Hippocampal extracellular matrix levels and stochasticity in synaptic protein expression increase with age and are associated with age-dependent cognitive decline

of individual mice at 20, 40, 50, 60, 70, 80, 90 and 100 weeks of age. show across diseases without obvious genetic heritability. Our data show that normal aging is associated with increased stochasticity in synaptic protein levels and dysregulation of protein networks that are known to be involved in neurodegenerative diseases, and may thus explain to some extent the highly prevalent sporadic forms of these diseases amongst aging individuals.


Introduction
As the proportion of aged individuals in our population continues to grow, we are also faced with an increase in age-related health problems. Brain aging invariably leads to functional decline and impairments in cognitive function and motor skills, which may seriously affect quality of life. A better understanding of the neurobiological mechanisms underlying age-related cognitive decline is crucial, to facilitate maintenance of cognitive health in the elderly, and to reveal potential causes of highly prevalent age-related forms of dementia, in particular Alzheimer's disease, in which cognitive decline is severely impaired by yet unknown mechanisms.
Several studies showed that normal brain aging is associated with subtle morphological and functional alterations in specific neuronal circuits (1,2), and that reduced cognitive function with increasing age is likely due to synaptic dysfunction (3). Increasing evidence supports that alterations in hippocampal activity are correlated with deficits in learning and memory in healthy aging humans (4, 5). In addition, rodent models of healthy aging demonstrate strong correlations between impaired performance in learning and memory tests on one hand, and disturbed hippocampal network activity 7 pelleted by ultracentrifugation at 80,000 x g for 30 min at 4 °C. Pellets were redissolved in 5 mM HEPES and protein concentrations were determined using a Bradford assay (Bio-Rad, Hercules, CA, USA). For each sample, 50 µg of protein was transferred to a fresh tube and dried in a SpeedVac overnight. Two-dimensional liquid chromatography (2DLC). The lyophilized iTRAQ labeled samples were separated in the first dimension by strong cation exchange column (2.1x150 mm polysulfoethyl A column, PolyLC), and the in second dimension on an analytical capillary reverse phase C18 column (150 mm x 100 µm i.d. column) at 400 nL/min using the LC-Packing Ultimate system. The peptides were separated using a linear increase in concentration of acetonitrile from 4 to 28% in 75 minutes, to 36% in 7 minutes and finally to 72% in 2 minutes. The eluent was mixed with matrix (7 mg of re-crystallized α-cyano-hydroxycinnaminic acid in 1 mL 50% acetronitrile, 0.1% trifluoroacetic acid, Protein identification. Protein identification and quantification are described in detail in (18). To annotate spectra, Mascot (MatrixScience, version 2.3.01) searches were performed against Mus musculus sequences in the Swissprot database (20/10/2010; 16,326 sequences searched) and in the larger but more redundant NCBI database (20/10/2010; 147,581 sequences searched). MS/MS spectra were searched with trypsin specificity and fixed iTRAQ modifications on lysine residues and Ntermini of the peptides and methylthio modifications on cysteine residues. Oxidation on methionine residues was allowed as a variable modification. Mass tolerance was 150 ppm for precursor ions and 0.4 Da for fragment ions, while allowing a single site of miscleavage. For each spectrum the best scoring peptide sequence and protein accession were selected. The false discovery rate (FDR) for peptides identification was calculated using a decoy database search (provided by the Mascot software) and the peptide list was limited to 5% false positives in order to improve protein identification confidence. Protein redundancy in the result files was removed by clustering the precursor protein sequences at a threshold of 90% sequence similarity over 85% of the sequence length. If present, a random Swissprot protein entry was chosen as representative for a protein cluster, if not, a random NCBInr sequence was chosen. Subsequently all peptides were matched against the protein clusters and only those peptides were included that mapped unique to one protein. Proteins were considered for quantification if at least three peptides were identified in four replicate iTRAQ sets and at least one peptide in the four other sets.
Protein quantification and identification of differentially expressed proteins. iTRAQ areas (m/z 113-121) were extracted from raw spectra and corrected for isotopic overlap. Peptides with iTRAQ signature peaks of less than 1,500 were not considered for quantification. To compensate for the possible variations in the starting amounts of the samples, the individual peak areas of each iTRAQ signature peak were log transformed to yield a normal distribution, and normalized to the mean peak area of every sample. No measures were taken to correct for iTRAQ compression. Next, iTRAQ reporter ions were standardized within each spectrum by subtracting the average intensities of all reporter ions in that spectrum. Protein abundances in every experiment were determined by taking the average normalized standardized iTRAQ peak area of all unique peptides annotated to a protein. To determine which proteins show significant differential expression, three types of statistical analysis were performed. First, the Two-Class Analysis option implemented in the Significance Analysis of Microarrays (SAM) package (19) was used to determine significant regulation at individual time points. Two-class SAM analysis was used both as an unpaired analysis, and as a paired analysis in which samples within iTRAQ sets were considered as paired with the 20 weeks time point as the base line.
Paired analysis was performed to reduce false positives resulting from technical variation between iTRAQ sets, and resulted in more stringent selection of differentially regulated proteins than unpaired analysis. In addition, the One-Class Time Course Analysis option in SAM was used to determine significant regulation over time. In all SAM analyses a threshold q-value of 10% was used to determine significant differential protein expression. The Time Course Analysis was used to determine the set of proteins for which there was strongest evidence for regulation over time. K-means clustering and Pearson correlation analysis were used to further analyze this set of proteins. The intersection of the paired and the unpaired Two-Class Analysis was used to determine the set of proteins for which there was evidence for regulation at one or more time points.

