Human thalamocortical structural connectivity develops in line with a hierarchical axis of cortical plasticity

Abstract

Human cortical development follows a hierarchical, sensorimotor-to-association sequence. The brain’s capacity to enact this sequence indicates that it relies on unknown mechanisms to regulate regional differences in the timing of cortical maturation. Given evidence from animal systems that thalamic axons mechanistically regulate periods of cortical plasticity, here we evaluate in humans whether the development of structural connections between the thalamus and cortex aligns with cortical maturational heterochronicity. By deriving a new tractography atlas of human thalamic connections and applying it to diffusion data from three youth samples (8–23 years; total n = 2,676), we demonstrate that thalamocortical connectivity matures in a generalizable manner along the cortex’s sensorimotor–association axis. Associative cortical regions with thalamic connections that take the longest to mature exhibit neurochemical, structural and functional signatures of protracted developmental plasticity as well as heightened sensitivity to the socioeconomic environment. This work highlights the role of the thalamus in the expression of hierarchical periods of cortical developmental plasticity and environmental receptivity.

Main

Human cortical development is a prolonged and temporally asynchronous process, with distinct cortical regions maturing at varying rates. During childhood and adolescence, interregional variability in maturational timing (or maturational ‘heterochronicity’) is organized by the cortex’s sensorimotor–association (S–A) axis¹. Accumulating evidence suggests that reductions in plasticity gradually progress with age from the sensorimotor to the association pole of this hierarchical cortical axis^{1,2,3,4,5,6,7,8,9}. The brain’s capacity to enact an S–A developmental sequence over decades indicates that intrinsic timing mechanisms exist that regulate the relative pace of cortical development in each region. Here we evaluate the hypothesis that thalamic connections with the cortex have a mechanistic role in coordinating timescales of cortical developmental plasticity. To conduct this evaluation noninvasively, we leverage diffusion magnetic resonance imaging (MRI) data from three independent youth samples and test whether thalamocortical structural connections exhibit a chronological maturational gradient that aligns with the S–A axis of cortical developmental heterochronicity.

The thalamus is a bilateral gray matter structure in the diencephalon that sends (typically reciprocal) axonal projections throughout the cortical mantle. These direct thalamic projections to cortex are established in prenatal and early postnatal development, at which point they begin to have a profound effect on the formation and refinement of cortical circuits. Rodent and nonhuman primate studies have shown that thalamic inputs to cortex partly control cortical region arealization^10,11, gene expression¹¹ and laminar architecture¹² from embryonic through juvenile developmental stages. Cortical sculpting by thalamic inputs has primarily been documented in primary cortices, with recent work extending this developmental phenomenon to association cortex in mice¹³ and in nonhuman primates¹⁴. Neuroimaging studies in humans have furthermore demonstrated that thalamocortical connectivity continues to evolve from infancy to adulthood, with structural and functional connectivity between the thalamus and association cortex increasing through adolescence^15,16,17,18 in a manner that facilitates hierarchical cortical dynamics^15,19,20. Cross-species data thus indicate that the thalamus densely interconnects with cortex, governs local cortical circuit maturation and organizes hierarchical cortical communication, making it well positioned to regulate a spatiotemporal maturational program that progresses along the S–A axis.

Axonal projections from the thalamus to cortex impact not only cortical area properties during development but also time windows of cortical developmental plasticity. In the mouse brain, critical periods of heightened cortical plasticity co-occur with the assembly²¹, reorganization²² and normative strengthening^23,24,25 of thalamocortical axons. Observed relationships between the expression of cortical plasticity and refinements in thalamocortical connectivity have origins in thalamic modulation of parvalbumin (PV) cortical interneurons—inhibitory cells that receive potent thalamic synapses in development^25,26,27,28 and exert strong control over the timing of windows of developmental plasticity^7,29,30. Previous work has shown that increases in the strength of glutamatergic thalamic inputs onto both pyramidal and PV neurons can enhance cortical plasticity²⁴, likely by shifting the cortex’s excitation/inhibition (E/I) balance to a plasticity-permissive state^24,29. In contrast, the stabilization of thalamocortical–PV interactions by perineuronal nets²⁴ or cell adhesion complexes²⁷ serves to restrict ongoing plasticity²¹. Multiple lines of evidence thus implicate the strengthening and stabilization of thalamocortical connectivity in the opening and closing of periods of cortical developmental plasticity.

Despite compelling findings linking thalamic input maturation to time windows of cortical malleability, prior studies documenting this developmental phenomenon were almost exclusively conducted in animal model sensory cortices during juvenile developmental stages. It is therefore not clear whether coordinated maturation of cortical regions and thalamic white matter pathways occurs in the human brain, with its evolutionarily expanded association cortices and uniquely protracted neurodevelopmental time course. This gap precludes a mechanistic understanding of whether age-related restructuring of thalamic connectivity could contribute to a core feature of human neurodevelopment—the existence of S–A gradients of cortical plasticity. This gap is partially attributable to the fact that prior diffusion MRI studies examining the development of thalamocortical structural connectivity have typically divided the cortex into less than ten cortical territories^{17,18,31,32,33}. While such studies have provided evidence of cortex-wide strengthening of thalamic structural connections with age^{17,18,31,32,33}, insight into whether distinct thalamocortical connections exhibit variability in maturational timing remains limited. This limited insight is due, in part, to a paucity of readily implementable diffusion tractography approaches for reliably reconstructing connections between the thalamus and localized regions of cortex.

In this study, we overcome prior methodological limitations by creating, curating and anatomically validating a comprehensive tractography atlas of human thalamic connections to over 200 regions of cortex. We then leverage this atlas to identify consistent and robust thalamocortical connections in n = 2,676 youth across three independent discovery and replication datasets. Using this rich connectivity data, we first test the hypothesis that the maturation of thalamocortical structural connectivity occurs asynchronously across cortical regions and progresses along the cortex’s S–A axis. Next, we assess whether the maturation of thalamic connectivity is temporally coordinated with developmental change in noninvasive markers of cortical plasticity. In a final set of analyses, we investigate whether thalamocortical connections may serve as a biological substrate for the embedding of environmental influences. Periods of developmentally enhanced plasticity allow neural circuits to adapt to environmental inputs; circuits that are plastic for the longest are expected to exhibit the greatest environmental adaptability and vulnerability. We therefore predicted that thalamocortical connections to association cortices that undergo protracted maturation would show evidence for enhanced sculpting by socioeconomic environmental exposures, linking the timeline of thalamic connection development to the degree of environmental sensitivity. As described below, we demonstrate that the development of thalamocortical structural connectivity is spatiotemporally synchronized with environmentally sensitive cortical developmental programs, thereby centering the thalamus in child and adolescent cortical development.

Results

We evaluated whether developmental and environmental associations with thalamocortical structural connectivity vary along the S–A axis of cortical developmental plasticity. To facilitate this evaluation, we leveraged one adult and three cross-sectional developmental datasets, advances in diffusion MRI tractography, multimodal maps of brain organization and cortical charts of in vivo plasticity marker maturation. We first used high-resolution, multishell diffusion MRI data from the Human Connectome Project (HCP) Young Adult (HCPYA) dataset (n = 1,065, ages = 22–37 years) to create an atlas composed of white matter connections between the thalamus and individual cortical regions. We then applied this atlas to delineate person-specific thalamocortical connections in three independent developmental datasets. We used the Philadelphia Neurodevelopmental Cohort (PNC; n = 1,145, ages = 8–23 years) and the HCP in Development (HCPD; n = 572, ages = 8–22 years) as primary discovery and replication datasets. PNC and HCPD are community-representative and typically developing samples that can be used to determine normative developmental trajectories. After obtaining the core set of findings in PNC and HCPD, we additionally studied the Healthy Brain Network (HBN; n = 959, ages = 8–22 years), a psychiatric sample, to assess whether findings generalized to a help-seeking population enriched for psychopathology.

To study thalamocortical connectivity development across samples, we quantified fractional anisotropy (FA) in every thalamic connection. FA is a measure that indexes the directional coherence of water diffusion and is sensitive to white matter microstructural properties, including axon packing density, orientational coherence and myelination^34,35. We elected to study FA as it has been widely applied in the child and adolescent literature, is highly sensitive to developmental change in white matter and has good test–retest reliability in youth samples³⁶. Furthermore, unlike many diffusion measures, FA can be appropriately derived from both single-shell (PNC) and multishell (HCPD and HBN) acquisitions. In what follows, we leverage diffusion MRI data from four samples to reproducibly characterize spatial gradients of thalamocortical structural connectivity features, temporal gradients of thalamocortical connectivity maturation and associations between thalamocortical connectivity and the developmental environment.

An atlas of regionally specific thalamocortical connections

Investigating whether the development of the cortex is constrained by the maturation of its structural connections with the thalamus necessitates the identification of white matter connections between the thalamus and localized cortical regions. However, reconstructing thalamocortical connections using diffusion tractography is technically challenging. It was therefore imperative to use a methodological approach that could identify accurate and consistently defined thalamocortical connections with known cortical endpoints across individuals. Our approach was executed in two key steps, including the creation of a tractography atlas of cortically annotated thalamocortical structural connections (step 1), which we subsequently leveraged as an anatomical prior to identify analogous connections in person-specific data (step 2).

Prior studies have shown that tractography atlases derived from group-average diffusion templates can capture population-representative white matter connection architecture while simultaneously enabling the delineation of homologous connections in individual diffusion images^37,38,39. The use of a group-average diffusion template to define connections of interest confers technical benefits—the template reduces noise, artifacts and modeling uncertainty, leading to high-quality tractography outputs that can be rigorously checked and curated. Subsequently, the use of a curated tractography atlas as a prior when reconstructing connections at the individual level offers additional advantages. In particular, the atlas facilitates the rejection of erroneous streamlines, increases tractography reproducibility and ensures high correspondence of identified connections across individuals.

To create a high-resolution tractography atlas of thalamocortical connections, we used a group-average diffusion template derived from the HCPYA dataset (see Methods for dataset selection rationale). The HCPYA diffusion template was used to track between the thalamus and individual ipsilateral cortical regions defined by the HCP–multimodal parcellation (MMP). Following manual inspection and curation of individual connections to ensure robustness (Methods), we identified connections between the left and right thalamus and >75% of the cortical mantle; identified connections spanned the entirety of the S–A axis (Fig. 1a; atlas connections colored by S–A axis rank). Cortical regions without identified thalamic connections had significantly smaller surface areas (anatomical enrichment test P_spin < 0.001) and significantly greater sulcal depth (anatomical enrichment test P_spin = 0.035) than the rest of the cortex. Unconnected regions thus mainly occupied small cortical territories closer to sulcal banks (Fig. 1b,c).

Fig. 1: An atlas of regionally specific thalamocortical structural connections.

a, All connections that constitute the population-level thalamocortical tractography atlas are shown in 3D views as well as in 2D slices. Connections exhibit broad cortical coverage and are colored by the cortical endpoint’s position along the S–A axis of brain organization. The color bar represents S–A axis ranks and ranges from rank 1 (sensorimotor pole) to rank 360 (association pole). b, Surface area z scores are plotted separately for cortical regions with and without connections included in the thalamocortical tractography atlas; 360 cortical regions defined by the HCP–MMP cortical parcellation were studied. c, The relative sulcal depth of cortical regions with and without connections included in the thalamocortical tractography atlas is shown based on the 360 regions that comprise the HCP–MMP parcellation. Sulcal depths ranged from 0% of maximal sulcal depth (gyral crowns) to 100% of sulcal depth (sulcal banks). In c and d, box plots summarizing data distributions are included (centerline—median; hinges—first and third quartiles; whiskers—1.5× interquartile range). d, Exemplar thalamocortical connections included in the tractography atlas are displayed. Connections are arranged from lowest (V1; yellow) to highest (TE1p; dark blue) position in the S–A axis. L, lateral; S, superior; P, posterior.

Source data

While no comprehensive and readily deployable atlas of structural connections between the thalamus and the cortex previously existed for the human brain, prior investigations have described the architecture of specific thalamocortical connections in humans and nonhuman primates. To anatomically validate the present tractography atlas, we therefore first compared individual atlas connections to those previously described in the literature. These comparisons are detailed in Extended Data Fig. 1 and the Supplementary Information. Critically, these comparisons confirm that the thalamocortical connections that comprise the current atlas have anatomical features that correspond to those identified in invasive tract tracing studies in macaques as well as noninvasive studies in humans that used advanced diffusion acquisitions and neuroanatomist-guided tracking pipelines. Using a Morel-guided parcellation of human thalamic nuclei, we additionally confirmed that the tractography atlas captures expected patterns of connectivity between individual thalamic nuclei and their main cortical projection zones (Supplementary Fig. 1). In sum, we developed a human tractography atlas of regionally specific thalamocortical connections that displays broad cortical coverage and anatomical accuracy and that can be used to guide tractography in individual brains.

