Results
To monitor the secreted signaling proteome in plasma, we manufactured glass-based microarrays with commercially available antibodies to measure the relative levels of close to 600 distinct secreted signaling proteins. Using these arrays, we obtained quantifiable results for 582 signaling proteins (Additional file 1: Figure S1A to D and Additional file 2) in archived blood plasma from 47 sporadic, cognitively impaired AD patients and 52 non-demented, closely age- and sex-matched controls obtained from two clinical centers (Additional file 1: Table S1). While these proteins do not encompass all secreted signaling proteins, they do provide a strong representation of all major signaling pathways and represent the largest dataset of this kind available today (Additional file 1: Figure S1A). Raw data were processed, normalized (Additional file 1: Figure S2), and then subjected to three parallel analyses, aimed at integrating both molecular and clinical data, followed by external and internal validation steps (Fig. 1a).
Fig. 1 The circulatory AD signaling proteome reveals changes in cellular communication. a Overview of the experimental and analysis workflow. Plasma samples were collected at clinical centers, relative protein abundance was determined by antibody microarray and three types of analyses were performed: Protein level, MMSE correlation (cognitive performance), and protein co-secretion analysis. The analyses results were then integrated in a network and pathway enrichment framework and finally subjected to internal and external validation. b Heat map representation of the protein level analysis showing the top 50 most different proteins after unsupervised clustering (q < 0.05), separating samples into AD (pink, right) and controls (blue, left) and proteins into higher in control (blue, top) and higher in AD (pink, bottom). c Volcano-plot showing the distribution of all proteins and naming those significantly different between AD and control subjects (p corr < 0.01). d A network representation of the most significantly changed proteins (p corr < 0.015; un-connected proteins omitted) after integration with known pathway and physical interaction data reveals many densely connected hits in pathways related to TGFβ/GDF/BMP, angiogenesis, and apoptosis signaling. e Example scatter plots of the six top changed proteins (see dashed box in e, mean ± s.e.m; all p-values are corrected for multiple hypothesis testing)
To identify signaling proteins with significantly changed plasma levels in AD we calculated corrected p-values for the quality controlled, centered, and normalized array data. Principal component analysis showed that our data were relatively free of obvious batch effects or confounding factors (Additional file 1: Figure S3). Clustering of the top 50 most different proteins illustrated clear differences between AD and control samples (Fig. 1b and c). Using the most significant proteins (FDR < 0.05, corresponding to pcorr < 0.015) as a starting point, we then queried known pathway or physical interaction databases to test the hypothesis that the signaling proteome could be mined to identify pathologically disturbed pathways (i.e., deregulated pathways should reveal themselves through changes in multiple receptors and/or ligands). This approach greatly reduces the chance of false-positive discoveries in contrast to following individually significant but unconnected leads. Using this methodology we identified highly interconnected clusters of receptors and ligands with growth factor activity (“TGFβ/GDF/BMP signaling” and “Angiogenesis”) or with activity linked to apoptosis (Fig. 1d). Reassuringly, the direction of changes was often coherent within each cluster/sub-cluster (Fig. 1d). Individual proteins can show highly significant differences between cohorts and, at the same time, exhibit large overlaps in the observed protein level ranges, highlighting the need for sufficient sample sizes and cohort stratification (Fig. 1e and Additional file 1: Figure S4).
To determine to what extent the observed changes in the AD signaling proteome are AD specific or the result of general neurodegeneration or other unrelated processes, we collected plasma samples from an additional 92 patients (Additional file 1: Table S2) suffering from semantic-variant primary progressive aphasia (svPPA), a sub-type of frontotemporal lobar degeneration (FTLD). SvPPA is almost always associated with Trans-activation response element (TAR) DNA-binding protein 43 (TDP-43)-aggregate pathology and appears to have weak genetic linkage [5–7]. This makes svPPA an ideal candidate to compare distinct neurological pathologies between two unrelated, sporadic, progressive dementias (svPPA vs. AD) via signaling proteome analysis [8]. The svPPA samples were prepared, handled, and analyzed in parallel to the AD samples to minimize experimental variations.
