Balanced deep learning on multi-omics networks identifies molecular subgroups of pathological brain aging.

Background: Neurodegenerative diseases, including Alzheimer's disease (AD), exhibit substantial clinical and molecular heterogeneity, complicating accurate diagnosis and development of effective therapies. Although multi-omics profiling provides unprecedented molecular resolution, systematic integration of high-dimensional, imbalanced data modalities with disease-relevant biological networks remains a methodological challenge. Methods: We developed a network-informed multi-omics integration framework that combines data-driven molecular networks with brain transcriptomic, proteomic, and metabolomic data from 356 participants in the Religious Orders Study and Rush Memory and Aging Project (ROS/MAP). Utilizing 25 functional, data-driven multi-omics groups (DAD-MUGs) derived by graph embedding from the AD Atlas, co-expression-guided feature extraction and systematic two-phase feature balancing were applied to derive representative molecular features, which were subsequently learned using DAD-MUG-specific autoencoders to generate compact multi-omics expression scores. These were then used to identify molecular subgroups via hierarchical clustering. Subgroup robustness was assessed in an independent ROS/MAP cohort (n=327) using a two-round nested classification strategy. Results: Subgroup identification based on DAD-MUG-derived expression scores resulted in five molecular subgroups exhibiting significant differences in cognitive performance and core neuropathological measures. Cross-validated nested classification using transcriptomic and proteomic data demonstrated reliable discrimination of subgroups. Applying these classifiers to the replication cohort, subgroup-trait association patterns showed strong agreement with discovery findings (Spearman ρ = 0.65). Differential expression analysis further revealed stage-dependent biological patterns of brain pathologies, ranging from early synaptic and immune activation to mitochondrial bioenergetic dysfunction at disease transition and proteostatic impairment in advanced stages. Conclusion: Using a balanced, network-informed multi-omics integration framework, we identified five molecular subgroups of brain aging, including a reference control subgroup and a distinct mixed subgroup characterized by amyloid, vascular pathology, and early-life adversity. Three additional subgroups formed a structured spectrum comprising molecularly Alzheimer's-like but cognitively and neuropathologically unimpaired At-risk controls, an intermediate stage, and typical Alzheimer's disease, with tau pathology differentiating advanced disease, underscoring the value of molecular subgroup identification beyond clinical diagnosis.