Identification of thalamocortical connections in individuals

We used the thalamocortical tractography atlas as a prior with an automated tractography procedure to identify person-specific thalamic connections in PNC and HCPD participants’ data. This procedure combines fiber tracking with trajectory-based connection recognition to delineate atlas-defined thalamocortical pathways in individual brains. Critically, this approach allows for subtle person-specific differences in anatomy to be retained while ensuring group-level correspondence in connection architecture and endpoints. As highlighted in Fig. 2, this approach allowed for the reliable identification of robust white matter pathways connecting the thalamus and specific cortical regions in participants of all ages in both PNC and HCPD. To quantitatively evaluate tractography performance, we measured the overlap between participants’ thalamocortical connections and connections included in the tractography atlas by calculating overlap sensitivity (that is, the true positive rate). Thalamocortical connections exhibited very high sensitivity values in both PNC (average sensitivity = 0.77 ± 0.15) and HCPD (average sensitivity = 0.80 ± 0.14), confirming stable and accurate reconstruction of thalamocortical pathways in individuals.

Fig. 2: Thalamocortical structural connections are consistently reconstructed in individual participants.

Exemplar thalamocortical connections included in the population-level tractography atlas (rows 1 and 4) are shown here, reconstructed in individual participants from PNC (rows 2 and 5) and HCPD (rows 3 and 6). Person-specific connections were robust and exhibited excellent reconstruction accuracy. Connection colors match those used in Fig. 1 and reflect the S–A axis rank of the thalamic connection’s cortical endpoint, ranging from the sensorimotor pole of the axis (yellow) to the association pole (blue). Each connection shown in the PNC and HCPD datasets is from a different participant; a random number generator was used to select which participant’s data to show. Data are shown from study participants that span the full age range in PNC (minimum age = 8.3 years, first quartile = 11.7 years, mean = 14.6 years, third quartile = 17.5 years, maximum age = 22.0 years) and in HCPD (minimum age = 8.9 years, first quartile = 11.0 years, mean = 14.9 years, third quartile = 18.6 years, maximum age = 21.8 years).

Connections reflect thalamocortical circuit anatomy

Before using the connections reconstructed in PNC and HCPD participants to understand thalamocortical connectivity development, we aimed to further establish their anatomical validity. To accomplish this goal, we surveyed whether connections delineated in these datasets adhered to both the core–matrix (C–M) structure of thalamic organization and the hierarchical arrangement of thalamocortical connection strength. Thalamic areas can be organized along a continuous core–matrix gradient based on the relative expression of more ‘core’-like or ‘matrix’-like neurons^20,40. Although the C–M organizational schema does not capture the full diversity of thalamic cell types^41,42, it provides a parsimonious characterization of different thalamic output projections to cortex. In particular, core thalamic neurons project in a topographically restricted manner to middle layers of sensory cortices, whereas matrix neurons innervate superficial layers of association cortices; neurons with intermediate phenotypes incorporate both core and matrix properties^20,41,42. Accordingly, we hypothesized that thalamocortical connections originating in thalamic areas characterized as relatively more core, intermediate and matrix would connect to different portions of the cortex’s S–A axis. We tested this differentiation by using a previously derived C–M thalamic (C–M_t) gradient²⁰ (Fig. 3a) to assign C–M_t values to each participant’s reconstructed connections based on where connection streamlines terminated within the thalamus.

Fig. 3: Identified structural connections reflect key features of thalamocortical circuit anatomy.

a, The C–M_t gradient derived in prior work²⁰ maps spatial variation in the relative distribution of different thalamic projection cell types. This gradient, defined and visualized within the thalamus, spans from areas with the highest relative proportion of core-like thalamic projection neurons (yellow voxels; lower C–M_t values) to areas with the highest proportion of matrix-like neurons (dark purple voxels; highest C–M_t values). C–M_t values range from −0.3 to 0.2. b, Thalamocortical connections were each assigned a C–M_t value, indexing their position in the C–M gradient, based on where their streamlines terminated within the thalamus. Connection-specific C–M_t values are plotted in PNC and HCPD. C–M_t values were nearly perfectly correlated between datasets, confirming that delineated pathways terminated in the same areas of the thalamus across datasets. c, Thalamic connection C–M_t values positively correlated with the S–A axis rank of the connection’s cortical partner in both PNC (left) and HCPD (right). Both datasets showed evidence of core-to-sensory and matrix-to-association thalamocortical connectivity motifs. The positive linear relationship between S–A axis ranks and C–M_t values is plotted with a 95% confidence interval. d, Thalamocortical connections are shown colored by mean FA (dark yellow—highest FA; dark purple—lowest FA). FA values range from 0.2 to 0.5 across connections. e, FA values are plotted for all thalamocortical connections in PNC and HCPD. FA robustly correlated between datasets, demonstrating that this microstructural connectivity measure exhibits highly reproducible variability across thalamic pathways. f, Thalamocortical connection FA values monotonically decreased along the S–A axis in PNC (left) and HCPD (right), revealing a continuum of connection strength and coherence that exhibits systematic hierarchical variation. The negative linear relationship between S–A axis ranks and FA is plotted with a 95% confidence interval. In b, c, e and f, Spearman’s correlations were used to calculate r values, and the significance of the correlation was determined using conservative spin-based spatial rotation tests.

Source data

Connection-specific C–M_t values were nearly perfectly correlated between PNC and HCPD (r = 0.99, P_spin < 0.001; Fig. 3b), confirming that the atlas-constrained tractography approach generates reproducible profiles of thalamic connectivity. As predicted, the distribution of C–M_t values was not homogeneous for pathways connecting to different portions of the cortical S–A axis. Thalamic connections to the S–A axis’s sensorimotor pole originated in areas of the thalamus enriched with core neurons (lowest C–M_t values). Connections that originated in matrix-like thalamic areas (higher C–M_t values) were linked to the axis’s association pole. A distribution of increasing C–M_t values across the S–A axis was observed in both PNC (r = 0.58, P_spin = 0.008) and HCPD (r = 0.50, P_spin = 0.040) and provides evidence that reconstructed thalamocortical connections exhibit core-sensory and matrix-association connectivity motifs (Fig. 3c).

Thalamic pathways that project to sensorimotor versus association cortices also exhibit differences in their morphological properties. Thalamic inputs to primary cortex are dense and project in a spatially focal manner, whereas projections to association cortex can be sparser with more tangentially spread or diffuse innervation^20,40,42. We therefore predicted that FA would be highest for thalamic connections with primary sensory regions and decrease along the S–A axis. Connection-specific FA values, visualized in Fig. 3d, were highly reproducible between PNC and HCPD (r = 0.97, P_spin < 0.001; Fig. 3e). In line with our prediction, connection-specific FA values negatively correlated with the S–A axis rank of the connection’s cortical partner (PNC—r = −0.42, P_spin = 0.045; HCPD—r = −0.42, P_spin = 0.061; Fig. 3f). This reduction in FA is suggestive of graded changes in connection microstructure along this organizational axis, likely reflecting decreases in axonal density and myelination or increases in the dispersion of axons for thalamic connections to association cortices. Together, the current set of findings confirms that reconstructed pathways intrinsically reflect thalamic cellular classifications and cortical connection profiles.

Thalamocortical connections exhibit maturational variability

Having now demonstrated that our atlas-based approach extracts connections with properties that capture thalamocortical circuit anatomy, we sought to investigate whether these connections exhibit hierarchically organized variability in developmental timing. We began by using generalized additive models (GAMs; accounting for sex and head motion) to characterize age-dependent trajectories of FA for all connections. FA significantly increased in the majority of thalamocortical connections during childhood and adolescence, with 90% (PNC) and 78% (HCPD) of connections showing significant developmental effects (P_FDR < 0.05). Although most connections showed a general increase in FA, a spectrum of developmental trajectories could be seen in both datasets (Fig. 4a), paralleling modes of developmental variability that typify the cortex¹. As a result of these variable trajectories, the magnitude of GAM-derived age effects (quantified as the partial R²) differed across connections (Fig. 4b). Connection-wise age effects, while relatively larger in PNC, were strongly correlated between datasets (r = 0.73, P_spin < 0.001; Fig. 4c).

Fig. 4: Variability in the magnitude and timing of thalamocortical structural connectivity development.

a, A spectrum of FA developmental trajectories manifests across thalamocortical connections in PNC (top) and HCPD (bottom). Developmental trajectories are GAM smooth estimates that are independently zero-centered along the y axis for each connection. Trajectories are colored by the connection’s age effect (partial R²) and represent the partial effect of age on FA conditional on model covariates. b, Thalamocortical connections from the tractography atlas are colored by their age effect, revealing the brain-wide distribution of developmental heterogeneity. c, A correlation plot confirming close correspondence between connection-specific age effects derived across datasets is shown; the linear fit (with 95% confidence interval) is plotted. The result of Spearman’s correlation with a spin test is indicated. d, Developmental trajectories for thalamic connections to primary motor, lateral parietal and superior prefrontal cortical regions are shown, overlaid on participants’ data, in PNC (top) and HCPD (bottom). Trajectories represent GAM-predicted FA values with a 95% credible interval. Corresponding color bars chart the rate of increase in FA during windows of significant developmental change. e, Connection-wise age effects and ages of maturation were correlated (PNC data and linear fit shown with a 95% confidence interval). The result of Spearman’s correlation with a spin test is indicated and confirms that the age of maturation metric provides insight into both the magnitude and timing of development. f, A brain map localizing cortical regions with the earliest-maturing thalamic connections (age of maturation first quartile; yellow) and latest-maturing thalamic connections (fourth quartile; blue) is shown. White and gray regions matured in middle age quantiles and were not included in the atlas, respectively. g, Results of a Neurosynth analysis used to functionally decode differences in thalamocortical connectivity maturational timing. Psychological terms linked to cortical regions with early-maturing thalamic connections are shown in yellow. Terms associated with cortical regions with the late-maturing thalamic connections are shown in blue. PNC data are presented; 11 terms that were also present in the HCPD Neurosynth analysis are written in bold font. L, lateral; M, medial; A, anterior; P, posterior.

Source data

Connections between the thalamus and sensorimotor cortical regions tended to show the smallest age effects and to exhibit the shortest windows of developmental change, as exemplified by the connection to primary motor area 4 (Fig. 4d, yellow). Thalamic connections with multimodal association cortices (for example, area PF of the supramarginal gyrus; Fig. 4d, purple) tended to develop for relatively longer. Thalamocortical connections exhibiting the largest and most protracted developmental increases involved regions of the lateral prefrontal and parietal association cortex, as seen for the thalamic connection to superior prefrontal region 8BL (Fig. 4d, blue). Developmental trajectories did not significantly differ by sex (0% and <2% of connections showed significant age-by-sex interactions in PNC and HCPD, respectively).

Next, to quantitatively study differences in developmental timing, we computed the age at which FA in each thalamocortical connection matured. Maturational age was operationalized as the age at which developmental change in FA (the first derivative of the age spline) was no longer significantly different from 0, denoting a plateau in the developmental trajectory. Connection-specific ages of maturation were highly similar between the two datasets (r = 0.57, P_spin < 0.001) and furthermore correlated with connection-specific age effects (partial R²) within each dataset (Fig. 4e). Critically, large differences in maturational timing emerged across thalamocortical connections—the difference in maturational age between the earliest- and latest-maturing connections was 10 years in PNC and 11 years in HCPD. Identifying thalamic connections that matured at the youngest ages (first quartile; Fig. 4f, yellow) versus the oldest ages (fourth quartile; Fig. 4f, blue) differentiated primary and early visual, somatomotor and auditory regions from lateral prefrontal and parietal regions.

