The genetic organization of longitudinal subcortical volumetric change is stable throughout the lifespan
Abstract
Development and aging of the cerebral cortex show similar topographic organization and are governed by the same genes. It is unclear whether the same is true for subcortical regions, which follow fundamentally different ontogenetic and phylogenetic principles. We tested the hypothesis that genetically governed neurodevelopmental processes can be traced throughout life by assessing to which degree brain regions that develop together continue to change together through life. Analyzing over 6000 longitudinal MRIs of the brain, we used graph theory to identify five clusters of coordinated development, indexed as patterns of correlated volumetric change in brain structures. The clusters tended to follow placement along the cranial axis in embryonic brain development, suggesting continuity from prenatal stages, and correlated with cognition. Across independent longitudinal datasets, we demonstrated that developmental clusters were conserved through life. Twin-based genetic correlations revealed distinct sets of genes governing change in each cluster. Single-nucleotide polymorphisms-based analyses of 38,127 cross-sectional MRIs showed a similar pattern of genetic volume–volume correlations. In conclusion, coordination of subcortical change adheres to fundamental principles of lifespan continuity and genetic organization.
Introduction
Cortical development follows a topographic organization through childhood and adolescence (Fjell et al., 2019; Krongold et al., 2017; Raznahan et al., 2011). This means that regions of the cortex with similar structural and functional properties tend to develop together (see Eickhoff et al., 2018 for a discussion of cortical topography in the context of neuroimaging). This topography is conserved through later development and aging (Fjell et al., 2019; Tamnes et al., 2013) and follows the genetic organization of the cortex, i.e. is controlled by overlapping sets of genes (Fjell et al., 2015). It is not known whether the same is true for subcortical structures. In contrast to the monotone thinning of the cerebral cortex (Storsve et al., 2014), lifespan trajectories of subcortical structures are more diverse and complex (Allen et al., 2005; Narvacan et al., 2017; Raznahan et al., 2014; Walhovd et al., 2005). This may be due to fundamental ontogenetic and phylogenetic differences between cortical and subcortical regions. The embryonic origin of the cortex is the pallium, while cerebellar and subcortical structures originate from the hindbrain, diencephalon, or subpallium (Tuller et al., 2008). These structures can be placed according to their position along the cranial vertical axis (see Table 2). Although the subcortex is evolutionary older than the cortex, it has a higher proportion of evolutionarily more recent genes, and a higher evolutionary rate, which is a basic measure of evolution at the molecular level (Tuller et al., 2008). It has also been argued that genes expressed in the subcortex generally are more region specific (Tuller et al., 2008; Zhang and Li, 2004). These mechanisms may be seen in human development and aging, with higher plasticity and potential for change in response to environmental impacts in phylogenetically older structures (Walhovd et al., 2016b), especially the hippocampus (Engvig et al., 2014; Eriksson et al., 1998). This combination of plasticity and vulnerability could contribute to the larger diversity in the lifespan trajectories of subcortical structures (Walhovd et al., 2005). On the other hand, a hypothesis is that genetically governed neurodevelopmental processes can be traced in the brain later in life (Chen et al., 2011; Satizabal et al., 2019). This would for instance entail that brain regions under shared genetic control in development continue to be influenced by the same genes and change together through life. This has been shown for the comparably less plastic cortex (Fjell et al., 2015). In light of the diverse age trajectories and high plasticity of subcortical structures, it is not known whether patterns of subcortical maturation in childhood can be traced back to principles of embryonic development, how developmental organization sets constraints on subcortical aging, and the degree to which this organization of change is under common genetic control. The aim of the present study was to address these unresolved issues about the organization of subcortical change across the lifespan. Specifically, we tested how subcortical developmental volumetric change clustered across different structures, how similar this organization was in development versus aging, and whether clusters of change were influenced by shared genetics. We hypothesized that volumetric changes in the developmental structures would tend to cluster according to embryonic principles, i.e. placement along the cranial vertical axis, that the pattern of change in aging would be similar to the pattern of change detected in childhood, and that structures changing together throughout the life would be governed by the same sets of genes.
Results
Clusters of change in development
First, we determined which regions showed correlated change in development. A single-center longitudinal dataset (Center for Lifespan Changes in Brain and Cognition [LCBC]), comprising 974 healthy participants from 4.1 to 88.5 years with a total of 1633 MRI examinations, was used. The sample was divided into development and adulthood/aging (development [<20 years], n = 644, 1021 MRIs, follow-up interval = 1.7 years [1.0–3.2]; adulthood/aging [≥20 years], n = 330, 612 MRIs, follow-up interval = 1.6 years [0.2–6.6]), see Fjell et al., 2015 for details (sample descriptives are provided in Table 1). Annual symmetrized percent change (APC) was calculated for each participant for each brain region, averaged over hemispheres, using the formula APC = (Vol Tp2 – Vol Tp1) / (Vol Tp2 + Vol Tp1) × 100. If more than two time points were available, the first and the last were used to calculate APC. These APCs were correlated across participants between each pair of brain regions. The Louvain algorithm for detecting communities in networks (Blondel et al., 2008a) was applied to derive clusters in the correlation matrix from the development sample, and the Mantel test run to compare the different matrices. Five clusters of coordinated developmental change were identified (Figure 1) (see Cluster stability analyses and Validation analyses below for a more detailed discussion and justification of the cluster solution). The optimized community-structure statistic Q, the so-called modularity, ranges between −1 and 1, and measures the relative density of connections within communities as compared to links between communities (Girvan and Newman, 2002). We compared the cluster solution’s Q value with the Q from 10,000 randomized networks preserving the signed degree distribution and rewiring each edge approximately five times using the randmio_und_signed function in the BCT. The community-structure solution was significantly more clustered than in the random networks (p<0.001, developmental change Q=0.44, the 2.5 and 97.5 percentile of the random Q distribution=0.36–0.40). Three large clusters consisted of the ventricles (Cluster 1); the brain stem, cerebellum white matter (WM) and cerebellum cortex, cortical WM, thalamus and hippocampus (Cluster 2); and cortex, putamen, amygdala, and nucleus accumbens (Cluster 3). Caudate (Cluster 4) and pallidum (Cluster 5) were represented by separate clusters.
Sample overview.
Sample | N | N longitudinal | Observations | Age Mean (range) | Sex Female/Male | Interval years Mean (range) |
---|---|---|---|---|---|---|
LCBC | 974 | 635 | 1633 | 25.8 (4.1–88.5) | 508/466 | 2.3 (0.2–6.6) |
VETSA* | 331 | 331 | 662 | 56.3 (2.6) | 0/331 | 5.5 (0.5) |
Lifebrainx | 756 | 756 | 1512 | 59.0 (19.3–89.0) | 330/426 | 2.2 (0.3–4.6) |
UKB | 38,127 | na | 38,127 | 63.6 (44–81) | 20,026/18,101 | na |
UKB long | 1337 | 1337 | 2674 | 62.5 (46–80) | 663/674 | 2.3 (2–3) |
-
*75 complete monozygote (MZ)/53 complete dizygote (DZ) pairs of male twins.
xNot including LCBC.

Volumetric change–change relationships.
Heatmaps represent pairwise correlation coefficients between volume change (annualized percent change) of the brain structures in development in the LCBC sample (left), aging in the LCBC sample (middle), and aging in the Lifebrain replication sample (right). The five clusters, delineated by the black lines, were derived from the developmental sample.
Pattern of subcortical change in adulthood and aging can be predicted from development
Next, we wanted to test whether the topographical organization of change detected in development was conserved in adulthood and aging. To this end, the pairwise change–change correlations between regions were calculated for the adult and aging sample (Figure 2), and the Mantel test was run to compare the developmental and the adult/aging matrices. The change–change matrices were more similar than expected by chance, r = 0.72 (p<0.0001), demonstrating substantial overlap of clusters of change in development and aging. The results were replicated using longitudinal data from the Lifebrain consortium (total n = 756, 1512 MRIs, mean follow-up interval = 2.3 years, age 19–89 years, mean 59.8 years), yielding almost identical results (r = 0.71, p<0.0001).

Genetic correlations.
Left: Change–change correlations in development used to generate clusters. Middle: Genetic change–change correlations, i.e.the genetic contribution to the relationships between change among any two structures, based on twin analysis (VETSA). Right: SNP genetic correlations from the UKB cross-sectional data. The five clusters, delineated by the black lines, were derived from the developmental sample.
Cluster stability analyses
We tested the stability of the identified developmental clusters. As different clustering approaches often yield different results, we ran a series of post hoc analyses to confirm the validity of the cluster solution. In our main analysis, we decomposed the correlation matrices into clusters or modules, where each module comprised regions that were more densely connected to each other – based on its correlation value – and sparsely connected to regions in other modules, by means of the commonly employed Louvain modularity algorithm (Blondel et al., 2008a). We proceeded to assess the stability and validity of the identified clusters using alternative approaches. As a sensitivity analysis, we performed consensus clustering (Lancichinetti and Fortunato, 2012; Romero-Garcia et al., 2018) combined with the versatility metric to aid selection of the resolution parameter γ for the Louvain algorithm (Shinn et al., 2017). In the main analysis, we used the default resolution parameter γ = 1, and using this for the consensus clustering yielded an identical solution. As the γ value increases from 0, the decomposition yields a progressively larger number of modules. Versatility provides an objective function to guide the choice of the resolution parameters. Specifically, the mean versatility across regions depends on how consistently the region affiliate with a specific module. An ‘optimal’ value of γ is therefore defined as a value for which the versatility is lowest, i.e. for which the decomposition is the least ambiguously defined. The mean versatility does not provide an objective global optimum of the resolution parameter γ; instead, it serves to guide optimization of γ to local minima within neighborhoods corresponding to the desired spatial resolution of the modules. Here, we calculated mean versatility across a range of resolution values, by re-running the Louvain algorithm (via the find_optimal_gamma_curve function from Shinn et al., 2017 and the consensus function therein) across the resolution range 0.01 ≤ γ ≤ 4.00, with increments of 0.01, and 1000 runs per γ value. Seven local minima of mean versatility were identified (γ = 0.66, 0.84, 0.93, 1.03, 1.12, 1.17, and 1.45). The final community partition at each γ was defined as a consensus across another 1000 runs of the Louvain modularity algorithm at the selected value of the resolution parameter. The results showed that beyond the local minima of γ ≤ 0.64, yielding ≤ three clusters, and γ ≥ 1.91 yielding 50% singleton modules (≥ eight clusters), the γ of 1.17–1.21 had the lowest versatility (versatility = 0). These γ levels all yielded the five-cluster solution of the main analysis. Also, γ = 0.7 and y = 0.96 yielded the same five-cluster solution. γ = 1.32 and γ = 1.35, the latter which also had a very low versatility of 0.001, yielded a six-cluster solution. γ = 1.60 and γ = 1.78 yielded a seven-cluster solution. Thus, the five-cluster solution was the most frequent and showed the lowest versatility over several γ levels. Specifically, the five-cluster solution appeared seven times compared with once for the three- and eight-cluster solutions and twice for the six- and seven-cluster solutions. This supported the stability of the initial solution. In the three-cluster solution, one cluster consisted of the ventricles and one consisted of pallidum, amygdala, and accumbens, while the remaining structures were included in the last cluster. Finally, we tested the similarities of the community structure (cluster solution) of the developmental and the adult/aging change–change matrices (Betzel and Bassett, 2017). We calculated the normalized mutual information (NMI) (Lancichinetti and Fortunato, 2009), variation of information (VI) (Meilă, 2003), and the z-score of the Rand coefficient (Red et al., 2011) using the Network Community Toolbox (http://commdetect.weebly.com). Null distributions were constructed by comparing the community structure of the developmental change–change matrix with the community structure derived from 10,000 randomized networks preserving the signed degree distribution of the adult/aging matrix. All of these metrics supported the conclusion that the developmental and the adulthood/ aging change matrices were more similar than expected by chance (NMI = 0.65, p=0.003; VI = 0.96, p=0.0002; zRand = 5.33, p<0.0001). Due to the nature of the research questions and data, which included gray matter (GM) and WM compartments as single structures, the null models generated were not spatially constrained (Alexander-Bloch et al., 2018; Burt et al., 2020). This may have increased the similarities between change matrices and partitions.
Patterns of change adhere to principles of genetic organization
An important part of the study was to assess whether regions within the same clusters showed shared genetic influences. Using multivariate latent change score models, we calculated the differences in subcortical volumes from baseline to follow-up in the Vietnam Era Twin Study of Aging (n = 331, mean follow-up interval = 5.5 years) and computed pairwise genetic correlations of the slope factor (i.e., change) between all regions (see Brouwer et al., 2017 for details on the statistical twin model). This yielded an estimate of how much of the change–change relationship between any two brain structures is due to common genetic influence. The latent change score model accounts for the relatedness between the twins. As the model is genetically informative, the relationship between the twins is fundamentally built into the model. The latent change score model was preferred over the simpler difference score model for the twin analyses, as it utilizes multiple sources of information within time to estimate change, restricts measurement error to the level of the observed variables, and allows for the estimation of covariance with intercept. The result is possibly a more precise estimation of change over the assessment window.
The matrix of shared genetic influences on change between each pair of brain structures (Figure 2) was tested against the developmental and the adult/aging change–change correlation matrices. The Mantel test confirmed that the shared genetic influences on the change–change relationships were statistically more similar to the pattern of correlated changes during development (r = 0.46, p=9.999e−05) and aging (r = 0.37, p<0.0002) than expected by chance. Replication was again run using Lifebrain data, yielding r = 0.37 (p<0.0004) between the matrix of shared genetic influences on change and the Lifebrain aging change–change correlation matrix.
In order to further explore the genetic contributions to coordinated subcortical change, we first attempted to calculate the pairwise single-nucleotide polymorphism (SNP)-based genetic correlation between change in each pair of structures by running a mega-analysis on 1337 participants with longitudinal MRIs from UK Biobank and 508 from LCBC. However, this initial analysis showed that statistical power – as could be expected – was too low to yield valid estimates. Thus, we instead based the SNP genetic analyses on the cross-sectional UKB data where power is much greater (n = 38,127, age 40–69 years), using age, sex, and the first 10 components of the genetic ancestry factor as covariates. A detailed overview of the pairwise genetic correlations is presented in Supplementary file 1 – SNP-based heritability estimates (Legend: pairwise co-heritability between brain structures derived from 38,127 participants from UKB). The Mantel test (r = 0.57, p<0.0005) demonstrated that the SNP genetic correlation matrix was more similar to the developmental change matrix than expected by chance. This showed that the genetic organization of subcortical structures in middle-age can be predicted from the organization of change during brain development in childhood. For completeness, we also compared the SNP genetic correlation matrix to the aging change matrix (r = 0.56, p<0.0003) and the heritability of coordinated change matrix from the VETSA sample (r = 0.38, p<0.0003), in both cases showing significantly higher similarity than expected by chance. The UKB SNP genetic correlation matrix also showed excellent coherence with the Lifebrain replication sample (r = 0.47, p<0.0006) and with the change–change matrix derived from the 1337 UKB participants with longitudinal MRI (r = 0.46, p<0.0002).
