structural disruption of the same. 91 In our data-driven, whole-brain ICA approach, the main finding was globally reduced within-network functional connectivity. Primary and association RSNs appeared to be similarly affected, in contrast to the findings of Smyser and colleagues, 9 who also used whole-brain correlation mapping, and found primary RSNs were less affected by prematurity. 10 This discrepancy may be due to differences in approach to RSN definition (adult-derived RSNs), network mapping (node-based) and inclusion criteria for the preterm group (530 weeks gesational age at birth). In another study investigating preterm-at-term infants with whole-brain ICA, the method comprised identification of 71 nodes by ICA followed by subject-specific network estimation and selection of discriminatory edges between cases and controls using machine-learning classifiers. 17 Figure 4 Relationship between postmenstrual age at scan and core network strength. Relationship between the residuals (after correcting for sex and motion) for PMA at scan and core network strength in term-born infants scanned at 37-43.5 weeks PMA. Core network strength was defined as the mean b-parameter value in each subject's RSN-specific spatial map after masking by the corresponding group-ICA network template thresholded at Z 4 3. Partial Spearman's correlation coefficients and associated P-values are displayed for the three RSNs significant at P 5 0.025. Example axial, coronal and sagittal slices for meaningful spatial patterns in the corresponding group-ICA network templates are shown for reference.
Connections to frontal and basal ganglia nodes were overrepresented among the discriminatory edges, indicating altered connectivity in preterm infants. Taken together, these different approaches provide complementary demonstrations of spatially widespread impaired RSN coherence in the preterm-at-term brain.

Modulation of parieto-motor connectivity
In the context of brain-wide deficiencies in functional connectivity in preterm-at-term infants, it was notable that there was also increased functional connectivity of the bilateral superior parietal lobule (Brodmann area 5) within the lateral motor network, both when prematurity was evaluated as a group effect (Fig. 6) and as a continuous variable ( Supplementary Fig. 5). The lateral motor network corresponds approximately to the primary somatotopic regions serving the upper limb, hand and face (Fig. 1D). Ex utero experience during the preterm period strongly influences the development of sensorimotor networks: bilateral functional responses in the peri-rolandic cortices to stimulation of the wrist increase with postnatal age, even after controlling for gesational age at birth. 57 Interestingly, connectivity with superior parietal regions appears to occur as a feature of normal development in the lateral motor network in older term-born infants (Fig. 3A). Brodmann area 5 comprises the somatosensory association cortex, which integrates visual and somatosensory inputs to encode limb configuration in space, enabling coordinated movements within the immediate environment. 97,98 Neural circuits are shaped by experience during critical periods of development; in animal studies, peak plasticity in somatosensory networks occurs earlier than peak plasticity in visual and auditory networks (reviewed in Reh et al. 99 ). It is intuitive that connectivity of area 5 with lateral motor cortex could be highly dependent upon ex utero experience, given the natural constraints upon limb movement and visuomotor integration in utero. We propose therefore that the experience of premature exposure to the extrauterine environment, occurring at a time of heightened sensorimotor plasticity, 57,100,101 modulates the normal development of parieto-motor connectivity and leads to an abnormal increase in connectivity at TEA, while acknowledging this mechanistic account as speculative at the present time.
Previous studies have identified increased functional connectivity of certain primary cortical regions in preterm-at-term infants compared to controls, specifically the lateral postcentral gyrus with the thalamus 19 and regional connectivity within occipital/visual networks. 87 This may occur at the expense of connectivity in other brain areas, and can persist into later life; analysis of language networks in preterm children scanned at 12 years of age showed increased connectivity with primary sensorimotor areas, but reduced connectivity with higher-order frontal areas. 102 Relatively conserved topology of core structural networks has been reported in preterm-born babies, 92 persisting into later childhood and adulthood. 103,104 Disruption of the normal balance of sensorimotor development may have persisting effects on later motor and cognitive development. In the mature brain, the superior parietal lobule supports not only the smooth execution of motor plans 105 but also more abstract visuospatial functions such as mental rotation. 106 The aberrant parietal connectivity we have identified at TEA could therefore be a prelude to specific difficulties occurring with high prevalence in preterm children, such as developmental coordination disorder, [107][108][109][110][111] inattention and intellectual impairment (reviewed in Rogers et al. 112 ). Long-term followup of the study population at school age will be required to confirm this hypothesis.

