Multi-organ AI endophenotypes chart the heterogeneity of brain, eye and heart pan-disease.

Disease heterogeneity within single disease entities, cross-disease commonalities, and etiologic overlap present significant challenges for precision medicine. In the case of single disease entities, artificial intelligence (AI) has been applied to brain magnetic resonance imaging (MRI) data, which revealed distinct disease subtypes or dimensions, highlighting the neuroanatomical heterogeneity of brain disorders, such as Alzheimer’s disease (AD). On the other hand, recent findings from genetics and transcriptomics have unraveled overlapping molecular and neuropathological signatures across brain disorders, highlighting their shared biological underpinnings. We argue that AI holds great potential to jointly model human aging and disease concurrently from the two abovementioned perspectives to advance precision medicine. Addressing disease heterogeneity and transcending traditional classifications, such as those based on International Classification of Diseases (ICD) codes, may reveal new insights into aging and disease. Recent initiatives have focused on identifying transdiagnostic disease subtypes and dimensions, particularly within brain disorders like psychiatric (depression and psychosis) and developmental conditions. This has provided new avenues for understanding polygenic and etiologically multi-faceted diseases. In our recent study, by leveraging a weakly- supervised learning framework (e.g., Surreal-GAN), we derived 9 AI-derived endophenotypes to capture the within-disease heterogeneity and cross-disease similarities in 4 brain diseases, including autism spectrum disorder, schizophrenia, late-life depression, and AD, as well as aging. Critically, the 9 AI-derived phenotypes were generated within a specific disease entity, such as AD. The observed neuroanatomical overlap underscores the need for new approaches that derive disease subtypes or dimensions that transcend traditional disease boundaries, enabling a more integrated understanding of disease mechanisms. Modeling human aging and disease necessitates a comprehensive, multi-scale approach spanning both spatial and temporal granularities. Recent advances in multi-organ research have opened up new possibilities for holistic modeling of human aging and disease. For example, Tian et al. used machine learning (ML) to calculate the biological age gap (BAG) in nine organ systems, linking these biomarkers to lifestyle and mortality in the UK Biobank (UKBB). In subsequent analyses, Wen et al. further investigated the genetic architecture of these multi-organ BAGs. Similarly, multi-omics approaches, such as combining brain MRI data with genetics and proteomics, offer enhanced diagnostic precision and granularity. For instance, Yang et al. recently showed that incorporating genetic data into imaging-based models (e.g., Gene-SGAN) enhances disease subtyping accuracy and outperforms approaches developed solely on imaging data (e.g., Smile-GAN). In another study, we established the brain-heart-eye axis using imaging-derived phenotypes (IDPs) from imaging, genetic, and proteomic data of these 3 organs, demonstrating that no organ system is an island. Taken together, this evidence underscores the importance of integrating multi-organ, multi-omics data across multiple human organ systems and omics data types to model disease heterogeneity.