We sought to put observed differences in thalamic connection maturational timing in a behavioral context. To do so, we used prior task-based functional MRI (fMRI) results—amassed and meta-analyzed via Neurosynth—to identify the core psychological functions subserved by cortical regions with thalamic connections that matured at younger versus older ages. We computed correlations between a cortical map indexing the age of thalamic connection maturation and Neurosynth meta-analytic maps derived for over 100 psychological terms. In this analysis, a negative correlation between the maturational map and a psychological term map indicates that the psychological function engages cortical regions with early-maturing thalamic connections. In contrast, positive correlations identify psychological functions that are subserved by cortical regions with late-maturing connections. We identified the 10 most negatively and positively correlated terms in both PNC and HCPD and found that 11 of these 20 terms overlapped between datasets (bolded terms in Fig. 4g; term-overlap P_spin = 0.004). Psychological terms linked to cortical regions with early-maturing thalamic connections predominantly described sensory and motor processing and object classification functions (Fig. 4g, negative correlations). Psychological terms associated with cortical regions with late-maturing thalamic connections involved executive control, decision-making and information retrieval functions (Fig. 4g, positive correlations). Overall, these results establish that thalamocortical structural connections exhibit different timescales of development. Maturational timing diverges most between connections to sensorimotor cortices that execute externally oriented functions and those to association cortices that are essential for higher-order cognitive control.

Thalamocortical connections mature along the S–A axis

If connections between the thalamus and the cortex help to organize differences in developmental timing across the S–A axis, observed variability in thalamocortical connection maturation should systematically align with this axis. To empirically study this alignment, we first demarcated age windows of significant developmental change for each connection. As shown in Fig. 5a, connection-specific windows of developmental change were staggered in time across the S–A axis and most protracted for connections to the axis’s association pole. We next calculated the correlation between each connection’s age of maturation and its S–A axis rank (Fig. 5b,c). Lending strong support to our primary developmental hypothesis, we found that the age of connection maturation got progressively older for thalamic connections to cortices ranked higher in the S–A axis. The correlation between S–A axis ranks and ages of maturation was similar in strength in PNC (r = 0.49, P_spin-FDR = 0.004) and HCPD (r = 0.51, P_spin-FDR = 0.008), underscoring that this spatiotemporal developmental pattern unfolds similarly in independent samples (Fig. 5d).

Fig. 5: Thalamocortical structural connections mature at progressively older ages along the S–A axis.

a, Developmental increases in FA persist for progressively longer for thalamic connections to cortical regions ranked higher in the S–A axis. Age windows during which thalamocortical connections showed significant increases in FA are highlighted for every connection. Connections are ordered along the y axis and colored by the S–A axis rank of the connection’s cortical endpoint. Windows of significant developmental change demarcate ages where the first derivative of the GAM smooth function for age was significantly greater than 0, as determined by a simultaneous 95% confidence interval. b, The S–A axis (left) exhibits shared spatial topography with a maturational map depicting the age at which each cortical region’s thalamic connection matured (right; PNC data). Light gray regions in these cortical maps were not analyzed or did not show significant developmental effects. c, Thalamocortical connections from the tractography atlas are colored by the connection’s age of maturation to further illustrate connection-wise differences in maturational timing. d, Ages of thalamic connection maturation systematically vary along the S–A axis in both PNC (left) and HCPD (right). Thalamocortical connections to the axis’s association pole tended to mature at the oldest ages. The linear relationship between S–A axis rank and maturational age is shown with a 95% confidence interval in both plots. The r value from Spearman’s correlation and the significance of the correlation (P_spin) as determined by a spin-based spatial rotation test are additionally provided for both plots. Significance values (P_spin) were corrected across all spatial axis correlations using the FDR correction. e, Results of the PNC analysis quantifying the alignment of thalamic connection maturational timing to the S–A axis as well as major cortical and thalamic axes. The correlation with maturational timing was stronger for the S–A axis than for A–P, D–V and M–L cortical axes, as well as the C–M_t thalamic gradient. Only the correlation with the S–A axis was significant, as determined by spatial null model testing. * indicates a P_spin-FDR value < 0.05 for the correlation between connection maturational timing and a spatial axis.

Source data

We conducted developmental sensitivity analyses to ensure that differences in thalamic connection maturational timing along the S–A axis were not driven by potential anatomical and methodological confounds. We considered the potential impact of (1) cortical region anatomy indexed by regional surface area, (2) diffusion data quality measured by temporal signal-to-noise ratio (tSNR), and (3) tractography reconstruction accuracy determined by atlas overlap sensitivity (true positive rate). We first tested whether connection-wise differences in these three variables (endpoint region area, tSNR and overlap sensitivity) were correlated with connection maturational timing at the group level. Correlations were not significant in PNC or HCPD (all P_spin > 0.05), confirming that these variables did not account for the observed maturational heterochronicity. We next refit developmental GAMs while covarying for these three variables at the individual level. When controlling models for these variables, maturational increases in FA remained significant and were robustly patterned along the S–A axis in both PNC and HCPD (Supplementary Fig. 2).

Finally, we conducted developmental specificity analyses wherein we tested whether across-connection differences in maturational timing were best captured by the S–A axis or could be equally or better explained by other large-scale cortical or thalamic axes. In particular, we assessed how connectivity maturational timing varied along anterior–posterior (A–P), dorsal–ventral (D–V) and medial–lateral (M–L) cortical axes and the C–M_t gradient of the thalamus (Fig. 5e). Correlations between the age of thalamocortical connection maturation and A–P (r = 0.31), D–V (r = 0.12) and M–L (r = 0.03) cortical axes and the C–M_t gradient (r = 0.29) were not significant in PNC (all P_spin-FDR > 0.05). Furthermore, statistical tests for comparing the magnitude of two dependent, overlapping correlations indicated that connection maturational ages were significantly more correlated with the S–A axis than with A–P, D–V and M–L axes and the C–M_t gradient (P_FDR < 0.001 for all four correlation comparisons conducted using a backtransformed averaged Fisher’s z⁴³). We observed the same pattern of results in HCPD, with the strongest alignment to the S–A axis as compared to all other cortical and thalamic neuroaxes (correlation comparison for A–P, P_FDR = 0.124; D–V, P_FDR < 0.001; M–L, P_FDR < 0.001; C–M_t, P_FDR = 0.029).

Thalamocortical connection maturation and cortical plasticity

The prior results indirectly relate thalamocortical connectivity maturation to temporal variation in the progression of developmental plasticity along the cortex’s S–A axis. We therefore next endeavored to more directly assess the correspondence between the maturation of thalamic connections and cortical regions by leveraging three noninvasive and biologically linked readouts of cortical developmental plasticity. Animal studies have shown that the maturation of PV inhibitory interneurons^29,30 and the formation of intracortical myelin⁴⁴ serve as biological regulators of critical periods of plasticity. As interneurons strengthen their outputs and myelin is formed, there is a reduction in the cortex’s E/I ratio and a consequent suppression of high-amplitude, intrinsic fluctuations in cortical activity^2,45,46. The transition from plastic to mature cortex can therefore be indexed by the following three signatures of decreasing circuit plasticity: a decline in the E/I ratio, an increase in cortical myelin content and a reduction in the amplitude of intrinsic fluctuations in cortical activity. We explored whether the timing of thalamocortical connection maturation was temporally coordinated with the sensorimotor-to-associative development of these three readouts of declining cortical plasticity.

We used developmental maps charting how in vivo measures sensitive to the cortical E/I ratio⁴, cortical myelin content³ and the amplitude of intrinsic fluctuations in cortical activity² change with age during childhood and adolescence. We discovered that cortical regions with thalamic connections that develop for longer also exhibit smaller developmental declines in the E/I ratio derived from computational modeling (Fig. 6a), experience a slower rate of intracortical myelin growth as indexed by the T1/T2 ratio (Fig. 6b) and show an initial decrease in the amplitude of intrinsic blood oxygen level-dependent (BOLD) activity at older ages (Fig. 6c). As such, protracted maturation of thalamocortical connections was associated with extended expression of neurochemical, structural and functional markers indicative of higher circuit plasticity. Relationships between thalamocortical connectivity maturational timing and developmental refinements in the cortical E/I ratio, cortical myelin and cortical activity amplitude were significant in all cases in both datasets: PNC (model-derived E/I ratio—r = 0.45, P_spin-FDR < 0.001; T1/T2 ratio—r = −0.45, P_spin-FDR = 0.001; BOLD amplitude—r = 0.30, P_spin-FDR = 0.030) and HCPD (model-derived E/I ratio—r = 0.45, P_spin-FDR < 0.001; T1/T2 ratio—r = −0.43, P_spin-FDR = 0.003; BOLD amplitude—r = 0.41, P_spin-FDR = 0.009; Fig. 6). These relationships provide evidence for coordinated development between thalamocortical structural connectivity and cortical plasticity.

Fig. 6: Thalamocortical structural connection maturation synchronizes with timescales of cortical plasticity.

Developmental refinements in cortical plasticity markers are coordinated with thalamocortical connection maturation in PNC (left column) and HCPD (right column). Cortical maps charting the in vivo development of the E/I ratio (a; magnitude of maturational decline), cortical myelin (b; rate of developmental growth) and intrinsic activity amplitude (c; age of decrease onset) are shown. In all three cortical maps, darkest blue brain regions are those that express signatures of protracted developmental plasticity. a, Cortical regions with thalamic connections that mature at older ages undergo smaller age-related declines in the E/I ratio during childhood and adolescence (less negative age slopes), implying that they remain in a less mature, plasticity-permissive state for longer. E/I ratio was estimated in developmental data in ref. ⁴ by applying a biophysically realistic computational circuit model to resting-state fMRI data. b, Cortical regions with thalamic connections that mature at older ages show a slower rate of T1/T2 ratio-indexed cortical myelin growth, suggesting that they exhibit slower formation of a structural feature that restricts developmental plasticity. T1/T2 ratio development data is from ref. ³. c, Cortical regions with thalamic connections that mature at older ages exhibit later-onset declines in the amplitude of intrinsic activity fluctuations, indicative of temporally delayed reductions in a potential functional signature of developmental plasticity. The age at which intrinsic activity amplitude began to decrease in each cortical region was determined by ref. ² through developmental modeling of age-related changes in BOLD fluctuation amplitude. In a–c, correlation plots include a linear fit with a 95% confidence interval band and provide the r value from Spearman’s correlation and the P_spin value from a spin test; P_spin values were FDR corrected across plasticity map correlations.

Source data

Thalamocortical connections exhibit environmental sensitivity

Ample prior work has documented associations between socioeconomic features of the environment and cortical properties during youth. During childhood and adolescence, across-cortex variation in environment associations has been found to be systematically patterned along the S–A axis^2,47. In a series of analyses, we aimed to ascertain whether a similar principle governs interactions between youths’ environmental exposures and thalamocortical structural connectivity. We studied relationships between thalamocortical connection FA and household and neighborhood socioeconomic indicators using GAMs that accounted for developmental effects. Household socioeconomic status was proxied by caregiver years of education in PNC and by both caregiver education and income-to-needs ratio in HCPD. Neighborhood-level socioeconomic information was only available in PNC and was summarized via factor analysis of geocoded census data; higher factor scores reflect greater socioeconomic advantage.

Over half (57%) of thalamocortical connections showed significant relationships between connection FA and neighborhood environment factor scores (P_FDR < 0.05) in PNC. In contrast to these robust effects, only 7% of connections showed a significant association (P_FDR < 0.05) with caregiver education in PNC, despite the fact that caregiver education and factor scores were correlated (r = 0.57). When caregiver education and neighborhood factor scores were entered into the same model as part of a specificity analysis, all significant caregiver education effects were abolished. Conversely, 91% of the significant neighborhood-level effects in the original analysis remained significant. Mirroring these null household-level findings in PNC, no thalamocortical connections showed a significant relationship between FA and either caregiver education or income-to-needs ratio in HCPD (all P_FDR > 0.05).

In the PNC, significant relationships between FA and the neighborhood environment factor score were widely distributed across thalamocortical connections and across the S–A axis (Fig. 7a). Statistical environmental effects (t values) were nearly exclusively positive (93% positive), indicating that more socioeconomically advantaged neighborhoods were associated with higher connection FA. To better understand the nature of these FA–environment associations, we modeled the maturation of thalamocortical connection FA for low and high factor scores for five quintiles of the S–A axis. Environmentally stratified developmental trajectories showed that lower neighborhood factor scores were associated with lower FA throughout the course of child and adolescent development (Fig. 7b).

