Figure 1 provides the data sources, analytical workflow and research methodology of this study. After the sample selection process (Appendix 1—figure 1, Appendix 1—table 1 and 2), longitudinal grey matter volume (GMV) trajectories in 40 ROIs (33 cortical and 7 subcortical ROIs, see Appendix 1—table 3) were estimated for each of the 37,013 healthy participants in UK Biobank. The first 15 principal components derived from dimensionality reduction via principal component analysis were used in the clustering analysis (see Methods; Alexander-Bloch et al., 2013; Whitwell et al., 2009). Two brain aging patterns were identified, where 18,929 (51.1%) participants with the first brain aging pattern (pattern 1) had higher total GMV at baseline and a slower rate of GMV decrease over time, and the remaining participants with the second pattern (pattern 2) had lower total GMV at baseline and a faster rate of GMV decrease (Figure 2a). Comparing the region-specific rate of GMV decrease, pattern 2 showed a more rapid GMV decrease in medial occipital (lingual gyrus, cuneus, and pericalcarine cortex) and medial temporal (entorhinal cortex, parahippocampal gyrus) regions (Figure 2b and c and Appendix 1—figure 2), which had the largest loadings in the second and third principal components (Appendix 1—table 4). These two patterns can be clearly stratified by both linear and non-linear dimensionality reduction methods, indicating distinct structural differences in brain aging between patterns (Appendix 1—figure 3). Sample characteristics of these 37,013 UK Biobank participants stratified by brain aging patterns are summarized in Appendix 1—table 5. Overall, participants with different brain aging patterns had similar distributions with regard to age, sex, ethnicity, smoking status, Townsend deprivation index (TDI), body mass index (BMI) and years of schooling.

To explore the relationships between structural brain aging and biological aging, we investigated the distribution of aging biomarkers, such as telomere length and PhenoAge (Levine et al., 2018), across brain aging patterns identified above (Figure 3 and Appendix 1—table 6). Compared to pattern 1, participants in pattern 2 with more rapid GMV decrease had shorter leucocyte telomere length (p=0.009, Cohen’s D=–0.028) and this association remained consistent after adjusting for sex, age, ethnic, BMI, smoking status and alcohol intake frequency (Demanelis et al., 2020). Next, we examined PhenoAge, which was developed as an aging biomarker incorporating composite clinical and biochemical data (Levine et al., 2018), and observed higher PhenoAge among participants with brain aging pattern 2 compared to pattern 1 (p=0.019, Cohen’s D=0.027). Again, the association remained significant after adjusting for sex, age, ethnic, BMI, smoking status, alcohol intake frequency and education years (p=3.05×10^–15, Cohen’s D=0.092). Group differences in terms of each individual component of PhenoAge (including albumin, creatinine, glucose, c-reactive protein, lymphocytes percentage, mean corpuscular volume, erythocyte distribution width, alkaline phosphatase and leukocyte count) were also investigated and results were consistent with PhenoAge (Appendix 1—figure 4).

Next, we conducted comprehensive comparisons of cognitive functions between participants with different brain aging patterns. In general, those with brain aging pattern 2 (lower baseline total GMV and more rapid GMV decrease) exhibited worse cognitive performances compared to pattern 1. Specifically, brain aging pattern 2 showed lower numbers of correct pairs matching (p=0.006, Cohen’s D=–0.029), worse prospective memory (OR = 0.943, 95% CI [0.891, 0.999]), lower fluid intelligence (p<1.00×10^–20, Cohen’s D=–0.102), and worse numeric memory (p=5.97×10^–11, Cohen’s D=–0.082). No statistically significant differences were observed in terms of the reaction time (p=0.99) and prospective memory (p=0.052) between these two brain aging patterns after FDR correction. Results were consistent when using models adjusted for sex, age, and socioeconomic status (TDI, education and income; Foster et al., 2018; Townsend et al., 2023; Figure 4). Full results demonstrating the associations between brain aging patterns and cognitive functions are presented in Appendix 1—table 7.

