In this study, we analysed plasma proteomics of early-mid stage AD cases (15 ≤ MMSE ≤ 30, age ≥ 65) cross-sectionally to identify molecular signatures that are not detectable with standard cognitive testing. For this, we used a graph-smoothed autoencoder with a STRING protein-protein interaction prior to derive deep-learning based latent representations followed by clustering. It revealed an Endolysosomal–Translation bipolar axis (a bidirectional axis of coordinated endolysosomal and translational protein changes) within a disease continuum. The proposed framework supports mechanism-based stratification for precision medicine and trial enrichment using minimally invasive methods.