nature neuroscience

Article https://doi.org/10.1038/s41593-024-01679-3

A shifting role of thalamocortical

connectivity in the emergence of
cortical functional organization

Received: 26 April 2023 Shinwon Park1,2, Koen V. Haak 3,4, Stuart Oldham 5,6, Hanbyul Cho1,
Kyoungseob Byeon7, Bo-yong Park1,8, Phoebe Thomson2, Haitao Chen ,
9,10
Accepted: 13 May 2024
Wei Gao 9,11, Ting Xu 7, Sofie Valk 12,13, Michael P. Milham 14,15,
Published online: xx xx xxxx Boris Bernhardt 16, Adriana Di Martino2 & Seok-Jun Hong 1,14,17,18,19

Check for updates

The cortical patterning principle has been a long-standing question
in neuroscience, yet how this translates to macroscale functional
specialization in the human brain remains largely unknown. Here we
examine age-dependent differences in resting-state thalamocortical
connectivity to investigate its role in the emergence of large-scale functional
networks during early life, using a primarily cross-sectional but also
longitudinal approach. We show that thalamocortical connectivity during
infancy reflects an early differentiation of sensorimotor networks and
genetically influenced axonal projection. This pattern changes in
childhood, when connectivity is established with the salience network,
while decoupling externally and internally oriented functional systems.
A developmental simulation using generative network models corroborated
these findings, demonstrating that thalamic connectivity contributes to
developing key features of the mature brain, such as functional segregation
and the sensory-association axis, especially across 12–18 years of age. Our
study suggests that the thalamus plays an important role in functional
specialization during development, with potential implications for studying
conditions with compromised internal and external processing.

The human brain begins its complex developmental process after con-
ception, initially as a simple neural tube that bends, folds and expands,
a part of which eventually becomes the characteristic wrinkles of the
neocortex. During pregnancy, the neural tube is patterned along the
rostrocaudal axis driven by a gradient of morphogens and transcription
factors1. Shortly after this, thalamocortical axons begin to migrate into
the cortical plate1,2. The awareness of this intermingled process gave rise
to two influential views—the Protomap and Protocortex theories—on
how the neocortex differentiates into distinct functional regions across
development, guided by genetic or thalamic influence, respectively3,4.
The current consensus is that both intrinsic (genes) and extrinsic (thala-
mus) mechanisms collectively drive cortical development, yet viewing
the role of the thalamus exclusively as a mediator of sensory experience
may be an oversimplification5–7.
Indeed, on the one hand, most sensory information passes through
the thalamus before reaching the neocortex, placing this subcortical
structure in a strategic location for mediating external signals influ-
encing cortical differentiation6,7. For instance, exposure to sensory
stimuli accelerates initial synaptic plasticity of the newborn brain and
gradually shapes functional specialization through intensive inter-
actions with the thalamus8,9. On the other hand, influenced by axon
guidance molecules and transcription factors, the thalamic nuclei also
form direct topographic projections to specific cortical areas during
development10, a process that is crucial for regional differentiation

A full list of affiliations appears at the end of the paper. e-mail: hongseokjun@skku.edu

Nature Neuroscience
Article https://doi.org/10.1038/s41593-024-01679-3

a Neuroimaging analysis
(i) Thalamocortical connectopic and neocortical projection maps (ii) Developmental effects and hierarchical network profiling

ROI Mask
Yeo 7 network atlas
…
Dimensionality
ult
reduction /young ad
Childhood
Similarity Laplacian Thalamocortical Neocortical Infancy
matrix eigenmaps connectopic map projection maps
Min Gradient values Max

b Transcriptomic analysis
(i) PLS analysis (ii) Assess significance (iii) GO enrichment and cell-type-specific expression analysis
Ranked gene list Biological processes Cell-specific profiling
NEOMAPs Gene expression (AHBA) SBP-spatial models

Th tum us
PLS gene Loadings

p
or llum
Axonogenesis

r ia m
er ala
Empirical

us
Probability density

St ca
… KCNC1 13.2426

Pspin < 0.05

am
d
Neuron

H x
po
association

yg

eb

te

al
differentiation

ip
Am
KCNS1 13.2349
Regions

Regions

Axon

C
~

C
KCNAB3 13.1989
CNS guidance

Developmental stage
. .
development
Maximize . .
association Cell-to-cell
. .
signaling

Pspin < 0.05

Association (null model) UCHL3 –12.3989
DPYSL3 –12.4319
FABP7 –12.9762 Cell
morphogenesis
c Computational simulation
P-values

Generative network modeling (simulated data) Functional brain network (experimental data)
fMRI timeseries

(ii) Evaluate fit Time

(i) Edges generated based on wiring rule during iteration Time

(iii) Estimate measures from simulated correlation matrices (iv) Evaluation of network formation
Wiring rules Measures
• Thalamocortical connectivity • Cortical gradients
• Gene expression • Network node characterization
• Spatial distance Thalamocortical Gene Spatial • Segregation index
connectivity expression distance

Fig. 1 | Analytical flows of neuroimaging, transcriptomics and computational
simulation. a, Connectopic mapping16, a dimensional reduction technique,
was used to summarize complex thalamocortical connectopic topologies (i).
Aging effects were then tested for the thalamocortical maps within infancy
and childhood/young adulthood groups (ii). The maps for both the thalamus
and neocortex were profiled according to the Yeo-Krienan 7 Network Atlas28.
b, Associations between the neocortical projection maps and spatial pattern
of gene expression derived from the Allen Human Brain Atlas (AHBA38) were
quantified using partial least squares (PLS) analysis (i-ii). Significance was tested
using phenotype-based null models that maintain spatial autocorrelation
(10,000 spins)77 (iii). The significant gene sets identified by PLS analysis
were ranked and then submitted to a Gene Ontology enrichment analysis
to characterize the underlying biological processes and cell-type specific
expressions across developmental stages based on the BrainSpan atlas39.
c, Generative network models44 were utilized to investigate the effects of
thalamocortical, genetic, and spatial constraints when estimating how wiring
costs shape functional connectome topology. SBP, spatial brain phenotype.

of the neocortex. Given its substantial role, the exact involvement of one end to stimulus-independent, internally oriented processing on
thalamocortical connectivity in neocortical functional specialization the other end11,18. Notably a recent study showed an age-dependent
in early life remains an open question. shift in this neocortical organization, in which the primary functional
A core tenet of neocortical functional specialization is that it fol- axis mainly anchored within the unimodal cortex during childhood
lows a sensory-association axis, where maturation of low-level sensory changes into the adult-like hierarchical gradient during adolescence12;
regions precede the expansion of higher-order association areas, such however, the influence of early thalamocortical interaction on the
as frontoparietal and default-mode networks11,12. This principle is also development of large-scale organization remains unmapped. Specifi-
reflected in the development of thalamocortical connectivity, where cally, it is unclear whether thalamocortical gradients mirror the hierar-
connections with sensorimotor and salience networks are already chical organization to cortico-cortical gradients (or exhibit a unique
established in neonates, whereas those with the frontal and default subcortex-specific connectivity pattern), and what computational
mode networks develop later13–15; however, as the thalamus consists principles underlie thalamus-mediated brain network generation.
of several dozen nuclei, understanding its global integration into the The current study investigated the macroscale organization of
macroscale cortical landscape, beyond the regional properties of indi- thalamocortical functional connectivity and its role in developing
vidual subnuclei, has proven challenging. Recent connectopic gradient cortical networks during infancy and childhood/young adulthood.
mapping techniques based on dimensionality reduction16–18 provide Continuous axes of thalamocortical connectivity were first extracted
a viable solution, as they summarize high dimensional connectome with connectopic gradient mapping and analyzed across different age
profiles of the brain regions (either the whole brain or specific cortical groups. Associations between cortical genes and thalamic neocorti-
areas such as the thalamus) into parsimonious axes representing their cal projection maps were then characterized through Gene Ontol-
core functional organization. ogy enrichment analyses. Finally, we utilized a generative network
Previous studies using this technique have reported a hierarchi- modeling approach to simulate the role of both thalamus and genetic
cal gradient from sensory-related, externally oriented processing on information in the assembly of large-scale functional networks.

Nature Neuroscience
Article
Through this multi-method approach, we found: (1) during
infancy, thalamocortical gradients could capture both preliminary
cortical hierarchy as well as a spatial distribution of developmen-
tal gene expression; (2) from childhood through young adulthood,
thalamocortical functional coupling undertakes a unique role in dif-
ferentiating between internally and externally oriented functional
networks. Indeed, the thalamic functional connectivity with the sali-
ence network serves as a pivotal point where the segregation between
the externally oriented (for example, visual/sensorimotor and dorsal
attention networks) and internally oriented (for example, default
mode network) systems gradually emerges during development; (3)
generative network modeling indicates that when synthetic networks
are constructed based on thalamocortical functional connectivity, they
show the highest resemblance to the networks derived from empirical
data. Taken together, initially there seems to be a balanced contribution
between thalamus and developmentally related genes in the functional
network specialization; however, this balance shifts when entering
into youth, where the thalamus contributes to functional segregation
through the decoupling of external and internal networks.
Results
We utilized two age-group datasets from the Human Connectome
Project (HCP19,20), consisting of individuals in infancy (n = 255, 149 male;
mean (s.d.) age = 39.7 (3.04) weeks; developing HCP (dHCP)21,22) and
childhood/young adulthood (n = 603, 417 male, mean (s.d.) age = 14.8
(3.88) years; HCP-Development (HCPD)23). Details on participant inclu-
sion and the following main analysis are provided in Methods and are
schematically summarized in Fig. 1.
Thalamocortical connectopic mapping
To delineate the macroscale organization of thalamocortical functional
connectivity, we applied the Congrads pipeline16 to the timeseries of
resting-state functional magnetic resonance imaging (fMRI). In short,
it applies the Laplacian Eigenmap dimensionality reduction to a simi-
larity matrix of the functional connectivity between the thalamus and
cortex. The resulting thalamocortical connectopic maps (CMAPs)
topographically represent how similar or dissimilar the thalamocorti-
cal connectivity patterns are across the thalamic voxels, whereas the
neocortical projection maps (NEOMAPs) show a cortical counterpart
to the CMAP (which thalamic area is functionally connected to which
neocortical area).
Thalamocortical connectopic gradients. The two group-level CMAPs
(‘functional gradients of the thalamus’; number of maps chosen
based on the variance explained; Supplementary Fig. 1) are shown for
infancy (29–44 weeks; Fig. 2a) and childhood/young adulthood (8–22
years; Fig. 2b). CMAP1 shows a gradient spanning from the anterior/
medial-to-posterior/lateral ends of the thalamus. This pattern is rela-
tively consistent across the two age groups (Fig. 2; top) and conveys an
approximated differentiation between low-level and high-order func-
tional networks according to the Yeo 7 network thalamic parcellation24
(Supplementary Fig. 1). Additionally, this pattern shows significant cor-
relations from the result of a recent study analyzing thalamic transcrip-
tomic and structural connectivity25 (Supplementary Fig. 2). CMAP2
(Fig. 2, bottom) comprises a gradient from superior-to-inferior axis
during infancy, and while the overall direction (superior-to-inferior)
remains the same in childhood/young adulthood, the inferior part of
the gradient tends to diverge into both anterior and posterior regions
in the latter age group. To test for robustness, we used an independent
dataset (Nathan Kline Institute-Rockland sample (NKI-RS); n = 70 (34
male), age range 6–19 years) and found similar CMAP patterns (corre-
lation between HCPD and NKI-RS, r > 0.70 for both CMAPs; Extended
Data Fig. 1a).
It should be noted that the HCP datasets that we analyzed for
infancy were obtained during sleep, whereas those for childhood/
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
young adulthood were collected during awake states. To investigate
this potential state-dependent effect in our findings, we further ana-
lyzed a separate dataset consisting of the scans taken during both
sleep and awake states within individuals (see Supplementary Note
‘Comparison of sleep–wake states’). We confirmed that the present
findings are not due to state differences but rather reflect differences
across developmental periods (Extended Data Fig. 2). Similarly, as
the thalamus is known for regulating arousal and activity levels26, we
confirmed that our main results were replicated after global signal
regression (Supplementary Fig. 3).
Thalamic neocortical projection maps. The NEOMAP (the neocorti-
cal counterpart of CMAP) of infancy and childhood/young adulthood
showed overall diverging patterns across developmental periods
(Fig. 2), which indicates a shift in the relationship between the thala-
mus and neocortical functional organization.
In infancy, the first NEOMAP showed anchors in the low-level
sensory and visual areas while the other end revealed less differenti-
ated patterns across the brain areas (Fig. 2a, top). This result, derived
based on an adult-based parcellation, was replicated when using
age-matched, fully data-driven infant parcellation27 (Supplementary
Fig. 4). The second NEOMAP portrayed a strong geometric pattern
that seems to directly reflect the dorsoventral stream of the second
thalamic CMAP (Fig. 2a, bottom). Meanwhile, in childhood/young
adulthood, the NEOMAPs show patterns that are more differentiated
(spatially confined to known functional parcellations), suggesting the
emergence of mature functional networks (Fig. 2b). Specifically, both
NEOMAPs commonly show that the salience network is situated at one
end of the gradient, whereas the opposite ends of the NEOMAPs are
distinct: NEOMAP1 extends toward externally oriented regions, such
as the dorsal attention and primary sensory networks (Fig. 2b, top),
whereas NEOMAP2 mainly is occupied by the internally oriented default
mode network (Fig. 2b, bottom). Taken together, the thalamocortical
projection seems to be involved in a large-scale organization in which
the salience network serves as a pivotal point for differentiation of
external and internal axes (Fig. 2c (ref. 28) and Extended Data Fig. 3).
This spatial representation between the two NEOMAPs along with their
patterns were replicated in the NKI-RS dataset (correlation between
HCPD and NKI-RS, r ≥ 0.70 for both NEOMAPs; Extended Data Fig. 1b).
Thalamic effects on external-to-internal axis segregation. To further
confirm the above hypothesis, we quantified several network-level
indices. First, we created a segregation index by calculating the dif-
ference of NEOMAP values between the salience network and exter-
nally oriented areas (SegregationSAL-EXT) as well as internally oriented
areas (SegregationSAL-INT) along five different age groups (Fig. 3a). As
expected, there was greater segregation between the salience network
and both externally and internally oriented networks during childhood/
young adulthood in comparison to infancy. This suggests that the
external–internal axis more prominently emerges during childhood/
young adulthood. To corroborate this observation, we adapted two
metrics that statistically estimate the degree of network segrega-
tion (silhouette scores and network overlap) and results confirmed
that functional networks in childhood-young adulthood are more
segregated compared to those in infancy (Extended Data Fig. 4 and
Supplementary Note ‘Quantitative analyses of network segregation’).
Finally, compared to externally oriented networks that tend to
be more regionally confined (visual and somatosensory networks),
those for internal processes seem to recruit widely distributed areas
across the cortex29. This differentiation may reflect distinct patterns
of underlying thalamic projections based on the relative density of
‘core’ and ‘matrix’ cells6,7. Parvalbumin-rich core cells, predominantly
located in sensory thalamic nuclei, project to specific areas of primary
sensory cortices, while calbindin-rich matrix cells, more common in
higher-order thalamic nuclei, project more diffusely across the cortex.
Article https://doi.org/10.1038/s41593-024-01679-3

a Infancy (29–44 wks) b Childhood – young adult (8–22 yrs) c Network profiling

CMAP 1 CMAP 1
Z = 18 Z=7 Z = 11 Z = –1

CMAP values
Visual

Max

CMAP values

Max
Somatomotor
A
Dorsal attention
A P P

Min
Salience

Min
A P L R L R A P L R L R

1.0 1.0
Limbic
0.8 NEOMAP 1 Frontoparietal
Default mode network
NEOMAP 1 0.6
0.5

NEOMAP values
NEOMAP values
0.4
0.2
0 0
–0.2
External
–0.4
–0.5
Projection map –0.6 Projection map
–0.8

Axis 1
–0.2 0.2 –1.0 –1.0 –0.5 0.5
Functional networks Functional networks

CMAP 2 CMAP 2

Thalamus
S S
Z = 18 Z=7 Z = 11 Z = –1

CMAP values

Max
Max
CMAP values
Salience

Min
Min
A P A P L R L R A P A P L R L R

Axis 2
1.0 1.0
0.8 NEOMAP 2
NEOMAP 2 0.6
0.5
NEOMAP values

0.4
NEOMAP values
0.2
0 0 Internal
–0.2
–0.4
–0.5 Projection map
–0.6
Projection map
–0.8
–1.0 –1.0
–0.5 0.5
–0.2 0.2 Functional networks Functional networks

Fig. 2 | Thalamic connectopic maps and its neocortical projection maps. extend to the highest value within 1.5 × IQR above the third quartile and to the
a,b, CMAP1 and 2 and NEOMAP1 and 2 are shown for infancy (a) and childhood/ lowest value within 1.5 × IQR below the first quartile. The IQR represents the range
young adulthood (b). Network profiles sorted out based on the Yeo-Krienan 7 between the 25th and 75th percentiles. c, A schematic of the external-to-internal
Network Atlas28 are shown in the box plots overlaid with jittered dots next to each axis division derived from the childhood/young adulthood NEOMAPs is depicted
NEOMAP (no. of vertices for VIS = 2,770, SOM = 3,784, DAN = 2,118, SAL = 2,250, along with the Yeo-Krienan 7 Network Atlas28 legend. VIS, visual network; SOM,
LIM = 1,437, FPN = 2,317 and DMN = 3,819). The box plots display the median somatosensory network; DAN, dorsal attention network; SAL, salience network;
(central line) and interquartile range (IQR; the extent of the box). Whiskers LIM, limbic network; FPN, frontoparietal network, DMN, default mode network.

