Molecular subtyping of Alzheimer's disease using RNA sequencing data reveals novel mechanisms and targets

Abstract

Alzheimer's disease (AD), the most common form of dementia, is recognized as a heterogeneous disease with diverse pathophysiologic mechanisms. In this study, we interrogate the molecular heterogeneity of AD by analyzing 1543 transcriptomes across five brain regions in two AD cohorts using an integrative network approach. We identify three major molecular subtypes of AD corresponding to different combinations of multiple dysregulated pathways, such as susceptibility to tau-mediated neurodegeneration, amyloid-β neuroinflammation, synaptic signaling, immune activity, mitochondria organization, and myelination. Multiscale network analysis reveals subtype-specific drivers such as GABRB2, LRP10, MSN, PLP1, and ATP6V1A We further demonstrate that variations between existing AD mouse models recapitulate a certain degree of subtype heterogeneity, which may partially explain why a vast majority of drugs that succeeded in specific mouse models do not align with generalized human trials across all AD subtypes. Therefore, subtyping patients with AD is a critical step toward precision medicine for this devastating disease.

Fig. 1. Identification of five stable molecular subtypes of AD.

(A and B) WSCNA clustering dendrogram and topological overlap matrix (TOM) heatmap, showing three major classes (A, B, and C) and five subtypes annotated as A, B1, B2, C1, and C2, corresponding to the yellow, red, blue, turquoise, and orange clusters, respectively. (C) Number of samples in each subtype, control (CDR = 0) and mild cognitive impairment (MCI) (CDR = 0.5). (D) Gene expression profiles of all the samples in the PHG from the MSBB-AD cohort. The samples on the columns are grouped by subtype, and the genes on the rows are grouped by WINA module. FC, fold change. (E) Change in mean expression level of various gene pathways for each AD subtype in comparison with the normal control samples. AD-related pathways, representing differential expression from previous AD studies, are derived from the MSigDB. Sets are grouped by major area of biological activity.

Fig. 2. Mean values of several clinical and pathologic traits across AD subtypes.

(A) Bar plots of mean CDR, Braak score, Aβ plaque density, tau NFT densities (measured in the entorhinal cortex, the medial superior temporal cortex, and the medial frontal cortex), APOE4 allele count, and APOE2 allele count across five subtypes, control, and MCI. (B) Stacked bar chart of inferred biological sex from transcriptomic data for all the PHG samples, across five subtypes, control, and MCI. (C) Natural log-transformed P values from the KW ANOVA test of clinical, pathologic, and demographic variables. Significant tests are greater than ~3.0, which corresponds to an α of 0.05. n.s., not significant.

Fig. 3. MEGENA and BN-based key drivers of the AD subtypes.

(A and B) Top down- and up-regulated MEGENA key drivers for each subtype plotted in its location in the network. Color of a node represents subtype (ties resolved; described below), while size of a node corresponds to the total number of genes in the two-hop network neighborhood around the gene, which are differentially expressed. Some genes are drivers for more than one subtype, and ties are resolved by coloring the node corresponding to the subtype with the smallest signature so as to preserve faint signals. (C) Heatmap of the top 20 down- and up-regulated MEGENA key drivers for each AD subtype, where size of the node represents KNR natural log P value (larger is more significant). (D) Heatmap of the top 20 down- and up-regulated BN key drivers for each AD subtype. (E) AD subtype key drivers in the MEGENA network supported by gene expression changes in other brain regions, which the region of overlap listed.

Fig. 4. Cell type–specific changes within each MSBB-AD subtype.

(A) Mean change in cell-type proportion in each AD subtype, computed by averaging SPVs for the samples after cell-type deconvolution by BRETIGEA. (B to F) Vector addition of squared expression levels of AD subtype key regulators (up- and down- regulated) across five different cell types in a brain cell type–specific sequencing experiment (52).

Fig. 5. Identification of the AD subtypes in the ROSMAP cohort.

(A) WSCNA Clustering dendrogram and TOM similarity heatmap of the AD samples in the ROSMAP cohort. (B) Heatmap of gene expression across all the ROSMAP AD and control samples. Genes are organized by the WINA gene modules identified in the gene expression data of the PHG in the MSBB cohort. (C) Correlations between the MSBB and ROSMAP subtypes. First, mean gene expression across each AD subtype is computed for each gene, resulting in a 10 by 13,982 matrix of expression levels for all the 10 subtypes in the two cohorts (lowly expressed genes are excluded). Pearson correlation is then computed between each pair of subtypes, resulting in a 10 by 10 correlation matrix. (D) Cell-type SPV mean by ROSMAP AD subtype class, inferred as in (A) and (B), for MMSE-normalized data. (E) APOE genotype proportion by ROSMAP subtype. (F) Mean prediction accuracy of prediction of the ROSMAP subtypes by a classifier trained from the subtypes and the data from the RNA-seq data from the PHG in the MSBB cohort. RF classifiers are trained on the basis of different numbers of key regulators of the AD subtype classes identified in the MSBB cohort. The subtypes for each AD case in the ROSMAP cohort are independently determined as described in fig. S7 and are thus used as ground truth for evaluating the prediction by the predictors trained by the MSBB-AD data. (G) Bar plots of the association of clinical and pathologic phenotypes with the predicted subtypes of each ROSMAP sample.

Fig. 6. Matching existing AD mouse models to the MSBB-AD subtypes.

(A) GSEA enrichment of differential expression signatures of the identified AD subtypes (up- and down-regulated) for the gene signatures of the AD mouse models. Positive scores indicate strong consistency. (B) Gene expression of the top subtype key regulators across the mouse models, with significant DEGs shown.