Embryonic origins of subcortical organization
We tested the hypothesis that development in childhood mimics the organization of earlier embryonic development. An overview of the clusters and their embryonic developmental origins is given in Table 2. Although a one-to-one correspondence between embryonic development and clustering of change in childhood was not expected, there were clear tendencies to conservation of embryonic developmental principles in later childhood development. The regions of Cluster 2 mainly emerged from rhombencephalon or the posterior prosencephalon, making up structures placed low on the cranial vertical axis, including brain stem (myelencephalon), cerebellum cortex and WM (metencephalon), and the thalamus (diencephalon). The exception to this was that the hippocampus and the cortical WM were also included in Cluster 2. The extensive connectivity between cerebellum and cerebrum, and the similarities in development of WM in cerebellum and cerebrum, may explain the latter finding. Clusters 3–5 comprised structures developed from subpallium/ventral telencephalon (caudate, pallidum, putamen) and pallium/dorsal encephalon (amygdala, cortex), also showing consistency with the major principles from embryonic development and placement along the cranial vertical axis.
The embryonic origins of the clusters and placement along the cranial vertical axis.
Brain structure | Cluster | Embryonic development | Cranial vertical axis | ||
---|---|---|---|---|---|
Third ventricle | 1 | Prosencephalon (posterior) | Diencephalon | ||
Fourth ventricle | 1 | Rhombencephalon | |||
Lat ventricle | 1 | Prosencephalon (anterior) | Telencephalon | ||
Inf lateral ventricle | 1 | ||||
Brainstem (medulla oblongata) | 2 | Rhombencephalon | Myelencephalon | 0 | |
Cerebellum cortex | 2 | Rhombencephalon | Metencephalon | 1 | |
Cerebellum WM | 2 | Rhombencephalon | Metencephalon | 1 | |
Thalamus | 2 | Prosencephalon (posterior) | Diencephalon | 2 | |
Hippocampus | 2 | Prosencephalon (anterior) | Telencephalon (dorsal) | Pallium (medial) | 4 |
Cortical WM | 2 | Prosencephalon (anterior) | Forebrain WM | ||
Caudate | 4 | Prosencephalon (anterior) | Telencephalon (ventral) | Subpallium | 3 |
Pallidum | 5 | Prosencephalon (anterior) | Telencephalon (ventral) | Subpallium | 3 |
Putamen | 3 | Prosencephalon (anterior) | Telencephalon (ventral) | Subpallium | 3 |
Accumbens | 3 | Prosencephalon (anterior) | Telencephalon (ventral) | Subpallium | 3 |
Amygdala | 3 | Prosencephalon (anterior) | Telencephalon (dorsal) | Pallium (lateral) | 4 |
Cortex | 3 | Prosencephalon (anterior) | Telencephalon (dorsal) | Pallium (dorsal) | 4 |
Directly to explore the relationship between principles of embryonic development and adult genetics, we applied the Louvain algorithm on the UKB SNP genetic correlation matrix. This allowed us to detail how shared genetic influences were distributed across structures in the cross-sectional UKB data. The results were then mapped according to the main stages of early brain development, from the primary brain vesicles through the secondary brain vesicles and to the developed structures. A two cluster solution yielded a trivial divide between a ventricular cluster and one cluster containing the remaining structures. Thus, we ran a separate analysis on the non-ventricular structures. This revealed a match between the adult genetic clusters and their embryonic origins (Figure 3). Pallidum, putamen, nucleus accumbens, and caudate clustered together, all originating from the subpallium (ventral telencephalon), which is the developmental origin of the basal ganglia. Amygdala, hippocampus, and the cerebral cortex clustered together, having the pallium (dorsal telencephalon) as common embryonic origin. Brainstem, cerebellum WM, and cerebellum cortex, all from the rhombencephalon, formed a third cluster, while thalamus (diencephalon) and the cerebral WM constituted the last. Although this analysis is auxiliary to the main change-analyses reported above, it demonstrates a common fundamental principle of close correspondence between embryonic brain development and the brain’s genetic architecture decades later.

Correspondence between SNP heritability and embryonic brain development.
Clustering of the non-ventricular structures was used to test how shared genetic variance were organized in the UKB sample, and the clusters were compared to the main organization of embryonic brain development. The heatmap shows the pairwise genetic correlations. The flow chart shows the main features of embryonic brain development and how the genetic clusters obtained from middle-aged and older adults follow the same organization.
Age trajectories
Each cluster and regional volume was expected to yield unique age trajectories during development and adulthood. To detail these, we fitted the developmental trajectory of each cluster by using the total volume of the structures within each cluster, with sex and intracranial volume (ICV) as covariates, by generalized additive mixed models (GAMM) (Wood, 2006; Figure 4, Table 3). Both Akaike information criterion (AIC) and Bayesian information criterion (BIC) were calculated to select among models and guard against over-fitting. These analyses were done to assess differences in the trajectories between clusters. Since the total volume was used, large structures would potentially influence the cluster trajectories more than would smaller structures. Cluster 1 increased linearly, although the rate of increase was modest. Cluster 2 showed a decreasing exponential function with volume increase leveling off after 15 years. Cluster 3 mimicked a cubic relationship, with a slight increase in volume until about 8 years, then steeper reductions, which were gradually smaller from 15 years. Cluster 4 (caudate) showed an inverted U-shaped trajectory with a sharp increase until about 9 years, and Cluster 5 (pallidum) showed a cubic trajectory with similarities to Cluster 3. Next, we calculated aging-trajectories for each of the clusters defined in the developmental sample and. The trajectories across the adult age-range differed qualitatively between clusters in terms of steepness and shape. The trajectory for each cluster represented a continuation of the developmental trend seen in childhood and adolescence. Cluster 1 showed an exponential increase, Cluster 2 an inverted U-shaped trajectory, Clusters 3 and 4 almost linear reductions, while Cluster 5 showed reductions until about 50 years and little or no change after that. For completeness, the age trajectories of the clusters were also fitted across the full age-range from 4.1 to 88.5 years (numeric results in SI).

Cluster age trajectories for each cluster, for development (left), adulthood (middle), and the full lifespan (right).
The trajectories are fitted with GAMM, and the shaded areas represent 95% CI. Note that the y-axes scales vary for easier viewing. The trajectories were estimated for development and adulthood separately to ensure that the analyses were fully independent.
Cluster age trajectories.
Numeric results for the trajectory analyses in Figure 4. Edf: effective degrees of freedom (signifying the complexity of the trajectory, where the value two approximates a quadratic shape, 3 a cubic shape, etc). The p-value is associated with the null hypothesis that there is no relationship to age.
Development | Adulthood and aging | Lifespan | |||||||
---|---|---|---|---|---|---|---|---|---|
Edf | F | p | Edf | F | p | Edf | F | p | |
Cluster 1 | 1.1 | 51.0 | 0.23e−12 | 6.0 | 67.1 | 2e−16 | 7.5 | 176.6 | 2e−16 |
Cluster 2 | 6.5 | 363.5 | 2e−16 | 6.8 | 16.0 | 2e−16 | 8.7 | 214.3 | 2e−16 |
Cluster 3 | 5.4 | 37.4 | 2e−16 | 6.7 | 85.5 | 2e−16 | 8.6 | 206.7 | 2e−16 |
Cluster 4 | 5.7 | 18.1 | 2e−16 | 1.0 | 79.4 | 2e−16 | 8.3 | 45.9 | 2e−16 |
Cluster 5 | 3.7 | 16.9 | 3.33e−12 | 3.8 | 15.8 | 1.23e−11 | 7.9 | 171.0 | 2e−16 |
Finally, we also estimated the age trajectories of each volume of 16 brain regions (Figure 5, Table 4).

Lifespan trajectories of brain volumes.
Age on the x-axis, volume in units of milliliters on the y-axis. The trajectories are fitted with GAMM, using both longitudinal and cross-sectional data, and the shaded areas represent 95% CI. Y-axis is in units of 1000 mm3. Ventricular volumes not shown.
Generalized additive mixed model fits LCBC lifespan.
Generalized additive mixed models (GAMM) were run with each neuroanatomical volume as dependent variable, and age, estimated total intracranial volume, and sex as covariates. Separate models were run with a linear age (age) term or a slope function (s(Age)). Except for cortex and caudate, the slope function yielded the lowest IC values. GM: Gray matter. WM: White matter. AIC: Akaike information criterion. BIC: Bayesian information criterion.
AIC | BIC | Effect of sex | |||
---|---|---|---|---|---|
Age | S(Age) | Age | S(Age) | p | |
Accumbens | 18,324 | 18,335 | 18,357 | 18,367 | 0.57 |
Amygdala | 20,360 | 20,048 | 20,392 | 20,081 | 0.13 |
Brainstem | 27,514 | 26,805 | 27,546 | 26,837 | 0.11 |
Caudate | 22,636 | 23,558 | 22,669 | 23,591 | 0.75 |
Cerebellum cortex | 30,079 | 29,946 | 30,111 | 29,979 | 0.94 |
Cerebellum WM | 27,430 | 27,017 | 27,463 | 27,050 | 0.16 |
Cortex | 37,525 | 37,895 | 37,557 | 37,927 | 0.72 |
Cortical WM | 36847 | 36,258 | 36,879 | 36,290 | 0.31 |
Hippocampus | 22,494 | 22,235 | 22,526 | 22,268 | 0.22 |
Pallidum | 20,937 | 20,741 | 20,969 | 20,774 | 0.17 |
Thalamus | 23,841 | 23,669 | 23,874 | 23,701 | 0.06 |
Total GM | 38,130 | 37,849 | 38,163 | 37,881 | 0.78 |
Lateral ventricles | 30,276 | 30,147 | 30,308 | 30,174 | 0.33 |
In flat vent | 21,081 | 20,931 | 21,113 | 20,964 | 0.23 |
Auxiliary analyses were done relating the clusters to general cognitive function (GCA) as measured by the Wechsler’s Abbreviated Scale of Intelligence (Wechsler, 1999) in the full LCBC sample, using sex and age as covariates. Corrected for five comparisons, two clusters showed significant (p<0.01) relationships to GCA, that is Cluster 1 (t = 3.74, p=0.00018) and Cluster 3 (t = 2.75, p=0.006). Low GCA score was associated with lower volumes. The relationships survived including intracranial volume (ICV) as covariate. ICV was not controlled for in the initial analyses since it is expected to follow brain volume development in the first part of life. For Cluster 1 was a significant interaction between GCA and age found (F = 4.59, p=0.01), while for Cluster 3, age trajectories did not differ significantly as a function of GCA (all p’s>0.46), thus showing stable relationships across life.
Validation analyses
A cluster solution will ultimately depend on which brain regions that are included and how different parameters for the clustering algorithm are defined. To test the validity of using the clusters defined in development across the adult and genetic samples, we ran two-sample Student’s t-tests to assess whether the mean intra-cluster correlation was larger than the mean extra-cluster correlation. The intra-cluster correlations were significantly higher than the extra-cluster correlations in all the tested samples, except for Cluster 3 in the VETSA sample (Table 5). As the cluster solution was defined in an independent dataset, this outcome supports the validity of each cluster, and complements the Mantel test which is based on the full correlation matrix. Second, as changes in different parts of the ventricular system were expected to be highly correlated, we re-ran the Mantel tests excluding the ventricles. As expected, the coherence between matrices was reduced, but was still significant for all comparisons, except UKB (Dev vs LCBC adulthood and aging r = 0.44, p<0.009; Lifebrain r = 0.41, p<0.0008; VETSA r = 0.44, p<0.0009; UKB r = −0.06, p=0.63). This was in line with the higher within-cluster than between-cluster correlations reported above for the non-CSF(cerebrospinal fluid) clusters. Finally, we ran the Mantel test comparing the patterns of developmental to adult and aging changes in 12 genetically defined cortical regions from the same LCBC participants (see Chen et al., 2013; Fjell et al., 2015). This yielded a correlation of r = 0.83, which suggests that the organization of developmental subcortical change is conserved through life, but to a somewhat lesser extent than the organization of cortical change.
Within- vs. outside cluster correlations.
Two-sample Student’s t-tests were run to test whether the mean correlation within the developmentally defined clusters (ri) was larger than the mean correlation between the variables in the cluster and the variables outside the cluster (re). Note that the clusters were defined in the developmental sample, which is independent from the other four samples.
Cluster 1 | Cluster 2 | Cluster 3 | |||||||
---|---|---|---|---|---|---|---|---|---|
Dataset | ri | re | p< | ri | re | p< | ri | re | p< |
UKB cross-sectional | 0.52 | −0.10 | 1e−8 | 0.30 | 0.05 | 1e−7 | 0.28 | 0.06 | 9e−7 |
VETSA (TWIN heritability) | 0.57 | −0.11 | 1e−8 | 0.40 | −0.05 | 3e−7 | 0.16 | −0.02 | 0.17 |
Lifebrain (Aging) | 0.47 | −0.16 | 1e−8 | 0.14 | −0.05 | 0.0002 | 0.14 | −0.01 | 0.004 |
LCBC (Aging) | 0.45 | −0.20 | 1e−8 | 0.30 | −0.06 | 1e−8 | 0.28 | 0.05 | 1e−7 |
-
ri: intra-cluster correlation. re: extra-cluster correlation.
