Browse the corpus

MRI-based multi-organ clocks for healthy aging and disease assessment. Biological aging clocks across organ systems and tissues have advanced understanding of human aging and disease. In this study, we expand this framework to develop seven magnetic resonance imaging-based multi-organ biological age gaps (MRIBAGs), including the brain, heart, liver, adipose tissue, spleen, kidney and pancreas. Using data from 313,645 individuals curated by the MULTI Consortium, we link the seven MRIBAGs to 2,923 plasma proteins, 327 metabolites and 6,477,810 common genetic variants. Genome-wide associations identify 53 MRIBAG-locus pairs (P < 5 × 10-8). Genetic correlation and Mendelian randomization analyses support organ-specific and cross-organ interconnection, including 24 non-MRI biological aging clocks and 525 disease endpoints. Through functional gene mapping and Bayesian co-localization multi-omics evidence, we prioritize nine druggable genes as targets for future anti-aging treatments. Furthermore, the seven MRIBAGs are linked to future risk of systemic disease endpoints (for example, diabetes mellitus) and all-cause mortality. Finally, participants with more youthful versus more aged brain profiles exhibited distinct cognitive decline trajectories over 240 weeks of treatment with the Alzheimer's disease drug solanezumab, although this heterogeneity cannot be fully attributed to the drug. In summary, we developed seven MRIBAGs that enhance the existing multi-organ biological aging framework, and we demonstrate their clinical potential to advance aging research.

Magnetic resonance imaging (MRI1) provides a non-invasive window into structural and functional changes occurring with age, enabling the development of personalized disease biomarkers. Brain MRI-based aging clocks, often referred to as ‘brain age’ and derived from artificial intelligence and machine learning (AI/ML) applied to MRI data, have been widely used as clinical biomarkers of neurological aging, cognitive decline and neurodegenerative disease risk2. However, although brain imaging has been extensively used for aging research, no studies have systematically extended this concept to other organ systems. With the advent of large-scale, multi-organ MRI datasets, such as those from the UK Biobank (UKBB), we now have an unprecedented opportunity to conduct a more holistic, system-wide investigation of biological aging across multiple organs.

rch, no studies have systematically extended this concept to other organ systems. With the advent of large-scale, multi-organ MRI datasets, such as those from the UK Biobank (UKBB), we now have an unprecedented opportunity to conduct a more holistic, system-wide investigation of biological aging across multiple organs. Beyond imaging-based aging clocks, researchers have also developed various omics-based aging clocks to capture different biological dimensions of aging. For example, plasma proteomics-based aging clocks3,4 have been introduced to enhance the granularity and coverage of multi-organ aging clocks. Plasma proteomics-based aging clocks capture the biological aging process by analyzing the abundance of proteins in circulation. Because proteins are the functional units of the cell, these clocks provide insights into the downstream effects of genetic regulation and environmental influences on aging. Metabolomics-based aging clocks5,6 capture biological aging by profiling small-molecule metabolites that represent the biochemical activity of cellular and systemic metabolism. Metabolomics examines the end products of metabolic pathways, making it highly dynamic and sensitive to external factors such as diet, microbiome composition, lifestyle and environmental exposures. Finally, epigenetics-based aging clocks7 focus on the analysis of DNA methylation patterns, particularly at specific CpG sites across the genome, to estimate biological age. Integrating these aging clocks into the multi-organ and multi-omics concept8–10 is essential for gaining a comprehensive understanding of aging biology, age-related diseases and longevity.

ocus on the analysis of DNA methylation patterns, particularly at specific CpG sites across the genome, to estimate biological age. Integrating these aging clocks into the multi-organ and multi-omics concept8–10 is essential for gaining a comprehensive understanding of aging biology, age-related diseases and longevity. In our previous efforts, we developed nine phenome-based biological age gaps (BAGs)11,12 (PhenoBAGs), 11 proteome-based BAGs (ProtBAGs13) and five metabolome-based BAGs (MetBAGs5). In the present study, we derived six additional MRIBAGs, on top of the brain MRIBAG14, for the heart, liver, spleen, adipose, kidney and pancreas using data consolidated from the MULTI Consortium (Method 1). We first benchmarked the age prediction performance of the six additional MRIBAGs (Method 2). Next, we associated the seven MRIBAGs with 2,923 plasma proteins for proteome-wide associations (ProWASs) and 327 metabolites for metabolome-wide associations (MetWASs) (Method 3). Through genome-wide association studies (GWASs) and post-GWAS analyses, we depicted their genetic architecture and identified potential druggable genes (Method 4). Finally, we assessed the clinical applicability and predictability of the MRIBAGs and their polygenic risk scores (PRSs) (Method 5). All results and pre-trained AI/ML models are publicly available at the MEDICINE portal: https://labs-laboratory.com/medicine/.

The brain MRIBAG was initially developed in our previous study14 using 119 gray matter volumes, and we re-trained the brain MRIBAG here using a consistent nested cross-validation procedure along with a within-distribution, holdout test dataset (Extended Data Fig. 1). Specifically, to rigorously evaluate the performance (that is, overfitting and generalizability) of biological age prediction models, we partitioned the healthy control (CN, without any pathologies) participants into the CN training/validation/test (3,573 < N < 6,327 as the sample sizes vary across organs) and the within-distribution, holdout test (N = 500) datasets (Method 2 and Supplementary Table 1).

izability) of biological age prediction models, we partitioned the healthy control (CN, without any pathologies) participants into the CN training/validation/test (3,573 < N < 6,327 as the sample sizes vary across organs) and the within-distribution, holdout test (N = 500) datasets (Method 2 and Supplementary Table 1). When fitting the organ-specific MRI features (Fig. 1a, Method 2 and Supplementary File 1), the two AI/ML models (least absolute shrinkage and selection operator (LASSO) regression and support vector regressor) exhibited slight variability in performance, with no single model consistently outperforming the others (Fig. 1b). The selected optimal models for the seven MRIBAGs demonstrated moderate Pearson’s r coefficients (ranging from 0.23 to 0.77) and the mean absolute errors (MAEs) of approximately 5 years in the within-distribution, holdout test dataset (Fig. 1c) before applying the age bias correction15,16. Supplementary Table 2 presents detailed statistics for the age prediction tasks before the age bias correction (Supplementary Fig. 1). Supplementary Fig. 2 presents the incremental statistical power, beyond age and sex, of the seven MRIBAGs in predicting mortality and disease categories. Model improvements may be possible for abdominal features by employing more advanced modeling approaches, such as deep neural networks that directly process voxel-wise MRI data17. Additionally, we found that high collinearity among imaging features from abdominal MRI led to poor generalizability in within-distribution, holdout test datasets (Supplementary Fig. 3). Supplementary Fig. 4 shows potential domain shift while applying the pre-trained model of the brain MRIBAGs to external datasets, exemplified by the A4 study18,19. Supplementary Fig. 5 presents a sensitivity analysis for the Alzheimerʼs disease drug differential responses. Supplementary Fig. 6 discusses sex differences in the seven MRIBAGs. Supplementary Figs. 7 and 8 discuss the feature importance of deriving these MRIBAGs and their biological interpretation. For all subsequent analyses, we used the age bias-corrected16 MRIBAGs.Fig. 1Model performance benchmarking for seven MRIBAGs.a, The seven MRIBAGs are derived from various imaging modalities available in the UKBB. The number (N) of IDPs used to derive the seven MRIBAGs is displayed.

and their biological interpretation. For all subsequent analyses, we used the age bias-corrected16 MRIBAGs.Fig. 1Model performance benchmarking for seven MRIBAGs.a, The seven MRIBAGs are derived from various imaging modalities available in the UKBB. The number (N) of IDPs used to derive the seven MRIBAGs is displayed. b, For the seven MRIBAGs, the MAE of age prediction models using LASSO regression and linear support vector regressor (SVR) is shown for both the training dataset (cross-validated training/validation/test; 3,582 < N < 6,325) and the within-distribution, holdout test dataset (N = 500). Cohen’s d quantifies the effect size of the difference between these datasets, reflecting potential model generalizability, assuming similar age and sex distributions. The optimal model (#) for each organ and tissue was selected based on the lower Cohen’s d value, and these models were used for all subsequent analyses. All MAEs are reported without age bias correction. c, Scatter plots display the optimal model for each organ/tissue in the within-distribution, holdout test dataset, with Pearson’s r and two-sided P values indicating the association between chronological age and predicted age. All results are shown before the correction for age bias was applied16. To evaluate the generalizability of the brain MRIBAG model, we validated it using two independent studies. In b, individual data points are overlaid on box plots summarizing the distribution. The center line indicates the median; the box bounds represent the first and third quartiles; and the whiskers extend to the most extreme values within 1.5 times the interquartile range. Ind., independent.

sing two independent studies. In b, individual data points are overlaid on box plots summarizing the distribution. The center line indicates the median; the box bounds represent the first and third quartiles; and the whiskers extend to the most extreme values within 1.5 times the interquartile range. Ind., independent. a, The seven MRIBAGs are derived from various imaging modalities available in the UKBB. The number (N) of IDPs used to derive the seven MRIBAGs is displayed. b, For the seven MRIBAGs, the MAE of age prediction models using LASSO regression and linear support vector regressor (SVR) is shown for both the training dataset (cross-validated training/validation/test; 3,582 < N < 6,325) and the within-distribution, holdout test dataset (N = 500). Cohen’s d quantifies the effect size of the difference between these datasets, reflecting potential model generalizability, assuming similar age and sex distributions. The optimal model (#) for each organ and tissue was selected based on the lower Cohen’s d value, and these models were used for all subsequent analyses. All MAEs are reported without age bias correction. c, Scatter plots display the optimal model for each organ/tissue in the within-distribution, holdout test dataset, with Pearson’s r and two-sided P values indicating the association between chronological age and predicted age. All results are shown before the correction for age bias was applied16. To evaluate the generalizability of the brain MRIBAG model, we validated it using two independent studies. In b, individual data points are overlaid on box plots summarizing the distribution. The center line indicates the median; the box bounds represent the first and third quartiles; and the whiskers extend to the most extreme values within 1.5 times the interquartile range. Ind., independent.

sing two independent studies. In b, individual data points are overlaid on box plots summarizing the distribution. The center line indicates the median; the box bounds represent the first and third quartiles; and the whiskers extend to the most extreme values within 1.5 times the interquartile range. Ind., independent. In our ProWAS analyses (Method 3), we identified 603 protein−MRIBAG significant associations (P < 0.05/2,923/7). Among these, the kidney MRIBAG exhibited the highest number of significant protein associations (N = 301; for example, NPDC1, IGFBP6 and TAFA5), followed by the spleen MRIBAG for 136 associations (for example, VCAM1, PTPRH and C1QA), the liver MRIBAG for 62 associations (for example, NCAN, SEZ6L and LEP), the adipose MRIBAG for 57 signals (for example, GDF15, CHI3L1 and CA14), the pancreas MRIBAG for 21 associations (for example, PLA2G1B, CTRC and CELA2A), the brain MRIBAG for 16 associations (for example, BCAN, NCAN and GDF15) and the heart MRIBAG for only REN (Fig. 2a).Fig. 2ProWAS and MetWAS with the seven MRIBAGs.a, ProWAS between the seven MRIBAGs and 2,923 plasma proteins via a linear regression model, accounting for a full set of covariates (P < 0.05/2,923/7). An online interactive webpage is available at https://labs-laboratory.com/medicine/mribag_prowas.html to ease visualization. Plots of individual organs are shown in Extended Data Fig. 2. b, MetWAS between the seven MRIBAGs and 327 plasma metabolites using a linear regression model while adjusting for a comprehensive set of covariates (P < 0.05/327/7). Plots of individual organs are shown in Extended Data Fig. 4. An online interactive webpage is available at https://labs-laboratory.com/medicine/mribag_metwas.html to ease visualization. The colors and shapes of the icons indicate the organ systems involved in the associations. We used the standardized β value to represent the effect size.

rgans are shown in Extended Data Fig. 4. An online interactive webpage is available at https://labs-laboratory.com/medicine/mribag_metwas.html to ease visualization. The colors and shapes of the icons indicate the organ systems involved in the associations. We used the standardized β value to represent the effect size. a, ProWAS between the seven MRIBAGs and 2,923 plasma proteins via a linear regression model, accounting for a full set of covariates (P < 0.05/2,923/7). An online interactive webpage is available at https://labs-laboratory.com/medicine/mribag_prowas.html to ease visualization. Plots of individual organs are shown in Extended Data Fig. 2. b, MetWAS between the seven MRIBAGs and 327 plasma metabolites using a linear regression model while adjusting for a comprehensive set of covariates (P < 0.05/327/7). Plots of individual organs are shown in Extended Data Fig. 4. An online interactive webpage is available at https://labs-laboratory.com/medicine/mribag_metwas.html to ease visualization. The colors and shapes of the icons indicate the organ systems involved in the associations. We used the standardized β value to represent the effect size.

rgans are shown in Extended Data Fig. 4. An online interactive webpage is available at https://labs-laboratory.com/medicine/mribag_metwas.html to ease visualization. The colors and shapes of the icons indicate the organ systems involved in the associations. We used the standardized β value to represent the effect size. VCAM1 was associated with the spleen MRIBAG (β = −0.89 ± 0.04; P = 7.05 × 10−81; r = −0.38); PLA2G1B was associated with the pancreas MRIBAG (β = −1.14 ± 0.07; P = 8.26 × 10−56; r = −0.30) and the adipose and spleen MRIBAGs. The two proteins are organ enriched in respective organ and tissue, as identified by the Human Protein Atlas (HPA; https://www.proteinatlas.org/), emphasizing their relevance to organ-specific aging processes, which has been widely used to define organ-enriched proteins in previous proteome-based aging clock papers4,13. Supplementary File 2a presents detailed statistics. Supplementary Fig. 9 and Supplementary File 2b explore the potential replication of UKBB Olink ProWAS signals in Baltimore Longitudinal Study of Aging (BLSA) SomaScan data. Extended Data Figs. 2 and 3 present individual volcano plots and protein set enrichment results via the STRING platform (version 12.0 (ref. 20)) for each MRIBAG (Supplementary Figs. 10−14). Supplementary Fig. 15 presents the sex-specific ProWAS results.