Using partial least squares (PLS) analyses we found significant asso- between the salience–internal (t = 2.35, P = 0.02) and external–
ciation between the NEOMAPs and cyto-architecture-based thalamic internal (t = 2.63, P = 0.009) networks, as well as a marginal trend for
projections of core and matrix cells (r = 0.65, Pspin = 3.05 × 10−25). Moreo- the salience–external (t = 1.25, P = 0.21; Extended Data Fig. 6a–c) net-
ver, correlation between the latent constructs of the functional and works across development. Second, we applied a supervised learning
cytoarchitectural thalamic projections reveals an external-to-internal framework (fivefold cross-validation) to the same longitudinal data
axis with the salience network situated in the middle of the to elucidate the early influence of thalamocortical connectivity devel-
distribution (Fig. 3b). opment on the network differentiation in the neocortex during later
developmental stages. We found that the change in thalamocortical
Age-dependent changes in CMAPs and NEOMAPs. The thalamocor- NEOMAPs between baseline and follow-up can significantly predict the
tical CMAPs and NEOMAPs showed significant age effects (Extended functional segregation between internal and external networks within
Data Fig. 5 and Supplementary Note ‘Age-related differences in thalam- a ‘cortical’ gradient (mean average error = 1.91; median correlation
ocortical connectopic gradient and projection maps’ for statistical coefficient = 0.33, P = 0.005; Extended Data Fig. 6d).
details of following results). Specifically, in infancy, at each end of
the CMAPs, an overall expansion (more functional differentiation) Transcriptomic analyses of thalamocortical projection maps
occurred with age, given that positive gradient values in the anterior Thus far, we have shown that the thalamocortical functional connectiv-
regions show positive age effects while negative gradient values in the ity displays widespread neocortical projections with distinct patterns,
posterior parts of the thalamus show negative age effects, indicating reflecting a large-scale functional organization of cortical networks.
diverging age effects (PFDR < 0.05). A similar pattern is also depicted In brief, the infancy NEOMAP1 revealed a preliminary differentiation
in the NEOMAPs (PFDR < 0.05). In childhood/young adulthood, fairly between low-level sensory and high-order transmodal regions, whereas
consistent age effects are found, albeit weaker compared to that of the NEOMAPs of childhood/young adulthood collectively depicted a
infancy. Contrary to infancy which showed mainly a gradient expan- segregation between functional networks involved in externally versus
sion, the childhood/young adulthood group presented patches of age internally oriented processes, both anchored by the salience network.
effect across the CMAPs as well as NEOMAPs, indicating both expansion Compared to these, infancy NEOMAP2 (Fig. 2a, bottom) was not clear in
and convergence with age (PFDR < 0.05). A cross-sectional progressive terms of its biological implication; however, its graded expression along
change of the gradient value along the age is shown in the Supplemen- the dorsoventral axis resembles the characteristic areal patterning of
tary Video. the developing neocortex, which is typically driven by morphogens and
This age-dependent trend was also observed using the longitudinal transcription factors. To corroborate this observation, we examined its
data from NKI-RS. First, in the mixed-effects model (where we employed association with cortical gene expression, leveraging the Allen Human
age as a fixed effect and participant as a random effect), we examined Brain Atlas (AHBA) and the PLS analysis, an unsupervised multivari-
the segregation indices among the salience, external and internal ate correlation method30,31. To contextualize the multivariate genetic
networks. Our analysis revealed a significant increase in segregation association with respect to biological processes, we conducted a Gene

Nature Neuroscience
Article https://doi.org/10.1038/s41593-024-01679-3

a Segregation index calculated for each age group b Relation with core-matrix thalamic projections
Infancy (29–44 wks)
External–internal axis 1.6 1.6
NEOMAP 1 NEOMAP 2

Segregation

Segregation
External NEOMAP 1 CALB map
PLS
NEOMAP 1
0 0
7
–3 7–4
0 40 41 –44 7
–3 7–4
0 40 41 4
–4 wks
wks
29 42 29 42 NEOMAP 2 PVALB map
Thalamus

3 3
Salience
Childhood – young adult (8–22 yrs) r = 0.65, External
20 VN
NEOMAP 1 NEOMAP 2 Pspin = 3.05 × 10–25
1.6 1.6 SN

NEOMAP scores
DAN

Segregation

Segregation
NEOMAP 2 SAL
Internal LN
FPN
0 0 DMN
10 13 16 20 22 yrs 10 13 16 20 22 yrs Salience
8– 10– 13– 16– 20– 8– 10– 13– 16– 20– –20

SegregationSAL-EXT SegregationSAL-INT Internal

–0.1 CALB/PVALB scores 0.1

Fig. 3 | Thalamocortical influence on cortical network segregation.
a, The external–internal segregation is quantified and profiled along the age
subgroups. The differentiation between external and internal axes does not
clearly emerge until childhood/young adulthood periods. The first NEOMAP
represents segregation of the salience network with externally oriented networks
while the second NEOMAP shows segregation with the internally oriented
networks. Box plots display the median (central line) and IQR (the extent of
the box). Whiskers extend to the highest value within 1.5 × IQR above the third
quartile and to the lowest value within 1.5 × IQR below the first quartile. The
IQR represents the range between the 25th and 75th percentiles. The number of
participants for each age group are as follows: infancy (29–37 weeks (w), n = 49;
37–40 w, n = 59; 40 w, n = 47; 41 w, n = 48; 42–44 w, n = 52) and childhood/young
adult (8–10 years (y), n = 81; 10–13 y, n = 125; 13–16 y, n = 173; 16–20 y, n = 141;
20–22 y, n = 83). b, Association between childhood/young adulthood NEOMAPs
and core-matrix thalamic projections were assessed using two-tailed partial least
squares (PLS) regression. Statistical significance was determined by comparing
the empirical correlation of latent scores against a null distribution, generated
from the spatial permutation test (10,000 iterations) of PLS to control for spatial
autocorrelation (Pspin = 3.05 × 10−25). NEOMAP, neocortical projection map; SAL,
salience; EXT, external; INT, internal; CALB, calbindin; PVALB, parvalbumin.

Ontology enrichment analysis with ShinyGO v.0.77 (ref. 32). Finally, thalamus and cortex, significantly enriched from fetal through infancy
we performed a cell-type specific expression analysis (CSEA)33, using (FDR < 0.05; Fig. 4b).
the BrainSpan dataset to compare developmental expression profiles
across multiple developmental periods. Computational modeling analysis
The associations between gene expression and infancy NEOMAP2 Finally, to gain a mechanistic understanding of whether and how
revealed a single latent variable that was significant, after a spin test thalamocortical connectivity contributes to the development of cor-
preserving the original spatial autocorrelation (r = 0.65; Pspin < 0.001). tical functional organization, we utilized a recently proposed gen-
The latent variable explained 41.7% of the variance in NEOMAP2. The erative network modeling approach. This model uses defined rules or
post hoc loading analysis revealed a total of 813 strongly contributing constraints to probabilistically establish connections between nodes,
genes with both negative and positive loadings. Further gene enrich- generating simulated networks that mirror the essential characteristics
ment analysis revealed that the candidate negative and positive genes of real brain networks. Using this approach allows for direct compari-
were largely involved in different underlying biological processes. sons between simulated and empirical cortical networks, aiding the
Notably, multiple processes involved in the development of the cen- elucidation of the principles underlying cortical organization40. Our
tral nervous system (false discovery rate (FDR) = 0.0008), ranging main interests were to compare the relative effects of the thalamocorti-
from head/brain development (FDR = 0.002) to cell morphogenesis cal connectivity and gene expression in functional network generation.
involved in neuron differentiation (FDR = 0.008), were associated with For model comparison, we tested the effects of relevant spatial and
the negatively associated genes (Fig. 4a and Supplementary Table 1). topological features, as per previous studies41,42. Model performance
These results were replicated using other Gene Ontology enrichment was assessed based on the Kolmogorov–Smirnov (KS) statistic that
tools, Enrichr34–36 and g:Profiler37 (Supplementary Fig. 5). To confirm quantified the difference between synthetic and empirical networks
the specificity of our findings, we employed a stringent null-brain-gene based on node degree, node clustering, node-betweenness and edge
model that generates null distributions based on random genes that length distribution41.
are over-expressed in brain tissues (n = 2,711 selected by qFDR < 0.05, In general, the models based on developmental factors (gene
in statistically comparing brain tissues and other body tissues)31. The expression and thalamocortical connectivity) performed better than
association between the genes related to developmental processes and the spatial and topological factors (Fig. 5a and Extended Data Fig. 7).
infancy NEOMAP2 was stronger than expected for randomly selected The thalamocortical connectivity rules particularly yielded the best
brain genes (β = 0.49, Pnull-brain-gene < 0.001). Meanwhile, the genes with fitting to the empirical data of childhood/young adulthood (mean
positive loadings belonged to various pathways related to transmem- KSthal = 0.09, KSgene = 0.16, KSspatial = 0.45; Fig. 5a). Contrarily, gene
brane transport processes (Supplementary Table 2). expression and thalamocortical connectivity rules showed similar per-
While the AHBA38 was used due to its extensive whole-brain-wide formance in infancy (KSthal = 0.20; KSgene = 0.19; KSspatial = 0.43; Fig. 5a).
spatial coverage, we must note that it is derived from six postmor- We further reconstructed individual adjacency matrices derived from
tem adult brains. Thus, given the potential variance in gene expres- the simulations and calculated group-level cortical functional gradi-
sion between infancy and adulthood, we incorporated data from the ents, network modularity and segregation index. Notably, the func-
BrainSpan atlas39, which encompasses a wider age span from eight tional gradients derived from the synthetic networks built according
post-conception weeks to 40 years, albeit less spatial coverage. Using to the thalamocortical connectivity rule generated remarkably similar
CSEA to examine the developmental expression profiles of these genes patterns of the canonical cortical gradients during childhood/young
across specific cell types, we found concentrated expression in the adulthood (Fig. 5b). Significantly higher network modularity was

Nature Neuroscience
Article
a Chart of enriched GO terms
Central nervous system development
Head development
Multicell. organismal response to stress
Brain development
Reg. of membrane potential
Cell morphogenesis in neuron differentiation
Heterophilic cell–cell adhesion
Forebrain development
Cellular component morphogenesis
Trans-synaptic signaling
Cell–cell signaling
GABA signaling pathway
Chemical synaptic transmission
Anterograde trans-synaptic signaling
Synaptic signaling
Ion transmembrane transport
Neuron projection morphogenesis
Ion transport
Neuronal action potential
Cation transport
0 1 2 3
–log10(FDR)
Fig. 4 | Top enriched Gene Ontology terms indicating biological processes
showing significant association with the negatively enriched genes of
NEOMAP2 in infancy. a, Chart showing the enriched GO terms derived using
ShinyGO 0.77 (ref. 32). Fold enrichment is a measure of the degree to which
a particular term is overrepresented in the genes of interest compared to a
background set of genes. FDR indicates how likely the enrichment represents
a false-positive finding. b, For the cell-type-specific expression analysis33, the
overlap between our gene list and transcripts enriched in a particular brain
found in the synthetic networks based on thalamocortical connectiv-
ity compared to the other rules during childhood/young adulthood
(modularity Q = 0.53; Fig. 5c). Finally, corroborating our experimental
finding, the thalamocortical model showed the largest segregation
index between the salience and externally oriented as well as the inter-
nally oriented networks during childhood/young adulthood (Fig. 5d).
This was replicated across different network thresholds (Supplemen-
tary Fig. 6), after controlling for potential circularity issues due to
high resemblance between NEOMAPs and cortico-cortical gradients
(Extended Data Fig. 8), as well as in an independent dataset (NKI-RS;
Extended Data Fig. 9).
Next, to further pinpoint a direct effect of thalamocortical con-
nectivity on the functional organization across different ages, we
performed perturbation to thalamo-salience connectivity in a series
of growth models (an approach informing the network simulation with
age-dependent developmental pattern; Fig. 6a). Perturbations were
applied to different developmental periods (8–12 years; 12–16 years;
16–22 years) or to all of them, then compared to a no-perturbation
growth model. As expected, perturbations compromised the seg-
regation of both internal and external axes, but notably, it affected
the salience–internal axis more dramatically, suggesting a higher
contribution of thalamocortical connectivity on the development
of internal processing areas during youth (Fig. 6b). In particular, the
perturbation had a profound impact at a later stage of development
(>12 years), when most higher-level cognitive functions involving the
default mode network are actively developed43. These patterns were
similarly observed in the simulated cortical gradients across differ-
ent developmental periods (Fig. 6c). Finally, a series of ‘null models’
validated the specificity of our findings (whether perturbations to
thalamo-salience connections have a more dramatic effect on network
segregation compared to perturbations to other [primary sensory or
randomly-chosen] connections). Results show that perturbations to
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
b Developmental cell-specific enrichment analysis
Amyg Cere Cortex Hipp Str Thalamus
Fetal
No. of genes
10
20
30
40
Infancy
Fold enrichment
2
Childhood
4
6
8 Adolescence
10
12 Young adult
0.10 0.05 0
P values
region and developmental period is determined using the Fisher’s exact test
with Benjamini–Hochberg correction. The specificity of transcript expression is
assessed with a specificity index P value (PSI) across four thresholds (0.05, 0.01,
0.001 and 0.0001), where lower PSI values indicate those genes more specific to
a particular combination between brain region and developmental period.
The size of the hexagons from outside to center correspond to the PSI thresholds
(0.05, 0.01, 0.001 and 0.0001), respectively. Colors indicate the FDR-adjusted
P values.
thalamo-salience connections yield more decreased network segrega-
tion and less clearly differentiated cortical gradients compared to the
null models (Extended Data Fig. 10).
Discussion
Large-scale brain networks undergo extensive changes in their func-
tional specialization in terms of connectome profiles across develop-
ment. In the current study, we demonstrated, both experimentally and
theoretically, an age-dependent role of the thalamus in associating
multiple functional subsystems, with a special focus on externally and
internally oriented networks. We found that during infancy, the thala-
mus provides a functional scaffold for the hierarchical differentiation
between low-level and higher-order regions, while also interacting with
cortical genes involved in developmental processes; however, once cor-
tical functional arrangement stabilizes into distinct networks, the thala-
mus contributes to a system-level segregation between externally and
internally oriented systems. These results were substantiated through
computational simulations of generative networks. Specifically, the
thalamocortical connectivity constraints not only produced the most
realistic cortical networks, but also contributed to the segregation
between the functional subsystems and yielded connectopic topologies
that closely resemble the cortico-cortical gradients18. Furthermore,
perturbations to the thalamo-salience connectivity had distinct effects
for the external-to-internal segregation and cortical gradients in dif-
ferent developmental stages. Taken together, these results provide an
expanding perspective for the role of the thalamus in the manifestation
of large-scale functional networks in the developing brain.
Development of hierarchical functional brain networks
during infancy
Previous work has established the effect of thalamus in the early
development of structural brain organization44. Indeed, the structural
Article https://doi.org/10.1038/s41593-024-01679-3

a Model performance c Comparison of network modularity

Infancy Childhood – young adult Infancy Childhood – young adult
(29–44 wks) (8–22 yrs) (29–44 wks) (8–22 yrs)
1 1 Mean Mean
* * 0.7
Median
0.7
Median
* * 0.6 * 0.6 *

