Alzheimer’s disease is a neurodegenerative disorder accounting for more than 50% of cases of dementia. Diagnosis of Alzheimer’s disease relies on cognitive tests and analysis of amyloid beta, protein tau, and hyperphosphorylated tau in cerebrospinal fluid. Although these markers provide relatively high sensitivity and specificity for early disease detection, they are not suitable for monitor of disease progression. In the present study, we used label-free shotgun mass spectrometry to analyse the cerebrospinal fluid proteome of Alzheimer’s disease patients and non-demented controls to identify potential biomarkers for Alzheimer’s disease. We processed the data using five programs (DecyderMS, Maxquant, OpenMS, PEAKS, and Sieve) and compared their results by means of reproducibility and peptide identification, including three different normalization methods. After depletion of high abundant proteins we found that Alzheimer’s disease patients had lower fraction of low-abundance proteins in cerebrospinal fluid compared to healthy controls (p<0.05). Consequently, global normalization was found to be less accurate compared to using spiked-in chicken ovalbumin for normalization. In addition, we determined that Sieve and OpenMS resulted in the highest reproducibility and PEAKS was the programs with the highest identification performance. Finally, we successfully verified significantly lower levels (p<0.05) of eight proteins (A2GL, APOM, C1QB, C1QC, C1S, FBLN3, PTPRZ, and SEZ6) in Alzheimer’s disease compared to controls using an antibody-based detection method. These proteins are involved in different biological roles spanning from cell adhesion and migration, to regulation of the synapse and the immune system.
Citation: Khoonsari PE, Häggmark A, Lönnberg M, Mikus M, Kilander L, Lannfelt L, et al. (2016) Analysis of the Cerebrospinal Fluid Proteome in Alzheimer's Disease. PLoS ONE 11(3): e0150672. doi:10.1371/journal.pone.0150672
Editor: Kristel Sleegers, University of Antwerp, BELGIUM
Received: October 7, 2015; Accepted: February 16, 2016; Published: March 7, 2016
Copyright: © 2016 Khoonsari et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant quantification and identification data is within the paper and its Supporting Information files. The raw mass spectrometry data has been submitted to “Dryad” and are accessible using the following DOI: doi:10.5061/dryad.8v2d0.
Funding: This research was supported by Uppsala Berzelii Technology Centre for Neurodiagnostics, with financing from the Swedish Governmental Agency for Innovation Systems, the Knut and Alice Wallenberg Foundation, KTH Center for Applied Proteomics, ProNova Vinn Excellence Center, Marianne and Marcus Wallenbergs Foundation, Swedish Alzheimer’s foundation, Swedish Brain Foundation, Gun och Bertil Stohnes stiftelse, Geriatriska fonden, the Swedish Research Council FORMAS, the Swedish Research Council P29797-1 JB grant (621-2011-4423) as well from SciLifeLab. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Alzheimer’s disease (AD) is an age-dependent neurodegenerative disorder and the most common form of dementia in the elderly population, accounting for more than 50% of all dementia cases . Epidemiological investigations have estimated that the numbers of AD patients will double every 20 years to more than 66 million worldwide in 2030 and 100 million by 2050 [2, 3].
Alzheimer’s disease is associated with multiple molecular characteristics including extracellular beta-amyloid (Aβ) plaque deposition and accumulation of intracellular neurofibrillary tangles composed mainly of hyperphosphorylated tau protein. These pathological findings are believed to mediate the extensive loss of neurons and synapses as well as the inflammatory processes . The diagnosis of AD is based on clinical examinations that can be complemented by analysis of Aβ42, total tau (t-tau), and hyperphosphorylated tau (p-tau) level in cerebrospinal fluid (CSF) (reviewed in [5, 6]). Despite having relatively high sensitivity and specificity, these biomarkers have limited value for monitoring disease progression [6–8].
Cerebrospinal fluid is a proximal fluid in direct contact with the brain interstitial fluid that potentially reflects biochemical changes related to central nervous system (CNS), making it a promising source of biomarkers in neurological disorders such as AD . CSF protein concentration can vary between 15 to 60 mg/dl and the protein level can be affected by age . Over the last decade, there has been a growing interest in applying proteomics to identify disease-specific biomarkers to increase our understanding of underlying pathogenesis of AD. Most CSF biomarker discovery studies have been performed using a classic proteomics platform based on two-dimensional gel electrophoresis (2-DE) in combination with mass spectrometry (MS) or tandem mass spectrometry (MS/MS) [11–13]. Although 2-DE provides high resolution protein separation, it has limitations regarding detection of low abundant proteins . As an alternative, gel-free shotgun MS in conjunction with quantitative proteomic technique, e.g. stable isotope labeling [15, 16] or label-free methods [17–20], have recently been used for identification and quantification of proteins involved in the pathogenesis of AD. Furthermore, combining gel-free shotgun MS approaches with protein depletion of high abundant proteins enables detection and quantification of low abundant proteins .