Fig. 7: Hierarchically organized relationships between the neighborhood environment and thalamocortical structural connectivity.

a, Thalamocortical connections that exhibited a significant, positive association between neighborhood socioeconomic advantage and connection FA are presented in purple. The cortical endpoints of these connections are also shown, colored by the statistical t value of the environment effect. The largest t values localized to the lateral frontal and temporal cortices. White and gray denote cortical regions that had nonsignificant environmental effects and that were not included in the tractography atlas, respectively. b, GAM-predicted trajectories of FA development are displayed for low (10th percentile) and high (90th percentile) neighborhood factor scores for thalamic connections to five quintiles of the S–A axis. Trajectories reveal that environment-related differences in connection FA persist from childhood to early adulthood. c, Results of the environment effect enrichment analysis are displayed for five quintiles of the S–A axis. Each plot shows the mean empirical t value in that quintile along with a null distribution of t values obtained from spin-based spatial null models (t_spin). The enrichment test was significant in the fifth quintile of the S–A axis, indicating that environmental effects were significantly stronger for thalamic connections to the association pole of the axis as compared to the rest of the cortex. d, Environment effect t values significantly increased in magnitude for thalamic connections to cortical regions ranked higher in the S–A axis, as shown by the positive linear fit between variables (with a 95% confidence interval) and the result of Spearman’s correlation with a spin test. Nine thalamic connections with a negative t value are not shown. e, Results of an analysis correlating connection-wise environment effects with the S–A axis as well as with D–V, M–L and A–P cortical axes and the C–M_t gradient. Only the correlation with the S–A axis was significant, as determined by spatial null model testing. * indicates a P_spin-FDR value < 0.05 for the correlation between connection environment effects and a spatial axis. 

Source data

Although neighborhood environment associations were widespread, the magnitude of significant effects was not homogeneous across connections—larger effects could be seen in thalamic connections to lateral prefrontal and lateral temporal cortices (Fig. 7a, bottom). We therefore conducted an analysis to test whether environment effects (t values) were statistically enriched for connections to the association end of the S–A axis. Enrichment tests for five quintiles of the S–A axis uncovered that the strongest neighborhood environment effects (highest t values) were indeed overrepresented for thalamic connections with the association pole of the S–A axis (fifth quintile enrichment analysis—P_spin = 0.025; Fig. 7c). Sensitivity analyses that partitioned the S–A axis into quartiles and deciles for enrichment testing confirmed that significantly larger environmental effects were always present in the most associative portion of the S–A axis (fourth quartile enrichment—P_spin = 0.032; tenth decile enrichment—P_spin = 0.009). A complementary analysis correlating S–A axis ranks with neighborhood environment effects across connections confirmed a significant, positive association (r = 0.30, P_spin-FDR = 0.028; Fig. 7d). Specificity analyses revealed that the correlation with the S–A axis was significantly stronger than with the A–P axis (correlation comparison P_FDR = 0.001) and the C–M_t thalamic gradient (correlation comparison P_FDR < 0.001). Furthermore, correlations between environment effect t values and A–P, M–L and D–V cortical axes and the C–M_t gradient were all nonsignificant (all P_spin-FDR > 0.05; Fig. 7e). These analyses demonstrate that neighborhood-level socioeconomic conditions relate to thalamocortical connectivity properties during youth, with connections that experience protracted development displaying the greatest environmental sensitivity.

Results generalize to a sample enriched for psychopathology

Thus far, we have demonstrated that developmental and environmental influences on thalamocortical connection properties vary depending on a connection’s position in the cortex’s S–A axis. In a final set of analyses, we investigated whether these findings generalize to the HBN, a clinical sample of youth with transdiagnostic forms of psychopathology. In contrast to the PNC and HCPD samples that were not specifically enriched for psychiatric symptomatology, HBN is a sample of help-seeking youth where approximately 85% of individuals meet criteria for a clinical diagnosis. In HBN, 74% of thalamocortical structural connections showed a significant developmental change in FA (P_FDR < 0.05), with developmental profiles substantially varying across connections (Fig. 8a). Connection-specific age effects obtained in HBN strongly and significantly correlated with those obtained in PNC (r = 0.73, P_spin < 0.001; Fig. 8b), revealing a generalizable developmental signature. Neurosynth-based decoding of connection maturational timing linked early-maturing thalamocortical connections to perceptual and motor functions and late-maturing connections to memory retrieval, decision-making and cognitive control (Fig. 8c). Thalamocortical connections exhibited a hierarchical maturational gradient—connection maturational age was significantly correlated with the S–A axis (r = 0.69, P_spin = 0.002; Fig. 8d). Furthermore, ages of connection maturation aligned with age-related change in the three neuroimaging-based readouts of cortical developmental plasticity (E/I ratio—r = 0.57, P_spin-FDR = 0.007; T1/T2 ratio—r = −0.58, P_spin-FDR = 0.015; BOLD fluctuation amplitude—r = 0.69, P_spin-FDR = 0.007; Fig. 8e).

Fig. 8: Developmental and environmental results are generalizable to youth with psychopathology.

An overview of key results from the HBN sample, a clinical sample of youth that is enriched for psychopathology. a, Structural connections between the thalamus and cortex exhibit heterogenous profiles of FA development. b, Connection-specific age effects derived in HBN correlate with those obtained in PNC; the linear relationship between dataset-specific age effects is plotted with a 95% confidence interval. c, Neurosynth-based contextualization of thalamocortical connection developmental timing reveals psychological functions associated with cortical regions with early-maturing thalamic connections (yellow; negatively correlated terms) and late-maturing thalamic connections (blue; positively correlated terms). d, The age at which thalamocortical connections mature progressively increased for connections to cortical regions located higher in the S–A axis, resulting in a positive correlation between ages of thalamic pathway maturation and S–A axis ranks. The linear relationship between these variables is shown with a 95% confidence interval. e, Thalamocortical connection maturation significantly correlated with noninvasively derived maps charting the development of cortical properties, including the development of the cortical E/I ratio, cortical T1/T2 ratio and cortical BOLD activity fluctuation amplitude. The strength and significance of each of these three correlations are indicated. f, A plot depicting the significant correlation between thalamic connection environment effects (statistical t values) and the S–A axis is shown (linear fit with a 95% confidence interval). Positive environment effects indicate that more socioeconomically advantaged neighborhood environments were associated with higher thalamocortical connection FA. Environmental effects strengthened toward the axis’s association pole. g, The environment enrichment analysis confirmed that neighborhood environment effects were significantly greater in magnitude for thalamic connections to the fifth quintile of the S–A axis as compared to connections with the rest of the cortex. * indicates a quintile enrichment P_spin value < 0.05. In b, d, e and f, the r values and P_spin values from independent Spearman’s correlations with spin tests are denoted.

Source data

We used the same geocoding-based factor analysis approach as in PNC to summarize information about each participant’s neighborhood socioeconomic circumstances in HBN. In HBN, 53% of thalamocortical connections exhibited a significant relationship between neighborhood environment factor scores and connection FA. All associations were positive, linking more advantaged neighborhoods (higher factor scores) to stronger thalamocortical connectivity. The magnitude of environment-connectivity associations increased in strength along the S–A axis (r = 0.31, P_spin = 0.035; Fig. 8f) and the largest effects were enriched at the S–A axis’s association pole (fifth quintile enrichment analysis—P_spin = 0.031; Fig. 8g). Together, these findings highlight the generalizability of developmental and environmental results linking thalamocortical structural connectivity to the S–A axis.

Discussion

During early stages of cortical neurodevelopment, thalamocortical axons exert powerful control over the arealization, lamination and neurobiological specification of cortical areas^10,11,12,14. In this work, we tested the hypothesis that the thalamus is also linked to child and adolescent windows of cortical plasticity and provided evidence of synchronized maturation between the cortex and thalamocortical connectivity. To overcome known obstacles associated with thalamic tractography, we first curated and anatomically validated a high-resolution diffusion atlas composed of cortically annotated thalamocortical structural connections. Applying this atlas to data from three youth cohorts, we reproducibly demonstrated that the development of thalamocortical connectivity is globally heterogeneous, temporally hierarchical and spatially aligned with variability in cortical developmental profiles. Specifically, we showed that the maturation of thalamocortical structural pathways progresses along the S–A axis and parallels the development of three noninvasive readouts of cortical developmental plasticity. In this maturational scheme, thalamocortical pathways that mature latest connect to transmodal association regions that are integral to higher-order cognition and that show signatures indicative of protracted circuit malleability. We furthermore defined relationships between thalamocortical connectivity and neighborhood environment conditions that increase in strength along the S–A axis, thus aligning with the dominant mode of brain–environment associations observed in the cortex during youth². Together, these results uncover that thalamocortical connections develop in a hierarchical and environment-dependent manner across the cortex, consistent with a role for the thalamus in organizing the S–A axis of child and adolescent cortical plasticity.

Mechanistic studies in animal models have shown that the thalamus regulates the temporal unfolding of developmental processes throughout the span of cortical ontogeny. In early embryonic development, thalamic axons influence the speed of cortical progenitor cell proliferation by releasing a diffusible factor that affects cell cycle kinetics⁴⁸. As development progresses, the rate at which thalamocortical axons grow determines the temporal emergence of regionally specific cortical maps (for example, somatotopic maps)⁴⁹. During early postnatal development, experience-dependent transfer of homeoproteins from the thalamus to cortical PV interneurons impacts the timing of sensory cortex critical periods⁵⁰. As maturation decelerates, the stabilization of thalamocortical synapses onto PV interneurons helps to terminate periods of developmental plasticity^21,27. Animal studies thus point to the thalamus as a timekeeper of cortical neurodevelopment. In the present work, we extend this line of research to humans by linking the timing of thalamocortical connection maturation to the hierarchical progression of developmental processes along the human cortex.

Prior diffusion imaging studies in humans have provided data consistent with a cortex-wide strengthening of thalamocortical connectivity with age^{17,18,31,32,33}. Yet, an understanding of cortical variability in thalamocortical connection maturation has remained elusive, in part due to the challenge of reliably delineating thalamic connections to localized regions of cortex. To address this challenge, we introduced a tractography atlas of regionally specific thalamocortical structural connections. We confirmed that the atlas reflects thalamic cellular classifications and connectivity motifs and can be applied to person-specific diffusion data acquired with different shells and sampling schemes. By using the tractography atlas in three independent developmental datasets, we uncovered a previously uncharted S–A axis of thalamocortical connection maturation that generalizes across typically developing and psychiatric samples. Furthermore, we showed that this maturational axis aligns with developmental refinements in the cortical E/I ratio, cortical myelin and intrinsic cortical activity, suggesting that thalamic innervation may influence the cascading expression of multiple cortical mediators of plasticity. These brain-wide correlations yield new mechanistic hypotheses as to how thalamocortical connectivity could both promote and restrict plasticity within the human cortex. In particular, during earlier stages of child and adolescent development, the strengthening of thalamic inputs onto both cortical pyramidal and PV cells may alter the local E/I ratio in a manner that enhances synchronous cortical activity and facilitates activity-dependent plasticity^21,24. Subsequently, the stabilization of thalamic inputs onto PV interneurons may trigger the consolidation of inhibitory circuitry by myelin⁵¹ and a consequent reduction of intrinsic activity and ongoing cortical plasticity.

The thalamus’s central position in global modes of brain connectivity and communication makes it well-suited to be a subcortical regulator of environment-driven cortical developmental plasticity. The thalamus is richly interconnected with a diverse mosaic of cortical regions as well as with peripheral sensory systems that process the external world, allowing it to serve as a nexus that can link evolving developmental schedules to physical, cognitive and social environmental demands. The thalamus has also been identified as a central ‘gate’ that supports indirect cortico–cortical activity propagation, specifically gating information transfer up hierarchically organized processing streams^19,52,53. Speculatively, transthalamic cortico–cortical interactions could serve to communicate information about a cortical region’s developmental state up the S–A maturational hierarchy. Communication could be achieved through cortico–thalamo–cortical activity transfer, given that cortical activity is refined in a stereotyped manner during development. Early in development, exposure to new environmental inputs (light, sound and touch) elicits a robust increase in excitatory activity in primary cortices that is relayed through thalamic axons and that initiates activity-dependent cortical plasticity. As primary sensory and motor circuits mature, there is a reduction and desynchronization of their intrinsically generated circuit activity^2,45,46. This well-characterized refinement of intrinsic activity that occurs during cortical maturation may alter functional signals relayed back to the thalamus via cortical–thalamic projections. This feedback, in turn, could lead to a consequent shift in thalamic inputs to higher-order cortex that initiates activity-dependent plasticity at the next stage of the processing hierarchy. In this system, cortical activity motifs communicated to the thalamus would provide a biological readout of maturational status, and the thalamus serves as a gate that propagates developmental sequences up the S–A axis.