KSmax
KSmax

Modularity (Q)
0.5 0.5

0 0 0.4 * 0.4
*
*

e
e

l
l
or

or

en
en

ia
ia
0.3 0.3

at
at
C

G
G

al

al

Sp
Sp
Th

Th
0.2 0.2
*
b Cortical gradients extracted from simulated affinity matrices
0.1 0.1

0 0
Infancy Childhood – young adult ThalCort Gene Spatial ThalCort Gene Spatial
(29–44 wks) (8–22 yrs)

d Comparison of segregation index

ThalCort

ThalCort

Gradient 1 Gradient 1 Infancy Childhood – young adult

0.5 0.5
(29–44 wks) (8–22 yrs)

SegregationSAL-EXT

SegregationSAL-EXT
0.4 0.4

Gradient 2 Gradient 2 0.3 0.3

0.2 0.2

0.1 0.1
Gene

Gene

Gradient 1 Gradient 1 0 0
0.5 0.5

SegregationSAL-INT

SegregationSAL-INT
0.4 0.4
Gradient 2 Gradient 2
0.3 0.3

0.2 0.2
Spatial
Spatial

Gradient 1 Gradient 1 0.1 0.1

Max
0 0

ThalCort Gene Spatial

Gradient 2 Gradient 2 Min

Fig. 5 | Generative network modeling. a, Model performance is measured
using KS statistics, where lower values imply greater similarity to empirical
data. Two-tailed t-tests, adjusted for multiple comparisons with Bonferroni
correction at a significance threshold of P = 0.001, were employed to compare
performance metrics. For infancy, significant differences were found for gene
versus spatial (t = −43.4, P = 2.2102 × 10−119) and thalcort versus spatial (t = −24.4,
P = 1.5961 × 10−68), at a sample size (n) of 255. There was no significant difference
for gene versus thalcort wiring rules (t = −1.4972, P = 0.14). For childhood/young
adulthood, significant group differences were found for all three contrasts:
thalcort versus gene (t = −52.6, P = 2.6144 × 10−227), thalcort versus spatial
(t = −288.8, P = 0) and gene versus spatial (t = −249.5, P = 0), with Bonferroni
correction (n = 603). b, The first two cortical gradients extracted from the affinity
matrices of the synthetic networks are shown for each different wiring rule.
c, Comparison of network modularity, calculated from each synthetic network
based on different wiring rules, were analyzed using ANOVA (n = 603) with
Bonferroni correction, resulting in a significant group difference for infancy
(F = 1,418.9, P = 1.2103 × 10−257) and childhood/young adulthood (F = 516.56,
P = 3.8966 × 10−178). d, Segregation indices computed from the individual
synthetic networks in infancy (n = 255) and childhood-adulthood (n = 603)
based on different wiring rules are shown with error bars indicating s.d. thalcort,
thalamocortical connectivity; gene, correlated gene expression; SAL, salience
network; EXT, externally oriented networks; INT, internally oriented networks;
w, weeks; y, years.

properties such as myelin and cortical thickness develop earliest in
the cortical regions that receive first-order thalamic relays45. Mean-
while, protracted maturation in the transmodal cortex, associated
with high-order thalamic relays, shows experience-dependent plas-
ticity11,46. Here, we extend previous findings and reveal that the thala-
mus in the infant brain shows functional coupling with the neocortex
primarily anchored in sensorimotor and visual areas while those
related to presumably higher-order regions are still not fully special-
ized13,47. This delay in development may be because the formation of
higher-order functional networks premises an integration of distrib-
uted regions across the cortex, which in turn requires the establishment
of long-range connections and efficient synaptic pruning that occurs
throughout childhood48,49. In addition, our findings are in line with the
established pattern of specialized connections between thalamus and
cortical regions. Infancy NEOMAP1 showed that low-level sensory corti-
cal regions have high connectivity with the posterior/lateral thalamic
regions, such as the pulvinar, medial geniculate nucleus and ventropo-
sterolateral nucleus, which are known to project to visual, auditory and
somatosensory regions, respectively6. On the other hand, the anterior
regions of the thalamus (for example, anterior, mediodorsal and ventral
anterior nuclei), which have been linked to higher-order functions6,
demonstrated strong connections with the prefrontal cortex and pos-
terior association regions.
During the perinatal period, there is an upsurge of strengthened
connectivity between neurons, subsequently followed by a more
prolonged selective pruning process that continues on until young
adulthood50,51. In this initial stage, various cellular mechanisms such
as synaptogenesis, axonogenesis and dendritic arborization underlie
the dynamic changes pertaining to postnatal surface expansion51–53. In
our study, infancy NEOMAP2 showed associations with genes enriched
for related processes such as axonogenesis, neuron differentiation
as well as cell projection morphogenesis and signaling. This is in line
with previous studies that support the critical role of thalamic axonal
projections in the formation and delineation of neocortical regions54,55.
Alternatively, the tilted inferior–superior axis depicted in NEOMAP2
reflecting the thalamocortical connectopic pattern may be interpreted
in light of the Dual Origin theory56. This theory conceptualizes the
sequential progression of cortical differentiation as originating from
the piriform cortex (paleocortex) and the hippocampus (archicor-
tex)56,57 and expanding to the sensorimotor and prefrontal cortical
areas, which results similarly in the inferior–superior axis of NEOMAP2.
Decoupling of external-to-internal axis during childhood/
young adulthood
A core aspect of cortical organization that underlies cognitive devel-
opment is captured in the organizing axis from sensory to association

Nature Neuroscience
Article https://doi.org/10.1038/s41593-024-01679-3

a Perturbation of thalamo-salience connectivity in the growth model c Simulated cortical gradients

Salience
8–12 yrs 12–18 yrs 18–22 yrs
network
Perturb Perturb Perturb

Thalamus Thalamo-salience
connectivity rule
No perturbation

Iterative edge generation based on wiring rules Perturb: 8–12 yr

b Comparison of segregation indices

0.5
s

Perturb: 12–18 yr
up

0.4
r
yr

y
ion

ro
2y

22
8

ll g
at

8–
2–
–1
rb

:a
:8

:1

1.6%
:1
rtu

rb

rb
rb
rb
Segregation index

pe

rtu

rtu
rtu
rtu

0.3
No

Pe

Pe
Pe
Pe

3.4% 26.6%
11.9%
16.0%
0.2 Perturb: 18–22 yr
45.5%

45.2%
0.1 0.03

97.8%

0
Salience - external Salience - internal Perturb: all groups –0.02

Fig. 6 | Perturbation of developmentally informed growth models.
a, A schematic of the growth model based on thalamo-salience connectivity
wiring rules that changed over the age span. In comparison to a non-perturbation
model, four models that perturbed the thalamo-salience connectivity in differing
developmental age groups were tested as follows: (1) perturbation applied to the
8–12 years age group; (2) 12–16 years age group; (3) 16–22 years age group; and (4)
all age groups. b, The segregation indices (salience–external; salience–internal)
were calculated for each growth model. The percentage indicates the amount
of difference compared to the non-perturbation model. c, Cortical gradients
extracted from the simulated affinity matrix of each growth model are shown.

functional areas. This principle encompasses a wide range of develop-
mental features related to the cortical organization of cyto-architecture,
structural connectivity and functional connectivity58. Our study found
that this principle is summarized in the ‘two main axes’ of thalamic
cortical function. NEOMAP1 ranges from salience network to other
cortical networks involved in externally oriented processes such as
the dorsal attention and low-level sensory regions. These networks
are externally driven and phasic in nature, implying a selective atten-
tional process to specific sensory input59,60. In NEOMAP2, an internally
oriented system (such as the default mode network), is situated at the
other end of the salience network. This gradient may represent how
the thalamus contributes to shaping internal mental model building,
or generative models of the environment61. In fact, the retrosplenial
cortex, a region closely associated with the default mode network,
shows strong connections with the anterior thalamic nuclei and hip-
pocampus, which are regions that may coordinate to update existing
mental representations62–64.
Of utmost importance, the salience network, a key component of
both NEOMAPs, shows integrative connectivity across wider distances
and diverse regions of the cortex, thus playing a critical role in facilitat-
ing communication between multi-level systems65. It is largely responsi-
ble for selecting the most homeostatically relevant information among
the complex sensory inputs from the external world to continuously
update our internal model66. This process entails the coordination of
brain network dynamics between externally and internally oriented
cognitive processing, which the salience network is most capable of,
given its position in the middle of the ‘sensory–fugal’ axis67. In sum,
thalamocortical projections encompassing these networks may engage
in the selection and gating of external inputs of high salience as a means
of allocating attentional resources and updating previous beliefs.
Implication of thalamocortical projections in the generative
model for developing brain networks
Building upon the neuroimaging findings, we further confirmed that
thalamocortical connectivity is indeed a core mechanism for large-scale
functional organization, based on generative network modeling. In pre-
vious studies, the synthetically generated networks have been shown
to faithfully recapitulate major characteristics of network organiza-
tion that are observed in the empirical data41,42. In our study as well,
when network generation was guided by thalamocortical connectiv-
ity, functional segregation of cortical networks naturally emerged,
and its patterns were strikingly similar with those from the real brain,
which was not the case in the models informed by different wiring
rules. Specifically, compared to the models based on spatial distance,
topological features or genetic correlations, the formation of a hier-
archical brain network with meaningful functional segregation was
driven only by the thalamocortical connectivity model. This pertains
to our modeling results from childhood/young adulthood and not
infancy, which confirms the diverging associations between CMAP
and NEOMAP between the two age groups. Using a growth model that

