MRI-based multi-organ clocks for healthy aging and disease assessment

Cross-organ biological aging clocks deepen our understanding of aging and disease. Building on that framework, we construct seven MRI-based multi-organ biological age gaps (MRIBAGs) covering brain, heart, liver, adipose, spleen, kidney, and pancreas. In 313,645 individuals integrated under MULTI, we connect the seven MRIBAGs to 2,923 plasma proteins, 327 metabolites, and ~6.47 million common genetic variants. Genome-wide association analysis yields 53 significant locus–MRIBAG pairs at P < 5×10⁻⁸. Genetic correlation and Mendelian randomization support organ-specific and cross-organ links involving 24 non-MRI aging clocks and 525 disease endpoints. Functional mapping plus Bayesian colocalization with proteomic/metabolic evidence prioritizes nine druggable genes as candidate targets for future anti-aging interventions. The seven MRIBAGs also relate to future systemic diseases (e.g., diabetes) and all-cause mortality. Finally, in the solanezumab Alzheimer’s trial (240 weeks), baseline brain MRI profiles that look “younger” vs “older” track different cognitive decline trajectories, but this heterogeneity cannot be fully attributed to drug assignment. Together, we develop seven MRIBAGs that strengthen the multi-organ aging framework and illustrate clinical research potential.

MRI non-invasively captures structural change with age. “Brain age” models are widely used, but systematic extension across other organs is rarer. Large-scale multi-organ MRI in UK Biobank plus MULTI Consortium integration makes it feasible to add organ-specific biological age gaps alongside proteomic (ProtBAG), metabolic (MetBAG), phenotypic (PhenoBAG), epigenetic, and other clocks—and to stitch a coherent aging–disease map.

• MULTI cohort (Methods 1): multi-organ MRI for seven systems/tissues, genotypes, UKB-PPP proteomics, Nightingale metabolomics; downstream meta-analyses with FinnGen, PGC, etc. Ethics span Columbia IRB (MULTI), UK Biobank, BLSA, A4 trial sites, and more.
• MRIBAG construction (Methods 2): refresh brain MRIBAG (119 gray-matter volumes, nested CV, 500-person holdout) and add six new MRIBAGs (heart, liver, adipose, spleen, kidney, pancreas); models include LASSO, linear SVR, elastic net, tuned neural nets, with attention to abdominal collinearity/overfitting. Biological age gap is predicted age minus chronological age; downstream analyses apply age-residualization.
• ProWAS / MetWAS (Methods 3): linear models linking each MRIBAG to 2,923 proteins and 327 metabolites with covariates for demographics, batches, LOD, organ-specific imaging covariates, 40 PCs, etc.; Bonferroni at P < 0.05/(2923×7) and P < 0.05/(327×7); interactive volcano portal.
• Genetics (Methods 4): European-ancestry fastGWA (~6.48M SNPs post-QC), FUMA annotation, GCTA SNP heritability, LDSC stratified heritability, genetic correlation with PhenoBAG/ProtBAG/MetBAG and 525 disease endpoints, bidirectional two-sample MR (TwoSampleMR), split-sample PRS-CS, gene mapping (positional, eQTL, chromatin interactions), Bayesian colocalization with ProtBAG/MetBAG, plus DGIdb mining for drug–gene repurposing hypotheses.
• Clinical prediction (Methods 5): Cox models for de novo ICD-10 diseases and all-cause mortality; in A4, stratify by baseline brain MRIBAG (accelerated vs decelerated aging) and model PACC to week 240 with spline mixed models consistent with the original trial.

On healthy controls, best per-organ models achieve Pearson r between predicted and chronological age ~0.23–0.77 on holdout data, with mean absolute error on the order of five years before residualization. Brain MRIBAG has additional external validation (including domain shift discussion for A4). Abdominal MRI phenotypes remain sparser and more collinear than brain—limited extrapolation; the paper discusses denser protocols and voxel-level deep models as future work.