A major goal of developmentally programmed plasticity is to allow neural circuits to adapt to environmental exposures. Here we identified associations between thalamic connectivity and variability in youths’ socioeconomic environments; environmental associations were significant at the neighborhood, but not household, level. A range of prior studies have identified differential associations between brain properties in youth and environmental exposures that operate at distinct levels of human ecology. In accordance with the present findings, many of these studies have reported that youth brain structure, activity and connectivity are more strongly related to neighborhood than to household socioeconomic indicators^2,54,55,56, underscoring the importance of enriching the neighborhood context for population-level youth health. This dissociation could indicate that neighborhood-level measures of the socioeconomic environment are more comprehensive or reliable indicators than self-reported education and income. Alternatively, this dissociation may reflect the tendency for neighborhood and household socioeconomic advantage to map onto distinct aspects of youths’ environmental exposures, resulting in distinct relationships with neural properties.

Notably, associations between thalamocortical connectivity and the neighborhood environment were not homogeneous across thalamic pathways. Rather, statistical estimates of environment effects were relatively larger for thalamocortical connections that develop for longer, consistent with an accumulation of environmental influences enabled by protracted developmental malleability. As a result, environmental associations increased in strength for thalamic connections to regions ranked highest in the S–A axis, mirroring the S–A patterning of associations between neighborhood disadvantage and regional intrinsic activity previously described in the cortex². The similarly patterned expression of environmental influences on thalamocortical and cortical properties indicates that thalamic signals help to promote environment-dependent sculpting of the cortex in youth. More broadly, these findings intimate that environmental impacts on the brain continue to accrue throughout development. They therefore suggest that interventions aimed at mitigating exposures to disadvantaged environments will still be effective at promoting healthy neurodevelopment during adolescence, given continued plasticity in association circuitry.

The present in vivo analysis of thalamocortical structural connectivity has several important limitations. First, white matter pathways identified with diffusion tractography are nondirectional; thus, the connections studied here inherently contain inputs from thalamus to cortex as well as from cortex to thalamus. Second, diffusion MRI and FA do not directly measure axonal pathways but rather aggregate directions of local water diffusion within a voxel. A portion of the signal ascribed to thalamocortical structural connections will therefore be influenced by diffusion induced by nearby connections in the same voxel, for example, by crossing axons in deep white matter or U-fibers in superficial white matter. Causal investigations in animal models will be needed to cellularly isolate thalamic inputs to cortex and study their impact on the timing of cortical development. Third, we aimed to study relationships between age-related change in thalamocortical connections and cortical properties by comparing group-level developmental hallmarks. Future work delineating within-individual longitudinal relationships between the development of thalamocortical connections and cortical plasticity will help to probe these relationships and their temporal precedence at the individual level. Fourth, we studied associations between thalamocortical connectivity and factor scores that index wealth, employment and educational inequality across geographic neighborhoods. Given that neighborhood socioeconomic privilege is linked to variation in many different environmental contexts and exposures (for example, access to material resources and healthcare, school and safety conditions, pollution and toxin exposure and sociocultural institutions), this approach precludes inference regarding which environmental factors most contribute to the associations with thalamic connectivity observed here. Additional studies designed to parse which facets of the environment sculpt brain development and plasticity will be important for informing environmental policies that support youth living across socioeconomic circumstances.

The maturation of the human cerebral cortex follows defined spatiotemporal sequences during its prolonged neurodevelopmental course. The current study bridges results from animal systems and approaches to generalizable developmental neuroimaging to provide evidence linking the thalamus to the dominant sequence of child and adolescent cortical development. It furthermore identifies thalamocortical connectivity as an environmentally responsive biological system that could influence individual differences in the expression of developmental plasticity. Further insight into how thalamocortical connectivity affects individualized trajectories of cortical plasticity could ultimately help to inform environmental interventions for youth that promote their capacity for adaptive cortical malleability.

Methods

Creation of a diffusion atlas of thalamocortical connections

Our analytic approach began with the creation of a population-level tractography atlas comprising spatially specific connections between the thalamus and ipsilateral regions of cortex. This population-level atlas was instrumental for the subsequent reliable delineation of regionally specific thalamocortical pathways in data from individual participants. To create the thalamocortical tractography atlas, we used a publicly available population-average diffusion MRI template that was constructed using data from the HCPYA cohort⁵⁷ (1,200-participant release, n = 1,065, ages = 22–37 years, 575 females). We primarily elected to use HCPYA diffusion data for the tractography atlas given its high-quality nature—55 min of diffusion data were acquired in three shells with dense directional sampling and a high spatial resolution. The diffusion data were furthermore acquired on scanners specifically designed to increase the maximum gradient amplitude to enhance the SNR of diffusion imaging⁵⁸. These acquisition features make these data well-suited to characterize diffusion signal in subcortical structures and capture detailed thalamocortical connection anatomy. The availability of these data from over 1,000 participants additionally ensures that derived connections reflect anatomical features present across many individuals. Notably, the use of a young adult sample for constructing the atlas is apposite given that entire thalamocortical connections are not expected to form or be pruned away following infancy; rather, only changes in microstructural indices of connectivity strength are expected. In the context of the present study, the HCPYA diffusion atlas could therefore be used as a reliable prior for delineating thalamic connectivity in youth while preventing findings from being biased to any single developmental sample. Beyond the current work, the use of a young adult sample makes future application of this tractography atlas appropriate for many lifespan investigations.

The construction of the HCPYA diffusion template has been described in detail^37,59. Briefly, high-resolution, multishell diffusion MRI scans were acquired from participants with 1.25 mm isotropic voxels at b values of 1,000, 2,000 and 3,000 s/mm² (90 directions per shell, each acquired in reverse phase-encoding directions). Preprocessed data were reconstructed using q-space diffeomorphic reconstruction (QSDR)⁶⁰, which performs generalized q-sampling imaging (GQI) in MNI ICBM152 2009a space. GQI is a model-free diffusion MRI reconstruction approach that estimates diffusion orientation distribution functions (ODFs) directly from the diffusion signal (the distribution of diffusion spins) to resolve complex fiber orientations⁶¹. GQI-based ODFs can be estimated in single-shell and multishell diffusion sampling schemes. QSDR outputs were aggregated across the 1,065 participants to build a population-averaged template of voxel-wise diffusion distributions. We used this QSDR diffusion template in DSI Studio for the construction of the thalamocortical connectivity atlas.

We first used DSI Studio to perform deterministic fiber tracking on the QSDR diffusion template to identify 2 million streamlines terminating in the left thalamus and 2 million streamlines terminating in the right thalamus. The left and right thalamus were identified in the QSDR template using FreeSurfer’s subcortical segmentation, which comes built into DSI Studio. Contralateral white matter masks were used as regions of avoidance during hemisphere-specific thalamic tracking to only extract streamlines between the thalamus and ipsilateral brain regions. Deterministic tractography in DSI Studio uses voxel ODFs and quantitative anisotropy measures to resolve crossing fibers, reduce partial volume effects, filter noisy fibers and define better tracking termination locations⁶². Deterministic tracking offers advantages for identifying cortically constrained thalamocortical connections with high validity and high termination specificity (as compared to probabilistic tracking approaches that typically achieve broader coverage of connection profiles with a trade-off of more false positives and lower endpoint specificity^63,64). Reflecting this advantage, in an international tractography challenge, the deterministic tractography approach implemented here reconstructed 92% validly connecting streamlines and additionally identified the lowest number of invalid white matter bundles⁶⁵.

The following parameters were used for thalamic tractography, which were chosen following parameter testing: threshold index = qa, otsu threshold = 0.45, minimum streamline length = 10 mm and maximum streamline length = 300 mm. In addition to these stable parameters, random parameter saturation was used to select the anisotropy threshold, turning angle, step size and smoothing level for each generated streamline. Random parameter saturation entails using a random combination of the aforementioned tracking parameters (each within a predefined, set range of appropriate values) to generate a broad array of viable streamlines. By sampling across the entire appropriate parameter space rather than arbitrarily selecting a single value in the space, this approach can resolve streamlines with varied properties and enhance both reconstruction accuracy and computational reproducibility⁶⁶. Following the identification of streamlines with endpoints in the thalamus, we extracted ipsilateral connections between the thalamus and individual cortical regions defined by the HCP–MMP⁶⁷, which was included with DSI Studio. These connections served as the basis of the thalamocortical tractography atlas.

All regionally specific thalamocortical connections extracted by this analytic pipeline underwent a process of visual quality assurance and manual curation before their inclusion in the final atlas. The atlas curation procedure was performed by a single author over several months following technical training by diffusion tractography experts and an in-depth review of the literature on macroscale thalamocortical projection anatomy. During the curation process, the anatomy of extracted connections was compared to prior thalamocortical tractography results and tract tracing accounts, when available, to confirm anatomical accuracy. All connections were additionally subject to potential manual editing to delete false-positive streamlines with biologically impractical architectures, as well as superfluous streamlines that diverged from the core connection. The manual checking process was also used to entirely remove n = 16 thalamocortical connections from the final atlas that were deemed overly sparse based on the template tracking or that could not be reconstructed in participant-level data in both the PNC and HCPD datasets. After curation of extracted thalamocortical connections, a subset of finalized connections was visually reviewed by researchers familiar with thalamocortical connectivity. Skeletonized versions of the final set of connections were then generated by deleting ‘repeat’ streamlines with redundant trajectories, operationalized here as streamlines within a distance of <3 voxels. Removal of redundant streamlines enhances the computational efficiency of the subsequent automated tracking procedure without compromising anatomical representation or coverage. Skeletonized connections were combined to create a new atlas of human thalamocortical structural connections (Fig. 1). This atlas was used as a custom atlas with DSI Studio’s automated tractography (replacing the built-in ICBM152_adult atlas) to study thalamocortical connection maturation in developmental datasets.

Developmental datasets

Developmental analyses were conducted independently in the following three large, cross-sectional youth datasets: PNC, HCPD and HBN. PNC and HCPD were used as the primary discovery and replication datasets for all study analyses. HBN, a sample of youth presenting with psychiatric concerns, was additionally included to assess whether key developmental and environmental findings replicated in a clinical sample. All subsequent methods concern these three datasets. In all three studies (PNC, HCPD and HBN), all participants over the age of 18 years gave informed consent before participating in the study. Participants under the age of 18 gave informed assent and were enrolled with written consent from their legal guardians. Participants in all three studies received monetary compensation for participation; HBN participants additionally received diagnostic evaluations and referral information. PNC study procedures were approved by the Institutional Review Boards of both the University of Pennsylvania and the Children’s Hospital of Philadelphia. HCPD study procedures were approved by a central Institutional Review Board at Washington University in St Louis. HBN study procedures were approved by the Chesapeake Institutional Review Board.

PNC

PNC⁶⁸ is a community sample of children and adolescents from a broad range of socioeconomic circumstances who were residing in the greater Philadelphia area. Initial exclusion criteria for the PNC were minimal and included inability to engage in psychiatric and cognitive phenotyping, impaired vision or hearing and medical problems that could impact brain function (described in detail in ref. ⁶⁸). Data from 1,145 PNC participants were included in the current study. Participants ranged in age from 8 to 23 years (mean age = 15.3 ± 3.5 years) and had a sex distribution of 608 females and 537 males (sex was self-reported; intersex was not assessed). Participants self-identified their race and ethnicity—0.3% of participants identified as American Indian or Alaskan Native, 0.9% as Asian, 42.4% as Black or African American, 10.5% as multiracial and 45.9% as White.

HCPD

HCPD⁶⁹ is a sample of children and adolescents who were recruited at four academic sites, including the University of Minnesota, Harvard University, Washington University in St Louis and the University of California, Los Angeles. HCPD was designed to be a US population-representative study of typical brain development that included participants from varied geographical, ethnic and socioeconomic backgrounds. Initial exclusion criteria for this study are detailed in a prior report⁶⁹; notable exclusions included premature birth, serious neurological and endocrine conditions, requiring special services at school, treatment for a psychiatric illness for >12 months and hospitalization for a psychiatric condition for >2 days. Data from 572 HCPD participants were included in the current study (lifespan 2.0 release). Participants ranged in age from 8 to 22 years (mean age = 14.8 ± 3.9 years) and had a sex distribution of 306 females and 266 males (sex was self-reported; intersex was not assessed). Participants self-identified their race and ethnicity—7.7% of participants identified as Asian, 10.3% as Black or African American, 14.0% as multiracial and 65.2% as White (data were missing for 2.8% of participants).