Nature Neuroscience
Article
allows for the tracking of brain network generation across the itera-
tions (analogous to developmental phases in the real brain)41, we found
that the thalamic influence on functional segregation is exceedingly
greater during later developmental stages. Particularly, compared to
the externally oriented functional regions, segregation of the internal
area is more strongly modulated by thalamo-salience connectivity,
implying its role in the maturation of the default mode network. Our
results provide a mechanistic clue of how thalamocortical connectivity
affects the whole-brain functional connecto-genesis.
Limitations and future directions
Several limitations should be noted. First, our results were based on a
static (‘non-dynamic’) model of thalamocortical connectivity analyses;
however, recent emphasis on non-stationary aspects of brain activ-
ity suggests that future research could explore the thalamus’s role in
cortical synchronization using different analytical approaches such
as co-activation pattern analyses68. Second, although different condi-
tions for data acquisition (sleep and awake) limits our interpretation,
we found that the underlying architecture remains unchanged in an
independent dataset of individuals scanned in both states. Future data
collection should consider a more systematic investigation for test/
retest reliability of the results across different arousal states. Finally,
as our transcriptomic findings are based on the AHBA, which utilizes
postmortem tissue, it is pertinent to mention a recent report highlight-
ing the differences in gene expression compared to living brain tissues69.
Although a noticeable caution on future gene-imaging analysis, their
findings were based on regional tissue sampled from the prefrontal
cortex of individuals with Parkinson disease, thus cannot be general-
ized to undermine the utility of postmortem analyses, especially those
based on spatial correlation across the whole brain70.
All together, we revealed age-dependent differences in thalamo-
cortical connectopic gradients and their corresponding neocortical
projection maps. Neuroimaging and transcriptomic analyses gave
us new insights into its shifting role across infancy and childhood/
young adulthood in macroscale functional organization, whereas
computational models confirmed the contribution of thalamocorti-
cal connectivity in functional network segregation. While our study
focused on the interactions between the thalamus and cortex, further
studies may consider the extensive connections that thalamus has with
other subcortical regions such as the basal ganglia and cerebellum71.
Additionally, considering a recent generative network model which
accounts for the strength of connectivity weights in the network
formation72, future research may incorporate weighted thalamo-
cortical connectivity to reveal a more nuanced picture of thalamus’
developmental role in the emergence of large-scale functional brain
organization. Finally, as the thalamus is a critical region for regulating
cortical networks, alterations in thalamocortical connectivity may
be a harbinger of aberrant whole-brain connectivity that underlies
various developmental conditions, including schizophrenia, epilepsy
and autism73–76.
Online content
Any methods, additional references, Nature Portfolio reporting sum-
maries, source data, extended data, supplementary information,
acknowledgements, peer review information; details of author contri-
butions and competing interests; and statements of data and code avail-
ability are available at https://doi.org/10.1038/s41593-024-01679-3.
References
1. Cadwell, C. R., Bhaduri, A., Mostajo-Radji, M. A., Keefe, M. G. &
Nowakowski, T. J. Development and arealization of the cerebral
cortex. Neuron 103, 980–1004 (2019).
2. O’Leary, D. D. M. & Nakagawa, Y. Patterning centers, regulatory
genes and extrinsic mechanisms controlling arealization of the
neocortex. Curr. Opin. Neurobiol. 12, 14–25 (2002).
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
3. Rakic, P. Specification of cerebral cortical areas. Science 241,
170–176 (1988).
4. O’Leary, D. D. Do cortical areas emerge from a protocortex?
Trends Neurosci. 12, 400–406 (1989).
5. Wolff, M., Morceau, S., Folkard, R., Martin-Cortecero, J. & Groh, A.
A thalamic bridge from sensory perception to cognition.
Neurosci. Biobehav. Rev. 120, 222–235 (2021).
6. Jones, E. G. The Thalamus (Springer Science & Business Media,
2012).
7. Müller, E. J. et al. Core and matrix thalamic sub-populations relate
to spatio-temporal cortical connectivity gradients. Neuroimage
222, 117224 (2020).
8. Fox, S. E., Levitt, P. & Nelson, C. A. 3rd How the timing and
quality of early experiences influence the development of brain
architecture. Child Dev. 81, 28–40 (2010).
9. Tooley, U. A., Bassett, D. S. & Mackey, A. P. Environmental
influences on the pace of brain development. Nat. Rev. Neurosci.
22, 372–384 (2021).
10. Molnár, Z. & Blakemore, C. How do thalamic axons find their way
to the cortex? Trends Neurosci. 18, 389–397 (1995).
11. Sydnor, V. J. et al. Neurodevelopment of the association cortices:
patterns, mechanisms, and implications for psychopathology.
Neuron 109, 2820–2846 (2021).
12. Dong, H.-M., Margulies, D. S., Zuo, X.-N. & Holmes, A. J. Shifting
gradients of macroscale cortical organization mark the transition
from childhood to adolescence. Proc. Natl Acad. Sci. USA 118,
e2024448118 (2021).
13. Alcauter, S. et al. Development of thalamocortical connectivity
during infancy and its cognitive correlations. J. Neurosci. 34,
9067–9075 (2014).
14. Fair, D. A. et al. Maturing thalamocortical functional connectivity
across development. Front. Syst. Neurosci. 4, 10 (2010).
15. Larivière, S. et al. Multiscale structure–function gradients in the
neonatal connectome. Cereb. Cortex 30, 47–58 (2020).
16. Haak, K. V., Marquand, A. F. & Beckmann, C. F. Connectopic
mapping with resting-state fMRI. Neuroimage 170, 83–94 (2018).
17. Vos de Wael, R. et al. BrainSpace: a toolbox for the analysis
of macroscale gradients in neuroimaging and connectomics
datasets. Commun. Biol. 3, 103 (2020).
18. Margulies, D. S. et al. Situating the default-mode network along a
principal gradient of macroscale cortical organization. Proc. Natl
Acad. Sci. USA 113, 12574–12579 (2016).
19. Van Essen, D. C. et al. The WU-Minn Human Connectome Project:
an overview. Neuroimage 80, 62–79 (2013).
20. Glasser, M. F. et al. The minimal preprocessing pipelines
for the Human Connectome Project. Neuroimage 80,
105–124 (2013).
21. Makropoulos, A. et al. The developing human connectome
project: a minimal processing pipeline for neonatal cortical
surface reconstruction. Neuroimage 173, 88–112 (2018).
22. Fitzgibbon, S. P. et al. The developing Human Connectome
Project (dHCP) automated resting-state functional processing
framework for newborn infants. Neuroimage 223, 117303 (2020).
23. Somerville, L. H. et al. The Lifespan Human Connectome Project
in Development: a large-scale study of brain connectivity
development in 5–21 year olds. Neuroimage 183, 456–468 (2018).
24. Raut, R. V., Snyder, A. Z. & Raichle, M. E. Hierarchical dynamics as
a macroscopic organizing principle of the human brain. Proc. Natl
Acad. Sci. USA 117, 20890–20897 (2020).
25. Oldham, S. & Ball, G. A phylogenetically-conserved axis of
thalamocortical connectivity in the human brain. Nat. Commun.
14, 6032 (2023).
26. Schiff, N. D. Central thalamic contributions to arousal regulation
and neurological disorders of consciousness. Ann. NY Acad. Sci.
1129, 105–118 (2008).
Article
27. Wang, F. et al. Fine-grained functional parcellation maps of
the infant cerebral cortex. eLife 12, e75401 (2023).
28. Yeo, B. T. T. et al. The organization of the human cerebral cortex
estimated by intrinsic functional connectivity. J. Neurophysiol.
106, 1125–1165 (2011).
29. Sepulcre, J. et al. The organization of local and distant functional
connectivity in the human brain. PLoS Comput. Biol. 6,
e1000808 (2010).
30. Fulcher, B. D., Arnatkeviciute, A. & Fornito, A. Overcoming
false-positive gene-category enrichment in the analysis of
spatially resolved transcriptomic brain atlas data. Nat. Commun.
12, 2669 (2021).
31. Wei, Y. et al. Statistical testing in transcriptomic-neuroimaging
studies: A how-to and evaluation of methods assessing spatial
and gene specificity. Hum. Brain Mapp. 43, 885–901 (2022).
32. Ge, S. X., Jung, D. & Yao, R. ShinyGO: a graphical gene-set
enrichment tool for animals and plants. Bioinformatics 36,
2628–2629 (2020).
33. Dougherty, J. D., Schmidt, E. F., Nakajima, M. & Heintz, N.
Analytical approaches to RNA profiling data for the identification
of genes enriched in specific cells. Nucleic Acids Res. 38,
4218–4230 (2010).
34. Kuleshov, M. V. et al. Enrichr: a comprehensive gene set
enrichment analysis web server 2016 update. Nucleic Acids Res.
44, W90–7 (2016).
35. Chen, E. Y. et al. Enrichr: interactive and collaborative HTML5
gene list enrichment analysis tool. BMC Bioinform. 14, 128 (2013).
36. Xie, Z. et al. Gene set knowledge discovery with Enrichr. Curr.
Protoc. 1, e90 (2021).
37. Kolberg, L. et al. g:Profiler-interoperable web service for
functional enrichment analysis and gene identifier mapping
(2023 update). Nucleic Acids Res. 51, W207–W212 (2023).
38. Hawrylycz, M. J. et al. An anatomically comprehensive atlas of the
adult human brain transcriptome. Nature 489, 391–399 (2012).
39. Kang, H. J. et al. Spatio-temporal transcriptome of the human
brain. Nature 478, 483–489 (2011).
40. Betzel, R. F. & Bassett, D. S. Generative models for network
neuroscience: prospects and promise. J. R. Soc. Interface 14,
20170623 (2017).
41. Oldham, S. et al. Modeling spatial, developmental, physiological,
and topological constraints on human brain connectivity. Sci.
Adv. 8, eabm6127 (2022).
42. Betzel, R. F. et al. Generative models of the human connectome.
Neuroimage 124, 1054–1064 (2016).
43. Paus, T. Mapping brain maturation and cognitive development
during adolescence. Trends Cogn. Sci. 9, 60–68 (2005).
44. Nakagawa, Y. Development of the thalamus: From early
patterning to regulation of cortical functions. Wiley Interdiscip.
Rev. Dev. Biol. 8, e345 (2019).
45. Ball, G. et al. Development of cortical microstructure in
the preterm human brain. Proc. Natl Acad. Sci. USA 110,
9541–9546 (2013).
46. Mitchell, A. S. The mediodorsal thalamus as a higher
order thalamic relay nucleus important for learning and
decision-making. Neurosci. Biobehav. Rev. 54, 76–88 (2015).
47. Toulmin, H. et al. Functional thalamocortical connectivity at term
equivalent age and outcome at 2 years in infants born preterm.
Cortex 135, 17–29 (2021).
48. Huttenlocher, P. R. & Dabholkar, A. S. Regional differences in
synaptogenesis in human cerebral cortex. J. Comp. Neurol. 387,
167–178 (1997).
49. Petanjek, Z., Judas, M., Kostović, I. & Uylings, H. B. M. Lifespan
alterations of basal dendritic trees of pyramidal neurons in the
human prefrontal cortex: a layer-specific pattern. Cereb. Cortex
18, 915–929 (2008).
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
50. Changeux, J. P. & Danchin, A. Selective stabilisation of developing
synapses as a mechanism for the specification of neuronal
networks. Nature 264, 705–712 (1976).
51. Huttenlocher, P. R. Synaptic density in human frontal cortex -
developmental changes and effects of aging. Brain Res. 163,
195–205 (1979).
52. Petanjek, Z. et al. Extraordinary neoteny of synaptic spines
in the human prefrontal cortex. Proc. Natl Acad. Sci. USA 108,
13281–13286 (2011).
53. Stiles, J. & Jernigan, T. L. The basics of brain development.
Neuropsychol. Rev. 20, 327–348 (2010).
54. Vue, T. Y. et al. Thalamic control of neocortical area formation in
mice. J. Neurosci. 33, 8442–8453 (2013).
55. Chou, S.-J. et al. Geniculocortical input drives genetic distinctions
between primary and higher-order visual areas. Science 340,
1239–1242 (2013).
56. Pandya, D., Petrides, M. & Cipolloni, P. B. Cerebral Cortex:
Architecture, Connections, and the Dual Origin Concept (Oxford
University Press, 2015).
57. Goulas, A., Margulies, D. S., Bezgin, G. & Hilgetag, C. C. The
architecture of mammalian cortical connectomes in light of
the theory of the dual origin of the cerebral cortex. Cortex 118,
244–261 (2019).
58. Hilgetag, C. C., Goulas, A. & Changeux, J.-P. A natural cortical axis
connecting the outside and inside of the human brain.
Netw. Neurosci. 6, 950–959 (2022).
59. Corbetta, M. & Shulman, G. L. Control of goal-directed and
stimulus-driven attention in the brain. Nat. Rev. Neurosci. 3,
201–215 (2002).
60. Vossel, S., Geng, J. J. & Fink, G. R. Dorsal and ventral attention
systems: distinct neural circuits but collaborative roles.
Neuroscientist 20, 150–159 (2014).
61. Friston, K. The free-energy principle: a unified brain theory? Nat.
Rev. Neurosci. 11, 127–138 (2010).
62. Corcoran, K. A., Frick, B. J., Radulovic, J. & Kay, L. M. Analysis of
coherent activity between retrosplenial cortex, hippocampus,
thalamus, and anterior cingulate cortex during retrieval of recent
and remote context fear memory. Neurobiol. Learn. Mem. 127,
93–101 (2016).
63. Halassa, M. M. & Kastner, S. Thalamic functions in
distributed cognitive control. Nat. Neurosci. 20, 1669–1679
(2017).
64. Wolff, M. & Vann, S. D. The cognitive thalamus as a gateway to
mental representations. J. Neurosci. 39, 3–14 (2019).
65. Menon, V. & Uddin, L. Q. Saliency, switching, attention and
control: a network model of insula function. Brain Struct. Funct.
214, 655–667 (2010).
66. Seeley, W. W. The salience network: a neural system for
perceiving and responding to homeostatic demands. J. Neurosci.
39, 9878–9882 (2019).
67. Keller, A. S. et al. Hierarchical functional system development
supports executive function. Trends Cogn. Sci. 27, 160–174
(2023).
68. Liu, X. & Duyn, J. H. Time-varying functional network information
extracted from brief instances of spontaneous brain activity. Proc.
Natl Acad. Sci. USA 110, 4392–4397 (2013).
69. Liharska, L. E. et al. A study of gene expression in the living human
brain. Preprint at medRxiv https://doi.org/10.1101/2023.04.21.2328
8916. (2023).
70. Collado-Torres, L. et al. Comparison of gene expression in living
and postmortem human brain. Preprint at medRxiv https://doi.org/
10.1101/2023.11.08.23298172 (2023).
71. Shine, J. M. The thalamus integrates the macrosystems of the
brain to facilitate complex, adaptive brain network dynamics.
Prog. Neurobiol. 199, 101951 (2021).
Article https://doi.org/10.1038/s41593-024-01679-3

72. Akarca, S. et al. A weighted generative model of the human
connectome. Preprint at bioRxiv https://doi.org/10.1101/
2023.06.23.546237 (2023).
73. Bernhardt, B. C., Bernasconi, N., Kim, H. & Bernasconi, A. Mapping
thalamocortical network pathology in temporal lobe epilepsy.
Neurology 78, 129–136 (2012).
74. Hwang, W. J. et al. Thalamic connectivity system across
psychiatric disorders: current status and clinical implications.
Biol. Psychiatry Glob. Open Sci. 2, 332–340 (2022).
75. Park, B.-Y. et al. Differences in subcortico-cortical interactions
identified from connectome and microcircuit models in autism.
Nat. Commun. 12, 2225 (2021).
76. Park, S. et al. Atypical integration of sensory-to-transmodal
functional systems mediates symptom severity in autism. Front.
Psychiatry 12, 699813 (2021).
77. Burt, J. B., Helmer, M., Shinn, M., Anticevic, A. & Murray, J. D.
Generative modeling of brain maps with spatial autocorrelation.
Neuroimage 220, 117038 (2020).
Publisher’s note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Springer Nature or its licensor (e.g. a society or other partner) holds
exclusive rights to this article under a publishing agreement with
the author(s) or other rightsholder(s); author self-archiving of the
accepted manuscript version of this article is solely governed by the
terms of such publishing agreement and applicable law.
© The Author(s), under exclusive licence to Springer Nature America,
Inc. 2024

1
Center for Neuroscience Imaging Research, Institute for Basic Science, Sungkyunkwan University, Suwon, South Korea. 2Autism Center, Child Mind
Institute, New York, NY, USA. 3Department of Cognitive Science and Artificial Intelligence, Tilburg School of Humanities and Digital Sciences, Tilburg
University, Tilburg, The Netherlands. 4Donders Centre for Cognitive Neuroimaging, Donders Institute, Radboud University, Radboud, The Netherlands.
5
Developmental Imaging, Murdoch Children’s Research Institute, Parkville, Victoria, Australia. 6The Turner Institute for Brain and Mental Health, School
of Psychological Sciences and Monash Biomedical Imaging, Monash University, Clayton, Victoria, Australia. 7Center for Integrative Developing Brain,
Child Mind Institute, New York, NY, USA. 8Department of Data Science, Inha University, Incheon, South Korea. 9Department of Biomedical Sciences and
Biomedical Imaging Research Institute, Cedars-Sinai Medical Center, Los Angeles, CA, USA. 10Department of Bioengineering, University of California
at Los Angeles, Los Angeles, CA, USA. 11Department of Medicine, University of California at Los Angeles, Los Angeles, CA, USA. 12Max Planck Institute
for Human Cognitive and Brain Sciences, Leipzig, Germany. 13Institute of Neuroscience and Medicine (INM-7), Brain and Behavior, Forschungszentrum,
Juelich, Germany. 14Center for the Developing Brain, Child Mind Institute, New York, NY, USA. 15Nathan S. Kline Institute for Psychiatric Research,
Orangeburg, NY, USA. 16McConnell Brain Imaging Centre, Montreal Neurological Institute and Hospital, McGill University, Montreal, Quebec, Canada.
17
Department of Biomedical Engineering, Sungkyunkwan University, Suwon, South Korea. 18Department of Intelligent Precision Healthcare Convergence,
Sungkyunkwan University, Suwon, South Korea. 19Department of MetaBioHealth, Sungkyunkwan University, Suwon, South Korea.
e-mail: hongseokjun@skku.edu