We found 39 proteins with significant changes in both AD and svPPA (Fig. 2a, inset; p = 7.3×10−5 by hypergeometric test). Intriguingly, when we compared significance and direction of protein changes between the two pathologies, we found a perfect correlative trend of up/up or down/down amongst the 39 overlapping proteins (p < 1.8×10−12 by binominal test), indicating that a more general disease-profile does exist (Fig. 2a, red boxes). Additionally, we were able to identify proteins with svPPA-specific (Fig. 2a, purple box) and AD-specific proteome signatures (Fig. 2a and b, yellow box), respectively. A more detailed analysis of the svPPA findings is published elsewhere [9]. Manual literature curation of the AD-specific hits yielded numerous proteins involved in TGFβ/GDF/BMP signaling, complement activation, apoptosis, or with otherwise strong AD literature, suggesting that those pathways could play a role in AD (Fig. 2b). These findings indicate that a mixture of both disease-specific and nonspecific signaling profiles can be obtained from blood and that comparative proteomics can be applied to identify disease-specific changes.
Fig. 2 The plasma proteome contains disease specific information. To assess the specificity of the proteins identified in the expression level analysis, AD samples were compared to another, unrelated progressive dementia (svPPA = semantic-variant primary progressive aphasia). a Plotting signed, log-transformed p corr-values (more extreme = greater significance) of the AD vs. svPPA analysis show preserved directionality (binominal test) and can be used to categorize proteins into four distinct groups: “General neurodegeneration” (p AD & p svPPA < 0.05, same direction of changes in both diseases; red box); “Non-significant” (p AD & p svPPA > 0.05; green box), “svPPA specific” (p AD > 0.05, p svPPA < 0.05; purple box); “AD specific” (p AD < 0.05, p svPPA > 0.05; yellow box). Venn diagram showing the overlap of significantly changed proteins in AD or svPPA samples (top-left inset; threshold p corr < 0.05; overlap significance by hypergeometric test). b Zooming into the “AD specific” box (see dashed box in a), many proteins can be identified as part of TGFβ/GDF/BMP, complement, or apoptosis signaling in addition to numerous proteins with strong supporting AD literature (manual curation)
Changes in protein levels provide a binary view on AD (significantly changed or not). However, disease is a gradual process with patient’s cognition becoming increasingly more impaired. To explore the relationship between cognitive performance and relative plasma protein levels, we correlated the levels of 582 proteins with the Mini-mental state [10] examination (MMSE) scores of the respective patient (Fig. 3a and b, and Additional file 2; Spearman rank correlation and p-value [prho] as well as explicit p-values through 1000-fold sample permutations [pperm]). Applying the same network exploration method as before, we identified again many proteins involved in TGFβ/GDF/BMP signaling and apoptosis to be positively or negatively correlated with cognitive performance (Fig. 3c), lending further support to our AD-specific findings above. While proteins involved in “Complement” activation were detected by the MMSE correlation analysis, they did not meet the cutoff criteria in the differential analysis, indicating that the complement driven effects are more subtle and gradual in nature (Fig. 3c). This finding highlights the fact that analyzing a continuous functional parameter such as MMSE can retrieve additional non-binary pathways of interest.
Fig. 3 Correlation of cognitive function with the circulatory AD plasma proteome. a Proteins that exhibit significant correlation between cognitive function (evaluated by MMSE = Mini-mental state examination score) and protein levels ranked by correlation (cutoff p rho < 0.05, Spearman rank correlation; p perm based on 1000 MMSE-score permutations; dashed red line indicates p perm = 0.05 threshold). Many proteins are part of TGFβ/GDF/BMP, complement, or apoptosis signaling. b Example scattergrams of the top positive and negative MMSE-correlated proteins (red line indicates regression with 95 % confidence intervals). c Network representation of significantly correlated proteins after integration with known pathway and physical interaction data reveals many densely connected hits in pathways related to TGFβ/GDF/BMP, complement, and apoptosis signaling
Testing for mean differences in plasma protein levels allows us to identify proteins with significantly different levels of expression between AD and controls. However, we hypothesized that additional insight could be gained from comparing changes in co-expression of signaling molecules as this might carry information on the pairwise relationship between the underlying signaling proteins and regulatory pathways, and on how this relationship is affected in disease [11].
To assess changes in co-expression, we calculated the Spearman correlation between each protein pair under healthy control and under AD conditions and then subtracted these correlations from each other, creating co-expression and differential co-expression networks respectively (Fig. 4a and b). It can be shown that these networks carry valuable biological information by comparing co-expression profiles and gene ontology (GO) biological process similarity between genes/proteins (Additional file 1: Figure S5). We found that protein pairs with high differential co-expression profile correlation (R > 0.35; Additional file 1: Figure S5) have significantly higher median GO semantic similarity scores than expected by chance, indicating that these protein pairs are functionally related.