als in Baltimore Longitudinal Study of Aging (BLSA) SomaScan data. Extended Data Figs. 2 and 3 present individual volcano plots and protein set enrichment results via the STRING platform (version 12.0 (ref. 20)) for each MRIBAG (Supplementary Figs. 10−14). Supplementary Fig. 15 presents the sex-specific ProWAS results. In our MetWAS (Method 3), we identified 758 metabolite−MRIBAG significant associations (P < 0.05/327/7) between the seven MRIBAGs and metabolites, showing relatively weak associations (Pearson’s r < 0.3). The spleen MRIBAG showed the highest number of significant metabolite associations (N = 199; for example, acetate, acetoacetate and albumin), followed by the adipose MRIBAG (N = 196; for example, L_HDL_PL_pct, L_HDL_C_pct and PUFA_pct), the heart MRIBAG (N = 139; for example, PUFA_pct, M_HDL_TG and HDL_TG), the brain MRIBAG (N = 97; for example, LA_pct, Omega_6_pct and ApoA1), the pancreas MRIBAG (N = 57; for example, GlycA, Glucose_Lactate and HDL_PL_pct), the liver MRIBAG (N = 51; for example, Gly, Tyr and L_LDL_PL_pct) and the kidney MRIBAG (N = 19; for example, creatinine, citrate and DHA) (Fig. 2b). Descriptions of these metabolites are provided in Supplementary Table 3.

pct and ApoA1), the pancreas MRIBAG (N = 57; for example, GlycA, Glucose_Lactate and HDL_PL_pct), the liver MRIBAG (N = 51; for example, Gly, Tyr and L_LDL_PL_pct) and the kidney MRIBAG (N = 19; for example, creatinine, citrate and DHA) (Fig. 2b). Descriptions of these metabolites are provided in Supplementary Table 3. Creatinine was linked to the kidney (β = 21.77 ± 0.65; P = 1.58 × 10−258; r = 0.28) and adipose (β = −4.37 ± 0.83; P = 1.48×10 −7; r = −0.18) MRIBAGs, which may suggest distinct roles of the kidney and adipose tissue in creatinine regulation. In addition to these small-molecule metabolites, other lipid complexes and subclasses also showed significant associations. Total concentration of phospholipids (Total_P) was negatively associated with liver (β = 51.57 ± 3.59; P= 2.52 × 10−46; r = 0.18). Phospholipids in small HDL (S_HDL_PL) were linked to the brain, heart, spleen and adipose MRIBAGs. Similarly, cholesteryl esters in very large HDL (XL_HDL_CE) associated with the adipose and heart MRIBAGs. Supplementary File 3 presents detailed statistics. Extended Data Fig. 4 presents individual volcano plots for each MRIBAG. Metabolite set enrichment analysis is presented in Extended Data Fig. 5. Supplementary Fig. 16 shows the genetic correlations among the seven MRIBAGs, 2,923 proteins and 327 metabolites. Extended Data Fig. 6 presents example associations for the ProWAS and MetWAS. Supplementary Fig. 17 presents the sex-specific MetWAS results.

te set enrichment analysis is presented in Extended Data Fig. 5. Supplementary Fig. 16 shows the genetic correlations among the seven MRIBAGs, 2,923 proteins and 327 metabolites. Extended Data Fig. 6 presents example associations for the ProWAS and MetWAS. Supplementary Fig. 17 presents the sex-specific MetWAS results. We conducted GWAS (Method 4) for the seven MRIBAGs (19,686 < N < 31,557 participants with European ancestries) and identified 53 (P < 5 × 10−8) genomic locus−BAG pairs. We denoted the genomic loci using their top lead single-nucleotide polymorphisms (SNPs) defined by FUMA, considering linkage disequilibrium; the genomic loci are presented in Supplementary Table 4. We visually present the shared genomic loci annotated by cytogenetic regions based on the GRCh37 cytoband (Fig. 3a). Supplementary Fig. 18 and Extended Data Fig. 7 detail the robustness of our GWASs and sex-specific GWAS analyses.Fig. 3The genetics of the seven MRIBAGs.a, Cytogenetic regions associated with the seven MRIBAGs, with significant genomic loci identified using the genome-wide significance threshold (two-sided P < 5 × 10−8). b, SNP-based heritability estimates of the seven MRIBAGs (sample size: 18,778 < N < 31,557). c, Partitioned heritability enrichment analysis using chromatin-specific multi-tissue chromatin data, multi-tissue gene expression profiles and cell-type-specific datasets. Only significant results (two-sided P < 0.05/697) are displayed (sample size: 18,778 < N < 31,557). d, Genetic correlation estimates via LDSC between the seven MRIBAGs and eight PhenoBAGs (excluding the multimodal brain PhenoBAG11), 11 ProtBAGs13 and five MetBAGs5 (two-sided P < 0.05/11) as well as 525 disease endpoints from FinnGen and PGC (P < 0.05/525). e, Potential causal relationships between the seven MRIBAGs (that is, number of instrumental variables > 7) and 525 disease endpoints were examined through two networks: BAG2DE, where the seven MRIBAGs serve as exposure variables and the 525 DEs as outcome variables, with two-sided P < 0.05/525, and DE2BAG, where the 214 disease endpoints serve as effective exposure variables (that is, number of instrumental variables > 7), with two-sided P < 0.05/214. Although multiple sensitivity checks were performed to evaluate potential violations of underlying assumptions, these findings should be interpreted with caution.

d DE2BAG, where the 214 disease endpoints serve as effective exposure variables (that is, number of instrumental variables > 7), with two-sided P < 0.05/214. Although multiple sensitivity checks were performed to evaluate potential violations of underlying assumptions, these findings should be interpreted with caution. The direction array represents the causal relationship from the exposure to the outcome variables, where ‘+’ denotes an odds ratio > 1 and ‘–’ indicates an odds ratio < 1. f, The bar plot shows the incremental R2 (that is, the R2 of the alternative model minus that of the null model) for the PRS of each MRIBAG. The PRS was calculated using the split2 target GWAS data, with split1 GWAS data serving as the training set for the PRScs model. Bar plots represent the mean (center of bar), with error bars indicating ±1 s.e.

emental R2 (that is, the R2 of the alternative model minus that of the null model) for the PRS of each MRIBAG. The PRS was calculated using the split2 target GWAS data, with split1 GWAS data serving as the training set for the PRScs model. Bar plots represent the mean (center of bar), with error bars indicating ±1 s.e. a, Cytogenetic regions associated with the seven MRIBAGs, with significant genomic loci identified using the genome-wide significance threshold (two-sided P < 5 × 10−8). b, SNP-based heritability estimates of the seven MRIBAGs (sample size: 18,778 < N < 31,557). c, Partitioned heritability enrichment analysis using chromatin-specific multi-tissue chromatin data, multi-tissue gene expression profiles and cell-type-specific datasets. Only significant results (two-sided P < 0.05/697) are displayed (sample size: 18,778 < N < 31,557). d, Genetic correlation estimates via LDSC between the seven MRIBAGs and eight PhenoBAGs (excluding the multimodal brain PhenoBAG11), 11 ProtBAGs13 and five MetBAGs5 (two-sided P < 0.05/11) as well as 525 disease endpoints from FinnGen and PGC (P < 0.05/525). e, Potential causal relationships between the seven MRIBAGs (that is, number of instrumental variables > 7) and 525 disease endpoints were examined through two networks: BAG2DE, where the seven MRIBAGs serve as exposure variables and the 525 DEs as outcome variables, with two-sided P < 0.05/525, and DE2BAG, where the 214 disease endpoints serve as effective exposure variables (that is, number of instrumental variables > 7), with two-sided P < 0.05/214. Although multiple sensitivity checks were performed to evaluate potential violations of underlying assumptions, these findings should be interpreted with caution. The direction array represents the causal relationship from the exposure to the outcome variables, where ‘+’ denotes an odds ratio > 1 and ‘–’ indicates an odds ratio < 1. f, The bar plot shows the incremental R2 (that is, the R2 of the alternative model minus that of the null model) for the PRS of each MRIBAG. The PRS was calculated using the split2 target GWAS data, with split1 GWAS data serving as the training set for the PRScs model. Bar plots represent the mean (center of bar), with error bars indicating ±1 s.e.

emental R2 (that is, the R2 of the alternative model minus that of the null model) for the PRS of each MRIBAG. The PRS was calculated using the split2 target GWAS data, with split1 GWAS data serving as the training set for the PRScs model. Bar plots represent the mean (center of bar), with error bars indicating ±1 s.e. We estimated the SNP-based heritability (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${h}_{{SNP}}^{2}$$\end{document}hSNP2) for the seven MRIBAGs using GCTA21 software (Method 4), with values ranging from 0.29 to 0.47 (Fig. 3b). Supplementary Table 5 presents detailed statistics. We validated the GWAS signals through partitioned heritability analyses (Method 4) using LDSC software, revealing strong organ-specific enrichment. Notably, the heart MRIBAG exhibited chromatin state-specific enrichment across four distinct heart tissues (Fig. 3c). For example, significant heritability enrichment was shown in the right atrium in the H3K4me3 region (P = 1.00 × 10−6) and in the left ventricle in the H3K27ac region (P = 2.00 × 10−5) for the heart MRIBAG (Supplementary File 4).

xhibited chromatin state-specific enrichment across four distinct heart tissues (Fig. 3c). For example, significant heritability enrichment was shown in the right atrium in the H3K4me3 region (P = 1.00 × 10−6) and in the left ventricle in the H3K27ac region (P = 2.00 × 10−5) for the heart MRIBAG (Supplementary File 4). We then assessed the genetic correlations (Method 4) between the seven MRIBAGs and 24 previously developed multi-organ, multi-omics aging clocks, including eight PhenoBAGs11 (excluding the brain PhenoBAG based on multimodal imaging), 11 ProtBAGs13 and five MetBAGs5. After Bonferroni correction, we found seven within-organ and inter-organ significant genetic correlations (P < 0.05/11), exemplified between the kidney MRIBAG and the renal PhenoBAG (gc = 0.23 ± 0.05; P < 2.00 × 10−5) and between the spleen MRIBAG and the hepatic MetBAG (gc = 0.23 ± 0.08; P < 7.00 × 10−4). We also performed genetic correlation between the seven MRIBAGs and 525 disease endpoints and found six significant signals (P < 0.05/525). For example, within-organ connections were demonstrated between the heart MRIBAG and two forms of hypertension, which is a major risk factor for cardiovascular diseases (for example, FinnGen code: I9_HYPTENS: gc = 0.28 ± 0.06; P < 1.00 × 10−5); cross-organ interactions were also identified between the adipose MRIBAG and substance abuse (gc = 0.27 ± 0.06; P < 5.00 × 10−5) (Fig. 3d, Supplementary Fig. 19 and Supplementary File 5).

which is a major risk factor for cardiovascular diseases (for example, FinnGen code: I9_HYPTENS: gc = 0.28 ± 0.06; P < 1.00 × 10−5); cross-organ interactions were also identified between the adipose MRIBAG and substance abuse (gc = 0.27 ± 0.06; P < 5.00 × 10−5) (Fig. 3d, Supplementary Fig. 19 and Supplementary File 5). We conducted bidirectional Mendelian randomization analyses (Method 4) between the seven MRIBAGs and 525 disease endpoints, ensuring at least seven valid linkage disequilibrium-considered SNPs as instrumental variables after quality control. Although the MRIBAGs were generally underpowered to meet this threshold, we identified six potential causal relationships. The within-organ interactions between the heart MRIBAG and hypertension were reaffirmed by a potential causal relationship from various forms of hypertension to the heart MRIBAG (for example, I9_HYPTENS: P = 2.12 × 10−7; odds ratio (95% confidence interval) = 1.13 (1.08−1.18); number of instrumental variables = 110). Additionally, we identified a potential causal relationship from Alzheimer’s disease to the liver MRIBAG (G6_AD_WIDE: P = 3.02 × 10−5; odds ratio (95% confidence interval) = 0.91 (0.88−0.96); number of instrumental variables = 8), consistent with our previous findings linking Alzheimer’s disease to the hepatic PhenoBAG11. Finally, we found that the kidney MRIBAG was negatively linked to two forms of type 2 diabetes mellitus (T2D) (for example, T2D: P = 8.40 × 10−5; odds ratio (95% confidence interval) = 0.78 (0.70−0.88); number of instrumental variables = 9) (Fig. 3e). Supplementary File 6 and Supplementary Figs. 20−25 present detailed statistics for our causal analyses and sensitivity check analyses.

two forms of type 2 diabetes mellitus (T2D) (for example, T2D: P = 8.40 × 10−5; odds ratio (95% confidence interval) = 0.78 (0.70−0.88); number of instrumental variables = 9) (Fig. 3e). Supplementary File 6 and Supplementary Figs. 20−25 present detailed statistics for our causal analyses and sensitivity check analyses. Using split-sample GWAS analyses, we also developed PRSs for the seven MRIBAGs, with the first split serving as the training/source dataset and the second split serving as the test/target dataset (Method 4). Overall, these PRSs exhibit limited predictive power. In the second split population, the brain PRS exhibited the highest incremental R2 (2.18%; P < 1.00 ×10−10), whereas the spleen PRS accounted for an additional 0.70% (P < 1.00 × 10−10) of the variance in the spleen MRIBAG (Fig. 3f and Supplementary Table 6). We conducted multiple post-GWAS analyses (Method 4) to identify potential druggable genes associated with the seven MRIBAGs. Our overarching hypothesis posits that these imaging-derived aging clocks, as AI-driven endophenotypes22 (see Supplementary Fig. 26 for a conceptual visualization), reside along the causal pathway of human aging and disease, linking genetic predisposition to multi-omics data, such as gene expression, proteomics and metabolomics, leading to disease outcomes and cognitive decline.

