Proteomic analysis of Alzheimer's disease cerebrospinal fluid reveals alterations associated with APOE ε4 and atomoxetine treatment

Abstract

Alzheimer's disease (AD) is currently defined by the aggregation of amyloid-β (Aβ) and tau proteins in the brain. Although biofluid biomarkers are available to measure Aβ and tau pathology, few markers are available to measure the complex pathophysiology that is associated with these two cardinal neuropathologies. Here, we characterized the proteomic landscape of cerebrospinal fluid (CSF) changes associated with Aβ and tau pathology in 300 individuals using two different proteomic technologies-tandem mass tag mass spectrometry and SomaScan. Integration of both data types allowed for generation of a robust protein coexpression network consisting of 34 modules derived from 5242 protein measurements, including disease-relevant modules associated with autophagy, ubiquitination, endocytosis, and glycolysis. Three modules strongly associated with the apolipoprotein E ε4 (APOE ε4) AD risk genotype mapped to oxidant detoxification, mitogen-associated protein kinase signaling, neddylation, and mitochondrial biology and overlapped with a previously described lipoprotein module in serum. Alterations of all three modules in blood were associated with dementia more than 20 years before diagnosis. Analysis of CSF samples from an AD phase 2 clinical trial of atomoxetine (ATX) demonstrated that abnormal elevations in the glycolysis CSF module-the network module most strongly correlated to cognitive function-were reduced by ATX treatment. Clustering of individuals based on their CSF proteomic profiles revealed heterogeneity of pathological changes not fully reflected by Aβ and tau.

Figure 1.. Study Overview.

Cerebrospinal fluid (CSF) was sampled from 140 controls and 160 AD participants as defined by their CSF tTau/Aβ_1–42 ratio. CSF proteomes for each person were obtained by tandem mass tag mass spectrometry (TMT-MS) and SomaScan 7000 assays. After data quality control and cross-platform comparison, data from both platforms were integrated to generate an AD CSF protein co-expression network. Genetic influence on the network was assessed by module quantitative trait locus analysis (modQTL). CSF AD network modules were compared to serum network modules previously described in Emilsson et al. (27), and association of CSF module proteins in blood with risk of AD was assessed in the AGES-Reykjavik and ARIC studies. The CSF AD network was used to assess the effect of pharmacologic intervention with atomoxetine on disease-relevant pathophysiology. Finally, key hub proteins from across the network were used to cluster participants into groups based on similarity in their CSF proteomic features regardless of their CSF Aβ and tTau status.

Figure 2.. Cross-Platform Proteomic Comparisons.

(A) Overlap of unique proteins measured by TMT-MS and SomaScan in CSF across 300 individuals after quality control filtering in both platforms. Numbers represent counts of gene symbols. Only unique gene symbol overlap was considered. (B) Distribution of within-subject correlation of protein measurements shared between platforms. Measurements were required to have a minimum of 74 total observations and 3 measurements per diagnostic group in each platform (n=1274). The vertical red line indicates the median correlation. (C-E) Illustration of AD-relevant proteins that are strongly correlated between platforms (C), poorly correlated between platforms (D), and variably correlated depending on the SOMAmer used for correlation (E). (F) Differential abundance of proteins in AD as measured by SomaScan (left) and MS (right). Proteins increased in AD are located in the upper right quadrant. (G) Overlap of significantly increased (top) or decreased (bottom) proteins in each platform. (H) The six different SOMAmers for C3 protein highlighted in the SomaScan differential abundance volcano plot shown in panel (F) (left), and correlation to the MS C3 measurement for each SOMAmer (right). Significance of differential abundance was determined at p<0.05. Correlations were performed using Pearson correlation with Student’s test for significance. APOE, apolipoprotein E; APP, amyloid-β precursor protein; C3, complement C3; CHI3L1, chitinase-3-like protein 1; CHIT1, chitotriosidase-1; GAP43, neuromodulin; HEPACAM, hepatic and glial cell adhesion molecule; NRGN, neurogranin; PARK7, Parkinson disease protein 7; rel. abund., relative abundance; RFU, relative fluorescence units; SERPINA3, α−1-antichymotrypsin; SMOC1, SPARC-related modular calcium-binding protein 1.

Figure 3.. Multi-Platform AD CSF Protein Co-Expression Network.