Functional protein group analysis
All proteins were assigned to one of 17 functional synaptic protein groups as previously defined (20), and overrepresentation of regulated proteins within functional groups was determined using a Fisher's exact test. Enrichment was only considered relevant when overrepresented functional groups contained at least 3 proteins. In addition, functional enrichment was determined using the DAVID functional annotation tool (http://david.abcc.ncifcrf.gov/) (21,22). The functional categories used were GO term related to Biological Process (BP), Cellular Component (CC), and Molecular Function (MF), as well as pathway annotations derived from KEGG. The entire set of detected proteins was used as the background set, and a modified Fisher's exact p-value <0.05 was considered significant. Enrichment was only considered relevant when enriched functional groups contained at least 5 proteins.

Protein variance analysis
For each detected protein, mean standardized peak areas and standard deviations (SD) of the mean standardized peak areas were calculated per time point for both the unpaired dataset and paired data set against the 20 week time point. Parametric one-way ANOVA and non-parametric Kruskal-Wallis (23) tests were used to determine whether the distribution of these sample means and SDs are dependent on time. Next, all SDs from all time points were collected and ranked, and the top 5% was selected and designated high variable proteins, whereas the bottom 5% was selected and designated low variable proteins. The intersection of the unpaired and paired data was used to determine the set of proteins with either high or low variability. The distribution of high and low variable proteins over time points was determined, and functional enrichment analysis was performed using DAVID (http://david.abcc.ncifcrf.gov/) (21,22), using as a background the whole set of detected proteins and a modified Fisher's exact p-value <0.05 as significance threshold.

Immunoblotting
Immunoblot analysis was performed on six synaptosome protein extracts (six biological replicates) per time point. Samples were treated with chondroitinase ABC (Sigma Aldrich, Zwijndrecht, The Netherlands; 0.5 U/50 mg protein) for 90 min at 37 °C. Of each sample, 2.5 µg protein was mixed with SDS sample buffer and heated to 90 °C for 5 min. Proteins were separated on a Criterion TGX Stain-Free Precast Gel (4-16% Tris-Glycine; Bio-Rad) in a Criterion Cell Electrophoresis System (Bio-Rad), and electroblotted onto PVDF membrane overnight at 4 °C. After blocking with 5% non-fat dry milk in TBS-T (TBS plus 0.5% Tween) for 1 h, blots were incubated with primary antibodies, followed by a horseradish peroxidase-conjugated secondary antibody (Dako, Glostrup, Denmark; 1:10,000). The following antibodies were used: anti-brevican (gift from Dr. C. Seidenbecher, Magdeburg, Germany; 1:2,000), anti-HAPLN1 (Abcam; 1:1,000), anti-neurocan (Sigma; 1:1000). Blots were incubated with ECL substrate (GE Healthcare, Pollards Wood, UK), scanned with an Odyssey Imager (LI-COR) and analyzed with Image Studio software (LI-COR, version 1.1.7) using background correction. To correct for differences in sample input, all gels were imaged before electroblotting and the total protein densitometric values were used for sample normalization, which is more reliable then normalizing to the levels of a single protein (24, 25). Significance was determined using a Student's t-test (one-tailed, independent samples). P-values <0.05 were considered significant.