Discussion
The results demonstrate that volumetric change of subcortical structures in development forms meaningful clusters. These clusters tend to follow the main cranial vertical axis from embryonic brain development, suggesting continuity from earlier to later stages of development. Although the lifespan trajectories of subcortical structures are more divergent than those for cortical regions (Fjell et al., 2014; Walhovd et al., 2011), the clusters were conserved through life. Thus, similar to what has been shown for the cerebral cortex (Fjell et al., 2015), regional subcortical volumetric changes in aging follow a similar pattern as developmental changes in childhood. This makes a strong case for the theory that early life sets the stage for aging (Jagust, 2016; Walhovd et al., 2016a; Yeatman et al., 2014). Importantly, the volumetric correlations within each cluster and the coordinated volumetric change of each cluster tended to be governed by common sets of genes. This conclusion was supported by the similarity of the volumetric change pattern during development with the pattern of genetic change–change correlations in the VETSA twin study and the cross-sectional UKB SNP co-heritability. Thus, coordination of volumetric change across subcortical structures adheres to similar principles of lifespan continuity and genetic organization previously seen for the cortex. This supports the hypothesis that genetically governed neurodevelopmental processes can be traced in subcortical structures throughout life.
Superordinate structures of change in childhood mimic embryonic brain development
The topographical development of cortical thickness is well described, with regional variability that follows certain functional, anatomical, and genetic principles of organization (Chen et al., 2013; Fjell et al., 2019; Fjell et al., 2015; Raznahan et al., 2011; Tamnes et al., 2010). Compared to the cerebral cortex, subcortical structures show highly divergent lifespan trajectories (Fjell et al., 2013; Ostby et al., 2009; Raznahan et al., 2014) and follow fundamentally different ontogenetic and phylogenetic principles (Tuller et al., 2008; Zhang and Li, 2004). On this background, the present results are intriguing. Meaningful clusters of developmental volumetric change were identified, with divergent developmental trajectories. Importantly, the different regions tended to cluster according to their embryonic origins. For instance, the majority of Cluster 2 structures emerged from rhombencephalon or the posterior prosencephalon, which is the origin of structures placed low on the cranial vertical axis. This included the brain stem (myelencephalon), cerebellum cortex, and cerebellum WM (metencephalon) and the thalamus (diencephalon). Clusters 3–5 comprise structures developed from subpallium/ ventral telencephalon (caudate, pallidum, putamen) and pallium/dorsal encephalon (amygdala, cortex). There were also departures from this principle, i.e. the inclusion of the hippocampus and the cortical WM in the same cluster. Hippocampus is part of the cerebral cortex, but develops from the medial pallium in contrast to the neocortex. Compared to many regions in the cerebral cortex, hippocampus has a more similar appearance across the range of mammal species (Bingman and Salas, 2009). Cortical WM was placed in the same cluster. Cortical WM originates from the anterior prosencephalon, but the major part of myelination occurs postnatally, so this compartment of the brain is not easily placed within the same embryonic developmental context. In addition, the WM label is anatomically gross with substantial regional differences in development (Tamnes et al., 2010). This cluster is also characterized by relatively high myelin content among several of its constituents, which may have contributed to the inclusion of hippocampus and cortical WM. Nevertheless, at a general level, change in the structures tended to cluster according to trends from embryonic development and placement along the cranial vertical axis. It must be noted as a limitation that the coherence between the developmental clusters and embryonic brain development is based on a qualitative judgment.
Consistency in patterns of change across the lifespan
We found that subcortical structures that developed together during childhood tended to change together in adulthood and aging. Mapping the developmental clusters to the adult part of the sample yielded highly different change trajectories. Except for Clusters 3 and 4, which were characterized by mostly linear reductions, differences in the shapes of the slopes were observed. This suggests that clusters identified in development continued to show independent trajectories of change through the rest of life. The lifespan trajectories showed expected shapes, with accelerated increases for the ventricular system (Cluster 1) and an inverse U-shape for cluster 2, consisting of structures such as WM, hippocampus, and the brain stem (Fjell et al., 2013; Walhovd et al., 2005). Clusters 3, 4, and 5 showed different variants of initial increase in childhood, followed by relatively linear decline through most of adulthood that leveled off at high age, with variations between clusters in terms of break points. Clusters 4 (caudate) and 5 (pallidum) consisted of only one structure each. Cluster 3, however, consisted of four structures (cortex, putamen, amygdala, and nucleus accumbens), mimicking the known trajectory of cortical volume with more or less linear decline after the peak is reached in late childhood or early adolescence.
Genetic organization of subcortical structure and change
We found that longitudinal volumetric change in regions that cluster together tend to show overlapping genetic influences. This conclusion was supported by the twin analysis, yielding higher genetic change–change correlations within versus than between clusters. This suggests that regions that develop and change together through life to a certain degree are influenced by shared sets of genes. Furthermore, the SNP analysis showed that also the cross-sectional volumetric relationships followed a similar organization. Thus, there seems to be a genetic basis for the consistent pattern of change in subcortical structures from development through the rest of life. It must be noted that although significant heritability estimates for brain changes have been demonstrated in ENIGMA, evidence for genetic variants specific for brain change was found for cerebellar GM and the lateral ventricles only (Brouwer et al., 2017). Although limited statistical power prevented strong conclusions, genetic influence on volumetric change rates and baseline volume tended to overlap for most structures. Thus, we used cross-sectional data to further explore the genetic contribution to subcortical volumetric organization. Although these data do not reflect change, they increased sample size for these analyses from 6000 to more than 38,000 MRIs. Using the expanded sample for the SNP analyses, we found further evidence that regions within each developmental cluster tended to show overlapping genetic influences. A previous multi-sample GWAS also reported significant genetic correlations between some of the subcortical structures tested in the present study (Satizabal et al., 2019). The similarity of the developmental volumetric change–change matrix and the SNP genetic volume–volume correlation matrix from middle-aged adults thus yielded further support for the hypothesis that genetically governed neurodevelopmental processes can be traced through life. Moreover, the results revealed close correspondence between the genetic organization of subcortical structures and their embryonic origins, which suggest a link from embryonic brain development to the brain’s genetic architecture in adulthood and aging. It must also be mentioned that high genetic correlations were identified between some structures from different clusters. This was especially true for hippocampus and amygdala, which showed high genetic change–change correlations in VETSA and genetic cross-sectional correlations in UKB.
Limitations: Caveats in interpreting brain changes from MRIs and further research
Similar to all studies based on in vivo imaging methods, this study provides approximations of the underlying neurobiology. The MRI-derived measures are estimations, and the segmentations are based on signal intensities and contrast properties that are prone to the influence of multiple factors (Walhovd et al., 2017). Underlying mechanisms of volume differences and change are complex and may involve events such as aborization of axons and dendrites, axonal sprouting and loss, dendritic degeneration, vascular elaboration, synaptic pruning, as well as growth and reductions of myelination (see Fjell and Walhovd, 2010 for a discussion of these issues). Many of these likely affect both contrast, signal intensity and volumetric estimations, but the relative effect of each is challenging to tease apart. We have previously shown age differences in cortical GM–WM contrast (Westlye et al., 2009) and T1w signal intensities (Westlye et al., 2010), which can also be detected longitudinally (Vidal-Piñeiro et al., 2016). The present results thus reflect effects of various neurobiological events on signal intensities and contrast. For instance, as mentioned above, myelin content has a major impact on T1w intensities, and myelin content is strongly related to age in development and aging (Grydeland et al., 2019). Thus, the clustering results will likely partly reflect different myelin content in the structures analyzed, as changes in myelin may be correlated across regions in the brain. A promising avenue for further research is to use multi-modal neuroimaging with different MRI sequences and analysis methods to yield more insight in the foundations for the volumetric changes. A second caveat is that although the clustering of regions is based on pairwise change–change correlations, this does not imply that each cluster consists of homogenous regions. Still, regions within a cluster showed more correlated volumetric change and higher genetic correlations with other regions within the cluster than with regions outside the cluster.
Conclusion
Subcortical childhood development can be described according to meaningful clusters, which are stable through life, tend to follow gradients of embryonic brain development, and tend to be influenced by shared sets of genes. Thus, the pattern of change in subcortical regions may best be understood in a lifespan perspective.
Materials and methods
Samples
Multiple independent samples were used (Table 1). Details for all samples are found in SI.
LCBC lifespan sample
Request a detailed protocolA total of 1633 valid scans from 974 healthy participants (508 females/466 males), 4.1–88.5 years of age (mean visit age 25.8, SD 24.1), were drawn from studies coordinated by the Research Group for Lifespan Changes in Brain and Cognition (LCBC), Department of Psychology, University of Oslo, Norway (Fjell et al., 2015). For 635 participants, one follow-up scan was available, while 24 of these had two follow-ups. Mean follow-up interval was 2.30 years (0.15–6.63 years, SD 1.19). Sample density was higher in childhood/ adolescence than adulthood, since we expected more rapid changes during that age period (1006 observations < 10 years, 378 observations ≥ 20 and < 60 years, and 249 observations 60–88.5 years). All participants’ scans were examined by a neuroradiologist and deemed free of significant injuries or conditions. The studies were approved by the Regional Committee for Medical and Health Research Ethics South, Norway (2010/2359; 2010/3407; 2009/200). Written informed consent was obtained from all participants older than 12 years of age and from a parent/guardian of volunteers under 16 years of age. Oral informed consent was obtained from all participants under 12 years of age.
VETSA
Request a detailed protocolThree hundred and thirty-one male twins (150 MZ/106 DZ paired twins/75 unpaired) were randomly recruited from the Vietnam Era Twin Registry and had imaging data at two time points. The study was approved by the Institutional Review Board at the University of California, San Diego. Written informed consent was obtained from all participants. Average age at baseline was 56.3 (2.6) years and follow-up interval 5.5 (0.5) years (see Kremen et al., 2013; Kremen et al., 2006). Based on demographic and health characteristics, the sample is representative of US men in their age range (Kremen et al., 2013; Schoeneborn and Heyman, 2009).
The Lifebrain Consortium
Request a detailed protocolSeven hundred and fifty-six participants with longitudinal MRI were included from the European Lifebrain project (1672 scans, baseline age 19–89 [mean = 59.8, SD = 16.4], mean follow-up interval 2.3 years, range 0.3–4.9, SD = 1.2) (https://www.lifebrain.uio.no/) (Fjell et al., 2019), including major European brain studies: Berlin Study of Aging-II (BASE-II) (Bertram et al., 2014; Gerstorf et al., 2016), the BETULA project (Nyberg et al., 2020), the Cambridge Centre for Ageing and Neuroscience study (Cam-CAN) (Shafto et al., 2014), and University of Barcelona brain studies (Abellaneda-Pérez et al., 2019; Rajaram et al., 2016; Vidal-Piñeiro et al., 2014). The study was approved by the Regional Committee for Medical and Health Research Ethics South, Norway (2017/653). Participants were screened to be cognitively healthy and in general not suffer from conditions known to affect brain function, such as dementia, major stroke, multiple sclerosis, etc. Exact screening criteria were not identical across sub-samples (see Fjell et al., 2021 for details).
UK Biobank
Request a detailed protocolThirty-eight thousand one hundred and twenty-seven participants with available MRIs and quality checked (QC) genetic information were included in the final analyses from UKB (40–69 years), see https://biobank.ndph.ox.ac.uk/. UKB has approval from the North West Multi-centre Research Ethics Committee (MREC). We received called genotypes for 488,377 subjects, of whom 40,055 had available MRIs pre-processed by FreeSurfer v6.0. We performed quality control of the genotype data at the participant level by removing participants failing genotyping QC (n = 550) or with abnormal heterozygosity values (n = 969). In addition, we removed 481 participants suggested to be removed for genetic analysis by the UK Biobank team. Ninety-one of these 481 participants had abnormal heterozygosity values, and the remaining were flagged out as outliers in heterozygosity/missing rate from the current QC files (ukb_sqc_VZ.csv) provided the most recent UK Biobank team. After excluding these subjects, we further remove related subjects by computing kinship coefficients using the program PLINK (Chang et al., 2015), with the option --kinship 0.0625. This amount to remove one subjects that are within the third degree of relatedness to any other participant. At variant level, we removed SNPs having minor allele frequency less than 0.01 or Hardy–Weinberg equilibrium test p-value < 10−6. In total, 784,356 SNPs were used in the subsequent analysis. We used the bivariate linear mixed model with genome-based restricted maximum likelihood methods implemented in the program GCTA (Yang et al., 2011) to compute genetic correlations for the volume measures for each pair of the 16 brain subcortical structures. The principal components were computed using the above quality-controlled genotypes, after removing correlated SNPs with the option –indep-pairwise 100 50 0.1 from PLINK. In addition, 1331 participants with genotyping and longitudinal MRIs available were used as an additional replication sample to test stability of the phenotypic change–change pattern.
Cognitive testing
Request a detailed protocolGCA was assessed by WASI (Wechsler, 1999) for participants aged 6.5–89 years of age, while scores for corresponding subtests (Vocabulary, Similarities, Block design, and Matrices) from the Wechsler Preschool and Primary Scale of intelligence – III (WPPSI-III) (Wechsler, 2008) were used for the youngest participants (<6.5 years) (see Walhovd et al., 2016a). All participants scored within normal IQ range (82–145) or normal range of scaled scores (mean of subtests, s = 6.67–17.33).
MRI data acquisition and analysis
Request a detailed protocolImaging data for the LCBC sample were acquired using a 12-channel head coil on a 1.5-Tesla Siemens Avanto scanner (Siemens Medical Solutions, Erlangen, Germany) at Oslo University Hospital Rikshospitalet and St. Olav’s University Hospital in Trondheim (see Walhovd et al., 2016a). See Table 6 for details regarding scanners and sequences.
Scanner and acquisition parameters.