(A) A protein co-expression network of 5242 protein assays from SomaScan and TMT-MS platforms measured across 296 individuals after outlier removal identified 34 co-expression modules capturing different biological processes (outer ring). The module eigenprotein, or first principal component of module expression, was correlated with AD endophenotypes including CSF total tau (tTau), tau phosphorylated at residue 181 (pTau181), amyloid-β 1-42 (Aβ_1–42), tTau/Aβ_1–42 ratio, and Montreal Cognitive Assessment (MoCA, higher scores reflect better cognitive function). Module eigenproteins were also correlated with age, sex, and number of APOE ε4 alleles (APOE ε4 dose). Correlations were considered significant at an absolute value of approximately r=0.1 (p=0.05, dashed lines in legend). Modules were tested for preservation in a previous CSF network generated on 36 individuals (15) (CSF36) by overrepresentation analysis (ORA) and network preservation statistics (preservation), as well as an AD brain network (5) (brain ORA, brain preservation). Green shading indicates preservation at varying degrees of significance. Module eigenproteins were tested for significant change in AD (AD-CT EP), significant change in AD in the CSF36 network (CSF36 synth EP), and significant change in AD in the brain network (brain synth EP). Color shading indicates direction and level of significance (blue=decreased in AD, green=increased in AD, gray=not calculated). Modules were also tested for overlap of protein membership with brain cell type specific protein markers for neurons, oligodendrocytes (oligo), astrocytes, microglia, and endothelia. Color shading indicates degree of significant overlap. L1CAM, neuronal cell adhesion molecule L1; ECM, extracellular matrix. (B) The top 100 proteins by intramodule correlation for the M4 Autophagy/Ubiquitination module. Larger circles represent stronger correlation, with the largest circles representing module “hub” proteins. Proteins are outlined by the proteomic platform by which they were measured. (C) M4 module eigenprotein in AD compared to controls, and correlation with AD endophenotypes. (D) Proteins by intramodule correlation for the M20 Glycolysis/Redox homeostasis module. Proteins are outlined by the proteomic platform by which they were measured. (E) M20 module eigenprotein in AD compared to controls, and correlation with AD endophenotypes. Correlations were performed with bicor. Differences between groups were assessed by t test or one-way ANOVA. For overlap and preservation tests, see Methods.

Figure 4.. CSF AD Network Modules Correlated with APOE ε4 Dose.

(A) The top proteins by intramodule correlation for the M26 neddylation module. (B) M26 eigenprotein in control and AD, correlation with number of APOE ε4 alleles and AD endophenotypes (MoCA score, pTau181, tTau and Aβ42). (C) Protein membership for the oxidant detoxification/MAPK signaling module (M33). (D) M33 eigenprotein in control and AD, correlation with number of APOE ε4 alleles and AD endophenotypes. (E) Protein membership for the mitochondrion module (M34). (F) M34 eigenprotein in control and AD and correlation with number of APOE ε4 alleles and AD endophenotypes. Correlations were performed with bicor. Differences between groups were assessed by t test and adjusted for age and sex.

Figure 5.. Overlap of CSF AD Network With Blood Network.

CSF AD network modules were tested for overlap with serum network modules previously described in Emilsson et al. (27) The red box highlights modules that overlapped with the blood M11 Lipoprotein module. The blue box highlights modules that overlapped with the blood M27 Axon development/Semaphorin complex and M26 Neuron development/Ephrin signaling modules. Significance of overlap was determined using overrepresentation analysis and Fisher exact test with Benjamini-Hochberg correction. Red indicates overrepresentation, blue indicates underrepresentation.

Figure 6.. AD CSF Network Modules Influenced by Treatment with Atomoxetine.

(A) Scheme for atomoxetine (ATX) trial design. All participants had MCI due to AD and baseline CSF sampling. One arm (n=20) was treated initially with ATX for 6 months, then moved to a washout phase, while the other arm (n=19) was treated initially with placebo for 6 months and then moved to ATX treatment. The total trial length was 12 months, with CSF sampling at baseline, 6 months, and 12 months. After dropout, a total of 34 ATX samples were compared to 18 ATX-naïve samples, using ratios to baseline or placebo to assess within-subject treatment effects. The ATX washout samples were not analyzed. (B) M20 Glycolysis/Redox homeostasis CSF eigenprotein levels in control and AD groups in the n=300 (discovery) cohort (Figure 1). (C) M20 abundance after ATX treatment compared to ATX-naïve samples in the ATX trial cohort. (D) Correlation of M20 response to ATX treatment with baseline or placebo abundance of M20. (E) M21 ECM/Vasculature eigenprotein in control and AD groups in the discovery cohort (Figure 1). (F) M21 after ATX treatment compared to ATX-naïve samples in the ATX trial cohort. (G) Correlation of M21 response to ATX treatment with baseline or placebo abundance of M21. ATX treatment effects for all CSF network modules are provided in Table S14 and Data file S4. Differences between groups were assessed by t test. Correlations were performed using bicor and Pearson test.

Figure 7.. Grouping Individuals Based on CSF AD Proteome Network Features.

z scored relative protein abundance heatmap of the top ten hub proteins in each CSF network module, with individuals (n=296) grouped into ten clusters based on the similarity across hub proteins (see Supplementary Figure 7). Figures at the top represent each group, with underlines highlighting the three superclusters. Individuals are shaded based on the number of AD cases within the group as defined by CSF Aβ and tau abundance. In addition to diagnostic category, measures of APOE ε4 allele number (ε4 Dose; 0, 1, or 2), APOE risk (−1, ε2/3; 0, ε3/3, 1, ε3/4; 2, ε4/4), Montreal Cognitive Assessment (MoCA, higher scores indicate better cognitive function), CSF pTau181, CSF tTau, and CSF Aβ_1–42 are provided for each participant.