Nature Neuroscience
Article
Methods
Data acquisition and preprocessing
Infancy. We analyzed preprocessed neuroimaging data from the
second release of the Developing Human Connectome Project
(dHCP) neonatal data (http://www.developingconnectome.org/
second-data-release/). Individuals with both structural and functional
data were included. A neuroradiologist provided a radiology score
for all anatomical scans. A score of 3 (incidental findings with unlikely
clinical relevance but possible analytical relevance) or 5 (incidental
findings with possible/likely relevance for both clinical and imaging
analysis) were excluded from further analyses. After quality control and
preprocessing, the final sample consisted of 195 healthy term infants
(107 male, mean (s.d.) gestational age at birth = 39.9 (1.26) weeks, mean
(s.d.) age at scan = 41.0 (1.67) weeks) and 60 preterm infants (42 male,
mean (s.d.) gestational age at birth = 32.6 (3.12) weeks, mean (s.d.) age
at scan = 35.5 (2.78) weeks).
MRIs were acquired on a 3T Philips Achieva scanner with a dedi-
cated neonatal imaging system and a neonatal 32-channel phased array
head coil at the Evelina Neonatal Imaging Centre, St Thomas’ Hospital,
London. All infants were scanned without sedation during natural sleep.
Structural images (T1w, T2w) were acquired with a 0.8-mm isotropic
resolution and a thickness of 1.6 mm, overlapped by 0.8 mm (T2w, TE/
TR = 156 ms/12 s; T1w, TE/TR/TI = 8.7/4,795/1,740 ms). Resting-state
fMRI were acquired using multiband accelerated echo-planar imag-
ing for approximately 15 min (TE/TR = 38 ms/392 ms, flip angle = 34°,
spatial resolution = 2.15 mm isotropic, 2,300 volumes).
Details of the specific preprocessing procedures can be found
elsewhere21,22. In brief, structural images were motion-corrected,
reconstructed, bias-corrected, brain extracted for the segmentation
procedure using the DrawEM algorithm (https://github.com/MIRTK/
DrawEM)78. Then, cortical surfaces were extracted, inflated, and pro-
jected to a sphere and aligned to the 40-week dHCP surface template
using Multimodal Surface Matching (https://github.com/ecr05/
MSM_HOCR ref. 79,80). Minimal preprocessing of the resting-state
fMRI data included susceptibility field distortion correction, slice
timing and motion correction, registration to both individual struc-
tural image and 40-week group template, temporal high-pass filtering
(150 s high-pass cutoff) and independent component analysis denois-
ing. The first ten volumes were discarded in subsequent analyses to
achieve steady-state magnetization of the fMRI signal, resulting in
2,290 volumes. We excluded participants with mean framewise dis-
placement (FD) > 0.4 mm, which enables a tradeoff between sample
size and data quality.
Childhood/young adult. Minimally preprocessed neuroimaging data
of typically developing participants aged 8–21 years were downloaded
from the Lifespan HCPD (release 2.0) at the National Institute of Mental
Health Data Archive (NDA)23. After quality control and preprocessing,
the final sample consisted of 603 healthy participants (417 male, mean
(s.d.) age = 14.8 (3.88) years). Participants were excluded for insufficient
English, premature birth, MRI contraindications or a lifetime history
of serious medical conditions, endocrine disorders, head injury or
psychiatric/neurodevelopmental disorders.
Participants were scanned on a Siemens 3T Prisma (32-channel
head coil) using matched protocols across four sites (Harvard Uni-
versity, University of California Los Angeles, University of Minne-
sota and Washington University in St. Louis) in the United States.
T1-weighted images were acquired using a 3D multi-echo MPRAGE
sequence (TR = 2,500, TI = 1,000 ms, TE = 1.8/3.6/5.4/7.2 ms, flip
angle = 8°, FOV = 256 × 240 × 166 mm, 320 × 300 matrix, 208 slices,
0.8 mm isotropic voxels). Resting-state fMRI data were acquired with
a multiband gradient-recalled EPI sequence (TR = 800 ms, TE = 37 ms,
flip angle = 52°, FOV = 208 mm, 104 × 90 matrix, 72 oblique axial slices,
2-mm isotropic voxels, multiband acceleration factor of 8, 1,912 vol-
umes). A total of four fMRI scans were acquired in an eyes-open with
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
passive crosshair viewing condition for approximately 26 min in both
the AP/PA phase-encoding directions.
Details of the specific preprocessing procedures (HCP Pipelines
v.3.22) can be found elsewhere19,20. In brief, for structural data, bias-field
and gradient distortion correction, alignment between T1w and T2w
images and registration to MNI space, was followed by the typical Free-
Surfer procedures including brain extraction, intensity normalization,
cortical parcellation and subcortical segmentation. For functional
data, after applying correction for gradient nonlinearity, motion,
fieldmap-based EPI distortion, boundary-based registration of EPI
to T1w, non-linear (FNIRT) registration to MNI152 space and intensity
normalization, the volumetric data were projected into standard gray-
ordinate space (which includes cortical surface vertices and subcortical
voxels) and then smoothed with a 2-mm FWHM Gaussian kernel. The
first ten volumes were discarded, resulting in 1,902 volumes. As with
the dHCP data, FD was calculated and participants with a mean-FD that
exceeded 0.4 mm were excluded from the final analyses.
To ensure replicability, we also analyzed independent,
cross-sectional and longitudinal pediatric/adolescence datasets from
the NKI-RS (n = 70; age range 6–19 years; mean ± s.d. age = 10.9 ± 2.9).
We selected participants without any previous diagnosis and who
had at least two time points of fMRI data. As this dataset did not meet
the HCP Pipelines’ acquisition requirements (lacking T2w images
and/or fieldmaps), alternative processing methods were employed.
First, we ran FreeSurfer’s recon_all81 and the ciftify_recon_all from the
ciftify toolbox82, which essentially mimics the HCP Anatomical pipe-
lines. Next, fMRI data were preprocessed with fMRIprep83, followed
by ciftify_subject_fmri which generated outputs equivalent to those
of the HCP pipeline, specifically the CIFTI files (dtseries.nii) that were
utilized for further analyses. Written consent was obtained from all
participants or a parent/legal guardian in the case of those under 18
years of age for all open-source datasets used in this study. All data
articles of the open-source datasets used in this study mention that
written consent was obtained from all participants or a parent/legal
guardian in the case of those under 18 years of age23,84,85.
Data analysis
Connectopic gradient analysis. We estimated thalamocortical
CMAPs, which topographically represent dominant modes of func-
tional connectivity change within the thalamus, based on the voxel-wise
connectivity between each thalamic voxel and the rest of the neocor-
tex. This procedure is described in detail elsewhere16. In brief, we can
extract the CMAP in largely two steps that consist of (1) constructing
a similarity matrix followed by (2) a dimensionality reduction pro-
cess. First, we prepared two time-by-voxel matrices using the fMRI
timeseries data from the thalamus and neocortex. Before computing
the correlation between these two matrices, we performed a lossless
dimensionality reduction in the latter due to its relatively large size,
using singular value decomposition. Then we used the η2 coefficient
to quantify the similarity among the correlations between thalamic
voxels and neocortex vertices. In the second step, we applied Laplacian
eigenmaps to the resulting similarity matrix, which yielded the CMAPs
within the thalamus that represent the dominant modes of change in
functional connectivity between the thalamus and neocortex16,86. We
initially extracted the CMAPs at the group level for each dataset by
using the group-averaged similarity matrix, to use as a template to
align the individual CMAPs using the Procrustes alignment algorithm.
NEOMAPs were calculated using the aligned CMAPs of each individual
(https://github.com/koenhaak/congrads). To translate the CMAP pat-
terns within the thalamus onto the neocortex, we first obtained the
similarity between the CMAP and thalamic activation pattern for each
TR (β). Next, using these β values we computed the dot product with the
neocortex timeseries activation to yield an average of the neocortical
fMRI maps across the TRs, weighted by the β values. In short, the NEO-
MAPs were color-coded according to that of the thalamic voxel showing
Article
the highest correlation (for example, red vertices in the NEOMAP are
most highly connected with the red voxels in the CMAP)87. The same
analyses were conducted for the NKI-RS dataset and their results were
correlated with those from HCPD to test for replicability.
CMAPs and NEOMAPs were statistically tested for significant age
effects, controlling for sex and mean-FD, using SurfStat, surface-based
linear models implemented in a MATLAB toolbox (http://www.math.
mcgill.ca/keith/surfstat/ and https://brainstat.readthedocs.io/en/mas-
ter/)88,89. To interpret these gradients, we profiled the CMAP and NEO-
MAPs using the Yeo-Krienan 7 Network Atlas24,28. To visualize the change
in gradient values across age, we divided the participants into five age
groups for each dataset. Our sample division followed the age ranges
representative of developmental stages reported in previous litera-
ture90–92, and also considered the distribution of participants for each
group. Additionally, we employed a fully data-driven infant-specific,
functional parcellation27 to test the robustness of network profiling
in infant data.
To measure the impact of thalamus on segregation between
functional networks we devised a segregation index as follows:
SegregationSAL-EXT indicates an absolute value of the difference
between NEOMAP values of the salience network and networks com-
prising externally oriented functional processes such as the dorsal
attention, visual and somatosensory networks. On the other hand,
SegregationSAL-INT quantifies the degree of isolation between the sali-
ence and default mode networks. Next, we computed silhouette scores
to quantitatively evaluate the variations in the overlap of internal, sali-
ence and external networks across age. The silhouette score, ranging
from −1 to 1, is a metric for evaluating the quality of clustering: a score
of 1 signifies clear distinction between clusters, 0 indicates no separa-
tion and −1 suggests misassigned clusters. We then correlated these
scores with participants’ ages to examine developmental patterns.
A positive correlation with age would suggest increasing segregation
within the targeted networks. Subsequently, we divided the sample
into three age groups (8–12, 12–18 and 18–22 years) and calculated the
network overlap between internally and externally oriented networks.
Finally, we utilized a longitudinal subset from the NKI-RS data-
set to perform two distinct analyses. The first analysis involved a
mixed-effects model, where age was treated as a fixed effect and par-
ticipant as a random effect to model the segregation indices of both
salience–external and salience–internal networks. Second, we aimed
to estimate the influence of thalamocortical connectivity development
on network differentiation of the neocortex during later developmen-
tal stages. Specifically, we investigated whether the difference of the
thalamocortical NEOMAPs between follow-up and baseline data can
predict the segregation index between internal and external networks
within the ‘cortico-cortical’ gradients. For this, we applied linear regres-
sion predictive analysis employing least squares and lasso regulariza-
tion in a fivefold cross-validation.
Genetic analysis. Microarray gene expression. Regional microarray
expression data were obtained from six postmortem brains (one
female, ages 24.0–57.0, mean (s.d.) = 42.50 (13.38)) provided by the
AHBA (https://human.brain-map.org ref. 38). Data were processed with
the abagen toolbox (v.0.1.3; https://github.com/rmarkello/abagen)93
using the Schaefer 400-region surface-based atlas94; however, as only
two of the brains in the AHBA sampled both hemispheres, we focused
our analysis on the left cortex.
First, microarray probes were reannotated using data provided
by95; probes not matched to a valid Entrez ID were discarded. Next,
probes were filtered based on their expression intensity relative to
background noise96, yielding 31,569 probes. When multiple probes
indexed the expression of the same gene, we selected the probe with
the most consistent pattern of regional variation across donors97.
The MNI coordinates of tissue samples were updated to those gener-
ated via non-linear registration using Advanced Normalization Tools
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
(ANTs98; https://github.com/chrisfilo/alleninf). Samples were assigned
to brain regions by minimizing the Euclidean distance between the MNI
coordinates of each sample and the nearest surface vertex. Samples
where the Euclidean distance was more than 2 s.d. above the mean
distance for all samples belonging to that donor were excluded. If a
brain region was not assigned a sample from any donor based on the
above procedure, the tissue sample closest to the centroid of that par-
cel was identified independently for each donor. The average of these
samples was taken across donors, weighted by the distance between
the parcel centroid and the sample, to obtain an estimate of the par-
cellated expression values for the missing region. This procedure was
performed for nine regions that were not assigned tissue samples. All
tissue samples not assigned to a brain region in the provided atlas were
discarded. Inter-subject variation was addressed by normalizing tissue
sample expression values across genes using a robust sigmoid func-
tion99 and rescaled to the unit interval. Gene expression values were
then normalized across tissue samples using an identical procedure.
Samples assigned to the same brain region were averaged separately
for each donor and then across donors, yielding a regional expres-
sion matrix with 200 rows (half of Schaefer 400 regions of interest),
corresponding to brain regions and 15,631 columns, corresponding
to the retained genes.
PLS analysis. We used PLS analysis to investigate the correlation
between the NEOMAPs and spatial pattern of gene expression derived
from AHBA. PLS is an unsupervised multivariate statistical technique
used to identify orthogonal sets of latent variables that maximize the
covariance between the two datasets (in our case, gene expression val-
ues and NEOMAPs). The general underlying model of PLS is as follows:
T
X = TP + E
T
Y = UQ + F
where X is a 200 × 15,631 AHBA regional gene expression matrix and Y
is a 200-element vector representing the neocortical projection map
of interest. T and U are, respectively, the latent scores that are derived
from the projections of X (200 × 1) and projections of Y (200 × 1), while
P (15,631 × 1) and Q (1 × 1) are the eigenvectors indicating the loading
weights for each latent score. PLS finds the P and Q that maximize the
covariance of U and T. The latent variables are obtained by applying
singular value decomposition on the data after it has been z-scored.
To test whether the observed relationship between gene expres-
sion and NEOMAPs revealed by the PLS model was statistically sig-
nificant, we used an ensemble of randomized null brain phenotypes
controlling for spatial autocorrelation (spatial brain phenotype
(SBP)-spatial models)30,31. The surrogate maps of spatially permuting
parcellated brain regions while conserving for spatial autocorrela-
tion were generated using the BrainSMASH Python toolbox (https://
brainsmash.readthedocs.io/en/latest/)77. We compared the empirical
correlation result between latent scores from real data to the null
distributions of 10,000 correlations derived from applying PLS to
the spatially permuted datasets. The same test procedure was used
to test for the significance of feature contribution of each gene to the
association between gene expression and NEOMAPs. Specifically, the
empirical loading value of each gene was statistically tested using the
null distribution 10,000 loadings obtained for each gene. Genes with
positive loadings indicate upregulation whereas the negative loadings
represent a downregulation in relation to the neocortical projection
map. Genes with both negative and positive weights that showed a
significance of Pspin < 0.05 were considered in the enrichment analysis.
While the implementation of SBP-spatial models reduces false-positive
findings by correcting for spatial autocorrelation, we further assessed
the specificity of our genes of interest. Specifically, to assess if the
effect is unique to the genes of interest we used the Gene Annotation
Article
by Macroscale Brain-imaging Association MATLAB toolbox31. Among
the null models proposed (null-random-gene, null-coexpressed-gene
and null-brain-gene), we used the most stringent null-brain-gene
model that generates null distributions based on random genes that
are over-expressed in brain tissues (n = 2,711 selected by q < 0.05, FDR
correction, one-sided two-sample t-test comparing brain tissues and
other body tissues).
Finally, by implementing a targeted gene approach, we examined
whether the NEOMAPs reflect distinct patterns of thalamic projections
based on the relative density of ‘core’ and ‘matrix’ cells. Each subnucleus
of the thalamus consists of a combination of excitatory neurons with dis-
tinct cell types that project differentially to the neocortex7,100. On the one
hand, granular-projecting ‘core’ cells are prevalent in sensory thalamic
nuclei and project to specific regions of the cortex such as low-level
primary sensory cortices. On the other hand, supragranular-projecting
‘matrix’ cells, which are typically dense in higher-order thalamic nuclei,
project more diffusively across the cortex. Using the data provided by
the AHBA, we conducted a PLS analysis between the NEOMAPs and
genetic expression of calcium-binding proteins, namely parvalbumin
and calbindin, which represent core and matrix cells, respectively.
Genetic expression of parvalbumin and calbindin was derived from
thalamic nuclei using the abagen toolbox, based on the thalamic parcel-
lation derived from the Morel stereotactic atlas93,101.
Gene enrichment analyses. To characterize the biological processes of
the significant gene sets identified by PLS analysis, we used ShinyGO
0.77 (ref. 32), a toolbox for testing the gene enrichment against Gene
Ontology Biological Processes (available at http://bioinformatics.
sdstate.edu/go/). To test the robustness of the results we also used
other Gene Ontology enrichment analysis toolboxes, Enrichr34–36 and
g:Profiler37. The 15,631 genes retained from preprocessing the AHBA
data were used as the background reference set. We explored the
biological processes with which the genes are significantly involved,
separately, based on their loading signs. Significant results are reported
after Benjamini–Hochberg (BH) FDR correction (q < 0.05). Finally,
we fed the significant gene list into the cell-type specific expression
analysis (CSEA) developmental expression tool (http://doughertylab.
wustl.edu/csea-tool-2/)33, which utilizes the BrainSpan dataset (http://
www.brainspan.org)39 for constructing developmental expression
profiles. The statistical significance of overlap between our candi-
date genes and cell-specific genes for each developmental stage was
tested with Fisher’s exact test with BH multiple testing correction for
the number of cell types assayed; however, it should be noted that
the significant genes were extracted from the AHBA which consists
of adult postmortem data, thus, the results shown with CSEA do not
directly reflect developmental processes, although many genes have
been found to be consistent across the lifespan and conserved during
evolution102,103. Of the 550 genes, 476 genes existed in the brain region
and development expression dataset.
Computational modeling analysis. Finally, we sought to confirm
whether our neuroimaging findings (for example, functional effects
of thalamocortical connectivity) do not merely describe phenom-
enological aspects of the developing brain but reflect an underly-
ing mechanism. To this end, we employed an advanced simulation
approach, called ‘generative network modeling’. Using this tool, we
were able to investigate the effects of thalamocortical, genetic, spatial
and topological constraints by synthetically creating functional con-
nectome topology within the neocortex. We used the cost-topology
tradeoff model, which has been widely implemented to study human
brain networks41,42,104–107 and adopted the analysis code made available
by Oldham et al.41 (https://github.com/StuartJO/GenerativeNetwork-
Model). We used the Schaefer 200 cortical parcellation and focused
only on the left cerebral hemisphere to match our genetic analysis and
also to reduce computational costs due to numerous model iterations.
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
A total of seven tradeoff models derived from a combination of the
thalamocortical, genetic, spatial and topological constraints as param-
eters were tested as follows: (1) spatial model; (2) spatial + thalamo-
cortical connectivity model; (3) thalamocortical connectivity model;
(4) spatial + correlated gene expression model; (5) correlated gene
expression model; (6) thalamocortical connectivity + correlated gene
expression model; and (7) a model based on a topological ‘matching’
rule with spatial constraints (Fig. 5 and Extended Data Fig. 7). For the
thalamocortical connectivity constraint, we used the similarity matrix
of the correlation between thalamus and cortex, derived from the
connectopic gradient analysis. To compute the correlated gene expres-
sion constraint, we calculated the Pearson correlation between 1,899
brain-specific genes108,109, using the AHBA transcriptomic data. While
there are data available for the developing brain (BrainSpan atlas39),
the spatial coverage is inadequate to fully estimate the correlated gene
expression across the whole brain. Oldham, et al. estimated both uncor-
rected, raw correlated gene expression as well as those corrected for
their intrinsic spatial autocorrelation, and found that the former shows
better performance, which motivated us to only use the uncorrected
estimates of correlated gene expression41.
For the spatial parameter, we calculated the physical distance
of a connection between brain regions using the Euclidean distance
algorithm, based on the assumption that longer physical distances will
require more metabolic resources to form a connection, thus reflecting
wiring cost110. The spatial factor imposes a distance-based constraint
that forms a connection between nodes of shorter rather than longer
distances. As per previous studies41,42, synthetic networks formed based
on this rule did not show high resemblance to the empirical network
because of the lack of central hub brain regions that facilitate con-
nections between nodes of long-distance. Thus, following previous
studies, we also included a topological factor that forms a connection
between nodes that share a certain topological feature41,42. The match-
ing index was included as the representative topological factor because
it showed the best performance among other topological factors in
previous studies41,42. This rule estimates the similarity between a node’s
neighborhood (whether it is connected to similar nodes) and forms a
connection based on this overlap41,42.
We evaluated the model performance by comparing the synthetic
networks derived from the models to empirically extracted networks
in terms of degree, clustering, betweenness, and edge length distribu-
tions. For each measure, the distributions were compared using the KS
statistic, a nonparametric test that quantifies the distance between
two distributions, where lower values imply greater similarity. For
conservative evaluation, we determined the performance of a given
model based on the maximum KS statistic (corresponding to the least