Histology
Wisteria floribunda agglutinin (WFA) staining was performed on free-floating brain sections obtained from animals at 3 and 18 months of age. Sections were quenched (10% methanol, 0.3% H 2 O 2 in PBS) for 30 minutes, blocked with 0.2% Triton X-100 and 5% fetal bovine serum in PBS, and were incubated overnight with fluorescein labeled WFA (Vector Laboratories; 1:400). Sections were then washed and coverslipped in Vectashield including DAPI as a nuclear dye (Vector Laboratories). WFApositive cells in the CA1 region of the hippocampus were counted in 12 separate sections per animal using the analyze particles option in ImageJ (version 1.40g). The mean number of WFA-positive cells per section was calculated and significance was determined using a Student's t-test (two-tailed, independent samples). P-values <0.05 were considered significant.

Aged mice have hippocampus-dependent learning deficits
To determine age-related hippocampal learning deficits, three age groups of C57BL/6J wildtype mice were tested in a Barnes maze. Mice were trained to find an escape hole amongst 24 holes that were located radially on the edge of a circular platform ( Figure 1A) guided by visual cues surrounding the platform. During a probe trial, the escape hole was indifferent from the other holes, and mice were tested for their ability to memorize its location. Taking the proportion of hole visits in the correct octant of the maze during first 15 hole visits as a measure for spatial memory, we observed a significant decrease in performance, both in 40-60 weeks old mice and in 70-100 weeks old mice compared with 20 weeks old mice (one-way ANOVA; F 2,29 = 4.43, p = 0.021) ( Figure 1B).

Proteomic analysis identifies 100 synaptic proteins that are significantly regulated at any time point
To identify changes in synaptic protein expression underlying age-related impairments in hippocampal learning, we next performed 8-plex iTRAQ proteomics of hippocampal synaptosomes isolated from  Table S9). All mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium (26) via the PRIDE partner repository with dataset identifier PXD001135 and DOI 10.6019/PXD001135.
To determine significant regulation at individual time points, paired and unpaired two-class analyses were performed using the Significance Analysis of Microarrays (SAM) package (19). In the paired setup, samples within each iTRAQ set were considered as paired with respect to the 20 weeks time point, which reduces technical variation between iTRAQ sets. We only considered proteins as regulated when FDR <10% in both the unpaired and the paired analysis. This approach resulted in 100 proteins being significantly differentially expressed at any time point (Supplemental Table S9). As an initial indicator of the validity of these findings we calculated the correlation coefficients for proteins that are part of multisubunit complexes and thus expected to show similar regulation patterns. Indeed, we observed highly correlated expression for proteins within complexes, whereas expression between complexes is often very different (Supplemental Figure S2).
To allow functional characterization of these proteins, and to reduce the impact of potential false negatives at individual time points, we first combined differentially expressed proteins into three different age groups: early-aged (weeks 40 and 50), middle-aged (weeks 60, 70 and 80), and old-aged (weeks 90 and 100) (Figure 2A; Supplemental Tables S10-S12). For functional analysis, proteins were categorized in 17 functional synaptic protein groups as previously defined (20, 27), and regulated proteins at each age group were separately tested for overrepresentation of functional protein groups using all 501 detected protein as the background set ( Figure 2B; Table 1). In early-aged mice, we  Table S13), indicating that combining data from adjacent time points did not bias towards false positive findings.
As a complementary approach, functional enrichment of differentially expressed proteins was also determined using gene ontology (GO) and cellular pathway (KEGG) databases (Supplemental Table S14). Early-aged mice show an enrichment of several GO classes that include cytoskeletal proteins, in particular tubulins. In middle-aged mice we observed a >15 times enrichment of GO classes related to the extracellular matrix (ECM). Old-aged mice show significant enrichment for GO classes that are related to the cortical cytoskeleton (actins, actin-regulatory proteins and neurofilament proteins), but enrichment is relatively low (<4 times) compared with early-and middle-aged animals.