Having observed cognitive decline among participants with accelerated brain aging pattern, we next investigated whether brain aging patterns were associated with genetic vulnerability to major neuropsychiatric disorders. Since current GWAS are under-powered for attention-deficit/hyperactivity disorder (ADHD) and autism spectrum disorders (ASD) and the difficulty in identifying genetic variants was likely due to their polygenic nature, we calculated the corresponding polygenic risk scores (PRS) using multiple p value thresholds. This approach enabled robust investigation of the association between genetic susceptibility of neuropsychiatric disorders and brain imaging phenotypes. PRS for major neuro-developmental disorders including attention-deficit/hyperactivity disorder (ADHD) and autism spectrum disorders (ASD), neurodegenerative diseases including Alzheimer’s disease (AD) and Parkinson’s disease (PD), neuropsychiatric disorders including bipolar disorder (BIP), major depressive disorder (MDD), and schizophrenia (SCZ), and delayed structural brain development (GWAS from an unpublished longitudinal neuroimaging study) (Shi et al., 2023) were calculated for each participant using multiple p value thresholds (from 0.005 to 0.5 at intervals of 0.005) and results were then averaged over all thresholds (Figure 5). The primary GWAS datasets used for calculating the PRS were listed in Appendix 1—table 8. Overall, we observed increased genetic susceptibility to ADHD (p=0.040) and delayed brain development (p=1.4810^–6) among participants with brain aging pattern 2 after FDR correction, while no statistically significant differences were observed for ASD, AD, PD, BIP, MDD, and SCZ (Figure 5). Details regarding the genetic liability to other common diseases and phenotypes using enhanced PRS from UK Biobank are displayed in Appendix 1tables 9 and 10.

Having observed significant associations between brain aging patterns and cognitive performances / genetic liabilities to major neurodevelopmental disorders, we further investigated if there exist genetic variants contributing to individualized brain aging phenotype. We conducted genome-wide association studies (GWAS) using estimated total GMV at 60 years old as the phenotype. This phenotype was derived by adding individual specific deviations to the population averaged total GMV, thus providing additional information compared to studies using only cross-sectional neuroimaging phenotypes.

Six independent single nucleotide polymorphisms (SNPs) were identified at genome-wide significance level (p<5×10^–8) (Figure 6) and were subsequently mapped to genes using NCBI, Ensembl and UCSC Genome Browser database (Appendix 1—table 11). Among them, two SNPs (rs10835187 and rs779233904) were also found to be associated with multiple brain imaging phenotypes in previous studies (Smith et al., 2021), such as regional and tissue volume, cortical area and white matter tract measurements. Compared to the GWAS using global gray matter volume as the phenotype, our GWAS revealed additional signal in chromosome 7 (rs7776725), which was mapped to the intron of FAM3C and encodes a secreted protein involved in pancreatic cancer (Grønborg et al., 2006) and Alzheimer’s disease (Liu et al., 2016). This signal was further validated to be associated with specific brain aging mode by another study using a data-driven decomposition approach (Smith et al., 2020). In addition, another significant loci (rs10835187, p=1.11×10^–13) is an intergenic variant between gene LGR4-AS1 and LIN7C, and was reported to be associated with bone density and brain volume measurement (Smith et al., 2021; Estrada et al., 2012). LIN7C encodes the Lin-7C protein, which is involved in the localization and stabilization of ion channels in polarized cells, such as neurons and epithelial cells (Bohl et al., 2007; Kaech et al., 1998). Previous study has revealed the association of both allelic and haplotypic variations in the LIN7C gene with ADHD (Lanktree et al., 2008).

Having observed significant associations between brain aging and genetic susceptibility to neurodevelopmental disorders, we are now interested in examining the mirroring patterns between brain aging and brain development in the whole population, and whether these mirroring patterns were more pronounced in those with accelerated brain aging. Adolescents in the IMAGEN cohort showed more rapid GMV decrease in the frontal and parietal lobes, especially the frontal pole, superior frontal gyrus, rostral middle frontal gyrus, inferior parietal lobule and superior parietal lobule, while those in their mid-to-late adulthood showed more accelerated GMV decrease in the temporal lobe, including medial orbitofrontal cortex, inferior parietal lobule and lateral occipital sulcus (Figure 7a). The mirroring patterns (with slower GMV decrease during brain development and more rapid GMV decrease during brain aging) were particularly prominent in inferior temporal gyrus, caudal anterior cingulate cortex, fusiform cortex, middle temporal gyrus and rostral anterior cingulate cortex (Figure 7b). The regional mirroring patterns became weaker when we focus on late brain aging at age 75 years old, especially in the frontal lobe and cingulate cortex. Further, mirroring patterns were represented more prominently in participants with brain aging pattern 2, where stronger mirroring between brain aging and brain development was observed in frontotemporal area, including lateral occipital sulcus and lingual gyrus (Figure 7c).