Sample | Scanner | Field strength (Tesla) | Sequence parameters |
---|---|---|---|
LCBC | Avanto Siemens | 1.5 | TR: 2400 ms, TE: 3.61 ms, TI: 1000 ms, flip angle: 8°, slice thickness: 1.2 mm, FoV: 240 × 240 m, 160 slices, iPat = 2 |
Avanto Siemens | 1.5 | TR: 2400 ms, TE = 3.79 ms, TI = 1000 ms, flip angle = 8, slice thickness: 1.2 mm, FoV: 240 × 240 mm, 160 slices | |
Barcelona | Tim Trio Siemens | 3.0 | TR: 2300 ms, TE: 2.98, TI: 900 ms, slice thickness 1 mm, flip angle: 9°, FoV: 256 × 256 mm, 240 slices |
BASE-II | Tim Trio Siemens | 3.0 | TR: 2500 ms, TE: 4.77 ms, TI: 1100 ms, flip angle: 7°, slice thickness: 1.0 mm, FoV: 256 × 256 mm, 176 slices |
Betula | Discovery GE | 3.0 | TR: 8.19 ms, TE: 3.2 ms, TI: 450 ms, flip angle: 12°, slice thickness: 1 mm, FoV: 250 × 250 mm, 180 slices |
Cam-CAN | Tim Trio Siemens | 3.0 | TR: 2250 ms, TE: 2.98 ms, TI: 900 ms, flip angle: 9°, slice thickness 1 mm, FoV: 256 × 240 mm, 192 slices |
UKB | Skyra Siemens | 3.0 | TR: 2000 ms, TI: 880 ms, slice thickness: 1 mm, FoV: 208 × 256 mm, 256 slices, iPAT = 2 |
VETSA baseline | Siemens | 1.5 | TR = 2730ms, TI = 1000 ms, TE = 3.31ms, slice thickness = 1.33 mm, flip angle = 7°, voxel size 1.3 × 1.0 × 1.3 mm. Acquisition in Boston and San Diego. |
VETSA follow-up (Boston) | Siemens Tim Trio | 3.0 | TE = 4.33 ms, TR = 2170 ms, TI = 1100 ms, flip angle = 7°, pixel bandwidth = 140, number of slices = 160, slice thickness = 1.2 mm. Acquisition in Boston. |
VETSA follow-up (San Diego) | GE Discovery 750x | 3.0 | TE = 3.164 ms, TR = 8.084 ms, TI = 600 ms, flip angle = 8°, pixel bandwidth = 244.141, FoV = 24 cm, frequency = 256, phase = 192, number of slices = 172, slice thickness = 1.2 mm. Acquisition in San Diego. |
-
TR: Repetition time, TE: Echo time, TI: Inversion time, FoV: Field of View, iPat: in-plane acceleration.
For all samples, subcortical volumes were obtained by use of FreeSurfer (http://surfer.nmr.mgh.harvard.edu/) (Dale et al., 1999; Dale and Sereno, 1993; Fischl et al., 2002), processed with the longitudinal stream (Reuter et al., 2012). Specifically an unbiased within-subject template space and image (Reuter and Fischl, 2011) is created using robust, inverse consistent registration (Reuter et al., 2010). Several processing steps, such as skull stripping, Talairach transforms, atlas registration as well as spherical surface maps and parcellations are then initialized with common information from the within-subject template, significantly increasing reliability and statistical power (Reuter et al., 2012). For children, the issue of movement is especially important, as it could potentially induce bias in the analyses (Reuter et al., 2015). All children MRIs were manually rated for movement on a 1–4 scale, and only scans with ratings 1 and 2 (no visible or only very minor possible signs of movement) were included in the analyses, reducing the risk of movement affecting the results. Also, all reconstructed surfaces were inspected and discarded if they did not pass internal quality control. This led to the exclusion of 46 participants from MoBa-Neurocog and nine from ND, reducing the total LCBC sample to the reported 1633 scans. FreeSurfer 5.3 was used for the LCBC and VETSA analyses, while Lifebrain and UKB MRI data were processed with FreeSurfer 6.0. UKB scans were QC by the UKB imaging team. Further details of the UKB imaging protocol (http://biobank.ctsu.ox.ac.uk/crystal/refer.cgi?id=2367) and structural image processing are provided on the UK Biobank website (http://biobank.ctsu.ox.ac.uk/crystal/refer.cgi?id=1977).
Genetic correlations
Request a detailed protocolGenetic change correlations were obtained by latent change score analyses on the VETSA one and VETSA two subcortical data (see Brouwer et al., 2017). All subcortical volumes were adjusted for site and ICV. Left and right volumes at baseline and follow-up for each subject were included in a variant of the ‘latent change model’ to characterize baseline subcortical volume and change in subcortical volume across the two assessments (McArdle and Plassman, 2009), with the extension of modeling genetic and environmental effects on the phenotypes (Panizzon et al., 2015). The model allows for the estimation of the means and variances of the intercept and slope factors, the relative genetic (i.e., heritability) and environmental contributions to those variances, as well as the phenotypic, genetic, and environmental correlations between the latent factors. A genetic correlation matrix was generated by estimating genetic correlations of slope factors between all pairwise combinations of subcortical structures in bivariate latent change models.
For the UKB SNP analyses, the volume measures of the 16 subcortical structures were corrected for ICV and normalized to have zero mean and one standard deviation, separately, before estimating genetic correlations. We used the bivariate restricted maximum likelihood methods implemented in the program Genome-wide Complex Trait Analysis (GCTA, Yang et al., 2011) to compute the genetic correlation for the volume measures for each pair of the 16 brain subcortical structures, including the first 10 principal components, sex and age as covariates. The likelihood ratio test from GCTA testing whether a genetic correlation is zero was used to compute p-values for estimated genetic correlations.
Experimental design and statistical analysis
Request a detailed protocolGAMM implemented in R (http://www.r-project.org) using the package ‘mgcv’ (Wood, 2006) was used to derive age trajectories for all structures based on the 1633 LCBC MRIs. Annual symmetrized percent change (APC) in volume was correlated across structures in each sample separately (development and adult/aging from LCBC and Lifebrain). To identify clusters of correlations that could be compared across matrices, the community structure or modules in the matrices were obtained using the Louvain algorithm (Blondel et al., 2008b), part of the Brain Connectivity Toolbox (http://www.brain-connectivity-toolbox.net Rubinov and Sporns, 2010). The optimal community structure is a subdivision of the network into non-overlapping groups of regions in a way that maximizes within-group connection strength and minimizes between-group strength. The community structure may vary from run to run due to heuristics in the algorithm pertaining to the order in which the nodes are considered, so 10,000 iterations of the algorithm were run, and each region assigned to the module it was most often associated with (by taking the mode of the module assignment across iterations). Negative values were treated asymmetrically (Rubinov and Sporns, 2011). To account for global brain changes, between-regional correlations were de-meaned before they were entered into the clustering analyses.
Data availability
The study comprises many different data sources. The PI does not have the legal right to share these data directly. UK Biobank data can be obtained from http://www.ukbiobank.ac.uk. The data repository for the Cambridge Centre for Ageing and Neuroscience (Cam-CAN) dataset can be found at http://www.cam-can.org/index.php?content=dataset. Access to BASE-II data can be obtained at http://www.base2.mpg.de/7549/data-documentation. Access to VETSA data can be obtained at https://medschool.ucsd.edu/som/psychiatry/research/VETSA/Researchers/Pages/default.aspx. Betula is described at http://www.umu.se/en/research/projects/betula---aging-memory-and-dementia/. For data from Barcelona brain studies, see http://www.neurociencies.ub.edu/david-bartres-faz/. For LCBC Lifespan sample, contact information can be found at https://www.oslobrains.no/presentation/anders-m-fjell/. Part of the developmental sample can be accessed through https://www.fhi.no/en/studies/moba/for-forskere-artikler/research-and-data-access/ (As of 2021, we are in the process of transferring MRI data to this repository). Please note that for all samples, data transfer agreements must be signed and proper ethical and data protection approvals must be in place, according to national legislation. Code used for data analysis accompany the submission as separate files. The correlation matrices constituting the basis for the Mantel tests are also uploaded.
-
Cam-CAN Data RepositoryID www.cam-can.org/index.php?content=dataset. Cam-CAN.
-
BASE-II DataID data-documentation. Berlin Aging Study II.
-
Umeå UniversityID aging-memory-and-dementia/. BETULA.
-
The Institute of NeurosciencesID www.neurociencies.ub.edu/david-bartres-faz/. Barcelona Brain studies.
-
LCBCID www.oslobrains.no/presentation/anders-m-fjell/. LCBC brain studies.
-
NIPHID data-access/. MoBa neurocognitive study.
-
imaging-dataID UKBiobank. UK Biobank.
References
-
The aging brain: the cognitive reserve hypothesis and hominid evolutionAmerican Journal of Human Biology 17:673–689.https://doi.org/10.1002/ajhb.20439
-
Cohort profile: the berlin aging study II (BASE-II)International Journal of Epidemiology 43:703–712.https://doi.org/10.1093/ije/dyt018
-
BookEvolution of the hippocampusIn: Binder M. D, Hirokawa N, Windhorst U, editors. Encyclopedia of Neuroscience. Berlin, Heidelberg: Springer. pp. 98–324.https://doi.org/10.1007/978-3-540-29678-2_3158
-
Fast unfolding of communities in large networksJournal of Statistical Mechanics: Theory and Experiment 2008:P10008.https://doi.org/10.1088/1742-5468/2008/10/P10008
-
Genetic topography of brain morphologyPNAS 110:17089–17094.https://doi.org/10.1073/pnas.1308091110
-
Improved localizadon of cortical activity by combining EEG and MEG with MRI cortical surface reconstruction: a linear approachJournal of Cognitive Neuroscience 5:162–176.https://doi.org/10.1162/jocn.1993.5.2.162
-
Effects of cognitive training on gray matter volumes in memory clinic patients with subjective memory impairmentJournal of Alzheimer's Disease 41:779–791.https://doi.org/10.3233/JAD-131889
-
Neurogenesis in the adult human HippocampusNature Medicine 4:1313–1317.https://doi.org/10.1038/3305
-
Critical ages in the life course of the adult brain: nonlinear subcortical agingNeurobiology of Aging 34:2239–2247.https://doi.org/10.1016/j.neurobiolaging.2013.04.006
-
Structural brain changes in aging: courses, causes and cognitive consequencesReviews in the Neurosciences 21:187–221.https://doi.org/10.1515/REVNEURO.2010.21.3.187
-
Genes, environment, and time: the vietnam era twin study of aging (VETSA)Twin Research and Human Genetics 9:1009–1022.https://doi.org/10.1375/twin.9.6.1009
-
VETSA: the vietnam era twin study of agingTwin Research and Human Genetics 16:399–402.https://doi.org/10.1017/thg.2012.86
-
Consensus clustering in complex networksScientific Reports 2:336.https://doi.org/10.1038/srep00336
-
Comparing clusterings by the variation of informationLearning Theory and Kernel Machines 2777:173–178.https://doi.org/10.1007/978-3-540-45167-9_14
-
Evolution of deep gray matter volume across the human lifespanHuman Brain Mapping 38:3771–3790.https://doi.org/10.1002/hbm.23604
-
Health characteristics of adults aged 55 years and over: united states, 2004-2007National Health Statistics Reports 16:1–31.
-
Versatility of nodal affiliation to communitiesScientific Reports 7:4273.https://doi.org/10.1038/s41598-017-03394-5
-
Effects of age on volumes of cortex, white matter and subcortical structuresNeurobiology of Aging 26:1261–1270.https://doi.org/10.1016/j.neurobiolaging.2005.05.020
-
BookWechsler Abbreviated Scale of IntelligenceSan Antonio: The Psychological Corporation.
-
BookWechsler Preschool and Primary Scale of Intelligence - IIINorwegian: Harcourt Assessment.https://doi.org/10.1007/978-0-387-79948-3
-
BookGeneralized Additive Models: An Introduction with RChapman and Hall/CRC.https://doi.org/10.1201/9781315370279
-
GCTA: a tool for genome-wide complex trait analysisThe American Journal of Human Genetics 88:76–82.https://doi.org/10.1016/j.ajhg.2010.11.011
-
Lifespan maturation and degeneration of human brain white matterNature Communications 5:4932.https://doi.org/10.1038/ncomms5932
-
Mammalian housekeeping genes evolve more slowly than tissue-specific genesMolecular Biology and Evolution 21:236–239.https://doi.org/10.1093/molbev/msh010
Decision letter
-
Alex FornitoReviewing Editor; Monash University, Australia
-
Tamar R MakinSenior Editor; University College London, United Kingdom
In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.
Acceptance summary:
This paper makes an important contribution to understanding human brain development and will be of note to readers with a particular interest in subcortical anatomy. The work sheds new light on principles of subcortical maturation and its genetic mechanisms.
Decision letter after peer review:
Thank you for submitting your article "The genetic organization of longitudinal subcortical volumetric change is stable throughout the lifespan" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, one of whom is a member of our Board of Reviewing Editors, and the evaluation has been overseen by Tamar Makin as the Senior Editor. The reviewers have opted to remain anonymous.
The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.
The Reviewers were all enthusiastic about the potential of the paper, but had a number of methodological concerns that will be important to address. We would be happy to consider a revision provided that you are able to comprehensively address all the points raised.
Essential revisions:
1. The analysis includes data from a wide age range spanning 4 to 88 years. T1 signal characteristics, particularly those affecting tissue contrast, can change significantly over this time. This will have an impact on the accuracy of any automated segmentation algorithm. How can we be sure that age-related differences are not simply due to variations in tissue contrast? This problem is compounded by the reliance on the Freesurfer aseg algorithm, which parcellates the brain using an adult training set. Thus, not only will there be a problem of age-related differences in tissue contrast, but also in the accuracy with which individual T1s can be spatially aligned to the template, which is likely to decline as a function of age difference from the young adults used to generate the templates/training set. The authors should demonstrate that such effects cannot explain their findings.
2. The authors conclude that "at a general level, change in the structures tended to cluster according to trends from embryonic development and placement along the cranial vertical axis, but with notable exceptions.". This is based on the unsupervised clustering solution from correlating inter individual variation in change for each structure (which they argue is reproduced in development aging and genetic analyses – see point 2 below). Looking at their clustering solution though, a more superficial explanation could easily explain the 3 main clusters they observe: Cluster 1 = fluid-filled ventricles, Cluster 2 = White matter (or white matter rich for thalamus, hippocampus and cerebellar cortex) ROIs, Cluster 3 = Most other ROIs, which are general non-white matter rich cortical or subcortical nuclei. I understand that linkage to embryological patterning is more profound – but there must be some effort to address the more concrete possibility that the clustering is basically saying CSF goes with CSF, white with white, and gray with gray (put crudely). Although thalamus, hippocampus and cerebellar cortex are classed as gray by tissue classification algorithms – it is well recognized that these structures also contain substantial white matter components. We suggest comparing mean MT or FA for Clusters 1 and 2 to get at this empirically. Evidence for the "tissue composition" hypothesis for clustering is not incompatible with the "embryological origins" hypothesis – but focusing on the latter to the extent that the authors do would be more justified if there were not evidence of the former.