To analyze data sets generated by mass spectrometry-based methods, specialized software programs are commonly required. Commercial solutions are widely popular, mostly because of providing user-friendly environments whereas, open source programs offer more flexibility in terms of possibilities to modify existing algorithms. Careful selection of proper programs for data processing is crucial, since different programs have been shown to produce different and in some cases contradictory results (for reviews on this topic, see and ). This inconsistency has been traced back to the application of different algorithms and improper choice of parameters (by the users) due to complex interface or lack of proper documentation (for more detailed description of different factors, see ). In addition to selection of software for initial data analysis, different methods of downstream processing and analysis such as normalization and statistical testing will influence the results. The basic assumption for many studies including CSF is that the protein concentrations in patients and healthy controls are similar; an assumption that is also reflected in global normalization methods used [15, 25, 26], however if this assumption is not correct it may have an negative impact on the final result.
The aim of this study was to detect novel protein markers that can be used to distinguish between AD and healthy elderly controls, to evaluate the consistency of software selection and impact of normalization methods used on result. We have employed a novel “Dot-it-Spot-it®” total protein assay to measure protein concentrations of small volumes of CSF from ten patients diagnosed with AD and ten non-demented subjects. The label free mass spectrometry data was processed using five different programs to evaluate quantification reproducibility and peptide identification performance. To evaluate the assumptions of normalization on the final results, we applied three different normalization algorithms and compared the results to that of using an affinity proteomics approach utilizing antibody suspension bead arrays of selected proteins. Finally, eight proteins were validated as different in expression between AD and controls.
Materials and Methods
This study was based on proteomic analysis of CSF from ten AD patients and ten non-demented controls. Samples were collected according to the recommended consensus protocol for CSF collection and biobanking  and obtained from the Uppsala Berzelii Technology Centre for Neurodiagnostics biobank at the Uppsala University Hospital. All patients underwent brain imaging, routine laboratory testing as well as neurological and cognitive examinations. The average Aβ42, tau, and p-tau in the AD patients were 420±117, 652±376, 132±112 (mean±SD ng/l), respectively. The control subjects had normal cognition according to their MMSE performance. The Regional Ethical Review Board in Uppsala, Sweden had approved the collection of CSF samples and the conducted research (Dnr 48–2005). The participants provided their written informed consent for research. The main clinical features of the patients are summarized in Table 1.
Sample collection and handling
CSF was collected at room temperature by aspiration lumbar puncture into polypropylene tubes. After collection, the samples were directly centrifuged at 1300 g for 10 minutes at +4°C to pellet any cell debris. After centrifugation, all CSF samples were visually inspected for blood contamination, frozen and stored at -80°C.
Chemicals and reagents
Acetonitrile (ACN), methanol (MeOH), acetic acid (HAc), formic acid (FA), ammonium bicarbonate (NH4HCO3) were obtained from Merck (Darmstadt, Germany). Acetone, ethylenediaminetetraacetic acid tetrasodium salt dihydrate (EDTA), protease inhibitor cocktail, phosphate buffered saline (PBS), trifluoroacetic acid (TFA), triethyl ammonium bicarbonate (TEAB), sodium citratemonobasic, sodium dodecyl sulphate (SDS), and chicken ovalbumin were purchased from Sigma Aldrich (St. Louis, MO, USA). For tryptic digestion, iodoacetamide (IAA), urea and dithiothreitol (DTT) were obtained from Sigma Aldrich and trypsin/Lys-C mixture (mass spectrometry grade; Promega, Mannheim, Germany). Ultrapure water was prepared by Milli-Q water purification system (Millipore, Bedford, MA, USA).