accuracy). Next, a three-step multi-resolution optimization procedure
that sampled 10,000 combinations of different parameters was used
to obtain the best-fitting parameters for each model, which were used
to construct the synthetic networks for each individual, described in
detail elsewhere41,42.
Next, we wanted to estimate how well the synthetic networks made
from thalamocortical connectivity models reflect the functional seg-
regation between cortical networks as seen in the neuroimaging data
analyses. The affinity matrices derived from the synthetic networks
showing the best performance were analyzed in terms of modularity
network measure, as well as for constructing cortical gradients. The
Brain Connectivity Toolbox111 and the BrainSpace toolbox17 were used
for these analyses, respectively. Notably, to avoid circularity in the
evaluation of thalamic contribution to cortical network formation,
measures such as cortical gradients and segregation index, which were
not used for parameter estimation were employed. Following previous
work41,42,105–107, we modeled the binary topology of the network, with an
average connectome density threshold of 10%. To test the potential
sensitivity of our results to the chosen threshold, we also conducted
additional analyses with other thresholds of 5% and 15%. This analysis
Article
verified that the observed patterns were not dependent on our ini-
tial threshold selection (Supplementary Fig. 6). Moreover, given the
resemblance between the cortical projections of the thalamocorti-
cal gradients (NEOMAP) and cortex-only gradients, we conducted a
control analysis of removing the influence of both NEOMAPs from
the thalamocortical connectivity matrix, and used the residual data
in generative network modeling (Extended Data Fig. 8). In addition, to
ensure the robustness and generalizability of our results, we performed
two cross-validation analyses: (1) We used the thalamocortical rules
obtained from the HCPD dataset to predict the network properties of a
second independent dataset (NKI-RS; Extended Data Fig. 9); (2) We ran
the generative network model in the two datasets separately, and then
examined the correlation between the simulated matrices derived from
the HCPD and NKI-RS datasets (see Supplementary Note ‘Sensitivity
analysis for generative network modeling’).
Finally, we tested a series of growth models based on perturba-
tions to thalamo-salience connectivity in different developmental age
groups (childhood, 8–12 years; adolescence, 12–18 years; and young
adulthood, 18–22 years) to examine the underlying mechanism of
how thalamocortical connectivity affects the functional organization
across development. A total of four perturbation models were tested:
(1) perturbation applied to the 8–12 years age group; (2) 12–16 years
age group; (3) 16–22 years age group; and (4) all age groups and com-
pared to a non-perturbation growth model. Contrary to the previous
analyses that evaluated the overall model performance across differ-
ent biological, spatial and topological constraints, here, we used the
growth models (https://github.com/StuartJO/GenerativeNetwork-
Model) to provide age-dependent developmental information of the
thalamocortical connectivity during network simulation. Perturba-
tions were made to the thalamo-salience connectivity by first identi-
fying the nodes within the salience network using the Yeo-Krienan 7
network atlas and then randomly mixing its connections to preserve
the overall amount of connectivity despite the disarranged connec-
tivity patterns. Additionally, to investigate whether disruptions to
thalamo-salience connections result in more notable effects than
perturbations of other connections, we further tested our simulations
with a set of ‘null models.’ These models included: (1) introducing
perturbations to random nodes within the primary sensory regions
particularly characterized for its well-known association with the
thalamus and showed strong connections with the thalamus during
infancy; and (2) creating disturbances in random nodes within cortical
connections. Similar to perturbing thalamo-salience connectivity,
we identified nodes within the primary sensorimotor regions or ran-
domly selected nodes across the whole network and then randomly
mixed their connections.
Reporting summary
Further information on research design is available in the Nature
Portfolio Reporting Summary linked to this article.
Data availability
The data used in this study are publicly available. The neuroimag-
ing data from dHCP are available at http://www.developingconnec-
tome.org/data-release/second-data-release/ and preprocessed with
available, recommended pipelines. Surface and volumetric neona-
tal atlas data used were downloaded from https://web.gin.g-node.
org/BioMedIA/dhcp-volumetric-atlas-groupwise and http://
brain-development.org/brain-atlases/atlases-from-the-dhcp-project/
cortical-surface-atlas-bozek/, respectively. The HCPD data can
be downloaded from the NDA website at https://nda.nih.gov/
general-query.html?q=query=featured-datasets:HCP%20Aging%20
and%20Development. The minimally preprocessed structural and
functional data from the public release HCP Pipelines v.3.22 (https://
github.com/Washington-University/HCPpipelines) that were
released as a part of the Lifespan HCPD release 2.0 were used. The
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
NKI-RS dataset can be downloaded from the International Neuroim-
aging Data-Sharing Initiative database (http://rocklandsample.org/
accessing-the-neuroimaging-data-releases). The sleep–wake com-
parison dataset is available from the NDAR database (https://nda.nih.
gov/; collection no. 2072). Regional microarray expression data were
provided by the AHBA (https://human.brain-map.org/static/download)
and BrainSpan (https://www.brainspan.org/static/download.html).
We downloaded the Yeo-Krienan 7 Network Atlas and the Schaefer
atlas from https://github.com/ThomasYeoLab/CBIG/tree/master/
stable_projects/brain_parcellation. The thalamic parcellation based on
the Morel stereotactic atlas was downloaded from Zenodo at https://
zenodo.org/records/5499504 (ref. 112). Source data are provided with
this paper.
Code availability
The codes used in this study are publicly available. Preprocessing
codes for the dHCP data are available at https://github.com/BioMedIA/
dhcp-structural-pipeline (structural data) and https://git.fmrib.ox.ac.
uk/seanf/dhcp-neonatal-fmri-pipeline (functional data). Scripts to
align dHCP native surfaces to template space can be found at https://
github.com/ecr05/dHCP_template_alignment. The NKI-RS dataset
was preprocessed using the ciftify toolbox from https://github.com/
edickie/ciftify. Connectopic mapping of resting-state functional data
was performed using https://github.com/koenhaak/congrads. The
‘procrustes’ function in MATLAB was used for alignment between
individual and group-level gradients. Statistical testing of age effects
was implemented using surface-based linear models in a MATLAB tool-
box, SurfStat (http://www.math.mcgill.ca/keith/surfstat/ and https://
brainstat.readthedocs.io/en/master/). The ‘silhouette’ function and
‘inpolygon’ function in MATLAB was used to compute silhouette scores
and network overlap, respectively. The ‘fitlme’ function in MATLAB was
used for mixed-effects analysis, while the ‘fitrlinear’ function in MAT-
LAB was used for predictive modeling. Regional microarray expression
data for the transcriptomic analysis were processed with the abagen
toolbox (v.0.1.3; https://github.com/rmarkello/abagen) and for the
spatial autocorrelation-preserving permutation tests, surrogate maps
were generated by spatially permuting parcellated brain regions using
the BrainSMASH Python toolbox (https://brainsmash.readthedocs.
io/en/latest/). Gene specificity was tested using the Gene Annotation
by Macroscale Brain-Imaging Association MATLAB toolbox (https://
github.com/dutchconnectomelab/GAMBA-MATLAB). Gene enrich-
ment analysis utilized ShinyGO 0.77 (http://bioinformatics.sdstate.
edu/go/) and a cell-type-specific expression analysis developmental
expression tool (http://doughertylab.wustl.edu/csea-tool-2/; v.1.1).
Alternative tools including Enrichr (https://maayanlab.cloud/Enrichr/;
using version available after updated on 8 June 2023) and g:Profiler
(https://biit.cs.ut.ee/gprofiler/gost; v.e111_eg58_p18_30541362)
were used for testing robustness. Generative network modeling was
implemented by adopting the analysis code made available at https://
github.com/StuartJO/GenerativeNetworkModel. For visualization of
NEOMAPs, including parcellation boundaries, the MATLAB codes are
available at https://github.com/StuartJO/plotSurfaceROIBoundary.
The following toolboxes were used for analyses of simulated data: Brain
connectivity toolbox (https://sites.google.com/site/bctnet/) and Brain-
Space toolbox (https://brainspace.readthedocs.io/en/latest/). All code
adapted and generated for analyses and for creating figures are availa-
ble at https://github.com/COMBINE-SKKU/NatNeurosci_2024_SWPark.
References
78. Makropoulos, A. et al. Automatic whole brain MRI segmentation
of the developing neonatal brain. IEEE Trans. Med. Imaging 33,
1818–1831 (2014).
79. Robinson, E. C. et al. MSM: a new flexible framework
for multimodal surface matching. Neuroimage 100,
414–426 (2014).
Article
80. Robinson, E. C. et al. Multimodal surface matching: fast and
generalisable cortical registration using discrete optimisation.
Inf. Process. Med. Imaging 23, 475–486 (2013).
81. Dale, A. M., Fischl, B. & Sereno, M. I. Cortical surface-based
analysis. I. Segmentation and surface reconstruction. Neuroimage
9, 179–194 (1999).
82. Dickie, E. W. et al. Ciftify: a framework for surface-based analysis
of legacy MR acquisitions. Neuroimage 197, 818–826 (2019).
83. Esteban, O. et al. fMRIPrep: a robust preprocessing pipeline for
functional MRI. Nat. Methods 16, 111–116 (2019).
84. Eyre, M. et al. The Developing Human Connectome Project:
typical and disrupted perinatal functional connectivity. Brain 144,
2199–2213 (2021).
85. Nooner, K. B. et al. The NKI-Rockland sample: a model for
accelerating the pace of discovery science in psychiatry. Front.
Neurosci. 6, 152 (2012).
86. Tian, Y., Margulies, D. S., Breakspear, M. & Zalesky, A. Topographic
organization of the human subcortex unveiled with
functional connectivity gradients. Nat. Neurosci. 23,
1421–1432 (2020).
87. Marquand, A. F., Haak, K. V. & Beckmann, C. F. Functional
corticostriatal connection topographies predict goal directed
behaviour in humans. Nat. Hum. Behav. 1, 0146 (2017).
88. Larivière, S. et al. BrainStat: a toolbox for brain-wide statistics
and multimodal feature associations. Neuroimage 266,
119807 (2023).
89. Worsley, K. J. et al. SurfStat: a MATLAB toolbox for the statistical
analysis of univariate and multivariate surface and volumetric
data using linear mixed effects models and random field theory.
NeuroImage https://doi.org/10.1016/S1053-8119(09)70882-1
(2009).
90. Kilford, E. J., Garrett, E. & Blakemore, S.-J. The development of
social cognition in adolescence: an integrated perspective.
Neurosci. Biobehav. Rev. 70, 106–120 (2016).
91. Güroğlu, B., van den Bos, W. & Crone, E. A. Sharing and giving
across adolescence: an experimental study examining the
development of prosocial behavior. Front. Psychol. 5, 291
(2014).
92. Keshavan, M. S. et al. Development of the corpus callosum
in childhood, adolescence and early adulthood. Life Sci. 70,
1909–1922 (2002).
93. Markello, R. D. et al. Standardizing workflows in imaging
transcriptomics with the abagen toolbox. eLife 10, e72129 (2021).
94. Schaefer, A. et al. Local-global parcellation of the human cerebral
cortex from intrinsic functional connectivity MRI. Cereb. Cortex
28, 3095–3114 (2018).
95. Arnatkeviciute, A., Fulcher, B. D. & Fornito, A. A practical guide
to linking brain-wide gene expression and neuroimaging data.
Neuroimage 189, 353–367 (2019).
96. Quackenbush, J. Microarray data normalization and
transformation. Nat. Genet. 32, 496–501 (2002).
97. Hawrylycz, M. et al. Canonical genetic signatures of the adult
human brain. Nat. Neurosci. 18, 1832–1844 (2015).
98. Avants, B. B. et al. Advanced normalization tools (ANTS). Insight J.
2, 1–35 (2009).
99. Fulcher, B. D., Little, M. A. & Jones, N. S. Highly comparative
time-series analysis: the empirical structure of time series and
their methods. J. R. Soc. Interface 10, 20130048 (2013).
100. Jones, E. G. Viewpoint: the core and matrix of thalamic
organization. Neuroscience 85, 331–345 (1998).
101. Saranathan, M., Iglehart, C., Monti, M., Tourdias, T. & Rutt, B.
In vivo high-resolution structural MRI-based atlas of human
thalamic nuclei. Sci. Data 8, 275 (2021).
102. Fontana, L., Partridge, L. & Longo, V. D. Extending healthy life
span—from yeast to humans. Science 328, 321–326 (2010).
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
103. Parikh, A. et al. Conserved developmental transcriptomes
in evolutionarily divergent species. Genome Biol. 11, R35
(2010).
104. Arnatkeviciute, A. et al. Genetic influences on hub connectivity of
the human connectome. Nat. Commun. 12, 4237 (2021).
105. Vértes, P. E. et al. Simple models of human brain
functional networks. Proc. Natl Acad. Sci. USA 109,
5868–5873 (2012).
106. Akarca, D., Vértes, P. E., Bullmore, E. T., CALM team & Astle, D. E.
A generative network model of neurodevelopmental
diversity in structural brain organization. Nat. Commun. 12,
4216 (2021).
107. Zhang, X. et al. Generative network models of altered structural
brain connectivity in schizophrenia. Neuroimage 225, 117510
(2021).
108. Yang, S. et al. Cortical patterning of morphometric similarity
gradient reveals diverged hierarchical organization in
sensory-motor cortices. Cell Rep. 36, 109582 (2021).
109. Burt, J. B. et al. Hierarchy of transcriptomic specialization across
human cortex captured by structural neuroimaging topography.
Nat. Neurosci. 21, 1251–1259 (2018).
110. Fornito, A., Zalesky, A. & Bullmore, E. Fundamentals of Brain
Network Analysis (Academic Press, 2016).
111. Rubinov, M. & Sporns, O. Complex network measures of
brain connectivity: uses and interpretations. NeuroImage 52,
1059–1069 (2010).
112. Saranathan, M. et al. Data for in vivo structural MRI-based
atlas of human thalamic nuclei. Zenodo https://zenodo.org/
records/5499504 (2021).
Acknowledgements
This work was supported by the Institute for Basic Science
IBS-R015-D1 (S.J.H. and B.Y.P.), the National Research Foundation of
Korea (NRF) grant funded by the Korea Government (MSIT) (NRF-
2022R1C1C1007095, R5-2023-00217361, RS-2024-00398768 to
S.J.H. and NRF-2022R1A5A7033499 to B.Y.P.), the Dutch Research
Council (VIDI grant no. 09150171910043; K.V.H.), NARSAD Young
Investigator Grant from the Brain & Behavior Research Foundation
(S.O.), NIMH RO1 (R01MH115363 and 5R01MH133334; A.D.M.),
National Science and Engineering Research Council of Canada
(NSERC Discovery-1304413; B.B.), the Canadian Institutes of Health
Research (FDN-154298, PJT-174995 and PJT-191853; B.B.), SickKids
Foundation (NI17-039; B.B.), Azrieli Center for Autism Research
(ACAR-TACC; B.B.), BrainCanada (Future-Leaders, McConnell Brain
Imaging Centre Platform; B.B.), Healthy Brains and Healthy Lives,
the Tier-2 Canada Research Chairs program (B.B.), as well as the
Montreal Neurological Institute and its Centre of Excellence in
Epilepsy (B.B.), National Institutes of Health grants RF1MH128696
(T.X.), U01DA055366 (W.G.) and the Institute for Information and
Communications Technology Planning and Evaluation funded by the
Korea Government (MSIT) (2022–0–00448, RS-2022–00155915 and
2021–0–02068; B.Y.P.).
Author contributions
S.P. conceived of the presented idea, performed all analyses and
took the lead in writing the manuscript. K.V.H. provided codes for
the gradient mapping and discussed the findings. S.O. provided
codes for the computational framework, data for comparison with
structural thalamic gradients and discussed the findings. H.C.
contributed to visualization of figures and advised on analytic
processes. K.B. provided assistance with preprocessing the
neuroimaging data and discussed the findings. B.P. provided
suggestions on idea-building input and revision. P.T. provided
preprocessed neuroimaging data for the sleep–wake state
comparison data and critical feedback for revising the manuscript.
Article
H.C. provided assistance with preprocessing the neuroimaging data
and revising the manuscript. W.G. provided critical feedback on the
infant brain imaging analyses. T.X. provided critical feedback on the
neuroimaging gradients analyses. S.V. encouraged the investigation
of the core and matrix thalamic projections and advised on the
implementation of the transcriptomic analyses. M.P.M. provided
consultation on interpreting the results and revising the manuscript.
B.B. provided consultation on interpreting the results and advised on
the implementation of the transcriptomic analyses. A.D.M. provided
critical consultation on interpreting the results and revising the
manuscript as well as the dataset for analysis of sleep–wake state
comparison. S.J.H. conceived of the presented idea, supervised the
project and provided funding. All authors provided critical feedback
and contributed to the final manuscript.
Competing interests
The authors declare no competing interests.
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
Additional information
Extended data is available for this paper at
https://doi.org/10.1038/s41593-024-01679-3.
Supplementary information The online version
contains supplementary material available at
https://doi.org/10.1038/s41593-024-01679-3.
Correspondence and requests for materials should be addressed to
Seok-Jun Hong.
Peer review information Nature Neuroscience thanks Danyal Akarca,
Monica Rosenberg, and the other, anonymous, reviewer(s) for their
contribution to the peer review of this work.
Reprints and permissions information is available at
www.nature.com/reprints.
Article
Extended Data Fig. 1 | Replication of neuroimaging findings using the Nathan
Kline Institute-Rockland Sample (NKI-RS). (a) Thalamic connectopic maps
(CMAP) derived from the NKI-RS dataset and scatter plots showing significant
correlation with those from HCPD dataset using two-tailed Pearson’s correlation
tests. (b) Neocortical projection maps (NEOMAP) from the NKI-RS dataset and
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
scatter plots showing significant correlation with those from HCPD dataset using
two-tailed Pearson’s correlation tests. The colorful scatter-plot in the bottom
right corner illustrates the distribution within the NEOMAP space based on the
NKI-RS dataset.
Article
Extended Data Fig. 2 | Associations between thalamic gradients obtained
from awake vs. sleep states collected in an independent sample. (a) Thalamic
gradients are visualized at the group level for both awake and sleep states. Strong
correlations are found for both gradient 1 (r = 0.9860, p = 0) and gradient
2 (r = 0.9747, p = 0) using two-tailed Pearson’s correlation. (b) Correlation values
between the awake and sleep state thalamic gradients for each individual are
sorted from the lowest to highest value. Majority of the individuals show a
correlation higher than 0.9 for both gradients. (c) Voxel-wise mixed-effects
analysis with state (sleep vs. wake) modeled as a fixed effect, with participant
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
variability accounted for as a random effect (N = 29), revealed significant
state-dependent variations in brain activity, surviving FDR correction (Q<0.05).
Regions outlined in white are indicative of elevated gradient values during sleep
while areas marked in black signify heightened gradient values when awake. Box
plots display the median (central line) and interquartile range (IQR, the extent
of the box). Whiskers extend to the highest value within 1.5 * IQR above the third
quartile and to the lowest value within 1.5 * IQR below the first quartile. The IQR
represents the range between the 25th and 75th percentiles.
Article
Extended Data Fig. 3 | Reference coordinate system (‘NEOMAP space’) of
the NEOMAP1 and 2 distributions. The two NEOMAPs from the childhood/
young adult group serve each as the X and Y axis, respectively. To quantitatively
describe this space, the top and bottom 5% of the NEOMAP values at each axis
were evaluated for which pivotal functional networks comprise the gradient
of each NEOMAP. As shown, the external (visual/dorsal attention) and internal
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
(default mode/limbic networks) brain areas are at one end of the axis at each
NEOMAP (the side bar graphs indicated by ➀, ➁), while the salience network is
commonly placed on the other end in both NEOMAPs (➂). This embedding of the
salience network places itself at a ‘joint’ position that links the two other network
chains (that is, internal and external systems), represented by differently coded
streamlines (dotted and solid lines, respectively).
Article
Extended Data Fig. 4 | Quantitative analyses of functional network
segregation across age. (a) The associations between silhouette scores and age
were examined using two-tailed Pearson’s correlation. Significant correlations
are reported across the networks: pertaining to external-internal in both
NEOMAPs (left; r = 0.26, p = 1.94 × 10−10), salience-external in NEOMAP1 (middle;
r = 0.14, p = 6.72 × 10−4), and salience-internal in NEOMAP2 (right; r = 0.10,
p = 0.01). (b) Differences in the NEOMAP space according to age groups,
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
highlighting a decreasing overlap (=more functional differentiation) between
external and internal networks across age groups. (c) Individual silhouette scores
across the internally and externally oriented functional networks in two age
datasets (dHCP vs. HCPD). A significant difference was found (two-tailed t = −19.8,
p = 5.107 × 10−72) between the two age groups, with higher silhouette scores in the
childhood/young adulthood group.
Article
Extended Data Fig. 5 | Age-related differences in (a) CMAPs and (b) NEOMAPs.
Significant age effects were found, controlling for sex and mean-FD, using
surface-based linear models (that is, SurfStat). Regions that showed significant
positive and negative aging effects at FDR-corrected p<0.05 were delineated
with black and white lines, respectively. Scatter plots illustrate the association
between age and regions showing significant aging effects. The range and
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
number of participants for the 5 age groups are as follows: i) Infancy (<37 weeks
(wks), n = 49; 37–40 wks, n = 59; 40 wks, n = 47; 41 wks, n = 48; 42–44 wks, n = 52);
and ii) Childhood/Young adulthood (8–10 years (yrs), n = 81; 10–13 yrs, n = 125;
13–16 yrs, n = 173; 16–20 yrs, n = 141; 20–22 yrs, n = 83). Abbreviation, CMAP,
thalamocortical connectopic map; NEOMAP, Neocortical projection map; mean-
FD, mean framewise displacement.
Article
Extended Data Fig. 6 | Results from a longitudinal analysis based on the
Nathan Kline Institute-Rockland Sample (n = 70). A mixed-effects model
(age as a fixed effect and participant as a random effect) was applied to
probe the segregation indices of (a) salience-external, (b) salience-internal,
and (c) external-internal networks. There was a significant age effect on
the salience-internal segregation index (t = 2.35, p = 0.02) and the external-
internal segregation index (t = 2.63, p = 0.009), as well as a marginal trend for
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
the salience-external segregation index (t = 1.25, p = 0.21). (d) Results from a
predictive analysis using linear regression with LASSO regularization, 5-fold
cross-validation, with least squares method as the learning algorithm, show that
changes in NEOMAPs between baseline and follow-up significantly predict the
segregation index between internal and external networks within the cortico-
cortical gradients (mean average error = 1.91; median correlation coefficient,
r = 0.33, p = 0.005). Error bands represent 95% confidence intervals.
Article https://doi.org/10.1038/s41593-024-01679-3