aging clocks, as AI-driven endophenotypes22 (see Supplementary Fig. 26 for a conceptual visualization), reside along the causal pathway of human aging and disease, linking genetic predisposition to multi-omics data, such as gene expression, proteomics and metabolomics, leading to disease outcomes and cognitive decline. Initially, we mapped independent GWAS signals to genes using three approaches—(1) positional mapping, (2) expression quantitative trait locus (eQTL) mapping and (3) chromatin interaction mapping—identifying a total of 1,164 unique genes. To further refine these genes, we prioritized genes that demonstrated genetic evidence from both eQTL mapping and chromatin interactions, resulting in a final set of 246 unique genes linked to the seven MRIBAGs. For instance, MAPT was prioritized in the brain MRIBAG through both eQTL and chromatin interaction mapping and was previously identified as a genetically supported druggable gene23. Subsequently, to further extend genetic evidence to other omics data types, we conducted Bayesian co-localization analysis on genomic loci associated with the seven MRIBAGs and aging clocks derived from plasma proteomics13 (ProtBAGs) and metabolomics (MetBAGs5). The key hypothesis is that a causal variant positionally or functionally linked to a gene influencing both MRIBAGs and other omics-based aging clocks is more likely to play a causal role in human aging and disease. Our co-localization analysis identified a total of 40 MRIBAG−MetBAG or MRIBAG−ProtBAG co-localization signals (PP.H4.ABF > 0.8) (Fig. 4a). By integrating these findings with the 246 prioritized genes, we identified 62 MRIBAG–gene pairs (comprising 62 unique genes) that share at least one causal SNP with either MetBAGs or ProtBAGs (Fig. 4b, Supplementary Table 7 and Supplementary File 7).Fig. 4Multi-level genetic evidence prioritizes 62 genes as potential drug repurposing genes.a, To prioritize potential druggable genes, leveraging aging clocks derived from plasma proteomics (ProtBAGs) and metabolomics (MetBAGs), we filtered and mapped genes to retain only those showing significant co-localization with any ProtBAG or MetBAG. As an example, we highlighted a specific genomic locus demonstrating co-localization among the spleen MRIBAG, the immune MetBAG and the pulmonary ProtBAG.

a proteomics (ProtBAGs) and metabolomics (MetBAGs), we filtered and mapped genes to retain only those showing significant co-localization with any ProtBAG or MetBAG. As an example, we highlighted a specific genomic locus demonstrating co-localization among the spleen MRIBAG, the immune MetBAG and the pulmonary ProtBAG. The shared causal variant (rs233721 on chromosome 12) exhibited a strong co-localization signal, with a posterior probability (PP.H4.ABF) of 0.99 between the spleen MRIBAG and the immune MetBAG, implying a single shared causal variant influencing both traits within this locus. To determine the significance of the H4 hypothesis, we set a threshold of PP.H4.ABF > 0.8. This variant was mapped to the TRAFD1 and ALDH2 genes through both eQTL mapping (for example, ALDH2 in the aorta from GTEx) and chromatin interaction mapping (for example, TRAFD1 in the left ventricle and liver). b, By integrating this multi-level genetic evidence, we identified 62 unique genes associated with the liver, spleen and pancreas MRIBAGs. LD, linkage disequilibrium.

gh both eQTL mapping (for example, ALDH2 in the aorta from GTEx) and chromatin interaction mapping (for example, TRAFD1 in the left ventricle and liver). b, By integrating this multi-level genetic evidence, we identified 62 unique genes associated with the liver, spleen and pancreas MRIBAGs. LD, linkage disequilibrium. a, To prioritize potential druggable genes, leveraging aging clocks derived from plasma proteomics (ProtBAGs) and metabolomics (MetBAGs), we filtered and mapped genes to retain only those showing significant co-localization with any ProtBAG or MetBAG. As an example, we highlighted a specific genomic locus demonstrating co-localization among the spleen MRIBAG, the immune MetBAG and the pulmonary ProtBAG. The shared causal variant (rs233721 on chromosome 12) exhibited a strong co-localization signal, with a posterior probability (PP.H4.ABF) of 0.99 between the spleen MRIBAG and the immune MetBAG, implying a single shared causal variant influencing both traits within this locus. To determine the significance of the H4 hypothesis, we set a threshold of PP.H4.ABF > 0.8. This variant was mapped to the TRAFD1 and ALDH2 genes through both eQTL mapping (for example, ALDH2 in the aorta from GTEx) and chromatin interaction mapping (for example, TRAFD1 in the left ventricle and liver). b, By integrating this multi-level genetic evidence, we identified 62 unique genes associated with the liver, spleen and pancreas MRIBAGs. LD, linkage disequilibrium.

gh both eQTL mapping (for example, ALDH2 in the aorta from GTEx) and chromatin interaction mapping (for example, TRAFD1 in the left ventricle and liver). b, By integrating this multi-level genetic evidence, we identified 62 unique genes associated with the liver, spleen and pancreas MRIBAGs. LD, linkage disequilibrium. Among the 62 genes with multi-level genetic evidence supporting their potential causal role in human aging and disease, we further explored existing drug−gene interaction data to identify potential drug repurposing opportunities and evidence. To achieve this, we queried the Drug Gene Interaction database (DGIdb) (https://dgidb.org/; query date: 21 March 2025) and identified nine genes (linked to the liver, spleen and pancreas MRIBAGs) with known interactions involving 122 unique drugs (Fig. 5a and Supplementary File 8). Among the 122 unique drugs, many indications are manifested, including immunosuppressant agents (for example, everolimus; DrugBank ID: DB01590), antipsychotic agents (for example, loxapine for the management of the manifestations of psychotic disorders such as schizophrenia; DrugBank ID: DB00408), antidiabetic agents (for example, bromocriptine for treating T2D; DrugBank ID: DB01200) and antiparkinson agents (lisuride for the management of Parkinson’s disease; DrugBank ID: DB00589), as well as many antineoplastic agents.Fig. 5Drug−gene interactions for nine potential drug repurposing genes.a, We queried the 62 prioritized MRIBAG genes in the DGIdb platform (https://dgidb.org/) to identify drug−gene interactions and potential drug repurposing opportunities. We show only the nine MRIBAG genes linked to existing drugs, whether approved or unapproved, using curated data via the DGIdb. For visualization, we annotated representative drug names; the complete results can be found at an online interactive webpage at https://labs-laboratory.com/medicine/mribag_gdi.html. b, Drug indications for existing drug−gene interactions related to the spleen MRIBAG-linked ALDH2 gene. c, The three-dimensional structure of the ALDH2 protein predicted by the AlphaFold model.

s; the complete results can be found at an online interactive webpage at https://labs-laboratory.com/medicine/mribag_gdi.html. b, Drug indications for existing drug−gene interactions related to the spleen MRIBAG-linked ALDH2 gene. c, The three-dimensional structure of the ALDH2 protein predicted by the AlphaFold model. a, We queried the 62 prioritized MRIBAG genes in the DGIdb platform (https://dgidb.org/) to identify drug−gene interactions and potential drug repurposing opportunities. We show only the nine MRIBAG genes linked to existing drugs, whether approved or unapproved, using curated data via the DGIdb. For visualization, we annotated representative drug names; the complete results can be found at an online interactive webpage at https://labs-laboratory.com/medicine/mribag_gdi.html. b, Drug indications for existing drug−gene interactions related to the spleen MRIBAG-linked ALDH2 gene. c, The three-dimensional structure of the ALDH2 protein predicted by the AlphaFold model.

s; the complete results can be found at an online interactive webpage at https://labs-laboratory.com/medicine/mribag_gdi.html. b, Drug indications for existing drug−gene interactions related to the spleen MRIBAG-linked ALDH2 gene. c, The three-dimensional structure of the ALDH2 protein predicted by the AlphaFold model. We highlighted the spleen MRIBAG-linked gene ALDH2 as exemplified in our co-localization analyses (Fig. 4a). Multi-level genetic evidence links the spleen BAG, ALDH2 and 15 drugs (loxapine, clozapine, rotigotine, sulpiride, prochlorperazine, terguride, sertindole, fenoldopam, pergolide, trifluoperazine, naltrexone, cabergoline, apomorphine, lisuride and bromocriptine) with multiple clinical indications. These indications include antipsychotic (for example, clozapine for treating severe forms of psychiatric conditions such as schizophrenia and for preventing suicide in patients who have this condition); antiparkinson (for example, rotigotine); antidepressive (for example, sulpiride, used to treat acute and chronic schizophrenia and major depressive disorder but potentially toxic with a high dose, with side effects for depression); antimigraine (for example, prochlorperazine, used to treat psychotic symptoms such as schizophrenia and non-psychotic anxiety but has off-label uses, including emergency use to treat migraine in adults and children); antihypertensive (for example, fenoldopam, a medication used to temporarily treat high blood pressure); antiemetics (for example, trifluoperazine, a medication used as an antipsychotic and an antiemetic); appetite suppressant; treatment of erectile dysfunction; antidiabetic; treatment of pulmonary arterial hypertension; treatment of substance abuse; analgesic; and treatment of sexual dysfunction in women (Fig. 5b). The 15 drugs linked to ALDH2 encompass a broad range of therapeutic areas, many of which are relevant to aging-related diseases such as neurodegeneration, metabolic disorders and immune system aging. As a central immune organ, the spleen’s aging may indicate immune dysfunction that contributes to neuroinflammation. ALDH2 could influence both immune cell activity in the spleen and neuroinflammation in the brain, especially concerning drugs for neurodegenerative and psychiatric disorders. Drugs targeting ALDH2 (or related pathways) for neurodegenerative diseases may also affect systemic aging, positioning them as potential candidates for modulating spleen or immune aging as well (Supplementary Fig. 27).

flammation in the brain, especially concerning drugs for neurodegenerative and psychiatric disorders. Drugs targeting ALDH2 (or related pathways) for neurodegenerative diseases may also affect systemic aging, positioning them as potential candidates for modulating spleen or immune aging as well (Supplementary Fig. 27). We highlight the clinical potential of the seven MRIBAGs and their PRSs in predicting (1) the incidence of 53 individual disease conditions (with at least 50 patients) using International Classification of Diseases, Tenth Revision (ICD-10) codes; (2) all-cause mortality risk through survival analyses; and (3) differentiation in cognitive decline trajectories in Alzheimer’s disease drug solanezumab (Method 5).

e incidence of 53 individual disease conditions (with at least 50 patients) using International Classification of Diseases, Tenth Revision (ICD-10) codes; (2) all-cause mortality risk through survival analyses; and (3) differentiation in cognitive decline trajectories in Alzheimer’s disease drug solanezumab (Method 5). We found that the brain, adipose and pancreas MRIBAGs and the pancreas PRS could significantly (P< 0.05/7) predict non-insulin-dependent diabetes mellitus (E119; 1.44 < hazard ratio < 1.70). The brain MRIBAG also predicted personal history of psychoactive substance abuse (Z864; hazard ratio = 1.27 (1.11−1.46); P = 6.70 × 10−4), anxiety disorders (F419; hazard ratio = 1.60 (1.20−2.11); P = 1.31 × 10−3) and gastrointestinal hemorrhage (K922; hazard ratio = 1.46 (1.11−1.91); P = 6.63 × 10−3). The heart MRIBAG was associated with an increased risk of hypertension (I10; hazard ratio = 1.16 (1.05−1.28); P = 4.80 × 10−3), reinforcing findings from genetic correlation and Mendelian randomization analyses (Fig. 6a and Supplementary File 9).Fig. 6Clinical utility of the seven MRIBAGs and their PRSs.a, The seven MRIBAGs and their PRSs show significant associations with the incidence of single disease entities (for example, I10 for hypertension). We only included disease entities with more than 50 cases (two-sided P < 0.05/7 = 0.0071 corrected for the number of organ systems), and non-cases are participants who do not show any disease diagnosis (disease free) after enrolling in the UKBB, so that all non-case populations remained the same reference population across all tasks. Age and sex were included as covariates in the Cox proportional hazard model. An asterisk (*) denotes results that remain significant after applying the Bonferroni correction for multiple comparisons (two-sided P < 0.05/53 = 0.00094, accounting for the number of disease endpoints), whereas a hash symbol (#) indicates significance under a more stringent threshold (two-sided P < 0.05/53/7 = 0.00013). b, The brain, adipose, spleen and liver MRIBAGs showed significant associations with all-cause mortality risk. To further validate the negative associations of the spleen and liver MRIBAGs with mortality, we conducted a disease-free analysis, including only participants without any prior disease diagnoses, to rule out potential confounders.

spleen and liver MRIBAGs showed significant associations with all-cause mortality risk. To further validate the negative associations of the spleen and liver MRIBAGs with mortality, we conducted a disease-free analysis, including only participants without any prior disease diagnoses, to rule out potential confounders. c, At week 240, PACC scores differed between the accelerated and decelerated aging groups—stratified by brain MRIBAG expression—among participants originally assigned to the drug group in the solanezumab trial. d, At week 240, PACC scores differed between the accelerated and decelerated aging groups among participants originally assigned to the placebo group in the solanezumab trial. e, At week 240 of the solanezumab trial, PACC scores in the accelerated aging group showed no difference between participants receiving the drug and those receiving a placebo. f, At week 240 of the solanezumab trial, PACC scores in the decelerated aging group showed no difference between participants receiving the drug and those receiving a placebo. Supplementary File 9 provides all statistical details, including sample size, P values, hazard ratios (HRs) and confidence intervals (CIs). In c–f, the shaded regions represent the 95% CI for the estimated PACC score. For a and b, each dot or shape represents the estimated HR for a given disease or mortality. The line extending from the dots or shapes shows the 95% CI for each HR. For c–f, all P values are two-sided, and the corresponding t values are reported for signals with P < 0.05.