The Allen Human Brain Atlas (AHBA) transcriptomic dataset (http://human.brain-map.org) were used to obtain the spatial correlation between gene expression profiles across cortex and structural brain development/aging via partial least square (PLS) regression. The first PLS component explained 24.7% and 53.6% of the GMV change during brain development (estimated at age 15y, r_spearman = 0.51, P_permutation = 0.03) and brain aging (estimated at age 55y, r_spearman = 0.49, P_permutation = 1.510^–4), respectively. Seventeen of the 45 genes mapped to GWAS significant SNP were found in AHBA, with LGR4 (r_spearman = 0.56, P_permutation <0.001) significantly associated with delayed brain development and ESR1 (r_spearman = 0.53, P_permutation <0.001) and FAM3C (r_spearman = –0.37, P_permutation = 0.004) significantly associated with accelerated brain aging. BDNF-AS was positively associated with both delayed brain development and accelerated brain aging after spatial permutation test (Appendix 1—table 12 and 13).

Next, we screened the genes based on their contributions and effect directions to the first PLS components in brain development and brain aging. 990 and 2293 genes were identified to be positively associated with brain development and negatively associated with brain aging at FDR corrected p value of 0.005, respectively, representing gene expressions associated with delayed brain development and accelerated brain aging. These genes were then tested for enrichment of GO biological processes and KEGG pathways. Genes associated with delayed brain development showed significant enrichment in ‘regulation of trans-synaptic signaling’, ‘forebrain development’, ‘signal release’ and ‘cAMP signaling pathway’ (Figure 8a), and genes associated with accelerated brain aging showed significant enrichment in ‘macroautophagy’, ‘establishment of protein localization to organelle’, ‘histone modification’, and ‘pathways of neurodegeneration – multiple diseases’ (Figure 8b). Full results of the gene set enrichment analysis were provided in Appendix 1—figure 5. In summary, the analyses from using the databases of GO biological processes and KEGG Pathways indicate synaptic transmission as an important process in the common mechanisms of brain development and aging, and cellular processes (autophagy), as well as the progression of neurodegenerative diseases, are important processes in the mechanisms of brain aging.

T1-weighted brain MRI images were obtained from 37,013 individuals aged 44–82 years old from UK Biobank (36,914 participants at baseline visit in 2014+, 4007 participants at the first follow-up visit in 2019+). All participants from UK Biobank provided written informed consent, and ethical approval was granted by the North West Multi-Center Ethics committee (https://www.ukbiobank.ac.uk/learn-more-about-uk-biobank/about-us/ethics) with research ethics committee (REC) approval number 16/NW/0274. Participants were excluded if they were diagnosed with severe psychiatric disorders or neurological diseases using ICD-10 primary and secondary diagnostic codes or from self-reported medical conditions at UK Biobank assessment center (see Appendix 1—tables 1 and 2). Data were obtained under application number 19542. A total of 1529 adolescents with structural MRI images were drawn from the longitudinal project IMAGEN (1463 at age 14, 1377 at age 19, and 1148 at age 23), of which the average number of MRI scans was 2.61 per adolescent. The lMAGEN study was approved by local ethics research committees of King’s College London, University of Nottingham. Trinity College Dublin, University of Heidelberg, Technische Universität Dresden, Commissariat à l'Énergie Atomique et aux Énergies Alternatives, and University Medical Center at the University of Hamburg in compliance with the Declaration of Helsinki (Association, 2013). Informed consent was given by all participants and a parent/guardian of each participant.

Quality-controlled T1-weighted neuroimaging data from UK Biobank and IMAGEN were processed using FreeSurfer v6.0. Detailed imaging processing pipeline can be found online for UK Biobank (https://biobank.ctsu.ox.ac.uk/crystal/crystal/docs/brain_mri.pdf) and IMAGEN (https://github.com/imagen2/imagen_mri; Schumann et al., 2010; Imagen, 2020). Briefly, cortical gray matter volume (GMV) from 33 regions in each hemisphere were generated using Desikan–Killiany Atlas (Desikan et al., 2006), and total gray matter volume (TGMV), intracranial volume (ICV) and subcortical volume were derived from ASEG atlas (Fischl et al., 2002 See Appendix 1—table 3). Regional volume was averaged across left and right hemispheres. To avoid deficient segmentation or parcellation, participants with TGMV, ICV or regional GMV beyond 4 standard deviations from the sample mean were considered as outliers and removed from the following analyses.