Proteomic analysis identifies 25 synaptic proteins that are significantly regulated over time
We next wanted to select differentially expressed proteins that are most consistently up-or downregulated over time. One-class time course analysis in SAM using the eight individual time points as input revealed a total of 17 proteins that were significantly upregulated over time, and 8 proteins that were significantly downregulated over time (Figure 3). K-means clustering was used to separate these 25 proteins into four different expression clusters (Figure 3). Clusters 1-3 contained proteins whose expression was increased or decreased already at early time points (40-50 weeks), and

Extracellular matrix proteins are progressively upregulated over time
Three ECM proteins, i.e., HAPLN1, BCAN and NCAN, showed a strong age-dependent increase in expression, with maximum log2-fold changes of 0.25-0.45 at 90 weeks of age ( Figure 4A).
Immunoblotting confirmed their age-dependent upregulation ( Figure 4B). NCAN and BCAN are chondroitin sulfate proteoglycans (CSPGs) that bind to the hyaluronic acid ECM backbone, whereas HAPLN1 is the principle link protein required for this binding. Increased expression of all three proteins is thus strongly indicative of an increase in ECM levels in the aged hippocampus. To test this, we next stained hippocampal sections of 3 months and 18 months old mice with Wisteria floribunda agglutinin (WFA), which specifically labels CSPG glycosaminoglycan side chains. Indeed, we observed a strong increase in WFA staining in the aged hippocampus, both in the number of WFA-positive perineuronal nets (PNNs) and in diffuse ECM staining ( Figure 4C). Quantification of the number PNNs in the CA1 region of the hippocampus revealed a 45% increase at old age compared with young control animals (n = 4-5; p = 0.02) ( Figure 4D). Together, these findings show that the most significant and characteristic age-dependent alteration in the hippocampal synaptic proteome is a progressive increase in ECM levels.

Variance in synaptic protein expression increases with age
Recent studies show that heterogeneity and stochastic deregulation of gene expression increase with age (28-30) and possibly predict lifespan (31). To test if synaptic protein expression is also subject to age-dependent loss of regulatory control, we measured stochasticity in protein levels as a function of age. We calculated the standard deviations (SD) of the average standardized peak areas per protein per time point (Supplemental Table S9), and indeed observed a significant increase in SD with age (unpaired data: one-way ANOVA test, p = 9.91x10 -16 ; Kruskal-Wallis test, p = 2.2x10 -16 ; Figure 5A). A similar age-dependent shift in SD distribution towards greater values was observed in the paired dataset (one-way ANOVA test, p = 1.07 -11 ; Kruskal-Wallis test, p = 9.99 -11 ). Peptide based peak areas did not change with age (one-way ANOVA test, p = 0.134; Kruskal-Wallis test, p = 0.257), which excludes variation in absolute protein levels as an explanation for the observed increase in SD values.
Next, we ranked the combined SD values derived from all time points, and selected the 5% proteins with the lowest SD values, which we designated low variability proteins, as well as the 5% proteins with the highest SD values, which we named high variability proteins ( Figure 5B). The intersection of the unpaired and paired data sets revealed 102 proteins with extreme SD values, either low variable or high variable, at different time points.
We next selected the low variability and high variability proteins that are specifically associated with the two oldest age groups (weeks 90 and 100) (Supplemental Tables S15-S16) and performed a GO term and KEGG pathway functional enrichment analysis ( Figure 5C; Supplemental Table S17). proteins. In addition, the neurodegeneration-related alpha-synuclein (SNCA) protein was found to be highly variable.