ProWAS yields 603 protein–MRIBAG associations after multiple testing; kidney MRIBAG has the most hits (e.g., NPDC1, IGFBP6, TAFA5), followed by spleen, liver, adipose, pancreas, brain, heart (example: REN). Organ-enriched proteins such as splenic VCAM1 and pancreatic PLA2G1B (also tied to adipose/spleen) illustrate tissue-specific aging biology (Human Protein Atlas context).

MetWAS reports 758 metabolite–MRIBAG links (many |r| < 0.3). Spleen and adipose MRIBAGs show the richest metabolic fingerprints (e.g., acetate, HDL subclasses); creatinine correlates positively with kidney MRIBAG and negatively with adipose MRIBAG, hinting at distinct regulatory roles. STRING pathway enrichments and sex-stratified analyses appear in supplements.

GWAS identifies 53 genome-wide significant locus–trait pairs. SNP heritability h² per MRIBAG is roughly 0.29–0.47. Stratified LDSC shows chromatin enrichments consistent with organ context (e.g., heart MRIBAG with cardiac H3K4me3/H3K27ac).

Genetic correlation links MRIBAGs to other multi-omic clocks and FinnGen/PGC endpoints (e.g., heart MRIBAG with hypertension). Bidirectional MR highlights paths that need cautious interpretation (e.g., hypertension → heart MRIBAG; AD → liver MRIBAG; bidirectional kidney MRIBAG and type 2 diabetes). Split-sample PRS explains limited incremental variance (brain PRS up to ~2.18% incremental R² on the target split).

Post-GWAS mapping plus colocalization with ProtBAG/MetBAG (PP.H4 > 0.8) supports 62 unique genes; DGIdb highlights nine genes with existing drug interactions (122 drugs), spanning immunosuppressants, antidiabetics, antineoplastics, etc. ALDH2 (spleen MRIBAG) exemplifies broad therapeutic annotations paired with AlphaFold structure discussion.

Cox models tie MRIBAGs and MRIBAG-PRSs to de novo ICD-10 diseases and all-cause mortality: e.g., brain/adipose/pancreas MRIBAGs and pancreas PRS with non–insulin-dependent diabetes; brain MRIBAG with substance use, anxiety, GI bleeding; heart MRIBAG with hypertension. Brain and adipose MRIBAGs skew toward risk for mortality, whereas liver and spleen MRIBAGs show protective-looking hazard patterns (interpreted cautiously in-text, including disease-free sensitivity analyses).

In A4, splitting baseline brain MRIBAG into “rejuvenated” vs “aged” profiles yields distinct PACC trajectories to week 240 in both drug and placebo arms; separation is stronger in placebo (permutation one-sided P≈0.03). At week 240 there is no significant drug–placebo difference within either stratum—cognitive heterogeneity therefore cannot be fully attributed to the investigational therapy.

MRIBAGs are positioned as imaging endotypes along genetic–multi-omic–disease axes, emphasizing both organ-specific and shared structure. Limitations include UK Biobank’s cross-sectional frame, less mature abdominal IDPs vs brain, limited environmental covariates, predominantly European GWAS, and biological ambiguity around “protective” liver/spleen associations—replication and mechanistic follow-up are needed. Future directions: longitudinal imaging, multi-ancestry GWAS, denser abdominal protocols, and tau/amyloid PET–informed stratification.

• MEDICINE knowledge portal — GWAS summary statistics and pretrained ML models
• Synapse (syn64923248)
• Zenodo — figures, extended data, MRIBAG derivation code, supplements
• Interactive browsers: ProWAS, MetWAS, Drug–gene (DGIdb)
• Software cited includes MLNI (MRIBAG generation), FUMA, GCTA/fastGWA, LDSC, TwoSampleMR, PRS-CS, lifelines, STRING, coloc—versions on Zenodo README.

Export the official reference from the Nature Medicine article page (open access). Short form:

The MULTI Consortium et al. MRI-based multi-organ clocks for healthy aging and disease assessment. Nature Medicine 32, 82–92 (2026). Open access; online first 16 Oct 2025. https://www.nature.com/articles/s41591-025-03999-8