3. The authors conclude that their findings show "substantial coherence in the pattern of change between development and the rest of the life". This major finding is based mainly on the cross-cell correlations and Mantel tests for compare different correlation matrices: (i) correlated intra-individual changes in development, (ii) correlated changes in adult/aging, (iii) genetic correlation in change, and (iv) SNP-based estimated of co-heritability. The dominant signal by far within any one of these matrices, and the main signal that appears to be recovered across matrices is the contrast between ventricular measures and non-ventricular measures. The reproducibility of this contrast is less meaningful than the reproducibility e.g. based on correlation across all non-ventricle edges between matrices. First and foremost, a fundamental issue with giving equal weight to ventricular vs. non-ventricular measures in their contribution to statements of similarity between matrices is that the ventricles are a single continuous fluid filled cavity, so that saying – eg lateral and lateral inf ventricles are correlated is considerably different from saying amygdala and thalamus are. Compounding this concern is that the weight given the ventricle-related edges in driving cross-matrix correspondences is somewhat arbitrarily defined by how many parts you want to chop the ventricles into. This level of arbitrariness does not apply for example for the distinction between the other main ROIs included. To address these issues, the cross-edge correlation between matrices should be presented and probed through scatterplots. Scatterplots for these edge-wise correlations between matrices should lay out the issue (using edge here to mean cell in the matrix). The ~0.8 value is likely to be inflated by dense point of edges to the top right and bottom left of the plot relating to inter ventricular edges and ventricle-nonCSF edges respectively. These scatterplots between matrices should be visualized with all edges, then with just the edges that do not include ventricles. The correlation should be recalculated for the latter, and a fit line shown to help readers identify those edges that show weakest correspondence between matrices. A v efficient way of looking at the sensitivity tests above is constructing a symmetric 5*5 matrix – each column/row is a matrix from Figures 2 and 4. The cells show inter-matrix correlations. Show this one with ventricles in, and once with them out. The core issue here is showing that the signals being discussed and interpreted aren't overly dependent in the simple contrast between tissue and non-tissue compartments.
4. The Mantel test is likely biased in the case of spatially autocorrelated data (Guillot and Rousset, 2013, Methods in Ecol and Evol). The authors should consider generating spatially constrained null models (e.g., Alexander-Bloch et al. NeuroImage, 2018; Burt et al. NeuroImage, 2020) for inference on the Mantel test.
5. It may be, in part, due to how you phrase some points but you seem to suggest that volumetric correlation is analogous to volumetric change correlation, which is not supported by the analyses you perform. The fact that the matrices of correlations between volume and volume change resemble each other does not prove they are driven by similar (genetic) factors. You would need to evaluate correlation between volume and volumetric change, in particular at a genetic level to support this claim. Here is a selection of the sentences I found misleading or problematic:
This was confirmed by single nucleotide polymorphisms-based analyses of 38127 cross-sectional MRIs.
Importantly, both the structure and the coordinated change of each cluster tended to be governed by common sets of genes
This means that regions that develop and change together through life tend to be governed by the same sets of genes
Rather, genetic influence on change rates and baseline volume overlapped for most structures. This finding allowed us to use cross-sectional data from UKB to further explore the genetic contribution to subcortical organization.
The similarity of the developmental change matrix and the SNP genetic correlation matrix obtained from middle-aged adults thus yielded further support for the hypotheses that genetically governed neurodevelopmental processes can be traced in subcortical structures through life.
6. You also seem to assume that the clusters defined from pairwise correlations would consist in homogeneous regions (e.g. with common genetics or determinants). I would argue that this is not necessarily the case, and it could be investigated at a genetic level using multivariate twin models (e.g. common pathway model).
7. I would find useful that you highlight the significant phenotypic and genetic correlations (e.g. in Figure 2 and 4). For example, I would find a cluster consisting of significant correlations more convincing. On the other hand, a significant positive correlation between clusters would suggest that the clusters are not independent, which is also interesting.
8. In my opinion, the description of age trajectories, albeit extremely interesting, are not central to the scientific question you are trying to tackle. I think it would improve readability to present them after the main results in the same section as the cluster trajectories. In particular, reading the result section led me to think that GAMM were central to your analysis, while they are only used for a visualisation of the age trajectories, which only serves a descriptive purpose.
9. More generally, the result section would benefit from being more linear (starting from the main results) and/or would benefit from a better correspondence between the figures and the text. At the moment, it feels like the figures and text have different progressions all together, which is confusing. For example, the results presented in Figure 2 are spread in different sections of the text that are not even following each other. Figure 4 would make sense near Figure 2, or even as an additional panel of Figure 2. An example of what I think would be a more linear progression of the results.
o E.g. Phenotypic correlations/clusters.
o Replication in adults.
o Sensitivity analysis.
o Genetic correlations clusters.
o Relationship with embryonic development.
o Visualisation of age trajectories of regions and clusters.
o Additional analyses (e.g. cognition).
10. Modularity analysis – There does not appear to be any assessment of the statistical significance of the cluster solution. This could be done by shuffling the correlation matrix using appropriate algorithms, depending on the degree to which the authors want to consider spatial effects in the data (e.g., Roberts et al. 2015, NeuroImage; Rubinov and Sporns, NeuroImage, 2011).
11. Modularity analysis – it appears the authors used a standard implementation of the Louvain algorithm. The default definition of Q defines a null expectation for within-cluster connectivity that is not appropriate for correlation matrices. With such data, the mean correlation can be appropriate.
12. Modularity analysis – the evaluation of different levels of γ is commendable. It is unclear why this is left to the end of the Results. This should either be placed at the beginning, as a way of justifying the 5-cluster solution, or the beginning should indicate what value of γ was used initially and contain a reference to subsequent analyses that investigate the issue of γ in more detail.
13. The similarity between the clusters and developmental origins is qualitative. Given that this is a core aim of the analysis, perhaps the authors could perform some inference by shuffling the labels and using a measure of partition similarity such as their normalized mutual information?
14. Please explain why the cortex is treated as a single structure? Cortical development is regionally heterogeneous.
15. Sex is known to modify trajectory shape for cortical and subcortical structures. It would be good to show that the observed clustering holds for male and female subsamples. This would provide an important sensitivity analysis and also potential evidence for a sort of split half reliability.
16. It would also be important to show how ROIs cluster when you use the first derivatives from the gam fits in Figure 1. That would provide a complimentary approach to the inter-individual change method on which current work is built, and also help unpack some of the potential concerns re "fluid vs. tissue" and "white right vs. other" as two big potential drivers for findings. I would add plots for the ventricular components in Figure 1 too, as well as dendrograms for the primary heat map clustering used to order other matrices. My guess would be split one is CSF vs. tissue, and split 2 is white/white-rich tissue vs. other tissue.
17. In estimating APCs – how did the authors deal with were people with more than 2 scans, and the potential for including mid-scan age as a factor before correlating given the non-linear volume changes in development?
18. I may have missed it but I did not expect the analysis of cognitive function. What purpose it is serving and does it really integrate in the theory of neurodevelopment? In addition, only one association seems significant after multiple testing and the difference seems to be localised in the 25-60 years old group for which you have the fewest observations (from Figure 1). Could this be due to a handful of outliers? Could you also add precision about which test was used? Maybe reporting several curves for difference quantiles of the distribution (instead of above/below average) would help visualise the effect?
19. Figures 4 and 5 need a color scale.
20. Figure 2 is the main figure of result and I strongly suggest you expand the caption to improve self-readability. For example, the fact that the first row are results obtained on the LCBC sample. Clarify that the clustering was not recalculated for each matrix but is that of the first panel. In addition, the titles of the correlation plots could be clearer (e.g. clearly stating phenotypic and genetic correlations). Also the 4th panel is not a volumetric change-change relationship, which is the title of the figure.
21. Abstract and introduction should announce that the focus is about volume of subcortical structures.
22. The discussion reads well but contains many statements that do not seem supported by the results or that sounded too definitive considering the analysis only focused on 16 volumetric measurements. It also lacks a limitation section.
23. Some aspects of the text could be clearer. For example:
Line 44 – it is unclear what the hypothesis is precisely.
Line 59 – what does a topographic organization correspond to in this context? regionally specific?
Line 61 – it is unclear what it means to follow the "genetic organization of the cortex".
Line 79 – it is very hard to understand what this hypothesis refers to and what it predicts.
Line 88 – it is unclear what a genetic anatomical architecture is.
Line 105 – please clarify: was the correlation across subjects?
Line 121 – it is unclear what "common factors" refers to here.
Line 217 – only a correlation rather than prediction has been shown.
Line 256 – please clarify: was consensus clustering run over 1000 runs at each γ?
“Although the lifespan trajectories of subcortical structures are much more divergent than those for cortical regions.”
Unclear to me how this is supported by results from the article, could you expand on how you conclude this?
“Mapping the developmental clusters to the adult part of the sample yielded highly different change trajectories.”
Here it may help to cite figures and tables in Discussion – also is this the case for all clusters?
“Clusters 1, 3 and to a lesser extent 4 were related to general cognitive function in a mostly age-invariant manner, which implies that the relationship between cognitive function and subcortical volumes is established early in life.”
This sounds like a huge overstatement. You studied only volume change of 16 regions – and the associations with cognition scores are not the most convincing.
“This means that regions that develop and change together through life tend to be governed by the same sets of genes.”
Beyond my previous criticism, this reads as a rather large generality.
“…and are governed by distinct sets of genes.”
What is this conclusion based on?
“It has also been argued that genes expressed in the subcortex generally are more region-specific and tend to evolve more rapidly than genes expressed in cortical regions”.
I am not sure what you mean by "evolve more rapidly" is it that it is under greater selection pressure?
The SNP genetic correlation matrix was highly similar to the developmental change matrix, as demonstrated by the Mantel test (r = 0.57, p <.0005, see Figure 2).”
I am not a big fan of qualitative judgments (e.g. highly). Especially that I assume r can vary between 0 and 1 so "highly" can be seen as an overstatement.
As much as the Mantel test statistics are different from 0, are they also different from 1?
Generalized
“Additive Mixed Models (GAMM)”
Could you add some details about the maximal order of the splines you considered? How was the best model – best order selected?
“This revealed a close to perfect match between the adult genetic clusters and their embryonic origins.”
24. You say grouping is extremely similar, but compared to the clusters presented in Figure 2, I feel that this is an overstatement. E.g. caudate, accumbens, putament, pallidum – two of those were in separate clusters previously. Also positive rG between cerebellum cortex and WM while this was negative before.
25. “In all cases yielded the slope function the lowest IC values”. This is not true – see caudate and cortex.
What to make from small AIC/BIC differences i.e. a marginally better fit? I wonder whether Table 3 is really adding anything, especially that the age trajectories are mostly descriptive?
26. “We fitted the developmental trajectory of each cluster…”.
I assume you took the total volume of each cluster? What is this analysis really adding? First it assumes the cluster is somewhat homogenous (see my second comment) and another problem is that the different volumes can have extremely different scales which makes interpreting the sum difficult.
27. “Using multivariate latent change score models, we calculated…”.
Why did you not also use the annual symmetrized percent change you used previously? Especially that you want to compare rG to the phenotypic correlations you previously studied.
28. rGs from twin models are corrected for ICV, which does not seem to be the case for the GREML approach. Is this simply missing in the text?
“…pair of the 16 brain sub-cortical structures, including the first ten principal components, sex and age as covariates.”
29. “log likelihood test”. I am more used to it being referred to as the likelihood ratio test.
30. “restricted maximum likelihood methods” – (very) minor detail but REML is the optimisation approach used to estimate the parameters of what is a bivariate mixed model. A compromise may be to talk about GREML, which is not more correct but quite commonly used to refer to the LMM implemented in GCTA. For an analogy, it is as if you referred to the GAMM as a Restricted Marginal Likelihood method.
31. “…due to heuristics in the algorithm…”. What are the heuristics here, is it random starting values?
32. “Also, all reconstructed surfaces were inspected, and discarded if they did not pass internal quality control.” Surely, this does not apply to the UKB analyses. Did you perform any QC on them beyond the UKB provided screening?
33. “In addition, we removed participants suggested to be removed for genetic analysis by the UK Biobank team.” Could you add a precision about why this exclusion is recommended?
https://doi.org/10.7554/eLife.66466.sa1Author response
Essential revisions:
1. The analysis includes data from a wide age range spanning 4 to 88 years. T1 signal characteristics, particularly those affecting tissue contrast, can change significantly over this time. This will have an impact on the accuracy of any automated segmentation algorithm. How can we be sure that age-related differences are not simply due to variations in tissue contrast? This problem is compounded by the reliance on the Freesurfer aseg algorithm, which parcellates the brain using an adult training set. Thus, not only will there be a problem of age-related differences in tissue contrast, but also in the accuracy with which individual T1s can be spatially aligned to the template, which is likely to decline as a function of age difference from the young adults used to generate the templates/training set. The authors should demonstrate that such effects cannot explain their findings.
We fully agree that there are substantial age-differences in T1 signal intensities and tissue contrast. In previous work, we have shown that GM/WM contrast in T1w images is lower in older than younger participants (Increased sensitivity to effects of normal aging and Alzheimer's disease on cortical thickness by adjustment for local variability in gray/white contrast: A multi-sample MRI study – ScienceDirect) and that T1w signal intensity is related to chronological age both within the cortex and in subcortical structures (Differentiating maturational and aging-related changes of the cerebral cortex by use of thickness and signal intensity – PubMed (nih.gov)). GM/WM contrast changes can reliably be measured longitudinally, and the rate of contrast decay seems related to the initial regional myelin content (Accelerated longitudinal gray/white matter contrast decline in aging in lightly myelinated cortical regions – Vidal‐Piñeiro – 2016 – Human Brain Mapping – Wiley Online Library). Almost certainly, such contrast differences will affect brain segmentations, as we showed in the first paper referenced above, and as we discuss in more detail in a developmental context in a separate paper (Through Thick and Thin: a Need to Reconcile Contradictory Results on Trajectories in Human Cortical Development | Cerebral Cortex | Oxford Academic (oup.com)). In short, we found that rather than contrast differences inflating structural age-relationships, correcting for them increased age differences and sensitivity to detect AD. Thus, it is highly unlikely that the reported age-related differences are inflated as a function of contrast and intensity differences. However, importantly, as with all in vivo imaging methods, our study provides merely representations of the underlying neurobiology and inherently require some level of interpretation. It is critical to be aware of the fact that the MRI-derived measures are merely our best current approximations, where segmentations are based on signal intensities and contrast properties that are prone to the influence of multiple factors, including, but not limited to, age. We have previously suggested that the term “apparent cortical thickness” should replace “cortical thickness”, and the same could apply to any MRI-segmented brain volume. Underlying mechanisms of volume differences and change are complex and may involve events such as growth, proliferation of dendrites, dendritic spines, axonal sprouting, vascular elaboration, synaptic pruning as well as myelination. Many of these will likely affect both contrasts, signal intensities and volumetric estimations, but the relative effect of each is challenging to tease apart.