HBN

HBN⁷⁰ is a sample of children and adolescents from the New York City area who were referred to the study due to concerns about psychiatric symptoms. The HBN initiative was created by the Child Mind Institute to sample a broad range of commonly encountered forms of clinically significant youth psychopathology; as a result, study exclusions were generally minimal⁵. Extensive information about the HBN sample is provided in ref. ⁵. Approximately 85% of participants in this sample meet criteria for a clinical disorder, including but not limited to anxiety, depressive, attention deficit and hyperactivity, conduct, impulse control, autism spectrum, learning, communication and stress disorders. Data from 959 HBN participants were included in the current study. Participants ranged in age from 8 to 22 years (mean age = 12.2 ± 3.1 years) and had a sex distribution of 329 females and 585 males (sex was self-reported; intersex was not assessed; 45 participants were missing a binary sex indication and were assigned a sex of other). Participants self-identified their race and ethnicity—2.5% of participants identified as Asian, 13.5% as Black or African American, 46.5% as White and 18.2% as not belonging to these three categories (data were missing for 19.3% of participants).

MRI acquisition

This study uses T1-weighted (T1w) structural images, diffusion-weighted images and field maps collected from the three developmental datasets.

PNC

MRI data collected from all PNC participants were acquired on the same 3T Siemens Tim Trio Scanner at the University of Pennsylvania. The T1w images were acquired with a magnetization-prepared rapid acquisition gradient-echo (MPRAGE) sequence with the following parameters: repetition time of 1,810 ms, echo time of 3.51 ms, inversion time of 1,100 ms, flip angle of 9 degrees, 160 slices and a voxel resolution of 0.94 × 0.94 × 1 mm. Single-shell diffusion scans were acquired with b = 1,000 s/mm² in 64 directions, with seven interspersed scans with b = 0 s/mm²; all volumes were acquired in the A–P direction. The collection of these 71 volumes was divided between two runs. The following parameters were used for the diffusion acquisition: repetition time of 8,100 ms, echo time of 82 ms and a voxel resolution of 1.875 × 1.875 × 2 mm. In addition to the structural and diffusion acquisitions, a phase-difference-based field map was acquired to facilitate susceptibility distortion correction of the diffusion data. Field maps were acquired with a double-echo, gradient-recalled echo (GRE) sequence with a repetition time of 1,000 ms, echo times of 2.69 ms and 5.27 ms, a flip angle of 60 degrees, 44 slices and a voxel resolution of 3.75 × 3.75 × 4 mm.

HCPD

HCPD MRI scans were acquired at four sites on 3T Siemens Prisma scanners (MRI derivatives were harmonized across sites as described below). T1w images were acquired with a 3D multi-echo MPRAGE sequence with an in-plane acceleration factor of 2 and the following additional parameters: repetition time of 2,500 ms, echo times of 1.8, 3.6, 5.4 and 7.2 ms, inversion time of 1,000 ms, flip angle of 8 degrees, 208 slices and a voxel resolution of 0.8 mm isotropic. Diffusion scans were acquired in four consecutive runs with two shells of b = 1,500 and 3,000 s/mm² with a multiband factor of 4. In total, 92–93 directions were acquired per shell (370 directions total) along with 28 total b = 0 s/mm² volumes for a total of 398 volumes. Of the 370 diffusion-weighted volumes, 185 distinct directions were each acquired twice with opposite phase-encoding directions (A–P and posterior–anterior). The following parameters were used for the diffusion acquisition: repetition time of 3,230 ms, echo time of 89 ms and a voxel resolution of 1.5 mm isotropic.

HBN

This study analyzes 3T MRI data from the HBN, which were acquired at three sites (MRI derivatives were harmonized across sites as described below). Data were collected at the Rutgers University Brain Imaging Center on a 3T Siemens Tim Trio scanner as well as at the CitiGroup Cornell Brain Imaging Center and the CUNY Advanced Science Research Center on 3T Siemens Prisma Scanners. T1w images were acquired with an MPRAGE sequence with the following parameters: repetition time of 2,500 ms, echo time of 3.15 ms, inversion time of 1,060 ms, flip angle of 8 degrees, 224 slices and a voxel resolution of 0.8 mm isotropic. Diffusion scans were acquired with a multiband factor of three in two shells of b = 1,000 and 2,000 s/mm² in the A–P phase-encoding direction. In total, 64 directions were acquired per shell (128 directions total) along with 1 b = 0 volume. The following parameters were used for the diffusion acquisition: repetition time of 3,320 ms, echo time of 100.2 ms and a voxel resolution of 1.8 mm isotropic. A reverse phase-encoding b = 0 was additionally acquired for use as an echo planar image (EPI)-based field map in susceptibility distortion correction.

Diffusion MRI preprocessing

Diffusion scans (and corresponding T1w images and field maps) from PNC, HCPD and HBN were preprocessed using QSIPrep⁷¹, which is based on Nipype and uses Nilearn, Dipy, ANTs, the FMRIB Software Library (FSL) and additional software tools described below. QSIPrep versions 0.14.2 and 0.16.1 were used in PNC/HBN and HCPD, respectively. The same sequence of preprocessing steps was performed in both datasets. However, given that slightly different QSIPrep versions were applied to PNC/HBN and HCPD, the software versions used by its internal operations differed. The following internal software versions were used in PNC/HBN and HCPD, respectively: Nipype 1.6.1 and 1.8.5, Nilearn 0.8.0 and 0.9.2, ANTs 2.3.1 and 2.4.0, and FSL 6.0.3 and 6.0.5.

Anatomical processing was also performed within QSIPrep. T1w images were corrected for intensity nonuniformity using N4BiasFieldCorrection⁷² (ANTs), skull-stripped using antsBrainExtraction (ANTs) with OASIS as a target template and spatially normalized to the nonlinear ICBM152 2009c template using nonlinear registration with antsRegistration. For diffusion data processing, a series of preprocessing steps was applied separately to data from each diffusion run (two runs in PNC, four runs in HCPD and one run in HBN); individual runs were then concatenated. Any images with b < 100 s/mm² were considered b = 0 volumes. First, MP–PCA denoising, as implemented in dwidenoise (MRtrix3 (ref. ⁷³)), was applied with a five-voxel window. After MP–PCA, Gibbs unringing (MRtrix3) was performed with mrdegibbs. Following unringing, B1 field inhomogeneity was corrected using dwibiascorrect (MRtrix3), which implements the N4 algorithm⁷². After B1 bias correction, the mean intensity of the diffusion-weighted series was adjusted so that the mean intensity of all b = 0 images matched across separate runs. QSIPrep was additionally used to correct for head motion, eddy currents and susceptibility distortions. FSL’s eddy was used for head motion and eddy current correction⁷⁴. In all datasets, eddy was configured with a q-space smoothing factor of 10, a total of 5 iterations and 1,000 voxels used to estimate hyperparameters. Eddy’s outlier replacement was run⁷⁵.

Given that PNC, HCPD and HBN acquired different versions of field maps (GRE and EPI), different approaches were taken to correct for susceptibility artifacts. In the PNC, field map-based susceptibility distortion correction was performed by creating a B0 map using a phase-difference image and a magnitude image from the GRE field map acquisition. In HCPD and HBN, reverse phase-encoding EPI-based field maps were collected, resulting in pairs of images with distortions going in opposite directions. For EPI-based susceptibility distortion correction, b = 0 reference images with reversed phase-encoding directions were used along with an equal number of b = 0 images extracted from diffusion scans. From these pairs, the susceptibility-induced off-resonance field was estimated. Field maps were ultimately incorporated into the eddy current and head motion correction interpolation. Final interpolation was performed using the Jacobian modulation (jac) method. After preprocessing, the concatenated diffusion data were resampled to AC–PC to be in alignment with the T1w data. The output resolution of the fully preprocessed diffusion data was 1.8 mm isotropic in PNC and HBN and 1.5 mm isotropic in HCPD.

Sample construction

PNC, HCPD and HBN participants with T1w images, field maps and identical-parameter⁷⁶ diffusion MRI scans (that is, nonvariant acquisitions) available were considered for inclusion in the present work. Exclusion criteria were applied to available scans in each dataset to ensure that only high-quality neuroimaging data from youth without serious medical conditions were analyzed. In all three datasets, participants were excluded from the initial neuroimaging sample for medical problems that could impact brain function or incidentally encountered abnormalities in their neuroimaging data (medical exclusion). Participants were additionally excluded if they had low-quality T1w images with motion, artifacts or poor surface reconstructions (T1w quality exclusion), if their raw diffusion scans were missing gradient directions (diffusion directions exclusion), if in-scan head motion during the diffusion scan exceeded a mean framewise displacement of 1 mm (diffusion motion exclusion) or if their preprocessed diffusion scans had a low neighborhood correlation (diffusion quality exclusion). The neighborhood correlation metric quantifies the average pairwise spatial correlation between pairs of diffusion volumes that sample similar points in q-space, with lower correlations reflecting reduced data quality. The range of neighborhood correlation values obtained will vary by the diffusion scan acquisition parameters and noise level; thus, dataset-specific exclusion thresholds were used (<0.9 in PNC, <0.6 in HCPD and <0.7 in HBN).

PNC was used as the discovery dataset in this study. In PNC, 1,358 individuals had all required neuroimaging data available. From this original neuroimaging sample, 213 total participants (15.7%) were ultimately excluded for the abovementioned criteria (applied successively), including n = 118 for medical exclusions, n = 25 for the T1w quality exclusion, n = 57 for the diffusion quality exclusion and a remaining n = 13 for the diffusion motion exclusion. HCPD was used as the main replication dataset. In HCPD, 640 individuals had neuroimaging data available. From this originally eligible sample, 54 total participants (10.6%) were excluded for the abovementioned criterion, including n = 7 for medical exclusions, n = 8 for the diffusion directions exclusion, n = 15 for the diffusion quality exclusion and then n = 24 for the diffusion motion exclusion. In addition, to directly compare developmental results from HCPD to PNC, children less than 8 years old were excluded from HCPD to match the PNC age range (n = 14; only 2.4% of the remaining eligible sample). This young age exclusion was additionally important in HCPD, as there is likely insufficient data available from individuals <8 years old for accurate developmental modeling in this dataset (the same age range exclusion has been applied in the past by independent groups for this reason³). HBN was used as a secondary replication dataset to assess if findings could be extended to a predominantly psychiatric sample. As with HCPD, we only considered data from participants who overlapped with the PNC age range. In total, 1,530 participants from data releases 1–9 in HBN had nonvariant (collected with the same acquisition) T1w and diffusion scans available. In this initial sample, 1,149 were between the ages of 8 and 22 years. From the 1,149 individuals, 190 total participants (16.5%) were ultimately excluded for the abovementioned criterion, including n = 53 for the T1w quality exclusion, n = 10 for the diffusion quality exclusion and then n = 127 for the diffusion motion exclusion.

Delineation of individualized thalamocortical connections

To consistently identify the same thalamocortical connections across PNC, HCPD and HBN participants, we used our population-level thalamocortical tractography atlas as an anatomical prior with DSI Studio’s trajectory-based automated tract recognition (auto-track) approach^59,66. To conduct auto-track, we first reconstructed participants’ preprocessed diffusion scans with GQI. GQI was executed using QSIPrep’s GQI reconstruction workflow (version 0.16.0RC3 in all datasets) with a ratio of mean diffusion distance of 1.25. Auto-track uses a distance-based pathway recognition approach to best match streamlines generated in individual participants’ data to connections included in the thalamocortical connectivity atlas. Specifically, the auto-track procedure involves nonlinearly warping thalamocortical atlas connections to participant diffusion data, seeding deterministic tractography in voxels that correspond to each connection in the atlas, comparing generated streamlines to atlas connections using the Hausdorff distance and discarding streamlines with a Hausdorff distance above a selected threshold (and thus those with trajectories that diverge from the atlas connection). We used a relatively strict Hausdorff distance threshold of 10 to ensure identification of connections with similar shapes and anatomical endpoints across individuals while still allowing for subtle person-specific differences in anatomy. The following specific parameters were used for deterministic tracking: threshold index = qa, otsu threshold = 0.50, a track-to-voxel ratio of 4, a yield rate of 0.0000001 and no topology-informed pruning. As with the tractography atlas, random parameter saturation was used to select the anisotropy threshold, turning angle, step size and smoothing level for each generated streamline.