esent the 95% CI for the estimated PACC score. For a and b, each dot or shape represents the estimated HR for a given disease or mortality. The line extending from the dots or shapes shows the 95% CI for each HR. For c–f, all P values are two-sided, and the corresponding t values are reported for signals with P < 0.05. a, The seven MRIBAGs and their PRSs show significant associations with the incidence of single disease entities (for example, I10 for hypertension). We only included disease entities with more than 50 cases (two-sided P < 0.05/7 = 0.0071 corrected for the number of organ systems), and non-cases are participants who do not show any disease diagnosis (disease free) after enrolling in the UKBB, so that all non-case populations remained the same reference population across all tasks. Age and sex were included as covariates in the Cox proportional hazard model. An asterisk (*) denotes results that remain significant after applying the Bonferroni correction for multiple comparisons (two-sided P < 0.05/53 = 0.00094, accounting for the number of disease endpoints), whereas a hash symbol (#) indicates significance under a more stringent threshold (two-sided P < 0.05/53/7 = 0.00013). b, The brain, adipose, spleen and liver MRIBAGs showed significant associations with all-cause mortality risk. To further validate the negative associations of the spleen and liver MRIBAGs with mortality, we conducted a disease-free analysis, including only participants without any prior disease diagnoses, to rule out potential confounders. c, At week 240, PACC scores differed between the accelerated and decelerated aging groups—stratified by brain MRIBAG expression—among participants originally assigned to the drug group in the solanezumab trial. d, At week 240, PACC scores differed between the accelerated and decelerated aging groups among participants originally assigned to the placebo group in the solanezumab trial. e, At week 240 of the solanezumab trial, PACC scores in the accelerated aging group showed no difference between participants receiving the drug and those receiving a placebo. f, At week 240 of the solanezumab trial, PACC scores in the decelerated aging group showed no difference between participants receiving the drug and those receiving a placebo. Supplementary File 9 provides all statistical details, including sample size, P values, hazard ratios (HRs) and confidence intervals (CIs). In c–f, the shaded regions represent the 95% CI for the estimated PACC score.

ging group showed no difference between participants receiving the drug and those receiving a placebo. Supplementary File 9 provides all statistical details, including sample size, P values, hazard ratios (HRs) and confidence intervals (CIs). In c–f, the shaded regions represent the 95% CI for the estimated PACC score. For a and b, each dot or shape represents the estimated HR for a given disease or mortality. The line extending from the dots or shapes shows the 95% CI for each HR. For c–f, all P values are two-sided, and the corresponding t values are reported for signals with P < 0.05. We further found that several MRIBAGs significantly (P < 0.05/7) predicted all-cause mortality (Fig. 6b). Among these, the brain and adipose MRIBAGs were identified as risk biomarkers (hazard ratio > 1) for all-cause mortality, whereas the liver and spleen MRIBAGs (hazard ratio < 1) appeared to have a protective effect. We conducted additional sensitivity analyses by excluding participants with any disease diagnoses to assess disease-free survival. Despite the reduced sample size, the protective effect remained evident. For instance, the spleen MRIBAG retained nominal significance under this approach (P = 0.04; hazard ratio = 0.62 (0.40−0.99)). Supplementary Table 8a,b and Supplementary Figs. 28 and 29 present a detailed discussion on additional sensitivity check results.

he reduced sample size, the protective effect remained evident. For instance, the spleen MRIBAG retained nominal significance under this approach (P = 0.04; hazard ratio = 0.62 (0.40−0.99)). Supplementary Table 8a,b and Supplementary Figs. 28 and 29 present a detailed discussion on additional sensitivity check results. Finally, we assessed the use of the brain MRIBAG for stratifying clinical trial participants into subgroups to (1) demonstrate differential cognitive decline trajectories across subgroups and (2) provide insights for future clinical trial design, using data from the A4 trial. As shown in Fig. 6c,d, participants exhibiting decelerated aging signatures (that is, youthful brain: MRIBAG below the median) showed a higher Preclinical Alzheimer Cognitive Composite (PACC) score than participants within the accelerated aging group (that is, aged brain), both in the drug group (P = 0.003 and t = 2.923) and the placebo group (P = 0.001 and t = 4.049). A permutation test showed that the cognitive differences between accelerated and decelerated agers, quantified by t values, were significantly more pronounced in the placebo group compared to the drug group (one-sided P = 0.03 based on 1,000 permutations). A one-sided P value was used, as our hypothesis was directional; we specifically aimed to test whether the placebo group exhibited stronger cognitive separation between aging subgroups, consistent with the observed pattern in cognitive trajectories (Fig. 6c,d). Furthermore, neither the decelerated nor the accelerated aging group exhibited significant drug effects (Fig. 6e,f) in preventing cognitive decline.

In our ProWAS analyses (Method 3), we identified 603 protein−MRIBAG significant associations (P < 0.05/2,923/7). Among these, the kidney MRIBAG exhibited the highest number of significant protein associations (N = 301; for example, NPDC1, IGFBP6 and TAFA5), followed by the spleen MRIBAG for 136 associations (for example, VCAM1, PTPRH and C1QA), the liver MRIBAG for 62 associations (for example, NCAN, SEZ6L and LEP), the adipose MRIBAG for 57 signals (for example, GDF15, CHI3L1 and CA14), the pancreas MRIBAG for 21 associations (for example, PLA2G1B, CTRC and CELA2A), the brain MRIBAG for 16 associations (for example, BCAN, NCAN and GDF15) and the heart MRIBAG for only REN (Fig. 2a).Fig. 2ProWAS and MetWAS with the seven MRIBAGs.a, ProWAS between the seven MRIBAGs and 2,923 plasma proteins via a linear regression model, accounting for a full set of covariates (P < 0.05/2,923/7). An online interactive webpage is available at https://labs-laboratory.com/medicine/mribag_prowas.html to ease visualization. Plots of individual organs are shown in Extended Data Fig. 2. b, MetWAS between the seven MRIBAGs and 327 plasma metabolites using a linear regression model while adjusting for a comprehensive set of covariates (P < 0.05/327/7). Plots of individual organs are shown in Extended Data Fig. 4. An online interactive webpage is available at https://labs-laboratory.com/medicine/mribag_metwas.html to ease visualization. The colors and shapes of the icons indicate the organ systems involved in the associations. We used the standardized β value to represent the effect size.

We conducted GWAS (Method 4) for the seven MRIBAGs (19,686 < N < 31,557 participants with European ancestries) and identified 53 (P < 5 × 10−8) genomic locus−BAG pairs. We denoted the genomic loci using their top lead single-nucleotide polymorphisms (SNPs) defined by FUMA, considering linkage disequilibrium; the genomic loci are presented in Supplementary Table 4. We visually present the shared genomic loci annotated by cytogenetic regions based on the GRCh37 cytoband (Fig. 3a). Supplementary Fig. 18 and Extended Data Fig. 7 detail the robustness of our GWASs and sex-specific GWAS analyses.Fig. 3The genetics of the seven MRIBAGs.a, Cytogenetic regions associated with the seven MRIBAGs, with significant genomic loci identified using the genome-wide significance threshold (two-sided P < 5 × 10−8). b, SNP-based heritability estimates of the seven MRIBAGs (sample size: 18,778 < N < 31,557). c, Partitioned heritability enrichment analysis using chromatin-specific multi-tissue chromatin data, multi-tissue gene expression profiles and cell-type-specific datasets. Only significant results (two-sided P < 0.05/697) are displayed (sample size: 18,778 < N < 31,557). d, Genetic correlation estimates via LDSC between the seven MRIBAGs and eight PhenoBAGs (excluding the multimodal brain PhenoBAG11), 11 ProtBAGs13 and five MetBAGs5 (two-sided P < 0.05/11) as well as 525 disease endpoints from FinnGen and PGC (P < 0.05/525). e, Potential causal relationships between the seven MRIBAGs (that is, number of instrumental variables > 7) and 525 disease endpoints were examined through two networks: BAG2DE, where the seven MRIBAGs serve as exposure variables and the 525 DEs as outcome variables, with two-sided P < 0.05/525, and DE2BAG, where the 214 disease endpoints serve as effective exposure variables (that is, number of instrumental variables > 7), with two-sided P < 0.05/214. Although multiple sensitivity checks were performed to evaluate potential violations of underlying assumptions, these findings should be interpreted with caution.

two forms of type 2 diabetes mellitus (T2D) (for example, T2D: P = 8.40 × 10−5; odds ratio (95% confidence interval) = 0.78 (0.70−0.88); number of instrumental variables = 9) (Fig. 3e). Supplementary File 6 and Supplementary Figs. 20−25 present detailed statistics for our causal analyses and sensitivity check analyses. Using split-sample GWAS analyses, we also developed PRSs for the seven MRIBAGs, with the first split serving as the training/source dataset and the second split serving as the test/target dataset (Method 4). Overall, these PRSs exhibit limited predictive power. In the second split population, the brain PRS exhibited the highest incremental R2 (2.18%; P < 1.00 ×10−10), whereas the spleen PRS accounted for an additional 0.70% (P < 1.00 × 10−10) of the variance in the spleen MRIBAG (Fig. 3f and Supplementary Table 6).

We conducted multiple post-GWAS analyses (Method 4) to identify potential druggable genes associated with the seven MRIBAGs. Our overarching hypothesis posits that these imaging-derived aging clocks, as AI-driven endophenotypes22 (see Supplementary Fig. 26 for a conceptual visualization), reside along the causal pathway of human aging and disease, linking genetic predisposition to multi-omics data, such as gene expression, proteomics and metabolomics, leading to disease outcomes and cognitive decline.

We highlight the clinical potential of the seven MRIBAGs and their PRSs in predicting (1) the incidence of 53 individual disease conditions (with at least 50 patients) using International Classification of Diseases, Tenth Revision (ICD-10) codes; (2) all-cause mortality risk through survival analyses; and (3) differentiation in cognitive decline trajectories in Alzheimer’s disease drug solanezumab (Method 5).

This study introduces seven MRI-based aging clocks by leveraging large-scale multi-organ and multi-omics data from 313,645 participants curated by the MULTI Consortium. Collectively, this work expands existing biological aging clocks and highlights their clinical potential for future research. Overall, the age prediction models highlight limitations in relying on abdominal MRI features, including a limited number of informative features and substantial multicollinearity, which reduced model performance and hindered generalizability to unseen data. Compared to the brain imaging community24, the development of imaging protocols and standardized imaging-derived phenotypes (IDPs) for other organs may be relatively underdeveloped. This disparity is reflected in large-scale biobanks such as the UKBB, where the majority of well-characterized IDPs are derived from brain MRI (N > 4,000) and heart MRI, whereas imaging-derived features, returned and made available to the community, for other organs are limited in both number and diversity. Several factors contribute to this imbalance, including the historical focus of neuroimaging research, the availability of standardized brain imaging pipelines25 and the widespread adoption of advanced computational tools tailored for brain MRI analysis. Imaging data from other organs should also be incorporated into this framework. For instance, integrating retina imaging26 with brain MRI, genetic data and systemic biomarkers could reveal mechanisms underlying aging and disease.

d the widespread adoption of advanced computational tools tailored for brain MRI analysis. Imaging data from other organs should also be incorporated into this framework. For instance, integrating retina imaging26 with brain MRI, genetic data and systemic biomarkers could reveal mechanisms underlying aging and disease. The ProWAS results highlight both organ-specific aging mechanisms and systemic inter-organ communication, providing a foundation for understanding multi-tissue aging dynamics. Such organ specificity of the brain MRIBAG was also reinforced by two previous studies that used plasma proteomics data to develop brain tissue-enriched aging clocks (that is, the brain ProtBAG) using both SomaScan and Olink platforms. Oh et al.4 identified the three brain tissue-enriched proteins based on Genotype-Tissue Expression (GTEx) data to construct the brain ProtBAG, whereas Wen et al.13 leveraged the HPA platform, which integrates more extensive curation at both RNA and protein levels, to define brain tissue-specific proteins. Other proteins include CHI3L1, a well-established neuroinflammatory marker elevated in Alzheimer’s disease and multiple sclerosis27, and GDF15, a stress-induced cytokine linked to cognitive decline, Alzheimer’s disease and aging28,29. The significant metabolites identified in the MetWAS analysis further reinforce both cross-organ and within-organ associations. Although certain proteins or metabolites, such as GDF15 and FABP4, serve as systemic aging markers, others, such as NCAN and IGFBP6, exhibit tissue-specific roles. Shared underlying biological pathways may include inflammation30, extracellular matrix dysregulation31 and metabolic dysfunction32.

thin-organ associations. Although certain proteins or metabolites, such as GDF15 and FABP4, serve as systemic aging markers, others, such as NCAN and IGFBP6, exhibit tissue-specific roles. Shared underlying biological pathways may include inflammation30, extracellular matrix dysregulation31 and metabolic dysfunction32. The results of our genetic analyses reveal that, although the MRIBAGs share certain genetic underpinnings, each organ also demonstrates a distinct genetic architecture shaped by its unique biological processes. Genetic correlations between the MRIBAGs and other multi-omics aging clocks, as well as the MRIBAGs’ associations with various disease endpoints, strengthen the notion of both shared and organ-specific genetic influences. For example, the correlation between the kidney MRIBAG and the renal PhenoBAG highlights how kidney-specific genetic factors can interact with broader multi-organ aging mechanisms at multiple omics levels33. Notably, although our analyses show that these organs share common genetic influences that link them to broader disease pathways, they also exhibit unique, organ-specific contributions that reflect their individual physiological roles and functions. Together, these findings provide strong support for the hypothesis that the MRIBAGs share genetic influences across organs while also demonstrating distinct genetic signatures that contribute to the specificity of each organ’s role in aging and disease. Supplementary Fig. 30 and Supplementary Table 9 performed mediation analyses to link the MRIBAGs to ProtBAGs and late-life depression34.

t the MRIBAGs share genetic influences across organs while also demonstrating distinct genetic signatures that contribute to the specificity of each organ’s role in aging and disease. Supplementary Fig. 30 and Supplementary Table 9 performed mediation analyses to link the MRIBAGs to ProtBAGs and late-life depression34. The broad spectrum of drug indications linked to the prioritized genes highlights the opportunity to repurpose existing therapies for targeting aging and age-related diseases in different organs, providing a promising direction for future therapeutic advancements. For example, everolimus, a rapamycin analog, has shown promise in extending lifespan and improving lifespan in humans35,36, particularly in delaying age-related diseases and improving immune function. It is known for its ability to suppress immune system activity, which can reduce chronic inflammation, a hallmark of aging. Studies have suggested that chronic low-grade inflammation, often referred to as ‘inflammaging’37, contributes to various age-related diseases, such as cardiovascular disease38 and Alzheimer’s disease39. Immunosuppressants may mitigate this process by modulating immune responses and promoting tissue repair and regeneration. Notably, the mTOR pathway35, targeted by drugs such as everolimus, has been shown to regulate cellular processes such as growth, survival and metabolism, and its inhibition has been linked to lifespan extension in model organisms, including yeast, worms and mice40.

esponses and promoting tissue repair and regeneration. Notably, the mTOR pathway35, targeted by drugs such as everolimus, has been shown to regulate cellular processes such as growth, survival and metabolism, and its inhibition has been linked to lifespan extension in model organisms, including yeast, worms and mice40. The counterintuitive finding that higher values of the liver and spleen MRIBAGs are associated with a lower risk of mortality could reflect their role in preserving organ function and overall physiological balance as people age. Potentially, the higher MRIBAG values likely reflect healthier, more resilient organs that maintain their metabolic and immune functions more effectively over time. Specifically, a higher liver MRIBAG may signal better metabolic health, efficient detoxification and regulation of systemic inflammation, which are crucial for maintaining overall homeostasis. Similarly, a higher spleen MRIBAG could indicate better immune function, including more effective immune surveillance and response to infections and other stressors. In the context of aging, these organs might act as ‘reserve systems’41, buffering the body against environmental insults and chronic diseases, which would reduce the risk of mortality.