On this background, it is very interesting that we find high stability in the organization of change-change patterns in development and aging. It may be that contrast changes are relatively minor since analyses are based on within-subject longitudinal changes covering only a few years. For the same reasons, effects of inter-subject registrations to the probabilistic atlas have less impact since the used metric is within-subject change.
We agree that these issues are interesting, and that they deserve a more in-depth discussion in the manuscript. We have added the following paragraph to the Discussion section:
“Limitations: Caveats in interpreting brain changes from MRIs and further research
Similar to all studies based on in vivo imaging methods, this study provides approximations of the underlying neurobiology. […] A promising avenue for further research is to use multi-modal neuroimaging with different MRI sequences and analysis methods to yield more insight in the foundations for the volumetric changes studied.”
2. The authors conclude that "at a general level, change in the structures tended to cluster according to trends from embryonic development and placement along the cranial vertical axis, but with notable exceptions.". This is based on the unsupervised clustering solution from correlating inter individual variation in change for each structure (which they argue is reproduced in development aging and genetic analyses – see point 2 below). Looking at their clustering solution though, a more superficial explanation could easily explain the 3 main clusters they observe: Cluster 1 = fluid-filled ventricles, Cluster 2 = White matter (or white matter rich for thalamus, hippocampus and cerebellar cortex) ROIs, Cluster 3 = Most other ROIs, which are general non-white matter rich cortical or subcortical nuclei. I understand that linkage to embryological patterning is more profound – but there must be some effort to address the more concrete possibility that the clustering is basically saying CSF goes with CSF, white with white, and gray with gray (put crudely). Although thalamus, hippocampus and cerebellar cortex are classed as gray by tissue classification algorithms – it is well recognized that these structures also contain substantial white matter components. We suggest comparing mean MT or FA for Clusters 1 and 2 to get at this empirically. Evidence for the "tissue composition" hypothesis for clustering is not incompatible with the "embryological origins" hypothesis – but focusing on the latter to the extent that the authors do would be more justified if there were not evidence of the former.
We agree with the reviewers that tissue composition contributes to the observed clustering, and also that this is not incompatible with the embryological origin hypothesis we have used to guide the interpretation of the results. Myelin content is almost certainly a factor that impacts the clustering. We acknowledged this in the original submission also (Discussion):
“This cluster is also characterized by relatively high myelin content among several of its constituents, which may have contributed to the inclusion of hippocampus and cortical WM.”
We have now expanded on this important point (Discussion):
“The present results reflect effects of various neurobiological events on signal intensities and contrast. […] Thus, the clustering results will likely partly reflect different myelin content in the structures analyzed, as changes is myelin may be correlated across regions in the brain.”
We appreciate the reviewers’ suggestion of running additional analyses by use of MT or FA. However, we believe the results of these will be challenging to interpret with regard to myelin. Although FA probably is related to myelin content, it is a nonspecific measure since there is not a direct correspondence between an FA value and a WM cellular component. For instance, regardless of myelin, high axon packing density and low axon diameter translate into high FA values due to high membrane density perpendicular to the axon. Further, high FA is seen in major tracts in newborns almost without cerebral myelin, and classic studies in rodents found reductions of only about 20% in FA in absence of myelin (shiver mice). Thus, if FA values were higher in cluster 2 than cluster 3, it would be difficult to ascribe this to myelin content. And vice versa, if there were no differences in FA values, this would still not rule out the tissue composition account. Similar considerations regard other MRI-based in vivo measures of white matter and myelin, such as MT, which measures myelination indirectly and can be influenced by for instance water content and neuroinflammation. This was also the main conclusion in a recent eLife paper (An interactive meta-analysis of MRI biomarkers of myelin | eLife (elifesciences.org)):
“Similarly to other qMRI biomarkers, MRI-based myelin measurements are indirect, and might be affected by other microstructural features, making the relationship between these indices and myelination noisy.”
Regardless of these caveats, we agree with the reviewer that combining different imaging modalities has potential to yield interesting information. We have therefore included a sentence stating this in the revised manuscript (Discussion):
“A promising avenue for further research is to use multi-modal neuroimaging with different MRI sequences and analysis methods to yield more insight in the foundations for the volumetric changes studied.”
Regarding CSF, please see comments to the point below.
3. The authors conclude that their findings show "substantial coherence in the pattern of change between development and the rest of the life". This major finding is based mainly on the cross-cell correlations and Mantel tests for compare different correlation matrices: (i) correlated intra-individual changes in development, (ii) correlated changes in adult/aging, (iii) genetic correlation in change, and (iv) SNP-based estimated of co-heritability. The dominant signal by far within any one of these matrices, and the main signal that appears to be recovered across matrices is the contrast between ventricular measures and non-ventricular measures. The reproducibility of this contrast is less meaningful than the reproducibility e.g. based on correlation across all non-ventricle edges between matrices. First and foremost, a fundamental issue with giving equal weight to ventricular vs. non-ventricular measures in their contribution to statements of similarity between matrices is that the ventricles are a single continuous fluid filled cavity, so that saying – eg lateral and lateral inf ventricles are correlated is considerably different from saying amygdala and thalamus are. Compounding this concern is that the weight given the ventricle-related edges in driving cross-matrix correspondences is somewhat arbitrarily defined by how many parts you want to chop the ventricles into. This level of arbitrariness does not apply for example for the distinction between the other main ROIs included. To address these issues, the cross-edge correlation between matrices should be presented and probed through scatterplots. Scatterplots for these edge-wise correlations between matrices should lay out the issue (using edge here to mean cell in the matrix). The ~0.8 value is likely to be inflated by dense point of edges to the top right and bottom left of the plot relating to inter ventricular edges and ventricle-nonCSF edges respectively. These scatterplots between matrices should be visualized with all edges, then with just the edges that do not include ventricles. The correlation should be recalculated for the latter, and a fit line shown to help readers identify those edges that show weakest correspondence between matrices. A v efficient way of looking at the sensitivity tests above is constructing a symmetric 5*5 matrix – each column/row is a matrix from Figures 2 and 4. The cells show inter-matrix correlations. Show this one with ventricles in, and once with them out. The core issue here is showing that the signals being discussed and interpreted aren't overly dependent in the simple contrast between tissue and non-tissue compartments.
We agree that different CSF compartments may be expected to correlate, probably due to general developmental growth or age-related atrophy, and this is also demonstrated by the relatively high change-change correlations for the different ventricular/ CSF variables. The analyses were also run without inclusion of the ventricles, and although the r-values were reduced, the matrices were still significantly more similar than what would have been expected by chance. To test how the clustering was affected by not including CSF, the following analyses were presented in the manuscript (Results section):
“Third, as changes in different parts of the ventricular system were expected to be highly correlated, we re-ran the Mantel tests excluding the ventricles. […] This was in line with the higher within-cluster than between-cluster correlations reported above for the non-CSF clusters.”
Thus, the reviewer is right in assuming that CSF compartments affected the similarities between the clusters, but the compared change-change matrixes are still significantly more similar than expected by chance when the CSF compartments are not included in the matrix comparisons. We also ran the clustering in the UKB SNP co-heritability matrix with and without CSF compartments included:
“A two cluster solution yielded a trivial divide between a ventricular cluster and one cluster containing the remaining structures. Thus, we ran a separate analysis on the non-ventricular structures. This revealed a match between the adult genetic clusters and their embryonic origins (Figure 3).”
Figure 3 also shows the cluster solution for this analysis not including CSF compartments. Finally, in the revised manuscript, we also describe the three-cluster solution. The following is added to the manuscript (Results, section on Cluster stability analyses):
“In the 3-cluster solution, one cluster consisted of the ventricles, one consisted of pallidum, amygdala and accumbens, while the remaining structures were included in the last cluster.”
Thus, we strongly believe the signals are not merely reflecting the simple contrast between tissue and non-tissue compartments.
4. The Mantel test is likely biased in the case of spatially autocorrelated data (Guillot and Rousset, 2013, Methods in Ecol and Evol). The authors should consider generating spatially constrained null models (e.g., Alexander-Bloch et al. NeuroImage, 2018; Burt et al. NeuroImage, 2020) for inference on the Mantel test.
We see the reviewer’s point. A challenge in our data, however, is that GM and WM cerebrum and cerebellum are included, in which distance measuring becomes problematic as such spatially extended structures do not lend themselves well to either three-dimensional Euclidean distance (subcortical) or surface-based geodesic distance (Burt et al. NeuroImage, 2020). We therefore consider it impractical to generate spatially constrained null models without deviating excessively from the original structures. To take this limitation of the Mantel test into account, in the revised manuscript we have added alternative analyses following Betzel and Bassett (2017) to compare the matrices, not vulnerable to spatial constraints (please see response to point 10 below).
5. It may be, in part, due to how you phrase some points but you seem to suggest that volumetric correlation is analogous to volumetric change correlation, which is not supported by the analyses you perform. The fact that the matrices of correlations between volume and volume change resemble each other does not prove they are driven by similar (genetic) factors. You would need to evaluate correlation between volume and volumetric change, in particular at a genetic level to support this claim. Here is a selection of the sentences I found misleading or problematic:
This was confirmed by single nucleotide polymorphisms-based analyses of 38127 cross-sectional MRIs.
Importantly, both the structure and the coordinated change of each cluster tended to be governed by common sets of genes
This means that regions that develop and change together through life tend to be governed by the same sets of genes
Rather, genetic influence on change rates and baseline volume overlapped for most structures. This finding allowed us to use cross-sectional data from UKB to further explore the genetic contribution to subcortical organization.
The similarity of the developmental change matrix and the SNP genetic correlation matrix obtained from middle-aged adults thus yielded further support for the hypotheses that genetically governed neurodevelopmental processes can be traced in subcortical structures through life.
We did not intend to imply that volumetric correlations are analogous to volumetric change correlations. Except for the UKB SNP-analyses, which are based on volume-correlations, all analyses in the paper are longitudinal (change). Unfortunately, the longitudinal UKB SNP analyses did not yield stable results due to lack of power (Results):
“In order to further explore the genetic contributions to coordinated subcortical change, we first attempted to calculate the pairwise single nucleotide polymorphism (SNP)-based genetic correlation between change in each pair of structures by running a mega-analysis on 1337 participants with longitudinal MRIs from UK Biobank and 508 from LCBC. […] Thus, we instead based the SNP genetic analyses on the cross-sectional UKB data where power is much greater (n = 38127, age 40-69 years), using age, sex and the first 10 components of the genetic ancestry factor as covariates.”
Thus, UKB results are cross-sectional only, and statements about genetic contributions to the change correlations are based on the twin analyses, for which power is greater to detect genetic contributions to change compared with SNP-bases analyses.
To avoid misunderstandings, we have re-formulated the statements highlighted by the reviewers as confusing:
“This was confirmed by single nucleotide polymorphisms-based analyses of 38127 cross-sectional MRIs.” “Single nucleotide polymorphisms-based analyses of 38127 cross-sectional MRIs showed a similar pattern of genetic volume-volume correlations.”
“Importantly, both the structure and the coordinated change of each cluster tended to be governed by common sets of genes” “Importantly, both the volumetric correlations within each cluster and the coordinated change of each cluster tended to be governed by common sets of genes.”
“This means that regions that develop and change together through life tend to be governed by the same sets of genes” “We found that longitudinal volumetric change in regions that cluster together are influenced by the same genes. […] Further, the SNP genetic correlation analysis showed that also the cross-sectional volumetric correlations followed a similar organization.”
“Rather, genetic influence on change rates and baseline volume overlapped for most structures. This finding allowed us to use cross-sectional data from UKB to further explore the genetic contribution to subcortical organization.” “Thus, we used cross-sectional data from UKB to further explore the genetic contribution to subcortical volumetric organization. Although the cross-sectional nature of these data prevents conclusions about change-change relationships per se, they increased sample size for these analyses from 6000 to more than 38000 MRIs.”
“The similarity of the developmental change matrix and the SNP genetic correlation matrix obtained from middle-aged adults thus yielded further support for the hypotheses that genetically governed neurodevelopmental processes can be traced in subcortical structures through life.” “The similarity of the developmental change matrix and the SNP genetic volume-volume correlation matrix obtained from middle-aged adults thus yielded further support for the hypotheses that genetically governed neurodevelopmental processes can be traced in subcortical structures through life.”
6. You also seem to assume that the clusters defined from pairwise correlations would consist in homogeneous regions (e.g. with common genetics or determinants). I would argue that this is not necessarily the case, and it could be investigated at a genetic level using multivariate twin models (e.g. common pathway model).
We agree with the reviewer that the clusters will not consist of homogenous regions. What we attempt to show is that there are meaningful clusters that can be identified, within which change is more highly correlated and genetic influence is shared to a larger degree than what it the case for regions outside the clusters. With this, however, we do not mean to imply that the regions within each cluster are homogenous. To make this clear, we have added a paragraph to the Limitation section of the Discussion:
“A second caveat is that although the clustering of regions is based on pairwise change-change correlations, this does not imply that each cluster consists of homogenous regions. Regions within a cluster show more correlated volumetric change with other regions within the cluster than with regions outside the cluster, and the genetic analyses show higher genetic correlations for change and absolute volume with other within-cluster regions than regions outside the cluster.”
7. I would find useful that you highlight the significant phenotypic and genetic correlations (e.g. in Figure 2 and 4). For example, I would find a cluster consisting of significant correlations more convincing. On the other hand, a significant positive correlation between clusters would suggest that the clusters are not independent, which is also interesting.
We understand the reviewers’ point. However, the focus of the paper is the patterns of change, not the pairwise relationships per se. Further, with our sample size, we believe the significance values are less relevant. For instance, in Lifebrain and for UKB genetic correlations, the critical r-value is below 0.075 for p <.05 (two-tailed), and for the cluster forming LCBC development sample, critical r = .10. The majority of the pairwise within-cluster correlations are significant at p <.05 (48 out of 60 for UKB, see supplemental information, 41 out of 60 for LB), but explained variance may be as low as < 1%, and we therefore believe the significance testing of the correlations presented in Table 5 is more appropriate than reporting p-values for each pairwise correlation. The p-values in Table 5 are not dependent on sample size, which is a further complicating issue when reporting p-values: two identical matrices may show substantial differences regarding which correlations are significant and which are not. As the analyses are based on the pattern of correlations, we are afraid that including pairwise significance levels in the figures will be confusing. Thus, we would prefer not to highlight significance levels of the individual cells in figures 1 and 2.
