Predicting organ age with multimodal data identifies ageing-associated plasma proteins and behavioural traits

Ageing progresses unevenly across organs, making it difficult to capture organ-specific physiological decline using a single organism-level measure.

We developed a multimodal Mixture-of-Experts framework integrating MRI and anthropometric measurements to estimate brain, heart, liver and pancreas age in 46 734 individuals. Organ age gap (OAG) was defined as predicted age minus chronological age. UK Biobank proteomic, behavioural, disease and phenotypic data were used to characterise OAGs.

Predicted and chronological age were strongly correlated across organs (average Pearson correlation coefficient [PCC] 0.861 ± 0.0245). OAGs were organ-specific yet clinically informative across 20 age-related diseases and 16 age-related phenotypes and improved prediction for both organ-matched and cross-organ outcomes. Comparison with plasma-protein-based proteomic age gap showed only weak correlations, supporting partial overlap rather than redundancy and integrated age models improved chronological-age prediction without consistently outperforming organ-specific OAG in downstream disease prediction. Plasma proteomic analyses identified 54-154 OAG-associated proteins per organ, including nine shared across all four organs that were also associated with all-cause mortality. Temporal analyses showed structured but non-uniform proteomic changes, with relative enrichment between ages 55 and 70. Behavioural analyses highlighted robust smoking-related associations across organs, together with more selective associations involving alcohol intake, sleep duration and physical activity.

This framework quantifies organ-specific age and supports systematic interpretation of OAGs across clinical, molecular, temporal and behavioural dimensions. Multimodal OAG and proteomic age gap capture partially overlapping but complementary ageing information.