BAG could indicate better immune function, including more effective immune surveillance and response to infections and other stressors. In the context of aging, these organs might act as ‘reserve systems’41, buffering the body against environmental insults and chronic diseases, which would reduce the risk of mortality. This study has several limitations. First, the analysis is based on cross-sectional data. Longitudinal studies will be crucial to better investigate the dynamic trajectories of MRIBAGs over time and their temporal relationships with health outcomes. Second, although the analysis focused primarily on genetic and imaging data, environmental factors such as lifestyle, diet and pollution were not considered, yet these may play important roles in aging and disease42. Third, the imaging features for the five abdominal organs and tissues were relatively underdeveloped, underscoring the need to implement more advanced feature engineering techniques and include external independent studies. Furthermore, the genetic analysis was focused primarily on individuals of European ancestry, and future studies should extend these findings to underrepresented populations to enhance their generalizability. Future refinement through larger, multi-ethnic GWASs and the inclusion of rare variants may improve their utility and biological interpretability. In addition, the current results derived from the UKBB study would benefit from confirmation in other large-scale cohorts to ensure their generalizability across populations and study designs. Moreover, as a future avenue, studies incorporating more Alzheimer’s disease-oriented and AI-derived imaging biomarkers (for example, subtypes informed by tau or amyloid positron emission tomography (PET) data) may enhance statistical power and clinical relevance to benefit future clinical trials. Finally, the observed protective associations for the liver and spleen MRIBAGs are biologically ambiguous and complex. As such, these findings should be interpreted cautiously.

The MULTI Consortium is an ongoing initiative to integrate and consolidate multi-organ data and multi-omics data, including imaging, genetics and proteomics. Building upon existing consortia and studies, such as those listed below, MULTI aims to curate and harmonize the data to model human aging and disease across the lifespan. In the present study, in total, individual-level data for 313,645 participants were analyzed, including multi-organ MRI data across seven organs and tissues (Category ID: 100003), genetics, plasma proteomics data and metabolomics data. GWAS summary statistics from FinnGen and the Psychiatric Genomics Consortium (PGC) were downloaded and harmonized for our post-GWAS analyses. Refer to Supplementary Table 1 for comprehensive information, including the complete list of data analyzed and their respective sample characteristics.

ata and metabolomics data. GWAS summary statistics from FinnGen and the Psychiatric Genomics Consortium (PGC) were downloaded and harmonized for our post-GWAS analyses. Refer to Supplementary Table 1 for comprehensive information, including the complete list of data analyzed and their respective sample characteristics. The MULTI Consortium received approval from the institutional review board at Columbia University (AAAV6751). Participants provided written informed consent for each study that included individual-level data, whereas FinnGen and PGC provided only summary-level statistics. Specifically, the UKBB obtained research ethics approval and received guidance from an Ethics Advisory Committee, with additional ethical input provided to the Access Committee, which oversees decisions regarding access to data and biological samples. The BLSA protocol was approved by the institutional review board of the National Institute of Environmental Health Science, National Institutes of Health (03AG0325). For the A4 trial, institutional review board approval was obtained at each trial site, and all participants provided written informed consent. The study was conducted in accordance with International Council for Harmonization Good Clinical Practice guidelines.

l Health Science, National Institutes of Health (03AG0325). For the A4 trial, institutional review board approval was obtained at each trial site, and all participants provided written informed consent. The study was conducted in accordance with International Council for Harmonization Good Clinical Practice guidelines. The UKBB43 is a population-based research initiative comprising approximtely 500,000 individuals from the United Kingdom between 2006 and 2010. Ethical approval for the UKBB study was secured, and information about the ethics committee can be found here: https://www.ukbiobank.ac.uk/learn-more-about-uk-biobank/governance/ethics-advisory-committee. This study collectively analyzed multi-organ MRI data from seven organ systems and tissues, including the brain, heart, liver, pancreas, spleen, adipose and kidney. For the genetic data, we conducted a quality check on the imputed genotype data43 for the entire UKBB population (approximately 500,000 individuals). Subsequently, we merged the processed data with the organ-specific populations for all genetic analyses. Finally, we also included Olink plasma proteomics data released by the UK Biobank Pharma Proteomics Project (UKB-PPP) and metabolomics data from the UKBB, which are detailed in the following sections.

individuals). Subsequently, we merged the processed data with the organ-specific populations for all genetic analyses. Finally, we also included Olink plasma proteomics data released by the UK Biobank Pharma Proteomics Project (UKB-PPP) and metabolomics data from the UKBB, which are detailed in the following sections. The main goal of the BLSA is to understand the normal aging process. Tracking physiological and cognitive changes over time aims to identify risk factors for age-related diseases, study patterns of decline and discover predictors of healthy aging. BLSA44,45 brain MRI and SomaScan proteomics data (https://www.blsa.nih.gov/) were used to compare and replicate the ProWAS results from the UKBB Olink data. After quality checks in this study, we included 1,114 brain MRI scans at baseline and measurements of 7,268 plasma proteins from 909 participants quantified with the SomaScan version 4.1 platform. Age (years), sex (male/female), race (white/non-white) and education level (years) were defined based on participant self-reports.

uality checks in this study, we included 1,114 brain MRI scans at baseline and measurements of 7,268 plasma proteins from 909 participants quantified with the SomaScan version 4.1 platform. Age (years), sex (male/female), race (white/non-white) and education level (years) were defined based on participant self-reports. The A4 study18,19 (https://atri.usc.edu/study/a4-study/) is a clinical trial study to test a specific way to prevent memory loss associated with Alzheimer’s disease (clinical trial number: NCT02008357). The A4 study focused on symptom-free adults at higher risk for Alzheimer’s disease to assess whether an investigational drug (that is, solanezumab) could slow memory decline linked to amyloid plaques in the brain. It also examined whether solanezumab could delay Alzheimer’s disease progression, measuring related brain changes using imaging, blood biomarkers and baseline PET scans to assess amyloid levels. This study analyzed 1,055 participants at baseline with brain MRI scans to derive the brain MRIBAG. Longitudinal outcomes from the clinical trial, with the PACC score as the primary measure over 312 weeks, were included. The PACC scores between groups were evaluated at week 240.

line PET scans to assess amyloid levels. This study analyzed 1,055 participants at baseline with brain MRI scans to derive the brain MRIBAG. Longitudinal outcomes from the clinical trial, with the PACC score as the primary measure over 312 weeks, were included. The PACC scores between groups were evaluated at week 240. The FinnGen46 study is a large-scale genomics initiative that has analyzed more than 500,000 Finnish biobank samples and correlated genetic variation with health data to understand disease mechanisms and predispositions. The project is a collaboration between research organizations and biobanks within Finland and international industry partners. For the benefit of research, FinnGen generously made their GWAS findings accessible to the wider scientific community (https://www.finngen.fi/en/access_results). This research used the publicly released GWAS summary statistics (version R9), which became available on 11 May 2022, after harmonization by the consortium. No individual data were used in the current study. FinnGen published the R9 version of GWAS summary statistics via REGENIE software (version 2.2.4)47, covering 2,272 disease endpoints, including 2,269 binary traits and three quantitative traits. The GWAS model encompassed covariates such as age, sex, the initial 10 genetic principal components and the genotyping batch. Genotype imputation was referenced on the population-specific SISu version 4.0 panel. We included GWAS summary statistics for 521 FinnGen disease endpoints in our analyses.

quantitative traits. The GWAS model encompassed covariates such as age, sex, the initial 10 genetic principal components and the genotyping batch. Genotype imputation was referenced on the population-specific SISu version 4.0 panel. We included GWAS summary statistics for 521 FinnGen disease endpoints in our analyses. PGC48 is an international collaboration of researchers studying the genetic basis of psychiatric disorders. PGC aims to identify and understand the genetic factors contributing to various psychiatric disorders, such as schizophrenia, bipolar disorder, major depressive disorder and others. The GWAS summary statistics were acquired from the PGC website (https://pgc.unc.edu/for-researchers/download-results/), underwent quality checks and were harmonized to ensure seamless integration into our analysis. No individual data were used from PGC. Each study detailed its specific GWAS models and methodologies, and the consortium consolidated the release of GWAS summary statistics derived from individual studies. In the present study, we included summary data for four brain diseases for which allele frequencies were present.

No individual data were used from PGC. Each study detailed its specific GWAS models and methodologies, and the consortium consolidated the release of GWAS summary statistics derived from individual studies. In the present study, we included summary data for four brain diseases for which allele frequencies were present. In our previous analysis of raw brain MRI data from the UKBB and other studies included in this work, we extracted 119 gray matter regions of interest from T1-weighted MRI to generate brain MRIBAGs through the iSTAGING consortium49. For the heart MRIBAG, we used 80 heart MRI traits from Bai et al.50 and used these imaging features in a previous study9. To develop the remaining five MRIBAGs based on abdominal MRI (Category ID: 105 (refs. 51–56)), we incorporated 16 imaging features for the adipose MRIBAG, four for the liver MRIBAG, three for the kidney MRIBAG, three for the spleen MRIBAG and three for the pancreas MRIBAG. Notably, in the development of adipose and kidney MRIBAGs, we observed a high degree of collinearity among certain imaging features, which led to overfitting. To mitigate this issue, we removed highly correlated features to improve model robustness.

MRIBAG, three for the spleen MRIBAG and three for the pancreas MRIBAG. Notably, in the development of adipose and kidney MRIBAGs, we observed a high degree of collinearity among certain imaging features, which led to overfitting. To mitigate this issue, we removed highly correlated features to improve model robustness. The abdominal MRI data underwent initial quality control procedures prior to public release, including the removal of low-quality images and biologically implausible values. As noted on the UKBB website (https://biobank.ndph.ox.ac.uk/ukb/label.cgi?id=158), different pipelines were used to generate imaging features from abdominal MRI, and data should not be combined without careful consideration. To account for this, we ensured that each abdominal feature was derived from a consistent pipeline, selecting the one with the largest available sample size across different pipelines. For liver MRI metrics specifically, we included only images acquired using the IDEAL protocol (Data-Field 40063, acquisition protocol 2). Following these additional quality control steps, our analyses incorporated abdominal MRI biomarkers from all available UKBB participants for the liver, kidney, spleen, pancreas and adipose MRIBAGs.

trics specifically, we included only images acquired using the IDEAL protocol (Data-Field 40063, acquisition protocol 2). Following these additional quality control steps, our analyses incorporated abdominal MRI biomarkers from all available UKBB participants for the liver, kidney, spleen, pancreas and adipose MRIBAGs. Our previous study systematically evaluated age prediction performance across various AI/ML models using multimodal brain MRI features14 for the brain MRIBAG. Applying the same framework, we assessed the performance of models in deriving the six additional MRIBAGs using several ML methods. Hyperparameter tuning was performed through nested, repeated holdout cross-validation57 with 50 repetitions (80% training/validation and 20% testing). Specifically, the hyperparameters of each model were tuned using grid search: LASSO regression’s alpha (α), linear support vector regressor’s C and both α and L1-ratio for the elastic net. Due to the flexibility and large hyperparameter space of neural networks (for example, number of layers and neurons), they were not included in the nested cross-validation pipeline. Instead, the neural network was manually tuned and evaluated on the cross-validation test dataset, and the final model was subsequently assessed on the within-distribution, holdout test set. The within-distribution, holdout test dataset was held out to unbiasedly evaluate model performance (Supplementary Fig. 31).

ine. Instead, the neural network was manually tuned and evaluated on the cross-validation test dataset, and the final model was subsequently assessed on the within-distribution, holdout test set. The within-distribution, holdout test dataset was held out to unbiasedly evaluate model performance (Supplementary Fig. 31). To rigorously train the AI/ML models, we first defined participants without any pathologies based on ICD code and clinical history as CN (N = 5,291 for the heart MRIBAG, varying across the seven organs/tissues). We further split the CN into the following datasets:CN within-distribution, holdout test dataset: 500 participants were randomly drawn from the CN population. Within-distribution, holdout test datasets are ideal for objectively evaluating ML model performance, especially in studies with large sample sizes, such as the present one.CN training/validation dataset: 80% of the remaining 4,791 CN (taking the heart MRIBAG for illustration) were used for the inner loop 10-fold cross-validation for hyperparameter selection.CN cross-validated test dataset: 20% of the remaining 4,791 CN were used for the outer loop 50 repetitions.Patient dataset: 29,785 patients who have at least one ICD-10-based diagnosis. CN within-distribution, holdout test dataset: 500 participants were randomly drawn from the CN population. Within-distribution, holdout test datasets are ideal for objectively evaluating ML model performance, especially in studies with large sample sizes, such as the present one.

To rigorously train the AI/ML models, we first defined participants without any pathologies based on ICD code and clinical history as CN (N = 5,291 for the heart MRIBAG, varying across the seven organs/tissues). We further split the CN into the following datasets:CN within-distribution, holdout test dataset: 500 participants were randomly drawn from the CN population. Within-distribution, holdout test datasets are ideal for objectively evaluating ML model performance, especially in studies with large sample sizes, such as the present one.CN training/validation dataset: 80% of the remaining 4,791 CN (taking the heart MRIBAG for illustration) were used for the inner loop 10-fold cross-validation for hyperparameter selection.CN cross-validated test dataset: 20% of the remaining 4,791 CN were used for the outer loop 50 repetitions.Patient dataset: 29,785 patients who have at least one ICD-10-based diagnosis. CN within-distribution, holdout test dataset: 500 participants were randomly drawn from the CN population. Within-distribution, holdout test datasets are ideal for objectively evaluating ML model performance, especially in studies with large sample sizes, such as the present one. CN training/validation dataset: 80% of the remaining 4,791 CN (taking the heart MRIBAG for illustration) were used for the inner loop 10-fold cross-validation for hyperparameter selection. CN cross-validated test dataset: 20% of the remaining 4,791 CN were used for the outer loop 50 repetitions. Patient dataset: 29,785 patients who have at least one ICD-10-based diagnosis.