Extended Data Fig. 7 | Model performance for all wiring rules. Model fitting performance for different wiring rules of generative network models for (a) infancy and
(b) childhood/young adulthood.

Nature Neuroscience
Article
Extended Data Fig. 8 | Sensitivity analyses for testing potential circularity.
(a) The first cortico-cortical gradient displayed a significant correlation with
the second NEOMAP (r = 0.59, p = 3.15 × 10−38) and the first NEOMAP (r = −0.37,
p = 2.37 × 10−14), using two-tailed Pearson’s correlation. These correlations
may indicate the possibility that the resemblance between the NEOMAPs and
cortico-cortical gradients might influence our simulation results derived
from generative network modeling. (b) Results from a sensitivity analysis that
regressed out both NEOMAPs from the thalamocortical connectivity adjacency
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
matrix and used the residual information as a wiring rule in generative network
modeling. The top panel (b. i) displays the simulated cortical gradients based
on the original thalamocortical connectivity rules, while the bottom panel (b. ii)
shows the results after regressing out the NEOMAPs. A high correlation between
both sets of simulated results (r = 0.78, p = 1.70 × 10−21), using two-tailed Pearson’s
correlation, indicates that the potential effects of circularity on our findings are
minimal.
Article
Extended Data Fig. 9 | Cross-validation analysis of generative modeling using
the independent NKI-RS dataset. Generative modeling results from NKI-RS
dataset using the predefined wiring rules derived from HCPD (a) Model fit and
network measures from the simulated data including (b) network modularity,
(c) cortical gradients and (d) segregation indices. Comparison of network
Nature Neuroscience
https://doi.org/10.1038/s41593-024-01679-3
modularity, calculated from each synthetic network based on different wiring
rules, were analyzed using ANOVA (N = 70) with Bonferroni correction, resulting
in a significant group difference (F = 31.04, p = 1.55 × 10−11). Segregation indices
computed from the individual synthetic networks based on different wiring rules
are shown with error bars indicating standard deviation (N = 70).
Article https://doi.org/10.1038/s41593-024-01679-3

Extended Data Fig. 10 | Sensitivity assessment for perturbation analysis across all regions. Cortical gradients derived from simulated matrices after
using null models. Differences in the segregation index resulting from the perturbations to random nodes (c) within primary sensory regions and (d) across
perturbation to random nodes (a) within primary sensory regions and (b) all regions.

Nature Neuroscience
 nature portfolio | reporting summary
Corresponding author(s): Seok-Jun Hong
Last updated by author(s): Apr 18, 2024

Reporting Summary
Nature Portfolio wishes to improve the reproducibility of the work that we publish. This form provides structure for consistency and transparency
in reporting. For further information on Nature Portfolio policies, see our Editorial Policies and the Editorial Policy Checklist.

Statistics
For all statistical analyses, confirm that the following items are present in the figure legend, table legend, main text, or Methods section.
n/a Confirmed
The exact sample size (n) for each experimental group/condition, given as a discrete number and unit of measurement
A statement on whether measurements were taken from distinct samples or whether the same sample was measured repeatedly
The statistical test(s) used AND whether they are one- or two-sided
Only common tests should be described solely by name; describe more complex techniques in the Methods section.

A description of all covariates tested

A description of any assumptions or corrections, such as tests of normality and adjustment for multiple comparisons
A full description of the statistical parameters including central tendency (e.g. means) or other basic estimates (e.g. regression coefficient)
AND variation (e.g. standard deviation) or associated estimates of uncertainty (e.g. confidence intervals)

For null hypothesis testing, the test statistic (e.g. F, t, r) with confidence intervals, effect sizes, degrees of freedom and P value noted
Give P values as exact values whenever suitable.

For Bayesian analysis, information on the choice of priors and Markov chain Monte Carlo settings
For hierarchical and complex designs, identification of the appropriate level for tests and full reporting of outcomes
Estimates of effect sizes (e.g. Cohen's d, Pearson's r), indicating how they were calculated
Our web collection on statistics for biologists contains articles on many of the points above.

Software and code

Policy information about availability of computer code
Data collection Open-source neuroimaging datasets from two cohorts of the Human Connectome Project (HCP), the developing HCP (dHCP) and
HCPDevelopment (HCPD), were used. A separate dataset from the Nathan Kline Institute - Rockland Sample (NKI-RS) was used for replication
purposes.
dHCP: MR images were acquired on a 3T Philips Achieva scanner equipped with a dedicated neonatal imaging system and a neonatal 32
channel phased array head coil at the Evelina Neonatal Imaging Centre, St Thomas’ Hospital, London.
HCPD: Participants are scanned on Siemens 3T Prisma (whole-body scanner with 80mT/m gradient coil, slew rate of 200T/m/s; Siemens 32-
channel head coil; software version: E11C) using matched protocols across 4 sites (Harvard University, University of California – Los Angeles,
University of Minnesota, and Washington University in St. Louis) in the United States.
NKI-RS: Scanned using a 3.0T Siemens TIM Trio located at the Nathan Kline Institute, with a resting state sequence of a multiband imaging
protocol (T= 645 ms/3 mm isotropic voxels).
For more details on data collection, please refer to the following references:
- dHCP: Edwards, A. David, et al. "The developing human connectome project neonatal data release." Frontiers in neuroscience 16 (2022):
886772.
- HCPD: Somerville, L. H. et al. The Lifespan Human Connectome Project in Development: A large-scale study of brain connectivity
development in 5–21 year olds. Neuroimage 183, 456–468 (2018).
- NKI-RS: Nooner, K. B. et al. The NKI-Rockland Sample: A Model for Accelerating the Pace of Discovery Science in Psychiatry. Front. Neurosci.
April 2023

6, 152 (2012).

Data analysis The codes used in this study are publicly available. Preprocessing codes for the dHCP data are available at https://github.com/BioMedIA/dhcp-
structural-pipeline (structural data) and https://git.fmrib.ox.ac.uk/seanf/dhcp-neonatal-fmri-pipeline (functional data). Scripts to align dHCP
native surfaces to template space can be found at https://github.com/ecr05/dHCP_template_alignment. The NKI-RS dataset was

1
 preprocessed using the ciftify toolbox from https://github.com/edickie/ciftify. Connectopic mapping of resting-state functional data was
performed using https://github.com/koenhaak/congrads. The 'procrustes' function in MATLAB was used for alignment between individual and

nature portfolio | reporting summary

group level gradients. Statistical testing of age effects was implemented using surface-based linear models in a Matlab toolbox, SurfStat
(http://www.math.mcgill.ca/keith/surfstat/; https://brainstat.readthedocs.io/en/master/). The 'silhouette' function and ‘inpolygon’ function in
MATLAB was used to compute silhouette scores and network overlap, respectively. The 'fitlme' function in MATLAB was used for mixed
effects analysis while the 'fitrlinear' function in MATLAB was used for predictive modeling. Regional microarray expression data for the
transcriptomic analysis was processed with the abagen toolbox (version 0.1.3; https://github.com/rmarkello/abagen), and for the spatial
autocorrelation-preserving permutation tests, surrogate maps were generated by spatially permuting parcellated brain regions using the
BrainSMASH python toolbox (https://brainsmash.readthedocs.io/en/latest/). Gene specificity was tested using the Gene Annotation by
Macroscale Brain-Imaging Association Matlab toolbox (https://github.com/dutchconnectomelab/GAMBA-MATLAB). Gene enrichment analysis
utilized ShinyGO 0.77 (http://bioinformatics.sdstate.edu/go/) and a cell-type specific expression analysis developmental expression tool
(http://doughertylab.wustl.edu/csea-tool-2/; version 1.1). Alternative tools including Enrichr (https://maayanlab.cloud/Enrichr/; using version
available after updated on June 8, 2023) and g:Profiler (https://biit.cs.ut.ee/gprofiler/gost; version e111_eg58_p18_30541362) were used for
testing robustness. Generative network modeling was implemented by adopting the analysis code made available at https://github.com/
StuartJO/GenerativeNetworkModel. For visualization of NEOMAPs including parcellation boundaries the matlab codes available at https://
github.com/StuartJO/plotSurfaceROIBoundary. The following toolboxes were used for analyses of simulated data: Brain connectivity toolbox
(https://sites.google.com/site/bctnet/) and Brainspace toolbox (https://brainspace.readthedocs.io/en/latest/).
All code generated for analyses and for creating figures are available at: https://github.com/COMBINE-SKKU/NatNeurosci_2024_SWPark.
For manuscripts utilizing custom algorithms or software that are central to the research but not yet described in published literature, software must be made available to editors and
reviewers. We strongly encourage code deposition in a community repository (e.g. GitHub). See the Nature Portfolio guidelines for submitting code & software for further information.