Discussion
Age is the primary risk factor for cognitive decline and age-related neurodegenerative disorders such as Alzheimer's disease. Age-related cognitive decline affects a significant proportion of the healthy aging population, but its causes remain to be determined. In mice, age-related cognitive decline can be observed as a decrease in hippocampus-dependent contextual fear memory and spatial learning (32). Here we confirm that aged mice have spatial memory impairments in a Barnes maze task. We then identified proteomic alterations that correlate with these age-dependent memory deficits. Agedependent hippocampal synapse loss is know to contribute to aging in rodents and humans (reviewed in 33), and to exclude the possibility that changes in relative protein abundances are due to synapse loss we focused our proteomics approach on biochemically enriched synaptosome preparations and quantified relative protein levels within these preparations. Two types of proteomic alterations were identified. Firstly, we identified synaptic proteins that significantly change in expression levels during aging. Secondly, we show that some proteins show increased stochasticity with aging, whereas others are more stably expressed. A schematic summary of these findings is provided in Figure 6.
The most significant age-dependent change in synaptic protein levels was observed for a small group of four proteins (cluster 4 in Figure 3), all of which were progressively upregulated with age. Three out of these four proteins are ECM proteins (NCAN, BCAN and HAPLN1 (41), and in the perirhinal cortex PNN depletion enhances object recognition memory (42). These data strongly suggest that high ECM levels limit plasticity in the adult brain.
At the cell physiological level, ECM structures may limit synaptic plasticity in different ways.
Firstly, ECM structures form physical barriers that restrict the lateral diffusion of AMPA receptors at postsynaptic sites (43,44). Secondly, activity-dependent local degradation of the ECM by matrix metalloproteinases unmasks protein fragments that bind to synaptic integrin receptors and regulate LTP (45,46). Finally, the ECM acts as a sink for growth factors and can either modulate or restrain their role in synaptic plasticity (47,48). An age-dependent increase in hippocampal synaptic ECM levels as observed here may thus contribute to age-related learning and memory deficits by reducing molecular and cellular signaling mechanisms that are normally required for synaptic plasticity to occur.
Several recent publications are in support of such a plasticity-limiting role of the ECM in he aging brain. Yamado and Jinno for instance reported age-dependent alterations in hippocampal ECM localization and composition up to one year of age (49). In addition, reducing cortical ECM levels in aged rats reactivates plasticity, both in sensorimotor cortex (50) and in visual cortex (51). Finally, increased expression of ECM proteins, in particular heparin sulphate proteoglycans, is associated with Alzheimer's disease pathology (reviewed in 52), and we recently showed that reducing hippocampal ECM levels restores plasticity and memory deficits in a mouse model of Alzheimer's disease (53).
In early-aged mice (40-50 weeks of age), we also observed alterations in the expression of proteins involved in structural plasticity. In particular, tubulins and microtubule-associated proteins (TUBB3, TUBB4, TUBB5, TPPP and MAP6) as well as intermediate filament proteins (NEFL, NEFM) were upregulated, whereas actin regulatory proteins (ADD2 and ARPC2) were downregulated.
Dysregulation of the cytoskeleton and of cytoskeleton-based transport is an important causative factor in age-related neurodegeneration. The microtubule stabilizing protein TAU for instance is causally linked with synaptic dysfunction and cognitive decline in Alzheimer's disease and in frontotemporal lobe dementia (54,55). Early-aged mice also showed an upregulation of proteins involved in cell metabolism. Interestingly, several of these proteins were found to have neuroprotective properties peroxides generated during metabolism (PRDX1) (58). Upregulation of these proteins could thus protect neurons against the harmful consequences of aging.
In old-aged mice (90-100 weeks of age), we observed increased heterogeneity or stochasticity in synaptic protein expression. An age-dependent increase in variance has previously been reported for gene expression in the human brain (29). Moreover, increased variance in gene expression selectively affects cellular pathways that contribute to Parkinson's disease (59). These authors conclude that increased variance in expression changes the predictability or robustness of gene networks and thereby dysregulates cellular states, and that this may be an essential element that is shared amongst age-related disorders in general. On the other hand, a significant reduction in gene expression variance was observed in stem cells from schizophrenia patients, indicating that increased network robustness and reduced plasticity are equally decremental to normal brain function (59). Our findings demonstrate that variance in protein expression levels also changes with age. Importantly, the increase in variance was specific for the old age group, and did not the result from a decrease in protein abundance with aging, since the distributions of the standardized peak areas (i.e., protein levels) did not significantly change with age. Moreover, we found that distinct functional protein groups are associated with either high or low variability, suggesting that increased stochasticity specifically affects particular synaptic proteins whereas others remain under tight control also at old age.
Low variability proteins in old-aged mice include proteins involved in synaptic vesicle transport, synaptic cell junction integrity, indicating high regulatory control with aging over mechanisms that play a basal role in synaptic neurotransmission. Highly variable proteins in old-aged mice on the other hand are specifically involved in mitochondrial respiration and are frequently associated with neurodegenerative diseases, in particular Parkinson's disease and Huntington's disease. Disease conditions may arise when an increase in variance changes the predictability of the network outcome, resulting in dysregulation of the preferred state (59). Interestingly, many complex diseases lack genetic variants associated with the disease (60, 61), and in such cases variability in gene or protein expression may contribute to the development of diseases without obvious genetic heritability. Our data show that normal aging is associated with increased stochasticity in synaptic protein levels and dysregulation of protein networks that are known to be involved in neurodegenerative diseases, and may thus explain to some extent the highly prevalent sporadic forms of these diseases amongst aging individuals.      lowly-constrained proteins as the top 5% high variability proteins (yellow box) (unpaired data). C, Functional enrichment analysis for high variability and low variability proteins in the old age group using GO term and KEGG pathway databases (see also Supplemental Table S17).      Végh et al. Figure 6 Early-aged (40-50 weeks) Old-aged (90-100 weeks) extracellular matrix cytoskeleton mitochondrion synaptic vesicle postsynaptic density