An Explainable MRI-Conditioned Generative Framework for Gene-Feature Representations in Alzheimer’s Disease

Magnetic resonance imaging (MRI) and peripheral-blood transcriptomics capture complementary aspects of Alzheimer’s disease (AD), but whether MRI can condition a meaningful molecular representation remains unclear. We developed and independently evaluated an explainable imaging-genetics conditional variational autoencoder (IG-CVAE) in Alzheimer’s Disease Neuroimaging Initiative (ADNI) and AddNeuroMed (ANM) cohorts using fixed 80% train-validation and 20% independent test sets with five-fold cross-validation. IG-CVAE comprises a genetic branch encoding expression into 128-dimensional gene-feature vectors (GFV128), an imaging branch encoding MRI into 128-dimensional MRI-feature vectors (MFV128), and a conditional variational autoencoder that generates GFV128 from MFV128. Paired analyses included 289 ADNI and 98 ANM participants. Test area under the receiver operating characteristic curve (ROC-AUC) values in ADNI/ANM were 0.646/0.851 for the genetic branch, 0.942/0.752 for the imaging branch, and 0.939/0.869 for generated GFV128. Generated-observed GFV128 agreement was moderate (cosine similarity, 0.590/0.589; coefficient of determination (R2), 0.161/0.211). Molecular importance was cohort-specific at the gene level, with no significant pathway-level convergence, whereas MRI perturbation showed cross-cohort convergence in cerebral white matter linked to distinct disease-informative molecular features. Overall, structural MRI retained disease-relevant information sufficient to generate disease-informative molecular representations, supporting IG-CVAE as a participant-paired hypothesis-generating imaging–transcriptomic framework while providing only limited evidence of participant-specific molecular correspondence without implying raw transcriptome reconstruction. The ANM test estimates should be interpreted cautiously given the small paired independent test set (n = 20).