Fig. 4 Protein co-expression analysis. a Schematic, hypothetical example of differential protein correlation: Proteins A to D exhibit a certain correlation pattern in control samples (top row) and a different pattern in AD samples (middle row). Subtracting the control correlations from the AD correlations yields the differential correlation “AD-Control” that captures the direction and magnitude of the correlation changes in disease (bottom row). b Zoomed-in correlation matrices for 50 proteins out of 582: Pairwise protein correlation in control samples (top left), AD samples (top right) and calculated differential correlation (bottom left; random subset of proteins in alphabetical order, Spearman rank correlation). “GO BP” represents the pairwise semantic similarity score of the protein pairs from ~0.1 (very different) to ~0.9 (very similar) in the “biological process” gene ontology as a measure for distance in the ontology tree and shared membership in biological processes. c Heat map of the differential profile correlations of all 582 proteins after unsupervised clustering with optimal leaf ordering. Positive correlations between two proteins indicate that these proteins change their correlations to many other proteins in a highly parallel fashion. Different clusters of proteins with similar profiles can be identified and are each significantly enriched for biological processes (boxes a-h, annotation below heat map; p-value based on a modified Fisher exact p-value; N = members with annotation/cluster size; three most significantly enriched clusters are underlined). d Cluster “a Regulation of growth” zoomed in with detailed sub-structure (same color scale as c)
Having established that meaningful biological information is contained within the co-expression profiles, we used hierarchical clustering [12] to arrange signaling proteins based on their co-expression profile correlations (Fig. 4c). Using DAVID [13, 14], we identified several clusters of proteins that were significantly enriched for a number of ontology terms (Fig. 4c). These protein clusters represent signaling molecules whose co-expression profiles change in a highly parallel fashion in AD, which in turn could indicate that the underlying regulatory pathways participate in AD pathogenesis. The most significantly enriched clusters were “Complement” (p = 9.7 × 10−10), “Regulation of growth” (p = 0.00018) and “Apoptosis” (p = 0.0037; all EASE score p-value based on 562 unique gene background). In this approach, the size (and number) of clusters identifiable is limited by the overall number of proteins probed: While it is apparent that cluster a-d and e-g could be grouped into larger “superclusters”, probing these superclusters for enrichment is not useful as they contain 25–45 % of all probed proteins (Fig. 4c).
Exploring the “Regulation of growth” cluster in more detail, we observed several sub-clusters that partially aligned with known biological relationships, often bridging regulatory pathways together (Fig. 4d). For example APOE deficient mice are prone to atherosclerosis, but C3 modulation of lipid metabolism can protect them [15], while VEGFC is a marker for advanced atherosclerosis and hypercholesterolemia in the same animals [16]. TGFβ and interferons (IFN) have antagonizing relationships in the control of inflammation [17], while interferons reduce ghrelin (GHRL) expression [18], and Follistatin-like 1 (FSTL1) is controlled by TGFβ [19]. We also noticed that numerous proteins were directly or indirectly associated with or regulated by TGFβ/GDF/BMP signaling (e.g., TGFBR1, FSTL1, THBS1, MMP11, GREM1, GDF3, GDF5, GDF9), making that pathway a lead candidate to test for its involvement in AD. These data suggest that co-expression analysis can be used to identify clusters of proteins and regulatory pathways that are linked through similar co-expression profiles, potentially indicating pathways that are affected by or affecting AD pathology simultaneously.
Because protein quantification is notoriously difficult [1, 2, 20, 21] we examined the existing published literature for suitable non-proteomics data that could be used to validate our experimental approach. Recently, a large post-mortem study examined the mRNA expression levels in tissue from late-onset AD and control patients (cerebellum, pre-frontal cortex, and visual cortex) and provided data on the correlation between transcript levels and brain atrophy and Braak staging, which provides a measure of tangle pathology [22]. Using this dataset, we asked whether proteins that we found to be significantly correlated to cognitive performance (MMSE) also show a significant correlation between pre-frontal cortex transcript level and brain pathology (Braak stage and atrophy). And indeed, 29.4 % of the proteins (15/51) we had identified using MMSE as trait also have significant transcript-Braak stage correlations (p = 0.0017, Chi2-test, Fig. 5a), while 33.3 % (17/51) exhibited transcript-atrophy correlations (p = 0.0195, Chi2-test, Fig. 5a). This supports the notion that our experimental approach focusing on circulatory signaling proteins is able to enrich for proteins linked to AD brain pathology.