Whole-brain GMV trajectory was estimated for each participant in 40 brain regions of interest (ROIs) (33 cortical regions and 7 subcortical regions), using mixed effect regression model with fixed linear and quadratic age effects, random intercept and random age slope. Covariates include sex, assessment center, handedness, ethnic, and ICV. Models with random intercept and with both random intercept and random age slope were compared using AIC, BIC and evaluation of intra-class correlation (ICC). Results suggested that random age slope model should be chosen for almost all ROIs (Appendix 1—table 14). Deviation of regional GMV from the population average was calculated for each participant at age 60 years and dimensionality reduction was conducted via principal component analysis (PCA). The first 15 principal components explaining approximately 70% of the total variations of regional GMV deviation were used in multivariate k-means clustering. Optimal number of clusters was chosen using both elbow diagram and contour coefficient (Appendix 1—figure 6). Rates of volumetric change for total gray matter and each ROI were estimated using generalized additive mixed effect models (GAMM) with fixed cubic splines of age, random intercept and random age slope, which incorporates both cross-sectional between-subject variation and longitudinal within-subject variation from 40,921 observations and 37,013 participants. Covariates include sex, assessment center, handedness, ethnic, and ICV. We also applied PCA and locally linear embedding (LLE; Roweis and Saul, 2000) to the adjusted GMV ROIs in order to map the high-dimensional imaging-derived phenotypes to a low-dimensional space for stratification visualisation. The GMV of 40 ROIs at baseline were linearly adjusted for sex, assessment center, handedness, ethnic, ICV, and second-degree polynomial in age to be consistent with the whole-brain GMV trajectory model.

Individuals with Z-standardized leucocyte telomere length (Codd et al., 2021) and blood biochemistry (which were used to calculate PhenoAge (Levine et al., 2018) that characterizes biological aging) outside 4 standard deviations from the sample mean were excluded for better quality control. A total of 11 cognitive tests performed on the touchscreen questionnaire were included in the analysis. More information about the cognitive tests is provided in Supplementary Information. Comparisons of biological aging (leucocyte telomere length, PhenoAge) and cognitive function were conducted among participants with different brain aging patterns using both unadjusted and adjusted multivariate regression models with Bonferroni / FDR correction. Polygenic Risk Scores (PRS) were calculated for autism spectrum disorder (ASD), attention deficit hyperactivity disorder (ADHD), Alzheimer’s disease (AD), Parkinson’s Disease (PD), bipolar disorder (BIP), major depressive disorder (MDD), schizophrenia (SCZ), and delayed brain development using GWAS summary statistics (Shi et al., 2023) at multiple p value thresholds (from 0.005 to 0.5 at intervals of 0.005, and 1), with higher p value thresholds incorporating larger number of independent SNPs. After quality control of genotype and imaging data, PRSs were generated for 25,861 participants on UK Biobank genotyping data. SNPs were pruned and clumped with a cutoff r² ≥ 0.1 within a 250 kb window. All calculations were conducted using PRSice v2.3.5 (Choi and O’Reilly, 2019). Enhanced PRS from UK Biobank Genomics for multiple diseases were also tested. Detailed instructions for calculating enhanced PRS in UK Biobank can be found in research of Thompson et al., 2022 Comparisons of neuropsychiatric disorders were conducted among participants with different brain aging patterns using t test with FDR correction. All statistical tests were two-sided.