CN training/validation dataset: 80% of the remaining 4,791 CN (taking the heart MRIBAG for illustration) were used for the inner loop 10-fold cross-validation for hyperparameter selection. CN cross-validated test dataset: 20% of the remaining 4,791 CN were used for the outer loop 50 repetitions. Patient dataset: 29,785 patients who have at least one ICD-10-based diagnosis. The CN training/validation/test datasets were used for model development and were employed with a nested cross-validation procedure for four AI/ML models (LASSO regression and support vector regressor, elastic net and neural network), whereas the within-distribution, holdout test set provided an unbiased assessment of model performance. Model evaluation metrics included MAE and Pearson’s r. Notably, consistent with our previous studies, only CN participants were included in the training/validation dataset for modeling and clinical interpretation considerations13. The BAG was determined by subtracting the participant’s chronological age from the AI/ML-predicted age. Age bias correction was applied using the approach outlined by Beheshti et al.16 and was systematically discussed in our previous comment13. To summarize, we applied an age bias correction method commonly used in neuroimaging studies, where a linear regression of the BAG on chronological age is used to estimate and remove systematic age-related bias via the model’s slope and intercept. Notably, to address potential domain shift in the patient data compared to the CN training set, we computed the regression parameters directly from the patient set to more accurately correct for this bias (refer to our comment13 for further details). Supplementary Fig. 32 details the feature importance across UKBB, BLSA and A4.

to address potential domain shift in the patient data compared to the CN training set, we computed the regression parameters directly from the patient set to more accurately correct for this bias (refer to our comment13 for further details). Supplementary Fig. 32 details the feature importance across UKBB, BLSA and A4. We used the original dataset (Category ID: 1838), which was analyzed and shared with the research community by the UKB-PPP58. The initial quality control procedures were described in the original study59, and we implemented additional quality control steps as outlined below. Our analysis focused on the first instance of the proteomics data (‘instance’ = 0). We then integrated Olink files containing coding information, batch numbers, assay details and limit of detection (LOD) data (Category ID: 1839) by matching them to the proteomics dataset ID. Finally, we excluded Normalized Protein eXpression (NPX) values that fell below the protein-specific LOD.

data (‘instance’ = 0). We then integrated Olink files containing coding information, batch numbers, assay details and limit of detection (LOD) data (Category ID: 1839) by matching them to the proteomics dataset ID. Finally, we excluded Normalized Protein eXpression (NPX) values that fell below the protein-specific LOD. We conducted ProWASs by linking the seven MRIBAGs to 2,923 unique plasma proteins measured in 53,016 participants (with sample sizes ranging from 10,018 to 39,489 per protein after quality control) using the Olink platform. The linear regression model adjusted for common covariates, including age (Field ID: 21003), sex (Field ID: 31), body weight (Field ID: 21002), height (Field ID: 50), waist circumference (Field ID: 48), body mass index (BMI) (Field ID: 23104), assessment center (Field ID: 54), protein batch number (Category ID: 1839), LOD (Category ID: 1839) and the first 40 genetic principal components. Additionally, organ-specific covariates were incorporated, such as brain positioning in the scanner (lateral, transverse and longitudinal; Field ID: 25756−25758), head motion (Field ID: 25741) and intracranial volume for the brain MRIBAG as well as diastolic (Field ID: 4079) and systolic (Field ID: 4080) blood pressure for the heart MRIBAG, among others. Multiple testing corrections were applied using Bonferroni adjustment (P < 0.05/2,923/7). To identify and exclude extreme outliers, we defined an upper threshold as the mean plus six times the s.d. for each protein/metabolite. Finally, ProWAS signals identified in the UKBB Olink dataset were compared with SomaScan data from the BLSA.

g corrections were applied using Bonferroni adjustment (P < 0.05/2,923/7). To identify and exclude extreme outliers, we defined an upper threshold as the mean plus six times the s.d. for each protein/metabolite. Finally, ProWAS signals identified in the UKBB Olink dataset were compared with SomaScan data from the BLSA. We used the original data (Category ID: 220), which were analyzed and made available to the community by Nightingale Health Plc. The original data (1) were calibrated absolute concentrations (or ratios) and not raw NMR spectra and (2), before release, had already been subjected to quality control procedures by Nightingale Health Plc60. Following the additional procedures described in Ritchie et al.61, we performed additional quality check steps to remove a range of unwanted technical variations, including shipping batch, 96-well plate, well position, aliquoting robo and aliquot tip. We focused our analysis on the first instance of the metabolomics data (‘instance’ = 0). The analysis included 327 metabolites (comprising both small molecules and lipoprotein measures), of which 107 were non-derived metabolites and the remainder were composite metabolites, across 274,247 participants. Descriptions of these metabolites are provided in Supplementary Table 3.

abolomics data (‘instance’ = 0). The analysis included 327 metabolites (comprising both small molecules and lipoprotein measures), of which 107 were non-derived metabolites and the remainder were composite metabolites, across 274,247 participants. Descriptions of these metabolites are provided in Supplementary Table 3. We conducted MetWASs by linking the seven MRIBAGs to 327 plasma metabolites. The linear regression model adjusted for common covariates, including age (Field ID: 21003), sex (Field ID: 31), body weight (Field ID: 21002), height (Field ID: 50), waist circumference (Field ID: 48), BMI (Field ID: 23104), assessment center (Field ID: 54) and the first 40 genetic principal components. Additionally, organ-specific covariates were incorporated, such as brain positioning in the scanner (lateral, transverse and longitudinal; Field ID: 25756−25758), head motion (Field ID: 25741) and intracranial volume for the brain MRIBAG, as well as diastolic (Field ID: 4079) and systolic (Field ID: 4080) blood pressure for the heart MRIBAG, among others. Multiple testing corrections were applied using Bonferroni adjustment (P < 0.05/327/7).

longitudinal; Field ID: 25756−25758), head motion (Field ID: 25741) and intracranial volume for the brain MRIBAG, as well as diastolic (Field ID: 4079) and systolic (Field ID: 4080) blood pressure for the heart MRIBAG, among others. Multiple testing corrections were applied using Bonferroni adjustment (P < 0.05/327/7). We used the imputed genotype data from the UKBB for all genetic analyses. Our quality check pipeline focused on European ancestry in the UKBB (6,477,810 SNPs passing quality checks), and the quality-checked genetic data were merged with respective organ-specific populations for GWAS. We summarize our genetic quality check steps. First, we skipped the step for family relationship inference62 because the linear mixed model via fastGWA63 inherently addresses population stratification, encompassing additional cryptic population stratification factors. We then removed duplicated variants from all 22 autosomal chromosomes. Individuals whose genetically identified sex did not match their self-acknowledged sex were removed. Other excluding criteria included (1) individuals with more than 3% of missing genotypes; (2) variants with minor allele frequency (MAF; dosage mode) of less than 1%; (3) variants with more than 3% missing genotyping rate; and (4) variants that failed the Hardy−Weinberg test at 1 × 10−10. To further adjust for population stratification,64 we derived the first 40 genetic principal components using FlashPCA software65. Details of the genetic quality check protocol are described elsewhere11,14,34,49,66,67.

e than 3% missing genotyping rate; and (4) variants that failed the Hardy−Weinberg test at 1 × 10−10. To further adjust for population stratification,64 we derived the first 40 genetic principal components using FlashPCA software65. Details of the genetic quality check protocol are described elsewhere11,14,34,49,66,67. We applied a linear mixed model regression to the European ancestry populations using fastGWA63 implemented in GCTA21. We used fastGWA to perform the seven MRIBAG GWASs, adjusting common variates, including age, dataset status (training/validation/test or within-distribution, holdout test), age-squared, sex, interactions of age with sex, BMI, waist circumference, standing height, weight and the first 40 genetic principal components, as well as organ-specific covariates, including the brain scan positions for the brain MRIBAG and systolic/diastolic blood pressure for the heart MRIBAG. We applied a genome-wide significance threshold (5 × 10−8) to annotate the significant independent genomic loci. We previously conducted GWAS of 2,923 plasma proteins and 327 metabolites using fastGWA5,68.

ncluding the brain scan positions for the brain MRIBAG and systolic/diastolic blood pressure for the heart MRIBAG. We applied a genome-wide significance threshold (5 × 10−8) to annotate the significant independent genomic loci. We previously conducted GWAS of 2,923 plasma proteins and 327 metabolites using fastGWA5,68. For all GWASs, genomic loci were annotated using FUMA69. For genomic loci annotation, FUMA initially identified lead SNPs (correlation r2 ≤ 0.1, distance < 250 kb) and assigned them to non-overlapping genomic loci. The lead SNP with the lowest P value (that is, the top lead SNP) represented the genomic locus. Further details on the definitions of top lead SNP, lead SNP, independent significant SNP and candidate SNP can be found in the FUMA documentation (https://fuma.ctglab.nl/). For visualization purposes in Fig. 3a, we mapped the top lead SNP of each locus to the cytogenetic regions based on the GRCh37 cytoband. We estimated the SNP-based heritability (h2) using GCTA21 with the same covariates as in GWAS. GCTA estimates the SNP-based heritability using a method called restricted maximum likelihood (REML) to quantify the proportion of phenotypic variance in a trait that the additive effects of all common SNPs can explain. The main steps involved in GCTA include constructing the genetic relationship matrix, modeling phenotypic variance and using REML to estimate the h2.

method called restricted maximum likelihood (REML) to quantify the proportion of phenotypic variance in a trait that the additive effects of all common SNPs can explain. The main steps involved in GCTA include constructing the genetic relationship matrix, modeling phenotypic variance and using REML to estimate the h2. The partitioned heritability analysis via stratified linkage disequilibrium score regression calculates the extent to which heritability enrichment can be attributed to predefined and annotated genome regions and categories70. Three sets of functional categories and cell-specific and tissue-specific types were considered. First, the partitioned heritability was calculated for 53 general functional categories (one including the entire set of SNPs). The 53 functional categories are not specific to any cell type and include coding regions, untranslated regions, promoter regions and intronic regions. The details of the 53 categories are described elsewhere70. Subsequently, cell-type-specific and tissue-type-specific partitioned heritability was estimated using gene sets from Cahoy et al.71 for three main cell types (that is, astrocyte, neuron and oligodendrocyte), multi-tissue chromatin state-specific data (Roadmap72 and ENTEx73) and multi-tissue gene expression data (GTEx version 8 (ref. 74)). Bonferroni correction was performed for all tested annotations and categories. The detailed methodologies for the stratified linkage disequilibrium score regression are presented in the original work70. The linkage disequilibrium scores and allele frequencies for the European ancestry were obtained from a predefined version based on data from the 1000 Genomes project.

annotations and categories. The detailed methodologies for the stratified linkage disequilibrium score regression are presented in the original work70. The linkage disequilibrium scores and allele frequencies for the European ancestry were obtained from a predefined version based on data from the 1000 Genomes project. We estimated the genetic correlation (gc) using LDSC75 software. We employed precomputed linkage disequilibrium scores from the 1000 Genomes of European ancestry, maintaining default settings for other parameters in LDSC. It is worth noting that LDSC corrects for sample overlap, ensuring an unbiased genetic correlation estimate76. Statistical significance was determined using Bonferroni correction. We constructed two-sample bidirectional Mendelian randomization by linking the seven MRIBAGs and 525 disease endpoints from FinnGen46 and PGC48. In total, two networks were established: MRIBAG2DE and DE2MRIBAG. The systematic quality-checking procedures to ensure unbiased exposure/outcome variable and instrumental variable selection are detailed below.

ional Mendelian randomization by linking the seven MRIBAGs and 525 disease endpoints from FinnGen46 and PGC48. In total, two networks were established: MRIBAG2DE and DE2MRIBAG. The systematic quality-checking procedures to ensure unbiased exposure/outcome variable and instrumental variable selection are detailed below. We used a two-sample Mendelian randomization approach implemented in the TwoSampleMR package77 to infer the causal relationships within these networks. We employed five distinct Mendelian randomization methods, including the inverse variance weighted (IVW) method, Egger, weighted median and simple mode and weighted mode estimators. The STROBE-MR Statement78 guided our analyses to increase transparency and reproducibility, encompassing the selection of exposure and outcome variables, reporting statistics and implementing sensitivity checks to identify potential violations of underlying assumptions. First, we performed an unbiased quality check on the GWAS summary statistics. Notably, the absence of population overlapping bias79 was confirmed, given that FinnGen and UKBB participants largely represent populations of European ancestry without explicit overlap with the UKBB. PGC GWAS summary data were ensured to exclude UKBB participants. Furthermore, the GWAS summary statistics from all consortia were based on or lifted to GRCh37. Subsequently, we selected the effective exposure variables by assessing the statistical power of the exposure GWAS summary statistics in terms of instrumental variables, ensuring that the number of instrumental variables exceeded seven before harmonizing the data. Crucially, the function ‘clump_data’ was applied to the exposure GWAS data, considering linkage disequilibrium. The function ‘harmonise_data’ was then used to harmonize the GWAS summary statistics of the exposure and outcome variables. Bonferroni correction was applied to all tested traits based on the number of effective disease endpoints.

n ‘clump_data’ was applied to the exposure GWAS data, considering linkage disequilibrium. The function ‘harmonise_data’ was then used to harmonize the GWAS summary statistics of the exposure and outcome variables. Bonferroni correction was applied to all tested traits based on the number of effective disease endpoints. Finally, we conducted multiple sensitivity analyses. First, we conducted a heterogeneity test to scrutinize potential violations of the instrumental variables’ assumptions. To assess horizontal pleiotropy, which indicates the instrumental variables’ exclusivity assumption80, we used a funnel plot, single-SNP Mendelian randomization methods and the Egger estimator. Furthermore, we performed a leave-one-out analysis, systematically excluding one instrument (SNP/instrumental variable) at a time, to gauge the sensitivity of the results to individual SNPs. PRS was computed using split-sample GWASs (split1 and split2) for the seven MRIBAG GWASs. The PRS weights were established using split1/discovery GWAS data as the base/training set, whereas the split2/replication GWAS summary statistics served as the target/testing data. Both base and target data underwent rigorous quality control procedures involving several steps: (1) excluding duplicated and ambiguous SNPs in the base data and (2) excluding high heterozygosity samples in the target data and (3) eliminating duplicated, mismatching and ambiguous SNPs in the target data.