Fig. 5 External Validation. To assess the biological validity of our findings top proteins were cross-referenced at the transcript or genomic level to external datasets containing AD brain mRNA transcriptome data or AD genome-wide association data, respectively. a Pre-frontal cortex transcripts in AD brain corresponding to proteins that correlated with MMSE in our study (see Fig. 3a) correlate strongly with Braak stage or brain atrophy (pre-frontal cortex) beyond what is expected by chance (chi2-test). b Based on meta-data from AD GWAS studies 114 genes which are part of “TGFβ/GDF/BMP signaling” exhibit higher than expected enrichment for significant SNPs (gene-wide p-value using VEGAS; cumulative curve comparison by Kolmogorov–Smirnov test). c Pre-fontal cortex transcripts for 120 genes of the TGFβ/GDF/BMP signaling pathway show many more significant transcript changes in AD than expected by chance (explicit p-value through sample permutation; cumulative curve comparison by Kolmogorov–Smirnov test). d Graphical representation of 92 proteins of TGFβ/GDF/BMP signaling pathway and integration with the findings from the various studies as indicated (only nodes with hits ≥ 1 are shown). Proteins are positioned relative to their location in the cells (extracellular, membrane bound, intracellular) and edges indicate physical or functional interactions (small border diagrams highlight proteins with associated significant changes). e Detailed pathway diagram for GDF-Activin receptor signaling. f Examples of protein (top row) and mRNA changes (bottom row) among the members of the GDF-Activin receptor signaling cascade (all corrected p-values)
The results thus far from the differential protein level analysis, the correlations with cognitive function, and the co-expression network analysis consistently implicate the complement pathway, apoptosis, and a signaling network around the TGFβ superfamily in AD. We were particularly intrigued by this latter network, and its GDF family members, many of which had not been linked to AD. To further validate the significance of this finding we asked whether any of the proteins in this pathway – including intra- or extracellular ones not measured with the array – were additionally linked to AD at the genomic or transcript level. We used the Gene Ontology to identify a list of 114 genes linked to the TGFβ/GDF/BMP signaling pathway (Additional file 2) and queried two large AD data sets consisting of meta-data from 10 genome-wide association studies with a total of 8,309 AD cases and 7,366 cognitively normal elders [23] and AD brain transcriptome data from 181 AD and 125 controls [24] for significant SNPs or mRNA changes within the TGFβ/GDF/BMP pathway associated with AD.
We found significantly more SNPs in genes associated with TGFβ/GDF/BMP signaling than what would be expected by chance (p = 8.0×10−17, Kolmogorov-Smirnov test, Fig. 5b). Specifically, several genes that are involved in GDF3/Activin-receptor signaling had significant associations (GDF3, ACVR1B, SMAD3; see Additional file 2). Similarly, we found significantly more mRNAs with altered expression in AD brains associated with TGFβ/GDF/BMP signaling than what would be expected by chance (p = 1.3×10−20, Kolmogorov-Smirnov test, Fig. 5c and Additional file 2), including CFC1 (Cripto, a co-factor of GDF3), and Activin-receptor subunits 2B and 1C (ACVR2B/ACVR1C, GDF3 receptors). Taken together, multiple layers of experimental evidence suggest that changes at the genome, the transcriptome, and the proteome level in TGFβ/GDF/BMP signaling are associated with AD (Fig. 5d-f).
Activin receptors and its ligands Inhibin A and B, GDF1, GDF3, GDF5, and the ligand binding proteins CFC1/cripto and gremlin were most prominently altered in AD throughout our study (Figs. 1d, 3c, 4d, 5f). While TGFβ and TGFβ-receptor signaling has been studied extensively in the context of AD and neuroinflammation [25, 26], and BMP9/GDF2 has been identified as a regulator of cholinergic neuronal development [27], Activin receptor signaling and GDFs have not been studied in AD. GDF3 represents a particularly intriguing candidate: it is highly expressed in the human brain [28] and mouse dentate gyrus (Fig. 6a), is part of the co-expression cluster a (Fig. 4c and d), exhibits significant SNP enrichment (Additional file 2), is positively correlated with cognitive performance (pperm = 0.038), yet its effects on neurons or the brain are unknown.