We performed Genome-wide association studies (GWAS) on individual deviations of total GMV relative to the population average at 60 years using PLINK 2.0 (Chang et al., 2015). Variants with missing call rates exceeding 5%, minor allele frequency below 0.5% and imputation INFO score less than 0.8 were filtered out after the genotyping quality control for UK Biobank Imputation V3 dataset. Among the 337,138 unrelated ‘white British’ participants of European ancestry included in our study, 25,861 with recent UK ancestry and accepted genotyping and imaging quality control were included in the GWAS. The analyses were further adjusted for age, age2, sex, assessment center, handedness, ethnic, ICV, and the first 10 genetic principal components. Genome-wide significant SNPs (p<510^–8) obtained from the GWAS were clumped by linkage disequilibrium (LD) (r² <0.1 within a 250 kb window) using UKB release2b White British as the reference panel. We subsequently performed gene-based annotation in FUMA (Watanabe et al., 2017) using genome-wide significant SNPs and SNPs in close LD (r² ≥ 0.1) using Annotate Variation (ANNOVAR) on Ensemble v102 genes (Wang et al., 2010).

To validate the ‘last in, first out’ mirroring hypothesis, we evaluated the structural association between brain development and brain aging. Longitudinal neuroimaging data from 1529 adolescents in the IMGAEN cohort and 3908 mid-to-late adulthood in the UK Biobank cohort were analyzed. Annual percentage volume change (APC) for each ROI was calculated among individuals with at least two structural MRI scans by subtracting the baseline GMV from follow-up GMV and dividing by the number of years between baseline and follow-up visits. Region-specific APC was regressed on age using smoothing spline with cross validated degree of freedom. Estimated APC for each ROI was obtained at age 15y for adolescents and at age 55y (early aging) and 75y (late aging) for participants in UK Biobank. Region-specific APC during adolescence (or mid-to-late adulthood) was then standardized across all cortical regions to create the brain development (or aging) map. Finally, the brain development map and brain aging map were compared to assess the mirroring pattern for each ROI in the overall population and across different aging subgroups.

The Allen Human Brain Atlas (AHBA) dataset (http://human.brain-map.org), which comprises gene expression measurements in six postmortem adults (age 24–57y) across 83 parcellated brain regions (Hawrylycz et al., 2012; Markello et al., 2021), were used to identify gene expressions significantly associated with structural brain development and aging. The expression profiles of 15,633 genes were averaged across donors to form a 8315,633 transcriptional matrix and partial least squares (PLS) regression was adopted for analyzing the association between regional change rate of gray matter volume and gene expression profiles. Specifically, estimated regional APC at 15 (obtained from IMAGEN cohort) and 55 years old (obtained from UK Biobank) were regressed on the high-dimensional gene expression profiles upon regularization. Associations between the first PLS component and estimated APC during brain development and brain aging were tested by spatial permutation analysis (10,000 times; Váša et al., 2018). Additionally, gene expression profiles of genes mapped to GWAS significant SNP were extracted from AHBA. The association between gene expression profiles of mapped genes and estimated APC during brain development and aging was also tested by spatial permutation analysis. Statistical significance of each gene’s contribution to the first PLS component was tested with standard error calculated using bootstrap (Li et al., 2021; Morgan et al., 2019; Romero-Garcia et al., 2020), and genes significantly associated with delayed brain development and accelerated brain aging were selected. Enrichment of Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways and gene ontology (GO) of biological processes for these selected genes were analyzed using R package clusterProfiler (Yu et al., 2012). All statistical significances were corrected for multiple testing using FDR.

R version 4.2.0 was used to perform statistical analyses. FreeSurfer version 6.0 was used to process neuroimaging data. lme4 1.1 in R version 4.2.0 was used to perform longitudinal data analyses. PRSice version 2.3.5 (https://choishingwan.github.io/PRSice/; Choi and O’Reilly, 2019) was used to calculate the PRS. PLINK 2.0 (https://www.cog-genomics.org/plink/2.0/) and FUMA version 1.5.6 (https://fuma.ctglab.nl/) were used to perform genome-wide association analysis, and ANNOVAR was used to perform gene-based annotation. AHBA microarray expression data were processed using abagen toolbox version 0.1.3 (https://doi.org/10.5281/zenodo.5129257). The rotate_parcellation code used to perform a spatial permutation test of a parcellated cortical map: https://github.com/frantisekvasa/rotate_parcellation (Váša, 2023; Váša et al., 2018). Code for PLS analysis and bootstrapping to estimate PLS weights are available at https://github.com/KirstieJane/NSPN_WhitakerVertes_PNAS2016/tree/master/SCRIPTS (Whitaker, 2016; Whitaker et al., 2016). clusterProfiler 4.6 in R version 4.2.0 was used to analyze gene-set enrichment.