8. In my opinion, the description of age trajectories, albeit extremely interesting, are not central to the scientific question you are trying to tackle. I think it would improve readability to present them after the main results in the same section as the cluster trajectories. In particular, reading the result section led me to think that GAMM were central to your analysis, while they are only used for a visualisation of the age trajectories, which only serves a descriptive purpose.
Thank you for this suggestion. We have now moved the description of the age trajectories as suggested by the reviewer, and present them after the cluster trajectories. We have also completely re-ordered the Results section according to the reviewer request #9 below.
9. More generally, the result section would benefit from being more linear (starting from the main results) and/or would benefit from a better correspondence between the figures and the text. At the moment, it feels like the figures and text have different progressions all together, which is confusing. For example, the results presented in Figure 2 are spread in different sections of the text that are not even following each other. Figure 4 would make sense near Figure 2, or even as an additional panel of Figure 2. An example of what I think would be a more linear progression of the results.
o E.g. Phenotypic correlations/clusters.
o Replication in adults.
o Sensitivity analysis.
o Genetic correlations clusters.
o Relationship with embryonic development.
o Visualisation of age trajectories of regions and clusters.
o Additional analyses (e.g. cognition).
We agree that this is a good idea. We have now re-ordered the whole Results section. We have also replaced Figure 1 and Figure 2 with new figures better suited to illustrate the new organization of the results. Figure 1 now contains all the three phenotypic change-change correlation matrices. Figure 2 contains the genetic matrices. We believe this re-organization has improved the structure of the manuscript considerably.
10. Modularity analysis – There does not appear to be any assessment of the statistical significance of the cluster solution. This could be done by shuffling the correlation matrix using appropriate algorithms, depending on the degree to which the authors want to consider spatial effects in the data (e.g., Roberts et al. 2015, NeuroImage; Rubinov and Sporns, NeuroImage, 2011).
As mentioned above, it is problematic to account for spatial effects in the current data given the inclusion of measures from the entire brain (GM and WM). We therefore tested the statistical significance of the cluster solution by shuffling the correlation matrix without considering the spatial effects. The following additions were made:
Results, section Clusters of change in development:
“Five clusters of coordinated developmental change were identified (Figure 1) (see Cluster stability analyses and Validation analyses below for a more detailed discussion and justification of the cluster solution). […] The community-structure solution was significantly more clustered than in the random networks (p<0.001, developmental change Q=0.44, the 2.5 and 97.5 percentile of the random Q distribution=0.36-0.40).”
Further, as an additional assessment of the cluster solutions, we tested the similarities of the community structure (cluster solution) of the developmental change and the adult/aging change. Following Betzel and Bassett 2017, we calculated the normalized mutual information (Lancichinetti et al., 2009), variation of information (Meilă, 2003), and the z-score of the Rand coefficient (Traud et al., 2011). The following was added to the manuscript (Results, section on Cluster stability analyses):
“Finally, we tested the similarities of the community structure (cluster solution) of the developmental and the adult/aging change-change matrices. […] Due to the nature of the research questions and data, including both GM and WM compartments as single structures, the null models generated were not spatially constrained (Alexander-Bloch et al., 2018; Burt, Helmer, Shinn, Anticevic, and Murray, 2020), which might have increased the similarities between change matrices and partitions.”
11. Modularity analysis – it appears the authors used a standard implementation of the Louvain algorithm. The default definition of Q defines a null expectation for within-cluster connectivity that is not appropriate for correlation matrices. With such data, the mean correlation can be appropriate.
We regret not describing our methods more clearly to avoid misunderstandings. We used an undirected weighted connection matrix with positive and negative correlations values (de-meaned) and the negative weights were treated asymmetrically (Rubinov and Sporns, 2011). This has now been clarified in the manuscript (in addition to a corrected reporting of the versatility curve, which was, erroneously, stemming from a non-de-meaned matrix in the original manuscript) (Materials and methods, section Experimental Design and Statistical Analysis):
“To identify clusters of correlations that could be compared across matrices, the community structure or modules in the matrices were obtained using the Louvain algorithm (V.D. Blondel, J-L., R., and E., 2008), part of the Brain Connectivity Toolbox (http://www.brain-connectivity-toolbox.net (Rubinov and Sporns, 2010)). […] To account for global brain changes, between-regional correlations were de-meaned before they were entered into the clustering analyses.”
And a further description is given in the section on Cluster stability analyses:
“As different clustering approaches often yield different results, we ran a series of post hoc analyses to confirm the validity of the cluster solution. […] Specifically, the 5-cluster solution resulted 7 times compared with once for the 3- and 8-cluster solutions, and twice for the 6- and 7-cluster solutions. Hence, this analysis supported the stability of the initial solution.”12. Modularity analysis – the evaluation of different levels of γ is commendable. It is unclear why this is left to the end of the Results. This should either be placed at the beginning, as a way of justifying the 5-cluster solution, or the beginning should indicate what value of γ was used initially and contain a reference to subsequent analyses that investigate the issue of γ in more detail.
We agree with the reviewers that this information should be presented earlier in the manuscript. In the re-organized Results section, this now follows just after the presentation of the cluster solution.
13. The similarity between the clusters and developmental origins is qualitative. Given that this is a core aim of the analysis, perhaps the authors could perform some inference by shuffling the labels and using a measure of partition similarity such as their normalized mutual information?
We understand the reviewer’s point, and agree that it could be good to have some kind of quantitative measure of overlap. We have discussed different ways of doing this, but have not come up with anything we believe is sufficiently solid. There may be approaches we are not aware of, but have kept the qualitative interpretations in the manuscript for now. Instead we acknowledge this directly in the Discussion:
“It must be noted as a limitation that the coherence between the developmental clusters and embryonic brain development is based on a qualitative judgement.”
14. Please explain why the cortex is treated as a single structure? Cortical development is regionally heterogeneous.
We certainly agree that cortical change is heterogeneous. The reason we did not include different cortical regions is that we assume that despite heterogeneity, within-cortical change is much more tightly integrated than cortical-subcortical change. Hence, including multiple cortical regions among the subcortical regions would likely have yielded an isolated cluster containing all cortical labels. We recently published a paper studying vertex-wise coordinated change of the cerebral cortex, please see: Continuity and Discontinuity in Human Cortical Development and Change From Embryonic Stages to Old Age – PubMed (nih.gov). This study is cited in the manuscript.
15. Sex is known to modify trajectory shape for cortical and subcortical structures. It would be good to show that the observed clustering holds for male and female subsamples. This would provide an important sensitivity analysis and also potential evidence for a sort of split half reliability.
We agree with the reviewer that multiple previous publications have focused on sex differences in brain development and aging. Running separate analyses for females and males would reduce power to approximately 50% and likely yield less stable solutions. In our experience, sex has minor, negligible effects on lifespan trajectories of subcortical volumes (see e.g. Minute effects of sex on the aging brain: a multisample magnetic resonance imaging study of healthy aging and Alzheimer's disease – PubMed (nih.gov)). Since we do not expect the major organizational principles of subcortical change to differ as a function of sex, we wish to report analyses for the full sample only. Still, we have included sex as a covariate in all analyses to account for possible confounding effects of sex.
16. It would also be important to show how ROIs cluster when you use the first derivatives from the gam fits in Figure 1. That would provide a complimentary approach to the inter-individual change method on which current work is built, and also help unpack some of the potential concerns re "fluid vs. tissue" and "white right vs. other" as two big potential drivers for findings. I would add plots for the ventricular components in Figure 1 too, as well as dendrograms for the primary heat map clustering used to order other matrices. My guess would be split one is CSF vs. tissue, and split 2 is white/white-rich tissue vs. other tissue.
The first derivatives would be based on group analyses, which we believe will address a potentially different question. It is conceivable that two regions that show very similar trajectories on a group level, and hence have similar derivatives, still show low within-subject change-change correlations. Although we agree that these can be interesting analyses, we believe they will not contribute to address the same questions as the original analyses in the manuscript. In light of the number of cluster analyses added to the revised manuscript, we are hesitant to further add to the complexity of the result. Regarding the question of contributions from WM and CSF, this is quite thoroughly treated in the revised manuscript, please see responses above.
17. In estimating APCs – how did the authors deal with were people with more than 2 scans, and the potential for including mid-scan age as a factor before correlating given the non-linear volume changes in development?
We apologize that this was not clearly explained in the manuscript. We have now added the following to the explanation of APCs were calculated (Results): “If more than two time points were available, the first and the last were used to calculate APC.” The reviewer is right that non-linear trajectories characterize brain development. Unfortunately, with relatively short follow up intervals (on average 1.7 years in our developmental sample), it is impossible to model non-linear changes, even in the very few cases where three timepoints were available.
18. I may have missed it but I did not expect the analysis of cognitive function. What purpose it is serving and does it really integrate in the theory of neurodevelopment? In addition, only one association seems significant after multiple testing and the difference seems to be localised in the 25-60 years old group for which you have the fewest observations (from Figure 1). Could this be due to a handful of outliers? Could you also add precision about which test was used? Maybe reporting several curves for difference quantiles of the distribution (instead of above/below average) would help visualise the effect?
We see the reviewers’ point. In response to this comment, we first decided to remove this section from the manuscript altogether, as it already contains a large number of comprehensive analyses. However, after further discussions among the co-authors, we came to the conclusion that it would be bad practice to remove these results at this stage, as the analyses were already conducted. Thus, we have removed the associated figure and kept the following shortened description in the main text (Results):
“Auxiliary analyses were done relating the clusters to general cognitive function (GCA) as measured by the Wechsler’s Abbreviated Scale of Intelligence (Wechsler, 1999) in the full LCBC sample, using sex and age as covariates. […] Importantly, only for Cluster 1 was a significant interaction between GCA and age found (F = 4.59, p = .01), suggesting that for the remaining clusters, age-trajectories did not differ significantly as a function of GCA (all p’s >.46).”
19. Figures 4 and 5 need a color scale
Yes, a color scale should have been included, thank you for spotting this. The previous Figure 4 is now part of Figure 1, which includes the color scale. A color scale has been added to Figure 3 (originally Figure 5).
20. Figure 2 is the main figure of result and I strongly suggest you expand the caption to improve self-readability. For example, the fact that the first row are results obtained on the LCBC sample. Clarify that the clustering was not recalculated for each matrix but is that of the first panel. In addition, the titles of the correlation plots could be clearer (e.g. clearly stating phenotypic and genetic correlations). Also the 4th panel is not a volumetric change-change relationship, which is the title of the figure.
We have now re-written the caption to the new Figures 1 and 2. The captions for the new figures read:
“Figure 1: Volumetric change-change relationships
Heat-maps represent pairwise correlations coefficients between volume change (annualized percent change) of the brain structures in development in the LCBC sample (left panel), aging in the LCBC sample (middle panel) and aging in the Lifebrain replication sample (right panels). The five clusters, delineated by the black lines, were derived from the developmental sample.”
“Figure 2 Genetic correlations
Left panel: Change-change correlations in development used to generate clusters. […] The five clusters, delineated by the black lines, were derived from the developmental sample.”
21. Abstract and introduction should announce that the focus is about volume of subcortical structures.
We agree that this should have been more clearly stated. In the Abstract, we now state:
“… we used graph theory to identify five clusters of coordinated development, indexed as patterns of correlated volumetric change in brain structures.”
In the Introduction, the following is included:
“Change was measures as annual percent change in the volume of a range of brain structures and areas. We hypothesized that volumetric changes in the developmental structures would tend to cluster according to embryonic principles, …”.
22. The discussion reads well but contains many statements that do not seem supported by the results or that sounded too definitive considering the analysis only focused on 16 volumetric measurements. It also lacks a limitation section.
We have gone through the discussion and modified the language. Specifically, we make sure that it is clear that we are discussing volumetric changes, and that statements do not exceed the empirical findings reported. We have also added a new section to the end of the Discussion: “Limitations: Caveats in interpreting brain changes from MRIs and further research”, cited above.
23. Some aspects of the text could be clearer. For example:
Line 44 – it is unclear what the hypothesis is precisely.
We have now reformulated the sentence to make the hypothesis clearer (Abstract):
“We tested the hypothesis that genetically governed neurodevelopmental processes can be traced throughout life by assessing to which degree brain regions that develop together continue to change together through life.”
Line 59 – what does a topographic organization correspond to in this context? regionally specific?
Yes. To clarify, we have slightly revised the sentence:
“Cortical development follows a topographic organization through childhood and adolescence (Fjell et al., 2018; Krongold, Cooper, and Bray, 2017; Raznahan et al., 2011), meaning that regions of the cortex that can be distinguished from neighboring regions by different criteria such as structural and functional properties tend to develop together (see (Eickhoff, Constable, and Yeo, 2018) for a discussion of cortical topography in the context of neuroimaging).”
Line 61 – it is unclear what it means to follow the "genetic organization of the cortex".
We have reformulated the sentence to make this clearer:
“This topography is conserved through later development and aging (Fjell et al., 2018; Tamnes et al., 2013), closely following the genetic organization of the cortex, i.e. being controlled by overlapping sets of genes (Fjell et al., 2015).”
Line 79 – it is very hard to understand what this hypothesis refers to and what it predicts.
We agree that this was not clear. Hence, we have reformulated the description of this:
“On the other hand, a hypothesis is that genetically governed neurodevelopmental processes can be traced in the brain later in life (Chen et al., 2011; Satizabal et al., 2019). […] This has been shown for the comparably less plastic cortex (Fjell et al., 2015).”
Line 88 – it is unclear what a genetic anatomical architecture is.
To make this clear, we have re-formulated the sentence:
“Specifically, we tested how subcortical developmental volumetric change clustered across different structures, how similar this organization was in development versus aging, and whether clusters of change were influenced by shared genetics.”
Line 105 – please clarify: was the correlation across subjects?
Yes, exactly, we now state: “These APCs were correlated across participants between each pair of brain regions.”
Line 121 – it is unclear what "common factors" refers to here.
We have reformulated: “The extensive connectivity between cerebellum and cerebrum and similarities in development of WM in cerebellum and cerebrum may explain the latter finding.”
Line 217 – only a correlation rather than prediction has been shown.
We have reformulated: “…regional subcortical volumetric changes in aging follow a similar pattern as developmental changes in childhood, …”
Line 256 – please clarify: was consensus clustering run over 1000 runs at each γ?