/testing data. Both base and target data underwent rigorous quality control procedures involving several steps: (1) excluding duplicated and ambiguous SNPs in the base data and (2) excluding high heterozygosity samples in the target data and (3) eliminating duplicated, mismatching and ambiguous SNPs in the target data. After completing the quality control procedures, PRS for the split2 and split1 groups was calculated using the PRS-CS81 method. PRS-CS applies a continuous shrinkage prior, which adjusts the SNP effect sizes based on their linkage disequilibrium structure. SNPs with weaker evidence are ‘shrunk’ toward zero, whereas those with stronger evidence retain larger effect sizes. This avoids overfitting and improves prediction performance. The shrinkage parameter was not set, and the algorithm learned it via a fully Bayesian approach. We conducted multiple analyses to establish genetic evidence for prioritizing potential druggable genes in future drug repurposing. Our central hypothesis is that genes with functional implications and causal roles validated across multiple omics layers, such as genomics, transcriptomics and proteomics, are more likely to be actionable for drug repurposing, as they offer a stronger and more robust foundation for identifying therapeutic targets82,83.

ur central hypothesis is that genes with functional implications and causal roles validated across multiple omics layers, such as genomics, transcriptomics and proteomics, are more likely to be actionable for drug repurposing, as they offer a stronger and more robust foundation for identifying therapeutic targets82,83. First, we performed SNP-to-gene mapping via three approaches: positional mapping, eQTL mapping and chromatin interaction mapping. For eQTL and chromatin interaction mapping, we used data from various resources, including GTEx, PsychENCODE, EyeGEx, TIGER, DICE, eQTLGen, Blood eQTL, MuTHER, CommonMind Consortium, BRAINEAC, FANTOM5, the Brain Open Chromatin Atlas and the Roadmap 111 epigenomes, all consolidated through the FUMA platform. By requiring both eQTL and chromatin interaction support, we enhance confidence in selecting functionally relevant genes because eQTL mapping identifies genetic variants that regulate gene expression. Chromatin interaction mapping links distal regulatory elements (for example, enhancers) to target genes, offering spatial genomic context. Together, these criteria ensure that the prioritized genes are not only associated with GWAS loci but also have strong regulatory and functional evidence, making them more biologically plausible as druggable targets.

distal regulatory elements (for example, enhancers) to target genes, offering spatial genomic context. Together, these criteria ensure that the prioritized genes are not only associated with GWAS loci but also have strong regulatory and functional evidence, making them more biologically plausible as druggable targets. Subsequently, we connected MRIBAG, as an imaging-derived endophenotype, to aging clocks derived from plasma proteomics and metabolomics to further strengthen the genetic evidence for prioritizing these druggable genes. To achieve this, we used a Bayesian co-localization approach to assess whether the genomic loci associated with the seven MRIBAGs share common causal variants with the 11 ProtBAGs or the five MetBAGs. We used the R package ‘coloc’ to investigate the genetic co-localization signals between two traits (for example, the brain MRIBAG versus the brain ProtBAG) at each genomic locus associated with the MRIBAG. This method (‘coloc.abf’) examines the posterior probability to evaluate hypothesis H4 (PP.H4.ABF), which suggests the presence of a single shared causal variant associated with both traits within a specific genomic locus. To determine the significance of the H4 hypothesis, we set a threshold of PP.H4.ABF > 0.8 (ref. 84). All other parameters (for example, the prior probability of p12) were set as the default. For each pair of traits, the genomic locus was defined by default from FUMA for one trait, and then the ‘coloc’ package extracted and harmonized the GWAS summary statistics within this locus for the other trait.

F > 0.8 (ref. 84). All other parameters (for example, the prior probability of p12) were set as the default. For each pair of traits, the genomic locus was defined by default from FUMA for one trait, and then the ‘coloc’ package extracted and harmonized the GWAS summary statistics within this locus for the other trait. Finally, we searched the DGIdb platform (https://dgidb.org/) for genes that demonstrated both functional mapping evidence and causal associations to investigate their drug−gene interactions, revealing existing drugs and their therapeutic indications. We investigated the clinical promise of the seven MRIBAGs and their PRSs (MRIBAG-PRS) in three sets of prediction analyses: (1) survival analysis for the incidence of single disease entities based on the ICD-10 code, (2) survival analysis for the risk of all-cause mortality and (3) differential response to the Alzheimer’s disease drug (solanezumab).

omise of the seven MRIBAGs and their PRSs (MRIBAG-PRS) in three sets of prediction analyses: (1) survival analysis for the incidence of single disease entities based on the ICD-10 code, (2) survival analysis for the risk of all-cause mortality and (3) differential response to the Alzheimer’s disease drug (solanezumab). We employed a Cox proportional hazard model while adjusting for covariates (that is, age and sex) to test the associations of the seven MRIBAGs with the incidence of ICD-based single disease entities. The covariates were included as additional right-side variables in the model. To train the model, the ‘time’ variable was determined by calculating the difference between the date of death (Field ID: 40000) for cases (or the censoring date for non-cases) and the date attending the assessment center (Field ID: 53). Participants who were diagnosed for a specific disease of interest after enrolling in the study were classified as cases; non-cases were defined by participants without any disease diagnoses.

h (Field ID: 40000) for cases (or the censoring date for non-cases) and the date attending the assessment center (Field ID: 53). Participants who were diagnosed for a specific disease of interest after enrolling in the study were classified as cases; non-cases were defined by participants without any disease diagnoses. We employed a Cox proportional hazard model while adjusting for covariates (that is, age and sex) to test the associations of the seven MRIBAGs and seven MRIBAG-PRSs with all-cause mortality. The covariates were included as additional right-side variables in the model. The hazard ratio, exp(βR), was calculated and reported as the effect size measure that indicates the influence of each biomarker on the risk of mortality. To train the model, the ‘time’ variable was determined by calculating the difference between the date of death (Field ID: 40000) for cases (or the censoring date for non-cases) and the date attending the assessment center (Field ID: 53). Participants who died after enrolling in the study were classified as cases. We also conducted a disease-free survival analysis, including only participants without any diagnosed conditions, to eliminate potential confounding effects from disease pathology.

) and the date attending the assessment center (Field ID: 53). Participants who died after enrolling in the study were classified as cases. We also conducted a disease-free survival analysis, including only participants without any diagnosed conditions, to eliminate potential confounding effects from disease pathology. To evaluate our hypothesis that individuals with varying aging clock paces may progress differently to cognitive decline, we analyzed clinical trial data from the A4 (ref. 18) study. Specifically, we investigated whether brain MRIBAG could be used to stratify participants into decelerated and accelerated aging groups and demonstrated different cognitive profiles at 240 weeks by taking solanezumab. The original trial did not demonstrate cognitive decline slowing compared to the placebo over 240 weeks, with the treatment group even showing a slight worsening in the PACC score. To evaluate this, we performed four comparisons using natural cubic spline modeling: (1) decelerated aging (for example, below the brain MRIBAG median) versus accelerated aging (for example, above the median) within the drug group and (2) within the placebo group; (3) between drug and placebo groups among participants with accelerated aging; and (4) between drug and placebo groups among those with decelerated aging.

d aging (for example, below the brain MRIBAG median) versus accelerated aging (for example, above the median) within the drug group and (2) within the placebo group; (3) between drug and placebo groups among participants with accelerated aging; and (4) between drug and placebo groups among those with decelerated aging. From a statistical perspective, this is to model repeated measures (that is, PACC as the primary trial outcome for global cognition) as a continuous outcome in a mixed-effect model. We used the same method proposed by the original work from Donohue et al.85, in which the authors proposed a constrained longitudinal data analysis with natural cubic splines that treated time as continuous and used test version effects to model the mean over time for PACC. Fixed effects included the following terms: (1) spline basis expansion terms (two terms); (2) interaction of the spline basis expansion terms with treatment (two terms); (3) the version of the PACC test implemented; (4) baseline age; (5) education; (6) APOE4 carrier status (yes/no); and (7) baseline florbetapir cortical standardized uptake value ratio (SUVR) value. At week 240, we compared the mean PACC values between each pair of the two groups. Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

The UKBB43 is a population-based research initiative comprising approximtely 500,000 individuals from the United Kingdom between 2006 and 2010. Ethical approval for the UKBB study was secured, and information about the ethics committee can be found here: https://www.ukbiobank.ac.uk/learn-more-about-uk-biobank/governance/ethics-advisory-committee. This study collectively analyzed multi-organ MRI data from seven organ systems and tissues, including the brain, heart, liver, pancreas, spleen, adipose and kidney. For the genetic data, we conducted a quality check on the imputed genotype data43 for the entire UKBB population (approximately 500,000 individuals). Subsequently, we merged the processed data with the organ-specific populations for all genetic analyses. Finally, we also included Olink plasma proteomics data released by the UK Biobank Pharma Proteomics Project (UKB-PPP) and metabolomics data from the UKBB, which are detailed in the following sections.

The main goal of the BLSA is to understand the normal aging process. Tracking physiological and cognitive changes over time aims to identify risk factors for age-related diseases, study patterns of decline and discover predictors of healthy aging. BLSA44,45 brain MRI and SomaScan proteomics data (https://www.blsa.nih.gov/) were used to compare and replicate the ProWAS results from the UKBB Olink data. After quality checks in this study, we included 1,114 brain MRI scans at baseline and measurements of 7,268 plasma proteins from 909 participants quantified with the SomaScan version 4.1 platform. Age (years), sex (male/female), race (white/non-white) and education level (years) were defined based on participant self-reports.

The A4 study18,19 (https://atri.usc.edu/study/a4-study/) is a clinical trial study to test a specific way to prevent memory loss associated with Alzheimer’s disease (clinical trial number: NCT02008357). The A4 study focused on symptom-free adults at higher risk for Alzheimer’s disease to assess whether an investigational drug (that is, solanezumab) could slow memory decline linked to amyloid plaques in the brain. It also examined whether solanezumab could delay Alzheimer’s disease progression, measuring related brain changes using imaging, blood biomarkers and baseline PET scans to assess amyloid levels. This study analyzed 1,055 participants at baseline with brain MRI scans to derive the brain MRIBAG. Longitudinal outcomes from the clinical trial, with the PACC score as the primary measure over 312 weeks, were included. The PACC scores between groups were evaluated at week 240.

The FinnGen46 study is a large-scale genomics initiative that has analyzed more than 500,000 Finnish biobank samples and correlated genetic variation with health data to understand disease mechanisms and predispositions. The project is a collaboration between research organizations and biobanks within Finland and international industry partners. For the benefit of research, FinnGen generously made their GWAS findings accessible to the wider scientific community (https://www.finngen.fi/en/access_results). This research used the publicly released GWAS summary statistics (version R9), which became available on 11 May 2022, after harmonization by the consortium. No individual data were used in the current study. FinnGen published the R9 version of GWAS summary statistics via REGENIE software (version 2.2.4)47, covering 2,272 disease endpoints, including 2,269 binary traits and three quantitative traits. The GWAS model encompassed covariates such as age, sex, the initial 10 genetic principal components and the genotyping batch. Genotype imputation was referenced on the population-specific SISu version 4.0 panel. We included GWAS summary statistics for 521 FinnGen disease endpoints in our analyses.

PGC48 is an international collaboration of researchers studying the genetic basis of psychiatric disorders. PGC aims to identify and understand the genetic factors contributing to various psychiatric disorders, such as schizophrenia, bipolar disorder, major depressive disorder and others. The GWAS summary statistics were acquired from the PGC website (https://pgc.unc.edu/for-researchers/download-results/), underwent quality checks and were harmonized to ensure seamless integration into our analysis. No individual data were used from PGC. Each study detailed its specific GWAS models and methodologies, and the consortium consolidated the release of GWAS summary statistics derived from individual studies. In the present study, we included summary data for four brain diseases for which allele frequencies were present.

In our previous analysis of raw brain MRI data from the UKBB and other studies included in this work, we extracted 119 gray matter regions of interest from T1-weighted MRI to generate brain MRIBAGs through the iSTAGING consortium49. For the heart MRIBAG, we used 80 heart MRI traits from Bai et al.50 and used these imaging features in a previous study9. To develop the remaining five MRIBAGs based on abdominal MRI (Category ID: 105 (refs. 51–56)), we incorporated 16 imaging features for the adipose MRIBAG, four for the liver MRIBAG, three for the kidney MRIBAG, three for the spleen MRIBAG and three for the pancreas MRIBAG. Notably, in the development of adipose and kidney MRIBAGs, we observed a high degree of collinearity among certain imaging features, which led to overfitting. To mitigate this issue, we removed highly correlated features to improve model robustness.

The abdominal MRI data underwent initial quality control procedures prior to public release, including the removal of low-quality images and biologically implausible values. As noted on the UKBB website (https://biobank.ndph.ox.ac.uk/ukb/label.cgi?id=158), different pipelines were used to generate imaging features from abdominal MRI, and data should not be combined without careful consideration. To account for this, we ensured that each abdominal feature was derived from a consistent pipeline, selecting the one with the largest available sample size across different pipelines. For liver MRI metrics specifically, we included only images acquired using the IDEAL protocol (Data-Field 40063, acquisition protocol 2). Following these additional quality control steps, our analyses incorporated abdominal MRI biomarkers from all available UKBB participants for the liver, kidney, spleen, pancreas and adipose MRIBAGs.