Limitations
The customized neonatal imaging system for the dHCP includes a close-fitting head coil sized specifically for the neonatal head, thus providing exceptional signal-to-noise at the cortical surface. 35 This bias towards surface-proximate sources is compounded by the use of highly accelerated multiband EPI. 37 As such, this has likely resulted in greater sensitivity to detect correlated signal fluctuations in the cerebral cortex compared to deeper sources such as the thalamus, basal ganglia and cerebellum. This may explain the relatively sparse involvement of subcortical regions in the identified RSNs ( Fig. 1A and C). Thalamocortical and cerebellar functional connectivity may be better appreciated with seed-based methods. 19,113 We also noted sparse involvement of inferior frontotemporal regions, even at Z 4 1 (Supplementary Fig. 1B). The dHCP functional pipeline includes advanced distortion-correction techniques, 37 but some signal loss related to air/tissue and bone/ tissue interfaces in this vicinity cannot be fully excluded, and our use of a single phase-encode direction (anterior-posterior) may also compress signal in the frontal regions. However, this sparsity may also reflect biological reality in these brain regions, which are the least myelinated at birth 61 and so may be the least able to participate in long-range phase-synchronous activity.
In this study we used a dense sampling strategy at TEA to infer longitudinal change in RSNs, but each infant was scanned on only one occasion. Gesational age at birth and PMA at scan were strongly correlated within the term-born group, which complicates the interpretation of these longitudinal analyses. Furthermore, as some potentially relevant neonatal characteristics, such as intracranial volume and postnatal days of life, are intrinsically associated to some of our variables of interest (i.e. PMA at scan, sex, gesational age at birth), it is difficult to disentangle their relative contributions to our results.
The optimized functional MRI dHCP pipeline includes multiple steps to control for motion and physiological confounds, thus minimizing data loss. However, while well-fed babies tend to fall asleep during the scan, subject motion is inherently correlated with the arousal and sleep state of the baby, which may have an effect in the reconstructed RSNs. 114 While our stringent control for high motion during the scan will minimize the potential effect of subject differences in arousal and sleep state, the specific measure that should be used as a surrogate to model arousal state is unclear. Future studies using simultaneous EEG-functional MRI could help to better understand the effect of different sleep states on RSNs. Differences in arousal in the scanner between infants and adults should also be considered when comparing RSN topology between these groups. 115 Our use of infants scanned at 43.5-44.5 PMA to define the group-ICA components may have missed some sources of structured noise occurring predominantly at younger ages, such as CSF signal in the cavum septum pellucidum. More fundamentally, the extent to which BOLD signal might be confounded by cerebrovascular factors differing between preterm and term-born infants 116 remains open to debate. Some of the spatial expansions in association RSNs depicted in Fig. 2 were localized to white matter (Supplementary Table 1), which may also reflect maturational changes in cerebrovascular factors, combined with other age-related changes in sulcal depth and white matter signalto-noise ratio. Some of the important temporal dynamics in functional networks may be missed by rs-fMRI, which predominantly Figure 6 Effect of preterm birth on functional connectivity. Group differences in functional connectivity between term and preterm-born infants scanned at 37-43.5 weeks PMA. Coloured t-statistic maps thresholded at P 5 0.025 (FWE corrected) show brain regions with reduced (blue) or increased (red-yellow) connectivity in preterm-born infants. Example sagittal, coronal, and axial slices for meaningful spatial patterns within each RSN are shown, overlaid on a T 1 structural template and displayed in radiological convention. White lines represent the outlines of the group-ICA RSNs, thresholded at Z 4 3. Box plots show group differences in core network strength after regressing out PMA at scan, sex and motion. Core network strength was defined as the mean b-parameter value in each subject's RSN-specific spatial map after masking by the corresponding group-ICA network template thresholded at Z 4 3. P-values relate to the term versus preterm group contrast in a GLM in which core network strength was the dependent variable and PMA at scan, sex and motion were controlled for as nuisance covariates.
identifies activity at 50.1 Hz. 3 Complementary approaches such as EEG may help to address this. 117,118 Conclusion Brain development occurs in a preprogrammed and spatially heterogeneous progression, modulated by environmental influence. As such, we observed different trajectories for different neural systems, obeying a generally primary-to-higher order sequence of maturation. At TEA we found already instated a complete set of adult-analogous unimodal RSNs corresponding to primary sensorimotor, visual and auditory cortices, with relatively little change from 37-43.5 weeks PMA. In contrast, association RSNs appear fragmented and incomplete compared to the adult repertoire, and are undergoing active maturation at this time. Connectivity within the visual association network in particular is highly associated with age, likely as a result of postnatal environmental experience, but also modified by the sex of the infant. Preterm birth is associated with profoundly reduced functional connectivity across all RSNs, but also with augmentation of parieto-motor connectivity, with possible implications for understanding certain neurocognitive sequelae of prematurity. In future we may be able to positively modulate RSN development in prematurity via targeted environmental manipulations. 119 Preterm birth is best conceptualized as a developmental perturbation that reconfigures, rather than simply diminishes, the organization of functional brain networks.