Molecular estimation of neurodegeneration pseudotime in older brains

Abstract

The temporal molecular changes that lead to disease onset and progression in Alzheimer's disease (AD) are still unknown. Here we develop a temporal model for these unobserved molecular changes with a manifold learning method applied to RNA-Seq data collected from human postmortem brain samples collected within the ROS/MAP and Mayo Clinic RNA-Seq studies. We define an ordering across samples based on their similarity in gene expression and use this ordering to estimate the molecular disease stage-or disease pseudotime-for each sample. Disease pseudotime is strongly concordant with the burden of tau (Braak score, P = 1.0 × 10^-5), Aβ (CERAD score, P = 1.8 × 10^-5), and cognitive diagnosis (P = 3.5 × 10^-7) of late-onset (LO) AD. Early stage disease pseudotime samples are enriched for controls and show changes in basic cellular functions. Late stage disease pseudotime samples are enriched for late stage AD cases and show changes in neuroinflammation and amyloid pathologic processes. We also identify a set of late stage pseudotime samples that are controls and show changes in genes enriched for protein trafficking, splicing, regulation of apoptosis, and prevention of amyloid cleavage pathways. In summary, we present a method for ordering patients along a trajectory of LOAD disease progression from brain transcriptomic data.

Fig. 1. Overview of manifold learning for unraveling staging in Alzheimer’s disease.

a Illustration of steps in manifold learning using reverse graph embedding DDRTree method. b Illustration of lineage inference process for LOAD. RNA-seq samples with different disease diagnoses were pooled, batch normalized, and a smooth manifold was learned for each brain region across individuals (each point is an individual). Total sample numbers are indicated across Mayo RNA-seq TCX and ROS/MAP DLPFC for the different diagnoses in parentheses.

Fig. 2. Manifold learning accurately infers disease states and stages from RNA-seq samples.

a Estimated disease progression trees from temporal cortex (TCX) and b dorsolateral prefrontal cortex (DLPFC) brain regions showing localization of identified LOAD samples on particular branches. c Distribution of pseudotime for AD cases and controls for both DLPFC and TCX for 218 independent samples from two independent studies. d Distribution of expression correlation with pseudotime for both LOAD GWAS genes and non-LOAD GWAS genes for 17446 genes from two independent studies. Box plots have lower and upper hinges at the 25th and 75th percentiles and whiskers extending to at most 1.5xIQR (interquartile range).

Fig. 3. Manifold learning replicates existing measures of staging in LOAD in DLPFC samples.

a Samples colored by three different external measures of LOAD staging, namely: Braak score (tau pathology), CERAD score (amyloid pathology), and cognitive diagnosis (clinical measure of disease severity). Black lines denote inferred lineages. b Distribution of samples by inferred stage for different distinct stages in each of the three methods of measuring LOAD severity for 338 independent samples from one study. Inferred disease stages generally corresponded with all methods, and cognitive diagnosis demonstrated the strongest alignment. Box plots have lower and upper hinges at the 25th and 75th percentiles and whiskers extending to at most 1.5xIQR (interquartile range).

Fig. 4. Cell-type gene expression signatures as a function of disease pseudotime.

a Mean expression of cell markers for astrocytes, neurons, microglia, and oligodendrocytes as a function of pseudotime for TCX brain region, b mean expression of cell markers for astrocytes, neurons, microglia, oligodendrocytes as a function of pseudotime for DLPFC (b) brain region.

Fig. 5. Disease resistant state.

a The inferred manifold from the TCX region with samples colored by their inferred disease subtype/state is shown in the left panel. State 5 (dots, circled) lies at the late end of the disease trajectory, indicating a strong disease-like transcriptomic phenotype, yet most samples in the group did not have pathologically diagnosed AD (Fig. 2a). We hypothesize this group represents a disease resistant state to the disease. b Biclustering results of average expression from each disease state, with increased expression of a gene cluster (Cluster 4) unique to State 5 is shown in the right panel.