Our previous study systematically evaluated age prediction performance across various AI/ML models using multimodal brain MRI features14 for the brain MRIBAG. Applying the same framework, we assessed the performance of models in deriving the six additional MRIBAGs using several ML methods. Hyperparameter tuning was performed through nested, repeated holdout cross-validation57 with 50 repetitions (80% training/validation and 20% testing). Specifically, the hyperparameters of each model were tuned using grid search: LASSO regression’s alpha (α), linear support vector regressor’s C and both α and L1-ratio for the elastic net. Due to the flexibility and large hyperparameter space of neural networks (for example, number of layers and neurons), they were not included in the nested cross-validation pipeline. Instead, the neural network was manually tuned and evaluated on the cross-validation test dataset, and the final model was subsequently assessed on the within-distribution, holdout test set. The within-distribution, holdout test dataset was held out to unbiasedly evaluate model performance (Supplementary Fig. 31).

We used the original dataset (Category ID: 1838), which was analyzed and shared with the research community by the UKB-PPP58. The initial quality control procedures were described in the original study59, and we implemented additional quality control steps as outlined below. Our analysis focused on the first instance of the proteomics data (‘instance’ = 0). We then integrated Olink files containing coding information, batch numbers, assay details and limit of detection (LOD) data (Category ID: 1839) by matching them to the proteomics dataset ID. Finally, we excluded Normalized Protein eXpression (NPX) values that fell below the protein-specific LOD.

We used the original data (Category ID: 220), which were analyzed and made available to the community by Nightingale Health Plc. The original data (1) were calibrated absolute concentrations (or ratios) and not raw NMR spectra and (2), before release, had already been subjected to quality control procedures by Nightingale Health Plc60. Following the additional procedures described in Ritchie et al.61, we performed additional quality check steps to remove a range of unwanted technical variations, including shipping batch, 96-well plate, well position, aliquoting robo and aliquot tip. We focused our analysis on the first instance of the metabolomics data (‘instance’ = 0). The analysis included 327 metabolites (comprising both small molecules and lipoprotein measures), of which 107 were non-derived metabolites and the remainder were composite metabolites, across 274,247 participants. Descriptions of these metabolites are provided in Supplementary Table 3.

We used the imputed genotype data from the UKBB for all genetic analyses. Our quality check pipeline focused on European ancestry in the UKBB (6,477,810 SNPs passing quality checks), and the quality-checked genetic data were merged with respective organ-specific populations for GWAS. We summarize our genetic quality check steps. First, we skipped the step for family relationship inference62 because the linear mixed model via fastGWA63 inherently addresses population stratification, encompassing additional cryptic population stratification factors. We then removed duplicated variants from all 22 autosomal chromosomes. Individuals whose genetically identified sex did not match their self-acknowledged sex were removed. Other excluding criteria included (1) individuals with more than 3% of missing genotypes; (2) variants with minor allele frequency (MAF; dosage mode) of less than 1%; (3) variants with more than 3% missing genotyping rate; and (4) variants that failed the Hardy−Weinberg test at 1 × 10−10. To further adjust for population stratification,64 we derived the first 40 genetic principal components using FlashPCA software65. Details of the genetic quality check protocol are described elsewhere11,14,34,49,66,67.

F > 0.8 (ref. 84). All other parameters (for example, the prior probability of p12) were set as the default. For each pair of traits, the genomic locus was defined by default from FUMA for one trait, and then the ‘coloc’ package extracted and harmonized the GWAS summary statistics within this locus for the other trait. Finally, we searched the DGIdb platform (https://dgidb.org/) for genes that demonstrated both functional mapping evidence and causal associations to investigate their drug−gene interactions, revealing existing drugs and their therapeutic indications.

We applied a linear mixed model regression to the European ancestry populations using fastGWA63 implemented in GCTA21. We used fastGWA to perform the seven MRIBAG GWASs, adjusting common variates, including age, dataset status (training/validation/test or within-distribution, holdout test), age-squared, sex, interactions of age with sex, BMI, waist circumference, standing height, weight and the first 40 genetic principal components, as well as organ-specific covariates, including the brain scan positions for the brain MRIBAG and systolic/diastolic blood pressure for the heart MRIBAG. We applied a genome-wide significance threshold (5 × 10−8) to annotate the significant independent genomic loci. We previously conducted GWAS of 2,923 plasma proteins and 327 metabolites using fastGWA5,68.

For all GWASs, genomic loci were annotated using FUMA69. For genomic loci annotation, FUMA initially identified lead SNPs (correlation r2 ≤ 0.1, distance < 250 kb) and assigned them to non-overlapping genomic loci. The lead SNP with the lowest P value (that is, the top lead SNP) represented the genomic locus. Further details on the definitions of top lead SNP, lead SNP, independent significant SNP and candidate SNP can be found in the FUMA documentation (https://fuma.ctglab.nl/). For visualization purposes in Fig. 3a, we mapped the top lead SNP of each locus to the cytogenetic regions based on the GRCh37 cytoband.

We estimated the SNP-based heritability (h2) using GCTA21 with the same covariates as in GWAS. GCTA estimates the SNP-based heritability using a method called restricted maximum likelihood (REML) to quantify the proportion of phenotypic variance in a trait that the additive effects of all common SNPs can explain. The main steps involved in GCTA include constructing the genetic relationship matrix, modeling phenotypic variance and using REML to estimate the h2.

The partitioned heritability analysis via stratified linkage disequilibrium score regression calculates the extent to which heritability enrichment can be attributed to predefined and annotated genome regions and categories70. Three sets of functional categories and cell-specific and tissue-specific types were considered. First, the partitioned heritability was calculated for 53 general functional categories (one including the entire set of SNPs). The 53 functional categories are not specific to any cell type and include coding regions, untranslated regions, promoter regions and intronic regions. The details of the 53 categories are described elsewhere70. Subsequently, cell-type-specific and tissue-type-specific partitioned heritability was estimated using gene sets from Cahoy et al.71 for three main cell types (that is, astrocyte, neuron and oligodendrocyte), multi-tissue chromatin state-specific data (Roadmap72 and ENTEx73) and multi-tissue gene expression data (GTEx version 8 (ref. 74)). Bonferroni correction was performed for all tested annotations and categories. The detailed methodologies for the stratified linkage disequilibrium score regression are presented in the original work70. The linkage disequilibrium scores and allele frequencies for the European ancestry were obtained from a predefined version based on data from the 1000 Genomes project.

We estimated the genetic correlation (gc) using LDSC75 software. We employed precomputed linkage disequilibrium scores from the 1000 Genomes of European ancestry, maintaining default settings for other parameters in LDSC. It is worth noting that LDSC corrects for sample overlap, ensuring an unbiased genetic correlation estimate76. Statistical significance was determined using Bonferroni correction.

We constructed two-sample bidirectional Mendelian randomization by linking the seven MRIBAGs and 525 disease endpoints from FinnGen46 and PGC48. In total, two networks were established: MRIBAG2DE and DE2MRIBAG. The systematic quality-checking procedures to ensure unbiased exposure/outcome variable and instrumental variable selection are detailed below.

n ‘clump_data’ was applied to the exposure GWAS data, considering linkage disequilibrium. The function ‘harmonise_data’ was then used to harmonize the GWAS summary statistics of the exposure and outcome variables. Bonferroni correction was applied to all tested traits based on the number of effective disease endpoints. Finally, we conducted multiple sensitivity analyses. First, we conducted a heterogeneity test to scrutinize potential violations of the instrumental variables’ assumptions. To assess horizontal pleiotropy, which indicates the instrumental variables’ exclusivity assumption80, we used a funnel plot, single-SNP Mendelian randomization methods and the Egger estimator. Furthermore, we performed a leave-one-out analysis, systematically excluding one instrument (SNP/instrumental variable) at a time, to gauge the sensitivity of the results to individual SNPs.

PRS was computed using split-sample GWASs (split1 and split2) for the seven MRIBAG GWASs. The PRS weights were established using split1/discovery GWAS data as the base/training set, whereas the split2/replication GWAS summary statistics served as the target/testing data. Both base and target data underwent rigorous quality control procedures involving several steps: (1) excluding duplicated and ambiguous SNPs in the base data and (2) excluding high heterozygosity samples in the target data and (3) eliminating duplicated, mismatching and ambiguous SNPs in the target data. After completing the quality control procedures, PRS for the split2 and split1 groups was calculated using the PRS-CS81 method. PRS-CS applies a continuous shrinkage prior, which adjusts the SNP effect sizes based on their linkage disequilibrium structure. SNPs with weaker evidence are ‘shrunk’ toward zero, whereas those with stronger evidence retain larger effect sizes. This avoids overfitting and improves prediction performance. The shrinkage parameter was not set, and the algorithm learned it via a fully Bayesian approach.

We conducted multiple analyses to establish genetic evidence for prioritizing potential druggable genes in future drug repurposing. Our central hypothesis is that genes with functional implications and causal roles validated across multiple omics layers, such as genomics, transcriptomics and proteomics, are more likely to be actionable for drug repurposing, as they offer a stronger and more robust foundation for identifying therapeutic targets82,83. First, we performed SNP-to-gene mapping via three approaches: positional mapping, eQTL mapping and chromatin interaction mapping. For eQTL and chromatin interaction mapping, we used data from various resources, including GTEx, PsychENCODE, EyeGEx, TIGER, DICE, eQTLGen, Blood eQTL, MuTHER, CommonMind Consortium, BRAINEAC, FANTOM5, the Brain Open Chromatin Atlas and the Roadmap 111 epigenomes, all consolidated through the FUMA platform. By requiring both eQTL and chromatin interaction support, we enhance confidence in selecting functionally relevant genes because eQTL mapping identifies genetic variants that regulate gene expression. Chromatin interaction mapping links distal regulatory elements (for example, enhancers) to target genes, offering spatial genomic context. Together, these criteria ensure that the prioritized genes are not only associated with GWAS loci but also have strong regulatory and functional evidence, making them more biologically plausible as druggable targets.

We investigated the clinical promise of the seven MRIBAGs and their PRSs (MRIBAG-PRS) in three sets of prediction analyses: (1) survival analysis for the incidence of single disease entities based on the ICD-10 code, (2) survival analysis for the risk of all-cause mortality and (3) differential response to the Alzheimer’s disease drug (solanezumab). We employed a Cox proportional hazard model while adjusting for covariates (that is, age and sex) to test the associations of the seven MRIBAGs with the incidence of ICD-based single disease entities. The covariates were included as additional right-side variables in the model. To train the model, the ‘time’ variable was determined by calculating the difference between the date of death (Field ID: 40000) for cases (or the censoring date for non-cases) and the date attending the assessment center (Field ID: 53). Participants who were diagnosed for a specific disease of interest after enrolling in the study were classified as cases; non-cases were defined by participants without any disease diagnoses.

d aging (for example, below the brain MRIBAG median) versus accelerated aging (for example, above the median) within the drug group and (2) within the placebo group; (3) between drug and placebo groups among participants with accelerated aging; and (4) between drug and placebo groups among those with decelerated aging. From a statistical perspective, this is to model repeated measures (that is, PACC as the primary trial outcome for global cognition) as a continuous outcome in a mixed-effect model. We used the same method proposed by the original work from Donohue et al.85, in which the authors proposed a constrained longitudinal data analysis with natural cubic splines that treated time as continuous and used test version effects to model the mean over time for PACC. Fixed effects included the following terms: (1) spline basis expansion terms (two terms); (2) interaction of the spline basis expansion terms with treatment (two terms); (3) the version of the PACC test implemented; (4) baseline age; (5) education; (6) APOE4 carrier status (yes/no); and (7) baseline florbetapir cortical standardized uptake value ratio (SUVR) value. At week 240, we compared the mean PACC values between each pair of the two groups.

We employed a Cox proportional hazard model while adjusting for covariates (that is, age and sex) to test the associations of the seven MRIBAGs with the incidence of ICD-based single disease entities. The covariates were included as additional right-side variables in the model. To train the model, the ‘time’ variable was determined by calculating the difference between the date of death (Field ID: 40000) for cases (or the censoring date for non-cases) and the date attending the assessment center (Field ID: 53). Participants who were diagnosed for a specific disease of interest after enrolling in the study were classified as cases; non-cases were defined by participants without any disease diagnoses.

We employed a Cox proportional hazard model while adjusting for covariates (that is, age and sex) to test the associations of the seven MRIBAGs and seven MRIBAG-PRSs with all-cause mortality. The covariates were included as additional right-side variables in the model. The hazard ratio, exp(βR), was calculated and reported as the effect size measure that indicates the influence of each biomarker on the risk of mortality. To train the model, the ‘time’ variable was determined by calculating the difference between the date of death (Field ID: 40000) for cases (or the censoring date for non-cases) and the date attending the assessment center (Field ID: 53). Participants who died after enrolling in the study were classified as cases. We also conducted a disease-free survival analysis, including only participants without any diagnosed conditions, to eliminate potential confounding effects from disease pathology.

To evaluate our hypothesis that individuals with varying aging clock paces may progress differently to cognitive decline, we analyzed clinical trial data from the A4 (ref. 18) study. Specifically, we investigated whether brain MRIBAG could be used to stratify participants into decelerated and accelerated aging groups and demonstrated different cognitive profiles at 240 weeks by taking solanezumab. The original trial did not demonstrate cognitive decline slowing compared to the placebo over 240 weeks, with the treatment group even showing a slight worsening in the PACC score. To evaluate this, we performed four comparisons using natural cubic spline modeling: (1) decelerated aging (for example, below the brain MRIBAG median) versus accelerated aging (for example, above the median) within the drug group and (2) within the placebo group; (3) between drug and placebo groups among participants with accelerated aging; and (4) between drug and placebo groups among those with decelerated aging.

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41591-025-03999-8.

Supplementary InformationAll supplementary figures, notes and tables are included. Reporting Summary Supplementary Data 1This is study-specific code to produce the seven MRIBAGs using MLNI software. Supplementary Data 2Supplementary Data, eFiles 1−9. Supplementary Data 3The pre-trained AI model was also shared via our public MEDICINE portal. Supplementary Data 4MULTI Consortium author list. All supplementary figures, notes and tables are included. Reporting Summary This is study-specific code to produce the seven MRIBAGs using MLNI software. Supplementary Data, eFiles 1−9. The pre-trained AI model was also shared via our public MEDICINE portal. MULTI Consortium author list.