<p>Infantile Neuroaxonal Dystrophy (INAD) is an extremely rare and paralyzing neurodegenerative disease that usually begins in the initial two years of life with acute regression of motor, cerebellar atrophy, hypotonia, and progressive loss of intellect. Traditional diagnostic devices do not combine the multiplicity of biological signals necessary for accurate classification and disease tracking. This work presents BioGAT INAD, a graph deep learning model that combines multi omics data (genomics, proteomics, neuroimaging (MRI), electrophysiological (EEG), clinical symptoms) into a single, interpretable model for early classification and prognosis prediction of INAD. The model builds a heterogeneous biological graph, in which the nodes are biologically meaningful entities and the edges represent curated interactions from clinical databases. A Graph Attention Network (GAT) modulates attention over the biological interactions dynamically, improving the domain specific feature relevance for INAD. To effectively train the model on small, noisy biomedical data characteristic of rare diseases, we use the Ranger optimizer, incorporating RAdam’s adaptive variance rectification with Look ahead’s stable weight interpolation. This allows it to achieve strong convergence and generalization in spite of the small size and noisiness of INAD concerned data. Additionally, the model uses a temporal graph module for disease trajectory prediction over time based on sequential omics and clinical data. Empirical results show that BioGAT INAD far outperforms standard classifiers and deep learning baselines in both early classification and long term disease prediction tasks.</p>