Fig. 6 GDF3 regulates neurogenesis and is reduced in AD brains. a To test whether GDF3 levels are also reduced in AD brains, Human AD and control cortical grey matter regions were lysed and the detergent soluble protein fraction was probed by western blot. Levels of active GDF3 (b) were quantified relative to neuron-specific enolase (NSE). c To identify areas in the brain where GDF3 may have a functional role, we referred to The Allen Brain Atlas, which showed strong RNA expression in the mouse hippocampus (blue = high expression). d Using qPCR, GDF3 mRNA expression was detected in non-differentiated adult mouse NPCs and NPCs cultured in differentiating conditions. e To determine whether GDF3 affects stem cell function, adult mouse NPCs were provided recombinant mouse GDF3 and NPC proliferation was assessed using BrdU. f Recombinant mouse GDF3 was also provided to dissociated adult mouse neurospheres and the number of newly formed neurospheres was subsequently quantified. g To investigate whether GDF3 promotes neurogenesis, human-derived NTERA cells stably transfected with Dcx promoter-controlled eGFP were provided recombinant human GDF3. Shown are representative images of DCX-GFP fluorescence expression from an entire well of NTERA cells treated with GDF3 or control for 30 days. h DCX-GFP fluorescence area was quantified relative to the cellular area detected by brightfield microscopy. Results were compared by a one-way ANOVA with a Dunnett's post-test (b,e), an unpaired Student’s t test (d), or a two-way ANOVA with a Bonferroni post-test (f,h) and are representative of at least 2 independent experiments (b n = 16 –18 per group, d,e,f,h; n = 3 per group). Values are mean ± s.e.m.. *p < 0.05, **p < 0.01, ***p < 0.001 compared to respective control groups
Given that we observed reduced GDF3 plasma levels, we first investigated whether these changes may be reflective of GDF3 levels in human AD brains. To do this, we measured the processed, active form of GDF3 in tissue extracts of AD and age matched control patients and found a significant reduction in the cortex (n = 16–18 subjects per group, p = 0.02, Fig. 6b), but not in the cerebellum, an area unaffected by AD (n = 8 subjects per group, p = 0.15, Additional file 1: Figure S6A and B). To evaluate our svPPA vs AD comparison, which predicted that GDF3 reduction is AD-specific, we measured activated GDF3 in cortical extracts from svPPA patients and controls and found no difference in GDF3 levels (n = 5 subjects per group, p = 0.08, opposite trend to AD, Additional file 1: Figure S6 C and D). While these studies support our plasma proteomic findings, the source and functional significance of GDF3 within the brain remain unclear.
Because GDF3 is highly expressed in the mouse dentate gyrus (a neurogenic area; Fig. 6c) and previous studies show a role for GDF3 in regulating embryonic and cancer stem cell fate and differentiation [29–31] we asked whether GDF3 regulates adult neurogenesis. We first cultured primary adult mouse neural progenitor cells (NPCs) under non-differentiating and differentiating conditions and measured GDF3 mRNA. Inducing NPC differentiation into neurons and astrocytes caused a marked upregulation of GDF3 (Fig. 6d), suggesting that, unlike in embryonic stem cells [32], adult NPCs are not the main source of secreted GDF3 and adult hippocampal GDF3 likely derives from mature cell types [28]. To assess whether GDF3 affects NPC function, we exposed primary adult NPCs to increasing concentrations of recombinant mouse GDF3 and measured proliferation. Indeed, GDF3 increased NPC proliferation as measured by BrdU incorporation (Fig. 6e) and changes in neurosphere number (Fig. 6f). Given that GDF3 has been implicated in embryonic stem cell fate [29, 30] and is capable of differentiating PC12 cells [33], we next investigated whether GDF3 promotes neuronal differentiation. To evaluate neuronal differentiation we utilized NTERA cells stably transfected with Doublecortin (Dcx) promoter-controlled eGFP, as previously described [34]. NTERA cells cultured with GDF3 showed a prominent increase in Dcx expression compared to control cells (Fig. 6g and h), which was also confirmed by western blot (Additional file 1: Figure S6E and F). Furthermore, a recent study demonstrated the potential importance of developmental factors such as the Repressor element 1-silencing transcription factor (REST) in AD pathology [35]. Intriguingly, GDF3 and REST appear to be controlled by shared transcription factor binding [36], potentially hinting at common upstream disturbances. Taken together these findings suggest that GDF3 plays an important role in NPC proliferation and neuronal differentiation and that it’s levels are altered in AD.