Therapeutic treatments for late-onset Alzheimer's disease (LOAD) are hindered by an incomplete understanding of the temporal molecular changes that lead to disease onset and progression. Here, we evaluate the ability of manifold learning to develop a molecular model for the unobserved temporal disease progression from RNA-Seq data collected from human postmortem brain samples collected within the ROS/MAP and Mayo Clinic RNA-Seq studies of the AMP-AD consortium. This approach defines a

more »

... ional ordering across samples based on their relative similarity in RNA-Seq profiles and uses this information to define an estimate of molecular disease stage - or disease progression pseudotime - for each sample. This transcriptional estimate of disease progression is strongly concordant with burden of tau pathology (Braak score, P = 1.0x10-5), amyloid pathology (CERAD score, P = 1.8x10-5), and cognitive diagnosis (P = 3.5x10-7) of LOAD. Further, the disease progression estimate recapitulates known changes in cell type abundance and impact of genes that harbor known AD risk loci. Samples estimated to reside early in disease progression were enriched for control and early stage AD cases, and demonstrated changes in basic cellular functions. Samples estimated to reside late in disease progression were enriched for late-stage AD cases, and demonstrated changes in known disease processes including neuroinflammation and amyloid pathology. We also identified a set of control samples with late-stage estimated disease progression who also showed compensatory changes in genes involved in affected pathways are protein trafficking, splicing, regulation of apoptosis, and prevention of amyloid cleavage. In summary, we present a disease specific method for ordering patients based on their LOAD disease progression from CNS transcriptomic data.