We ran the consensus clustering 1000 times at the γ used in the main analysis, this is now clearly stated in the manuscript (see response above).
“Although the lifespan trajectories of subcortical structures are much more divergent than those for cortical regions.”
Unclear to me how this is supported by results from the article, could you expand on how you conclude this?
This refers to previous literature, so we should have provided citations. We have slightly reformulated the sentence to make this clear: “Although the lifespan trajectories of subcortical structures have been shown to be much more divergent than those for cortical regions (Fjell et al., 2014; Walhovd et al., 2011), …”
“Mapping the developmental clusters to the adult part of the sample yielded highly different change trajectories.”
Here it may help to cite figures and tables in Discussion – also is this the case for all clusters?
We agree, this now reads:
“Mapping the developmental clusters to the adult part of the sample yielded highly different change trajectories (see Figure 4). Except for clusters 3 and 4, which were characterized by mostly linear negative trajectories, differences in the shapes of the slopes were observed, suggesting that clusters identified in development continued to show independent trajectories of change through the rest of life.”
“Clusters 1, 3 and to a lesser extent 4 were related to general cognitive function in a mostly age-invariant manner, which implies that the relationship between cognitive function and subcortical volumes is established early in life.”
This sounds like a huge overstatement. You studied only volume change of 16 regions – and the associations with cognition scores are not the most convincing.
We have now removed this section from the manuscript (see above).
“This means that regions that develop and change together through life tend to be governed by the same sets of genes.”
Beyond my previous criticism, this reads as a rather large generality.
We have modified the statement: “This means that regions that develop and change together through life tend to be influenced by shared sets of genes.”
“…and are governed by distinct sets of genes.”
What is this conclusion based on?
It is primarily based on the finding from the twin analyses that genetic change-change correlations show similar clustering to the developmental change-change correlations. We have modified the sentence somewhat in the revised manuscript:
“Subcortical change during childhood development can be organized in meaningful clusters, which are stable through life, tend to follow gradients of embryonic brain development and tend to be influenced by shared sets of genes.”
“It has also been argued that genes expressed in the subcortex generally are more region-specific and tend to evolve more rapidly than genes expressed in cortical regions”.
I am not sure what you mean by "evolve more rapidly" is it that it is under greater selection pressure?
It means that they tend to change more during evolution, and thus that genes expressed in subcortical structures tend to be evolutionary more recent compared to genes expressed in the cortex. We have reformulated the text slightly for clarity:
“Although the subcortex is evolutionary older than the cortex, it has a higher proportion of evolutionarily more recent genes, and a higher evolutionary rate, which is a basic measure of evolution at the molecular level (Tuller et al., 2008). It has also been argued that genes expressed in the subcortex generally are more region-specific (Tuller et al., 2008; Zhang and Li, 2004).”
The SNP genetic correlation matrix was highly similar to the developmental change matrix, as demonstrated by the Mantel test (r = 0.57, p <.0005, see Figure 2).”
I am not a big fan of qualitative judgments (e.g. highly). Especially that I assume r can vary between 0 and 1 so "highly" can be seen as an overstatement.
As much as the Mantel test statistics are different from 0, are they also different from 1?
Agree! We have reformulated:
“The SNP genetic correlation matrix was more similar to the developmental change matrix than expected by chance, …”
We have done the same change to a second sentence also containing “highly similar”.
Generalized
“Additive Mixed Models (GAMM)”
Could you add some details about the maximal order of the splines you considered? How was the best model – best order selected?
Degree of smoothness was estimated as part of model fitting, and model selection was guided by minimizing AIC and BIC. This information was “hidden” in the table legend to Table 3. We have now moved it to the main text, stating: “Both Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) were calculated to select among models and guard against over-fitting.”
24. “This revealed a close to perfect match between the adult genetic clusters and their embryonic origins.”
You say grouping is extremely similar, but compared to the clusters presented in Figure 2, I feel that this is an overstatement. E.g. caudate, accumbens, putament, pallidum – two of those were in separate clusters previously. Also positive rG between cerebellum cortex and WM while this was negative before.
This specific statement referred to the match between the non-CSF clustering of the SNP-coheritability estimates from UKB and the main divisions of embryonic brain development. Thus, we expect certain differences from the plots presented in Figure 2. Still, we have moderated the statement:
“This revealed a match between the adult genetic clusters and their embryonic origins (Figure 3).”
25. “In all cases yielded the slope function the lowest IC values”. This is not true – see caudate and cortex.
What to make from small AIC/BIC differences i.e. a marginally better fit? I wonder whether Table 3 is really adding anything, especially that the age trajectories are mostly descriptive?
You are right – thanks for spotting this! We have now corrected the statement. We agree with the reviewer that Table 3 is not critical to the manuscript. We think it is relevant to include it still, to yield some numbers to accompany Figure 5. However, we are open to remove it.
26. “We fitted the developmental trajectory of each cluster…”.
I assume you took the total volume of each cluster? What is this analysis really adding? First it assumes the cluster is somewhat homogenous (see my second comment) and another problem is that the different volumes can have extremely different scales which makes interpreting the sum difficult.
Yes, we used the total volume of each. We agree with the reviewer that there certainly are differences within clusters, but as we show in the change-change matrix analyses, the differences within are smaller than the differences between the clusters. The individual differences between the clusters can be inspected in Figure 5, showing the age-trajectories of the individual structures. We understand the reviewers’ points, but still believe that showing the variable developmental and lifespan trajectories for these different clusters yield useful information about their differences. To make our view clearer, and to acknowledge the reviewers’ points, the text has been modified, so that it now reads:
“We fitted the developmental trajectory of each cluster by using the total volume of the structures within each cluster, […] Since the total volume was used, large structures will potentially influence the cluster trajectories more than smaller structures.”
27. “Using multivariate latent change score models, we calculated…”.
Why did you not also use the annual symmetrized percent change you used previously? Especially that you want to compare rG to the phenotypic correlations you previously studied.
The main reason for this choice was that it was done in this way in the previous longitudinal subcortical twin study from ENIGMA (Bouwer et al.), of which the VETSA data used in the present study was drawn.
28. rGs from twin models are corrected for ICV, which does not seem to be the case for the GREML approach. Is this simply missing in the text?
“…pair of the 16 brain sub-cortical structures, including the first ten principal components, sex and age as covariates.”
Sorry, this was an omission in the methods description. We have now corrected this:
“For the UKB SNP analyses, the volume measures of the 16 sub-cortical structures were corrected for ICV and, …”
29. “log likelihood test”. I am more used to it being referred to as the likelihood ratio test.
We have replaced “log likelihood test” with “likelihood ratio test”.
30. “restricted maximum likelihood methods” – (very) minor detail but REML is the optimisation approach used to estimate the parameters of what is a bivariate mixed model. A compromise may be to talk about GREML, which is not more correct but quite commonly used to refer to the LMM implemented in GCTA. For an analogy, it is as if you referred to the GAMM as a Restricted Marginal Likelihood method.
We have slightly reformulated the sentence to make it more accurate: “We used the bivariate linear mixed model with genome-based restricted maximum likelihood methods implemented in the program GCTA”.
31. “…due to heuristics in the algorithm…”. What are the heuristics here, is it random starting values?
Yes, the algorithm loops over all nodes in an order that is random for each run, and we now clarify this point (Methods, section Experimental Design and Statistical Analysis):
“The community structure may vary from run to run due to heuristics in the algorithm pertaining to the order in which the nodes are considered, […]”.
32. “Also, all reconstructed surfaces were inspected, and discarded if they did not pass internal quality control.” Surely, this does not apply to the UKB analyses. Did you perform any QC on them beyond the UKB provided screening?
No, UKB scans have already undergone some QC. To make this clear, we have added the following sentence (Methods, section on MRI data acquisition and analysis):
“UKB scans were quality checked by the UKB imaging team.”
33. “In addition, we removed participants suggested to be removed for genetic analysis by the UK Biobank team.” Could you add a precision about why this exclusion is recommended?
We have now updated the text with more detailed information:
“Ninety-one of these 481 participants had abnormal heterozygosity values, and the remaining were flagged out as outliers in heterozegosity/missing rate from the current QC files (ukb_sqc_VZ.csv) provided the most recent UK Biobank team.”.
https://doi.org/10.7554/eLife.66466.sa2Article and author information
Author details
Funding
European Research Council (283634)
- Anders Martin Fjell
- Kristine Beate Walhovd
Horizon 2020 (732592)
- Kristine Beate Walhovd
Knut and Alice Wallenberg Foundation
- Lars Nyberg
Norwegian Research Council
- Anders Martin Fjell
- Kristine Beate Walhovd
Spanish Ministry of Science and Innovation (MICIU/FEDER/RTI2018-095181-B-C21)
- David Bartres-Faz
California Walnut Commission (NCT01634841)
- David Bartres-Faz
Federal Ministry of Education and Research (16SV5537/16SV5837/16SV5538/16SV5536K/01UW0808/01UW0706/01GL1716A/01GL1716B)
- Ulman Lindenberger
European Research Council (677804)
- Simone Kühn
Biotechnology and Biological Sciences Research Council
- Rogier Andrew Kievit
Medical Research Council
- Rogier Andrew Kievit
U.S. National Institute on Aging (AG022381)
- William S Kremen
European Research Council (725025)
- Anders Martin Fjell
- Kristine Beate Walhovd
European Research Council (313440)
- Anders Martin Fjell
- Kristine Beate Walhovd
U.S. National Institute on Aging (AG050595)
- William S Kremen
Institució Catalana de Recerca i Estudis Avançats (ICREA Academia-2019)
- David Bartres-Faz
The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.
Acknowledgements
The Lifebrain project is funded by the EU Horizon 2020 Grant: ‘Healthy minds 0–100 years: Optimising the use of European brain imaging cohorts (‘Lifebrain’)’. Grant agreement number: 732592. Call: Societal challenges: Health, demographic change and well-being. In addition, the different sub-studies are supported by different sources:
LCBC: The European Research Council under grant agreements 283634, 725025 (to AMF), and 313440 (to KBW), as well as the Norwegian Research Council (to AMF, KBW). Betula: a scholar grant from the Knut and Alice Wallenberg (KAW) foundation to LN University of Barcelona: Partially supported by a Spanish Ministry of Science, Innovation and Universities (MICIU/FEDER; RTI2018-095181-B-C21) to DB-F, which was also supported by an ICREA Academia 2019 grant award.; by the Walnuts and Healthy Aging study (http://www.clinicaltrials.gov; Grant NCT01634841) funded by the California Walnut Commission, Sacramento, California. BASE-II has been supported by the German Federal Ministry of Education and Research under grant numbers 16SV5537/16SV5837/16SV5538/16SV5536K/01UW0808/01UW0706/01GL1716A/01GL1716B, and SK has received support from the European Research Council under grant agreement 677804. Cam-CAN: Initial funding from the Biotechnology and Biological Sciences Research Council (BBSRC), followed by support from the Medical Research Council (MRC) Cognition and Brain Sciences Unit (CBU). VETSA is supported by U.S. National Institute on Aging grants R01s AG022381, and AG050595. Part of the research was conducted using the UK Biobank resource under application number 32048.
Ethics
Human subjects: The studies were approved by the Norwegian Regional Committee for Medical and Health Research Ethics South. Written informed consent was obtained from all participants older than 12 years of age and from a parent/guardian of volunteers under 16 years of age. Oral informed consent was obtained from all participants under 12 years of age. Non-Norwegian samples were approved by the relevant ethical review board for each country. Norway (2010/2359; 2010/3407; 2009/200).
Senior Editor
- Tamar R Makin, University College London, United Kingdom
Reviewing Editor
- Alex Fornito, Monash University, Australia
Publication history
- Received: January 12, 2021
- Accepted: June 26, 2021
- Accepted Manuscript published: June 28, 2021 (version 1)
- Version of Record published: July 6, 2021 (version 2)
Copyright
© 2021, Fjell et al.
This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.
Metrics
-
- 1,143
- Page views
-
- 137
- Downloads
-
- 3
- Citations
Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.
Download links
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
Further reading
-
- Computational and Systems Biology
- Neuroscience
Inhibition is crucial for brain function, regulating network activity by balancing excitation and implementing gain control. Recent evidence suggests that beyond simply inhibiting excitatory activity, inhibitory neurons can also shape circuit function through disinhibition. While disinhibitory circuit motifs have been implicated in cognitive processes including learning, attentional selection, and input gating, the role of disinhibition is largely unexplored in the study of decision-making. Here, we show that disinhibition provides a simple circuit motif for fast, dynamic control of network state and function. This dynamic control allows a disinhibition-based decision model to reproduce both value normalization and winner-take-all dynamics, the two central features of neurobiological decision-making captured in separate existing models with distinct circuit motifs. In addition, the disinhibition model exhibits flexible attractor dynamics consistent with different forms of persistent activity seen in working memory. Fitting the model to empirical data shows it captures well both the neurophysiological dynamics of value coding and psychometric choice behavior. Furthermore, the biological basis of disinhibition provides a simple mechanism for flexible top-down control of the network states, enabling the circuit to capture diverse task-dependent neural dynamics. These results suggest a biologically plausible unifying mechanism for decision-making and emphasize the importance of local disinhibition in neural processing.
-
- Medicine
- Neuroscience
The available treatments for depression have substantial limitations, including low response rates and substantial lag time before a response is achieved. We applied deep brain stimulation (DBS) to the lateral habenula (LHb) of two rat models of depression (Wistar Kyoto rats and lipopolysaccharide-treated rats) and observed an immediate (within seconds to minutes) alleviation of depressive-like symptoms with a high-response rate. Simultaneous functional MRI (fMRI) conducted on the same sets of depressive rats used in behavioral tests revealed DBS-induced activation of multiple regions in afferent and efferent circuitry of the LHb. The activation levels of brain regions connected to the medial LHb (M-LHb) were correlated with the extent of behavioral improvements. Rats with more medial stimulation sites in the LHb exhibited greater antidepressant effects than those with more lateral stimulation sites. These results indicated that the antidromic activation of the limbic system and orthodromic activation of the monoaminergic systems connected to the M-LHb played a critical role in the rapid antidepressant effects of LHb-DBS. This study indicates that M-LHb-DBS might act as a valuable, rapid-acting antidepressant therapeutic strategy for treatment-resistant depression and demonstrates the potential of using fMRI activation of specific brain regions as biomarkers to predict and evaluate antidepressant efficacy.