Data
Policy information about availability of data
All manuscripts must include a data availability statement. This statement should provide the following information, where applicable:
- Accession codes, unique identifiers, or web links for publicly available datasets
- A description of any restrictions on data availability
- For clinical datasets or third party data, please ensure that the statement adheres to our policy

The data used in this study are publicly available. The neuroimaging data from dHCP is available at http://www.developingconnectome.org/data-release/second-
data-release/ and preprocessed with available, recommended pipelines. Surface and volumetric neonatal atlas used were downloaded from https://web.gin.g-
node.org/BioMedIA/dhcp-volumetric-atlas-groupwise and http://brain-development.org/brain-atlases/atlases-from-the-dhcp-project/cortical-surface-atlas-bozek/,
respectively. The HCPD data can be downloaded from the NDA website at https://nda.nih.gov/general-query.html?q=query=featured-datasets:HCP%20Aging%
20and%20Development. The minimally preprocessed structural and functional data from the public release HCP pipelines v3.22 (https://github.com/Washington-
University/HCPpipelines) that were released as a part of the Lifespan HCP-Development Release 2.0 were used. The NKI-RS dataset can be downloaded from the
International Neuroimaging Data-Sharing Initiative database (https://fcon_1000.projects.nitrc.org/indi/enhanced/sharing_neuro.html). The sleep/wake comparison
dataset is available from the NDAR database (https://nda.nih.gov/; Collection # 2072). Regional microarray expression data were provided by the Allen Human Brain
Atlas (https://human.brain-map.org/static/download) and BrainSpan (https://www.brainspan.org/static/download.html). We downloaded the Yeo-Krienan 7
Network Atlas and the Schaefer atlas from https://github.com/ThomasYeoLab/CBIG/tree/master/stable_projects/brain_parcellation.

Research involving human participants, their data, or biological material

Policy information about studies with human participants or human data. See also policy information about sex, gender (identity/presentation),
and sexual orientation and race, ethnicity and racism.
Reporting on sex and gender This study reports on self-reported participant sex (as a biological attribute). For dHCP, 107 males and 88 females, for HCPD,
417 males and 186 females and for NKI-RS, 33 males and 37 females were included. Study findings apply to both male and
female sexes. Sex was considered as a potential covariate in all statistical analyses.

Reporting on race, ethnicity, or The dHCP dataset does not provide information on race, ethnicity, or other socially relevant groupings. The HCPD dataset
other socially relevant does report on race/ethnicity, however we did not use this information as a part of our analyses. We did not use any
groupings information related to race or ethnicity for the NKI-RS dataset.

Population characteristics
dHCP: All anatomical scans were reviewed by a neuroradiologist who provided a radiology score. Scans with a score of 3
(incidental findings with unlikely clinical significance but possible analysis significance) or 5 (incidental findings with possible/
likely significance for both clinical and imaging analysis) were excluded from further analyses. After quality control and
preprocessing, the final sample consisted of 195 healthy term infants (107 male, mean [SD] gestational age at birth=39.9
[1.26] weeks, age at scan=41.0 [1.67] weeks) and 60 pre-term infants (42 male, gestational age at birth=32.6 [3.12] weeks, at
scan=35.5 [2.78] weeks).
HCPD: After quality control and preprocessing, the final sample consisted of 603 healthy participants (417 male, mean [SD]
age=14.8 [3.88] years). Participants were excluded for insufficient English, premature birth, MRI contraindications, or a
lifetime history of serious medical conditions, endocrine disorders, head injury, psychiatric/neurodevelopmental disorders.
NKI-RS: To ensure replicability, we utilized the cross-sectional and longitudinal pediatric data from an independent open-
source dataset, the Nathan Kline Institute / Rockland Sample (NKI-RS; n=71 (34 male); age range 6 - 19 years; mean [SD]
age=10.9±2.9). Initially, we selected participants without any prior diagnosis and who had at least two timepoints of fMRI
data.
April 2023

Recruitment Various recruitment methods were employed to gather participants, including targeting related siblings for the HCP data,
which were not directly related to the objectives of the current study. However, our findings are not likely to be affected by
the recruitment approach used.

2
 Ethics oversight All participants provided written informed consent statements before participation in the study. The dHCP data were

nature portfolio | reporting summary

reviewed and approved by United Kingdom Health Research Authority (Research Ethics Committee reference number: 14/
LO/1169). Written informed consent to participate in this study was provided by the participants’ legal guardian/next of kin.
The HCP data were acquired using protocols approved by the Washington University institutional review board. For the NKI-
RS, Institutional Review Board Approval was obtained at the Nathan Kline Institute (Phase I #226781 and Phase II #239708)
and at Montclair State University (Phase I #000983A and Phase II #000983B). All data articles of the open-source datasets
used in this study mention that written consent was obtained from all participants or a parent/legal guardian in the case of
those under 18 years of age. Information regarding the participant compensation is currently not available on the websites
and/or data papers.

Note that full information on the approval of the study protocol must also be provided in the manuscript.

Field-specific reporting
Please select the one below that is the best fit for your research. If you are not sure, read the appropriate sections before making your selection.

Life sciences Behavioural & social sciences Ecological, evolutionary & environmental sciences
For a reference copy of the document with all sections, see nature.com/documents/nr-reporting-summary-flat.pdf

Life sciences study design

All studies must disclose on these points even when the disclosure is negative.
Sample size We included all available data from the relevant datasets, after applying several exclusion criteria explained below.

Data exclusions The dHCP second release dataset consists of a developmental sample of n=558 (224 female). Among individuals with both structural and
functional MRI scans (n=461) individuals with a neuro-radiological score of 3 (incidental findings with unlikely clinical significance but possible
analysis significance) or 5 (incidental findings with possible/likely significance for both clinical and imaging analysis) were excluded (n=374).
The final sample consisted of a total of 255 infants (106 females), after the MRI scans finished preprocessing without errors, survived motion-
related quality control (mean framewise displacement [FD] < 0.4) as well as showed strong resemblance between the individual thalamic
gradient and group-level gradient.
The HCPD dataset consists of a total of n=652. Likewise, among individuals with both structural and functional MRI scans, those that survived
the preprocessing pipeline and motion-related quality control of Mean FD < 0.4 (n=615), as well as showed strong spatial correlation to the
group-level thalamic gradient were included resulting in a final sample of n=603 (327 females).
NKI-RS: NKI-RS: To ensure replicability, we utilized the cross-sectional and longitudinal pediatric data from an independent open-source
dataset, the Nathan Kline Institute / Rockland Sample (NKI-RS; n=71 (34 male); age range 6 - 19 years; mean [SD] age=10.9±2.9). We selected
participants without any prior diagnosis and who had at least two time-points of all imaging data modalities.

Replication The results were reproduced over several relevant biological and image-processing factors, including 1) an independent dataset (i.e., NKI-RS)
utilizing both cross-sectional and longitudinal samples, 2) global signal regression, 3) sleep-wake state dependency, 4) different gene
enrichment analysis tools.

Randomization As this study did not entail distinct experimental groups or conditions, randomization was not performed.

Blinding Each participant in the study underwent an identical study protocol, and there was no allocation to distinct experimental groups. Thus
blinding was unnecessary.

Reporting for specific materials, systems and methods

We require information from authors about some types of materials, experimental systems and methods used in many studies. Here, indicate whether each material,
system or method listed is relevant to your study. If you are not sure if a list item applies to your research, read the appropriate section before selecting a response.

Materials & experimental systems Methods

n/a Involved in the study n/a Involved in the study
Antibodies ChIP-seq
Eukaryotic cell lines Flow cytometry
Palaeontology and archaeology MRI-based neuroimaging
Animals and other organisms
Clinical data
Dual use research of concern
April 2023

Plants

3
 nature portfolio | reporting summary
Plants
Seed stocks NA

Novel plant genotypes NA

Authentication NA

Magnetic resonance imaging

Experimental design
Design type Resting-state functional MRI

Design specifications NA

Behavioral performance measures NA

Acquisition
Imaging type(s) Resting-state functional MRI

Field strength 3 Tesla

Sequence & imaging parameters
dHCP: High-resolution T2w and T1w structural images were acquired with 0.8 mm isotropic spatial resolution and
thickness of 1.6mm, overlapped by 0.8 mm (T2w: TE/TR=156 ms/12 s; T1w: TE/TR/TI=8.7/4795/1740 ms). Resting-state
fMRI data were acquired using multiband accelerated echo-planar imaging for approximately 15 minutes (TE/
TR=38ms/392ms, flip angle=34°, spatial resolution=2.15 mm isotropic, 2300 volumes). Single-band reference scans
were also collected with bandwidth matched readout and additional spin-echo EPI acquisitions with both AP/PA
phaseencoding
directions.
HCPD: T1-weighted images were acquired using a 3D multi-echo MPRAGE sequence (TR=2500, TI=1000 ms,
TE=1.8/3.6/5.4/7.2 ms, flip angle=8°, FOV=256×240×166 mm, 320×300 matrix, 208 slices, 0.8 mm isotropic voxels).
Resting-state fMRI data were acquired with a multiband gradient-recalled (GRE) EPI sequence (TR=800 ms, TE=37 ms,
flip angle=52°, FOV=208 mm, 104×90 matrix, 72 oblique axial slices, 2 mm isotropic voxels, multiband acceleration
factor of 8, 1912 volumes).
NKI-RS: A 3.0-T Siemens MAGNETOM Trio-Tim scanner was used to collect resting state fMRI data using the following
imaging parameters: TR =645 ms, TE =30 ms, 40 slices, flip angle = 60° and voxel size = 3.0 × 3.0 × 3.0 mm.

Area of acquisition whole brain

Diffusion MRI Used Not used

Preprocessing
Preprocessing software We refer the reader to the original articles (and related reporting summaries) for a detailed description: dHCP MRI
preprocessing - Makropoulos et al., 2018 (NIMG) & Fitzgibbon et al., 2020 (NIMG); and HCPD MRI preprocessing – Somerville
et al., 2018 (NIMG). A summary can be found in the article methods section.

Normalization We refer the reader to the original articles (and related reporting summaries) for a detailed description: dHCP MRI
preprocessing - Makropoulos et al., 2018 (NIMG) & Fitzgibbon et al., 2020 (NIMG); and HCPD MRI preprocessing – Somerville
et al., 2018 (NIMG). A summary can be found in the article methods section.

Normalization template We refer the reader to the original articles (and related reporting summaries) for a detailed description: dHCP MRI
preprocessing - Makropoulos et al., 2018 (NIMG) & Fitzgibbon et al., 2020 (NIMG); and HCPD MRI preprocessing – Somerville
et al., 2018 (NIMG). A summary can be found in the article methods section.

Noise and artifact removal We excluded participants with mean framewise displacement (FD) > 0.4 mm, which enables a trade-off between sample size
April 2023

and data quality.

Volume censoring The first 10 volumes were discarded to achieve steady-state magnetization of the fMRI signal

4
Statistical modeling & inference
Model type and settings
Effect(s) tested
Specify type of analysis:
Statistic type for inference
(See Eklund et al. 2016)
Correction
Models & analysis
n/a Involved in the study
Functional and/or effective connectivity
Graph analysis
Multivariate modeling or predictive analysis
Functional and/or effective connectivity
Graph analysis
Multivariate modeling and predictive analysis Two distinct analyses were performed on the longitudinal dataset. The first analysis involved a mixed effects
nature portfolio | reporting summary
Neuroimaging analysis: We estimated thalamocortical connectopic maps, which represent dominant modes of functional
connectivity change within the thalamus, based on the voxel-wise connectivity between each thalamic voxel and the
neocortex, and its neocortical projection maps. This procedure is described in Haak et al., 2018 (NIMG).
Transcriptomic analysis: Partial least squares (PLS) analysis was used to investigate the correlation with gene expression
patterns. An ensemble of randomized null phenotypes controlling for spatial autocorrelation (SBP-spatial models) were used
to verify the statistical significance of these correlations. The surrogate maps of spatially permuting parcellated brain regions
while conserving for spatial autocorrelation were generated using the BrainSMASH python toolbox (https://
brainsmash.readthedocs.io/en/latest/). We further assessed the specificity of our genes of interest using the most stringent
null-brain-gene model that generates null distributions based on random genes that are over-expressed in brain tissues
(n=2,711 selected by q < 0.05, FDR correction, one-sided two-sample t-test comparing brain tissues and other body tissues)
via the Gene Annotation by Macroscale Brain-imaging Association Matlab toolbox. The significant gene sets were analyzed
for biological processes and cell-type specificity using ShinyGO and cell-type specific expression analysis (CSEA), respectively.
For the former, significant results are reported after Benjamin-Hochberg FDR correction (q<0.05). For the latter, statistical
significance of overlap between our candidate genes and cell-specific genes for each developmental stage was tested with
Fisher’s exact test with Benjamin-Hochberg multiple testing correction for the number of cell types assayed. Genes enriched
in a particular cell type were identified by specificity index p-value (pSI), where lower pSI values indicate more specific genes.
Computational modeling: Generative network modeling based on the cost-topology trade-off model was used to investigate
the effects of thalamocortical, genetic, spatial, and topological constraints by synthetically creating functional connectome
topology. Model performance was evaluated by comparing the synthetic networks derived from the models to empirically
extracted networks in terms of degree, clustering, betweenness, and edge length distributions. For each measure, the
distributions were compared using the Kolmogorov-Smirnov (KS) statistic, a non-parametric test that quantifies the distance
between two distributions, where lower values imply greater similarity. Detailed descriptions for generative network
modeling are in Oldham et al., 2022.
Neuroimaging analysis: These maps were statistically tested for significant aging effects at FDR-corrected p<0.05, controlling
for sex and mean FD, using surface-based linear models implemented in SurfStat.
Transcriptomic analysis: PLS analysis with an ensemble of randomized null phenotypes controlling for spatial autocorrelation
(SBP-spatial models) were used to verify the statistical significance of correlations with gene expression patterns.
Computational modeling: Model performance was evaluated by comparing the synthetic networks derived from the models
to empirically extracted networks in terms of degree, clustering, betweenness, and edge length distributions. For each
measure, the distributions were compared using the Kolmogorov-Smirnov statistic.
Whole brain ROI-based Both
Neuroimaging analysis: Voxelwise, massive univariate analysis with an FDR-corrected p<0.05
Neuroimaging analysis: False discovery rate correction was used
Transcriptomic analysis: spin tests and Benjamin-Hochberg FDR correction was used
Computational modeling: Kolmogorov-Smirnov statistic and a conservative 3-step multi-resolution optimization procedure
that sampled 10,000 combinations of different parameters was used to obtain the best-fitting parameters.
Neuroimaging analysis: Thalamocortical connectopic maps were calculated by conducting dimensionality
reduction on a similarity matrix which was derived from a Pearson correlation matrix between the functional
connectivity of thalamus and neocortex.
Computational modeling: To estimate how well the synthetic networks made from thalamocortical
connectivity models reflect the functional segregation between cortical networks as seen in the
neuroimaging data analyses, the affinity matrices derived from the synthetic networks showing best
performance were analyzed in terms of modularity network measure, which was derived using the Brain
Connectivity Toolbox.
model, where we employed the 'fitlme' function in MATLAB to execute the mixed effects analysis. In the
model, age was treated as a fixed effect and participant as a random effect to model the segregation indices
of both salience-external and salience-internal networks. Secondly, we aimed to estimate the influence of
thalamocortical connectivity development on network differentiation of the neocortex during later
April 2023
developmental stages. For this, we applied a supervised learning framework using the 'fitrlinear' function in
MATLAB, with a 5-fold cross-validation, employing least squares and lasso regularization. Specifically, we
investigated whether the difference of the thalamocortical NEOMAPs between follow-up and baseline data
can predict the segregation index between internal and external networks within the ‘cortico-cortical’
gradients. The mean average error and median correlation coefficient were used for evaluating model fit.