Nearly all thalamocortical structural connections represented in the population-level atlas could be clearly identified in diffusion data from all youth participants. Few connections, however, were either not identified or only sparsely reconstructed in participant data. These unsuccessful reconstructions at the participant-level provide important evidence that the thalamocortical tractography pipeline used here was not overfitting participants’ data to identify connections that were not faithfully represented in the diffusion signal. To ensure that only well-represented thalamocortical connections were analyzed in the present work, we applied exclusion criteria to connections at individual and dataset levels. For each participant, we excluded connections that had <5 streamlines from consideration. After applying this participant-level exclusion, we then only analyzed connections for which at least 90% of participants had a valid (≥5 streamlines) connection. As a result, 15 connections (6.3%) present in the original thalamocortical tractography atlas were not studied in PNC, 8 connections (3.4%) in HCPD and 13 connections (5.5%) in HBN.

To quantitatively index tractography performance in all participants, we computed a measure that parameterizes the voxel-wise overlap between a participant’s thalamocortical connection and the same connection in the tractography atlas. Specifically, we computed overlap sensitivity, which is calculated as the number of voxels included in the atlas connection that are additionally included in a participant’s connection, divided by the total number of voxels included in the atlas connection. This measure thus reflects the connection’s true positive rate in each individual. Overlap sensitivity values were calculated for all participants’ thalamocortical connections by deriving masks of the connections in atlas space.

Derivation of thalamocortical connectivity properties

For all reconstructed thalamocortical connections, we computed metrics that provide information about connection microstructure and about where the connection’s streamlines terminated within the thalamus. Specifically, we calculated two metrics, FA and C–M_t gradient position, in all analyzed connections from all youth datasets. We used DSI Studio to compute the average FA along each thalamocortical connection. FA was calculated based on the single shell of b = 1,000 s/mm² in PNC and shells <1,750 s/mm² in HCPD and HBN. High b value shells (>1,750) were not used for the calculation of FA, as high b values sensitive to non-Gaussian diffusion do not meet the assumptions of the tensor fit and can introduce noise or inaccuracies into DTI values. It is important to note, however, that all shells were used with auto-track to enhance tractography performance. Given that HCPD and HBN diffusion data were collected across different scanners, we used ComBat to harmonize thalamocortical FA measures within each dataset to mitigate scanner-specific effects. ComBat has previously been shown to remove unwanted scanner effects from FA measures when age is the biological phenotype of interest and to do so more effectively than alternative harmonization approaches⁷⁷. ComBat was implemented via the neuroCombat package in R (version 1.0.13). Age was protected as a biological variable of interest in ComBat along with developmental model covariates of sex and in-scan head motion.

We additionally derived a C–M_t value for every delineated thalamocortical connection, which quantifies where the thalamic terminations of a connection fall in a within-thalamus C–M gradient. To accomplish this derivation, we used the C–M_t gradient derived in ref. ²⁰ in MNI152 space. The study conducted by ref. ²⁰ produced a C–M cell-type gradient by using Allen Human Brain Atlas microarray data to determine the relative expression of human PV (gene symbol—PVALB) and calbindin (gene symbol—CALB1) genes within thalamic voxels (defined by the Morel atlas). Thalamic neurons that have a core-like cortical projection pattern express higher levels of PVALB, whereas thalamic cells that have a matrix-like cortical projection motif express higher levels of CALB1 (refs. ^20,42). The relative weighting of PVALB versus CALB1 genetic expression thus provides a proxy for the local proportion of core to matrix thalamic cell types. This weighting was quantified by ref. ²⁰ as the difference between normalized CALB1 and PVALB gene expression levels (CALB1–PVALB = C–M_t value). We used this gene expression-derived gradient in MNI152 space to assign every thalamocortical connection delineated in this study a numerical C–M_t value based on where its streamlines connected with the thalamus. We first transformed a mask of every individual thalamocortical connection from subject diffusion space to MNI space using person-specific warps computed with antsRegistration’s symmetric normalization during preprocessing in QSIPrep. We then calculated the average C–M_t value of all thalamic voxels that corresponded to a given connection’s thalamic termination area.

General approach to hypothesis testing

Throughout this work, we evaluate the statistical significance of GAM-derived smooth terms and linear terms as well as the significance of spatial correlations between brain maps. Here we note general approaches used in significance testing, which will be described in further detail below. When assessing significance with GAMs, which were applied individually to each thalamocortical connection, the false discovery rate (FDR) correction was used to correct P values (P_FDR) for multiple comparisons across all connections within a dataset and statistical significance was set at P_FDR < 0.05. When correlating brain maps, we used nonparametric Spearman’s rank-based correlations throughout to assess monotonic relationships between variables, except when comparing the same metric derived independently in two datasets (for example, age effects in PNC and HPCD), as linear relationships were expected. To compute the significance of each correlation comparing two brain maps, we used spin-based, spatial rotation tests or ‘spin tests’⁷⁸. Spin tests compute a P value (P_spin) by comparing the empirically observed correlation between two brain maps to a null distribution of correlations obtained by spatially rotating (spinning) spherical projections of the brain maps. Accordingly, rather than creating a null by permuting individual data points in the brain maps, spin tests create spatial null models that maintain the native spatial covariance structure of the empirical map. The P_spin value is computed as the number of times the rotation-based correlation value is greater than (for positive empirical correlations) or less than (for negative empirical correlations) the empirical correlation value, divided by the total number of spatial rotations (spins). We used a null of 10,000 spatial rotations throughout. To carry out spin tests with thalamocortical connectivity data, we assigned each HCP–MMP cortical region the value ascribed to its corresponding thalamic connection. Region-based spin tests were implemented using the rotate_parcellation algorithm in R. Spin tests were additionally used to create spatial null models and a null distribution of values for anatomical and environmental enrichment tests (described further below). All statistics were performed in R version 4.2.3. All statistical tests were two-sided. Statistical analyses were conducted separately in each of the three developmental datasets.

Maps of cortical organization, function and development

We integrated brain maps of cortical feature variability, functional diversity and developmental heterochronicity into the present study to provide insights into thalamocortical connection anatomy and development. To facilitate this, the HCP–MMP atlas was used to parcellate data from the brain maps described below.

The S–A axis of cortical organization

Every HCP–MMP region was assigned a specific rank in the S–A axis based on its position in this dominant motif of cortical feature organization. S–A axis ranks were previously computed in ref. ¹ by averaging rank orderings of ten cortical feature maps that exhibit systematic and concerted spatial variation between primary sensory and motor cortices and transmodal association cortices. These maps include the principal gradient of functional connectivity⁷⁹, the cortical T1/T2 ratio⁸⁰, macroscale MRI measurements of cortical thickness, allometric scaling calculated as local area scaling with changes in total brain size⁸¹, metabolic demand indexed by aerobic glycolysis⁸², cerebral perfusion estimated by arterial spin labeling⁸³, functional diversity determined by spatial variation in Neurosynth meta-analytic decodings, cytoarchitectural similarity measured using the BigBrain atlas⁸⁴, a dominant mode of gene expression proxied by the first principal component of brain-expressed genes⁸⁰ and macaque-to-human evolutionary cortical expansion⁸⁵. Accordingly, the S–A axis represents a large-scale, prominent axis of cortical organization that captures the stereotyped patterning of numerous macrostructural, microstructural, functional, metabolic, transcriptomic and evolutionary features across the human cerebral cortex. The rank ordering of all HCP–MMP regions along the S–A axis was obtained from https://github.com/PennLINC/S-A_ArchetypalAxis. Low-ranking regions at the axis’s sensorimotor pole are nearly exclusively primary and early unimodal cortices. Mid-ranking regions span from multimodal cortices involved in multisensory integration and language to those involved in attention, working memory, and decision-making capacities. Regions at the axis’s association pole are linked to cognitive control, social cognition, self-referential thought and emotion regulation.

Neurosynth meta-analytic association maps

Neurosynth⁸⁶ version 0.7 was used to create HCP–MMP meta-analytic maps for a diversity of psychological terms capturing both lower-order and higher-order psychological processes. In accordance with prior work⁸⁷, we obtained meta-analytic activation maps for 123 cognitive terms that were present in both the Neurosynth database and the Cognitive Atlas⁸⁸; maps were obtained using the Neuroimaging Meta-Analysis Research Environment (NiMARE)⁸⁹. Term-specific maps were computed in volumetric space using multilevel kernel density chi-square analysis, mapped to the fslr surface and parcellated with the HCP–MMP parcellation. Values in term-specific maps are association test z scores that quantify the extent to which activation in a cortical region occurred more consistently in prior fMRI studies that mentioned a given term as compared to studies that did not.

Brain charts of developmental heterochronicity

We used previously published maps characterizing across-cortex variability in the development of the E/I ratio, myelin and intrinsic activity amplitude as indexed by noninvasive MRI measures sensitive to these cortical properties. The cortical E/I ratio was noninvasively estimated in ref. ⁴ by fitting a biophysically plausible circuit model to resting-state fMRI data from youth in the PNC (n = 855, ages = 8–23 years). The magnitude of maturational decline in the E/I ratio was quantified across the cortex through a linear regression between model-estimated E/I ratio and age, producing an age slope map that was parcellated with the HCP–MMP atlas. Age-related change in cortical myelin content, as studied in ref. ³, was assessed by calculating the T1/T2 ratio in data from HCPD participants (n = 628, ages = 8–22 years). These authors determined the annualized rate of cortical myelin growth in each HCP–MMP region by fitting GAMs with a smooth function for age and quantifying the difference between the model-estimated T1/T2 ratio at ages 21 and 8 years, divided by the number of years in this range (a point estimate for the rate was estimated by posterior sampling). Intrinsic cortical activity amplitude was studied in ref. ² using resting-state fMRI data from the PNC sample (n = 1,033, ages = 8–23 years). The amplitude of intrinsic fluctuations in cortical activity was quantified as the average power of low-frequency fMRI fluctuations, given that increases in power in the frequency domain are mathematically proportional to increases in signal amplitude in the time domain. The age at which activity amplitude began to decrease was evaluated in every cortical region by fitting GAMs with thin plate regression splines to model developmental trajectories and identifying the youngest age at which the first derivative of the developmental spline was significantly negative.

Analysis of thalamocortical connection atlas coverage

The final diffusion tractography atlas comprised structural connections between the thalamus and cortical regions that occupy the majority of cortical surface area. However, a subset of cortical regions did not have thalamic connections included in the atlas. We hypothesized that regions without thalamic connections represented in the atlas would be harder to perform streamline tracking in and therefore would be smaller in size and located at greater sulcal depths than other cortical regions. To statistically test this hypothesis, we first computed the average surface area of cortical regions without thalamic connections. We then compared this empirical value to a null distribution of surface areas obtained by spatially rotating the alignment between the surface area map and the designation of regions excluded from the atlas. A spin testing-based P_spin value was obtained by calculating the number of times the empirical surface area value was less than the null value computed from spatial nulls of excluded connections. The same spin testing procedure was implemented to determine whether regions without thalamic connections in the atlas had significantly greater sulcal depth than other cortical regions.

Analysis of thalamocortical connection circuit anatomy

To study whether thalamocortical connections exhibited different patterns of connectivity with sensory versus association cortices, we assessed whether thalamocortical connection FA and C–M_t values exhibited ordered variation along the S–A axis. As described above, FA and C–M_t values were calculated for every thalamocortical connection reconstructed in every individual. To study the alignment of these connectivity measures with the S–A axis, the average (across-participant) value was then computed for each connection separately in each dataset. Spearman’s correlations between connection FA or C–M_t values and S–A axis ranks were then calculated, and the significance of these correlations was determined with spin tests.

Analysis of thalamocortical connection development

Developmental modeling

To quantitatively characterize diversity in the development of structural connections between the thalamus and individual cortical regions, we fit connection-specific GAMs with FA as the dependent variable, age as a smooth term and sex and diffusion scan head motion (mean framewise displacement) as linear covariates. Models were fit separately for each individual thalamocortical connection using thin plate regression splines as the smooth term basis set and the restricted maximal likelihood approach for selecting smoothing parameters. All models were fit using the mgcv package⁹⁰ in R. We used GAMs for developmental modeling, given that they are capable of capturing a broad array of both linear and nonlinear age-dependent relationships and can furthermore be harnessed to characterize age windows of significant developmental change and change offset. Each GAM produces a smooth function for age that is generated from a linear combination of weighted basis functions (splines) and that represents the connection’s developmental trajectory. As in our prior work², to prevent overfitting of the flexible smooth function, we set the maximum basis complexity (k) to 3. A value of k = 3 was selected as model fits were relatively noncomplex (requiring a smaller number of knots) and the k index indicated this basis complexity was sufficient. Higher values of k were, moreover, visually deemed to result in overfitting in multiple connections.

After fitting connection-specific GAMs, we interrogated GAM-derived properties, including the significance of the age effect, the magnitude of the age effect and the age at which the connection stopped exhibiting significant developmental change. The significance of the age effect, which denotes the significance of the smooth function quantifying the relationship between FA and age, was determined by an analysis of variance F test that compared the full GAM model to a nested, reduced model with no age term (in the form of a generalized likelihood ratio test). A significant F test (FDR corrected; P_FDR) indicates that including the smooth term for age in the model significantly improved the model fit as compared to a model with only sex and diffusion scan head motion as predictors. The magnitude of the age effect was calculated as the partial R² between the full GAM model and the reduced model without age. To resolve the age at which each connection stopped exhibiting significant developmental change during childhood and adolescence—which can be interpreted as the connection’s age of maturation—we first calculated the first derivative of the age smooth function using finite differences, which provides an age-specific rate of developmental change (ΔFA/Δ age). We then identified the youngest age at which the rate of developmental change was not significantly different from 0 by obtaining a simultaneous 95% confidence interval for these first derivatives and determining when the confidence interval first included 0. Derivative analyses used the gratia package⁹¹ in R. After fitting main developmental GAMs, we additionally fit models with a sex by age (factor smooth) interaction to assess whether developmental trajectories of FA significantly differed by sex. The significance of the interaction term was extracted and corrected for multiple comparisons (P_FDR).

Developmental characterization

A primary goal of this research was to empirically evaluate the hypothesis that the development of thalamocortical structural connections is coupled to the heterochronous patterning of developmental change across the cortex’s S–A axis. We assessed this hypothesis by characterizing how differences in ages of thalamocortical connection maturation were related to (1) cortical function, (2) cortical and thalamic axes, and (3) cortical developmental heterogeneity. To facilitate these three analyses, we computed the age of maturation for each connection and projected these ages onto HCP–MMP regions.

First, to functionally decode cortical regions with earlier and later maturing thalamic connections, we computed Spearman’s correlations between the cortex-projected thalamocortical age of the FA maturation map and each of the 123 psychological term meta-analytic maps made using Neurosynth. In this analysis, psychological terms with the most negative correlation values exhibit high fMRI activation-based z scores in cortical regions with the earliest-maturing thalamic connections. Terms with the most positive correlations display the highest z scores in cortical regions with the latest-maturing thalamic connections. To provide insight into the types of functions ascribed to regions with early versus late-maturing thalamocortical connections, we therefore identified the ten most negatively correlated and ten most positively correlated terms that emerged in a given dataset, extracting a list of 20 developmentally pertinent terms in total. We furthermore calculated (1) the number of overlapping terms between PNC and HCPD lists (n/20) and (2) a term-overlap P_spin value based on the comparison of this number to a null distribution of overlapping term counts obtained using the spatial rotation (spin) testing procedure. More explicitly, we spun the thalamocortical maturation map derived in HCPD, recomputed Spearman’s correlations between this spatial null map and the 123 Neurosynth term maps, and identified how many of the ten most negatively and positively correlated terms were present in the empirical PNC term list (n/20). This procedure was repeated 10,000 times, creating a null distribution of term-overlap counts used to compute a term-overlap P_spin value.

Second, to further characterize how differences in thalamocortical connection development were instantiated across the cortex and thalamus, we correlated thalamocortical connection maturational ages with the S–A axis, with the cortex’s three primary anatomical axes (A–P, D–V and M–L axes) and with the C–M_t gradient using Spearman’s correlations. We ascertained the significance (FDR corrected) of these individual correlations with spin tests. We additionally assessed whether the correlation with the S–A axis was significantly greater in magnitude than correlations with A–P, D–V and M–L cortical axes and the C–M_t gradient by implementing statistical tests (FDR corrected) for comparing two correlations with overlapping variables derived from dependent groups. These tests were executed using the cocor package⁹² in R with the hittner2003 test⁴³, which implements a modified version of Steiger’s z test⁹³ using a backtransformed average Fisher’s z procedure.

Third, we aimed to study the correspondence between the spatiotemporal patterning of maturational variability across cortical regions and across thalamocortical structural connections. To test for correspondence between cortical and thalamocortical development, we used the three previously described cortical charts of developmental heterochronicity. Specifically, we used Spearman’s correlations and spin tests to relate the thalamocortical age of maturation map to cortical developmental maps of the E/I ratio, the T1/T2 ratio and fMRI-based intrinsic activity amplitude.

Developmental sensitivity analyses

We conducted three sensitivity analyses to ensure that developmental findings were robust to potential confounds. The first sensitivity analysis considered cortical surface area. To execute this analysis, FreeSurfer 6.0.1 was used to quantify the surface area of HCP–MMP regions using participants’ T1w images. FreeSurfer’s recon-all analysis stream was first run; subsequently, the HCP–MMP atlas was resampled from the fsaverage template to participants’ individualized cortical surfaces using spherical registration. The second sensitivity analysis examined diffusion data quality as measured by the tSNR of all b = 0 s/mm² volumes. A voxel-wise map of b = 0 shell tSNR was calculated by FSL’s eddy during preprocessing with QSIPrep. Connection-specific tSNR was then calculated within the mask of each participant’s personalized thalamocortical connections. The third sensitivity analysis focused on thalamocortical connection reconstruction success. Reconstruction quality was indexed for all thalamocortical connections using the atlas overlap sensitivity value (true positive rate) described above.

Each of the three sensitivity analyses had two parts. For the first part of each analysis, we averaged surface area, tSNR and overlap sensitivity for each connection across all participants in a dataset. We then used Spearman’s correlations to assess whether each of these variables was correlated with connections’ ages of maturation. The goal of these correlations using group-averaged variables was to ensure that connection-wise differences in cortical endpoint anatomy, data quality and reconstruction quality did not account for differences in thalamic connection maturational timing. In the second part of each sensitivity analysis, we refit connection-specific GAMs while including individual-level measures of cortical surface area, connection tSNR or connection overlap sensitivity as additional model covariates. GAMs thus included a smooth term for age as well as linear terms for sex, diffusion scan head motion and the sensitivity variable. After refitting GAMs, developmental trajectories of FA were visualized, and developmental alignment with the S–A axis was tested. The goal of the sensitivity model GAMs was to ensure that maturational increases in FA were robust to potential age-related differences in regional surface area, diffusion signal quality and atlas-based delineation of thalamic connectivity. Note, six participants did not have valid FreeSurfer outputs in HCPD and were excluded from the surface area analysis.

Analysis of environmental effects

Measures of youths’ socioeconomic environments

We examined relationships between thalamocortical connection FA and socioeconomic features of youths’ household environments (education and income) and neighborhood environments (geocoding-derived factor scores). Not all datasets collected the same environmental information; thus, the most relevant variables available were studied here. Household socioeconomic position was proxied by the average years of education obtained by both (when data were available) or one caregiver in both PNC and HCPD. Caregiver years of education were directly reported in PNC and were missing for eight participants who were excluded from this analysis. In HCPD, caregiver education information was reported as the highest educational level achieved and was therefore recoded to a numerical value to allow for continuous analyses. Grade levels were directly recoded (for example, ‘eighth grade’ to 8 years, ‘tenth grade’ to 10 years), and degrees were recoded based on average years to degree completion in the United States (for example, ‘bachelor’s degree’ to 16 years, ‘professional degree’ to 19 years, ‘doctoral degree’ to 22 years). Household socioeconomic position was additionally estimated by the income-to-needs ratio in HCPD (income was not provided in PNC). This ratio was calculated by dividing total annual family income by the federal poverty line based on the participant’s total family size. A city cost of living-adjusted ratio was then computed to account for geographic area (St Louis = 1, Minneapolis = 1.175, Los Angeles = 1.558, Boston = 1.624). To mitigate the potential impact of extreme income outliers and account for the right-skewed distribution of these data, the city-adjusted income-to-needs ratio was winsorized (lowest and highest 1% of the data), and the natural logarithm was taken, in accordance with prior studies^94,95.

Information about the neighborhood, rather than household, environment was obtained using geocoding and was not available in HCPD, given that participant addresses were not released. In PNC, data about each individual’s neighborhood (block level) environment were extracted using home addresses and the census-based American Community Survey. Census-based variables were factor analyzed in ref. ⁹⁶ to derive a factor score for each participant that captured multivariate features of the neighborhood environment a child lived. The first factor from an exploratory two-factor analysis performed in ref. ⁹⁶ was used here, which has the following factor loadings: the percentage of residents who were married (loading = 0.85), median family income (0.82), the percentage of residents with a high school education (0.74), the percentage of residents who were employed (0.68), median age (0.61 years), the percentage of residents who were female (−0.26), the percentage of houses that were vacant (−0.60), population density (−0.71) and the percentage of residents experiencing poverty (−0.86).

Although neighborhood environment information could not be obtained in HCPD, participant addresses were collected in HBN, allowing us to assess whether neighborhood environment results replicated in a second sample. Mirroring the approach taken in the PNC, a neighborhood environment factor score was calculated for HBN participants by factor analyzing geocoded variables at the level of census block groups. Specifically, variables from the American Community Survey and Environmental Protection Agency were factor analyzed using exploratory structural equation modeling, and five factors were extracted. The neighborhood socioeconomic factor included here had the following loadings: median family income (loading = 0.89), median home value (0.80), median owner upkeep cost (0.80), median rent (0.78), the percent of residents with a high school education (0.57), median rooms per dwelling (0.52), the percentage of residents who were married (0.50), the percent of residents without health insurance (−0.49) and the percent of residents in poverty (−0.61). Address information and associated neighborhood environment scores were missing for 13 individuals in HBN who were excluded from this analysis.

Statistical testing of environmental associations

To model relationships between thalamocortical connection FA and environmental features, we refit developmental GAMs with environment variables included as linear continuous covariates. Model terms included a smooth term for age and linear terms for sex, diffusion scan head motion and the environment variable under study (caregiver education, income or neighborhood factor score). This model formulation allowed us to identify age-independent main effects of the environment across developmental stages. Significance of the linear environment term was FDR corrected (P_FDR) across all connection-specific GAMs in each analysis. The direction and magnitude of the relationship between connection FA and the environment variable were determined by the t value of the linear environment term, which we refer to as the statistical environmental effect. In addition to the GAMs used to quantify main effects of the environment, we furthermore fit GAMs with an age-by-environment varying-coefficient interaction to model how associations between the environment variable and connection FA vary over the smooth function of age. These varying-coefficient interaction GAMs were used to predict connection-specific FA developmental trajectories for varying levels of neighborhood environmental (dis)advantage.

Environmental characterization

In a final series of analyses, we explored how connection-specific environmental effects varied across the S–A axis during youth; the below analyses considered only statistically significant environment effects. These analyses were only conducted for neighborhood environmental effects, given that caregiver education and income effects were not statistically significant. We conducted two complementary analyses designed to assess whether neighborhood environment relationships were strongest for thalamic connections with the associative end of the S–A axis. The first was an enrichment analysis that divided the S–A axis into quintiles and tested whether the distribution of environmental effects (t values) was homogeneous across connections or significantly smaller or larger in magnitude for connections to certain portions of the S–A axis. The second was a correlation analysis using continuous S–A axis ranks. To execute the enrichment analysis, we first computed the average environment t value in each of the five quintiles of the S–A axis. We then compared this empirical t value to a null distribution of quintile-specific t values obtained by applying 10,000 spherical spatial rotations to the cortex-projected environment effect map, producing a spin-based enrichment P_spin. Quintiles were originally used for the enrichment analysis to differentiate it from the continual correlation while ensuring that sections of cortex with different hierarchical properties and developmental timelines were distinguishable. However, to ensure that significant enrichment results were not dependent on the specific partitioning of the S–A axis used, we repeated the enrichment analysis using S–A axis quartiles and deciles. Following enrichment testing, we used Spearman’s correlations (with FDR-corrected spin tests) to assess the relation of significant environment effects with the S–A axis and with A–P, D–V and M–L cortical axes and the C–M_t gradient. We compared the magnitude of correlations between the S–A axis and each of these four axes using the previously described cocor test (FDR corrected) for comparing two dependent, overlapping correlations.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.