To enrich the low abundant proteins prior to nanoLC-MS/MS analysis, each CSF sample was depleted of the seven highly abundant proteins (albumin, IgG, alpha-1-antitrypsin, IgA, haptoglobin, transferrin, and fibrinogen) using a human Multiple Affinity Removal System (MARS Hu-7) 4.6 mm×50 mm LC column (Agilent Technologies, Palo Alto, CA, USA) connected to ÄKTA Explorer 100 HPLC system (Pharmacia Biotech, CA, USA) in the order according to S1 Table. Depletion was performed according to the instructions provided by manufacturer with the exception that Buffer A in the supplied kit was replaced with PBS buffer (10 mM NaH2PO4/Na2HPO4, 3 mM KCl and 137 mM NaCl, pH 7.4) and Buffer B was replaced with 50 mM citrate buffer, pH 2.3.
Briefly, an aliquot of 500 μL of each CSF sample was dried using an ISS110 Speedvac system ISS110 (Thermo Scientific, Waltham, MA, USA). The dried samples were reconstituted in 100 μL of Buffer A and injected at 0.25 mL/min into the column equilibrated at room temperature with Buffer A. Fractions of 1 mL were collected and the depleted CSF was obtained in fractions 3 and 4, which were pooled together. After freezing 10 μL for protein estimation the remaining volume of the 2 mL pool was concentrated down under vacuum prior to protein digestion and MS analysis. After 5 mL washing with Buffer A, the eluent was changed to the pH 2.3 buffer and the flow rate was increased to 1 mL/min. The captured CSF proteins were released from the column and collected in fractions 10–12 and pooled. The pools were adjusted to neutral pH with 1 M NaOH to a final volume of 4.2 mL, and used only for protein estimation. Afterwards, the column was washed with Buffer A before next sample was injected.
Total protein concentration of the two pools from each CSF sample was estimated by an ultrasensitive total protein assay, the Dot-it-Spot-it® protein assay (http://dot-it-spot-it.com; Maplestone AB, Knivsta, Sweden) according to the instructions provided. Aliquots from the pools were diluted 1/10-1/40 in 0.75% SDS, 10 mM TRIS buffer pH 7.5, 0.15 M NaCl and 0.02% NaN3. Human albumin (Sigma) was used for calibration in the range 0.32–10 ug/mL. The total protein content of the crude CSF samples was measured after 200-times dilution of the sample. The diluted samples were dispensed in 1 μL aliquots on the detection sheets in 3 replicates and rapidly dried. The sheets were then placed in a large well with 1 mL of detection solution and incubated for 4 min, followed by 4 min incubation in 1 mL of washing solution. The absorbent sink was removed from the sheets. The sheets were then dried and mounted on the scanning template, which was detected by an Epson Expression 1600 Pro scanner (Epson, Long Beach, California, USA). The blackness intensity was quantified in each dedicated spot on the image with Image J (http://rsbweb.nih.gov/ij/). Protein concentrations were estimated by comparing the sample results with the outcome of the human albumin calibration curve using Rodbard curve fitting in Image J. The percentage of protein in the depleted fraction was calculated as the %-ratio between proteins in the depleted fraction/proteins in (depleted + released fraction).
The entire amount of protein in the depleted CSF sample was digested with trypsin. Briefly, the proteins were re-dissolved in 50 μL of digestion buffer (6 M urea, 100 mM TEAB). A volume of 10 μL of chicken ovalbumin solution (0.05 μg/μL) was added to each CSF sample. A volume of 15 μL of 45 mM aqueous DTT was added to all samples and the mixtures were incubated at 37°C for 2 hours to reduce the disulfide bridges. The samples were cooled to room temperature and 15 μL of 100 mM aqueous IAA was added before incubating the mixtures for an additional 40 min at room temperature in darkness to carabamidomethylate the cysteines. Afterwards, a volume of 50 μL of 100 mM TEAB was added to all the samples. Finally, trypsin/Lys-C mixture dissolved in 100 mM TEAB was added to the samples, yielding a final trypsin/protein concentration of 5% (w/w). The tryptic digestion was performed at 37°C overnight. Prior to mass spectrometry analysis, the peptides were purified and desalted on Isolute C18 solid phase extraction (SPE) columns (1 mL, 50 mg capacity, Biotage, Uppsala, Sweden) using the following schedule: The column was first wetted in 3×500 μL of 100% ACN and equilibrated with 3×500 μL 1% HAc. The tryptic peptides were adsorbed to the media using five repeated cycles of loading. The column was washed using 3×1 mL of 1% HAc and finally the peptides were eluted in 300 μL 50% ACN, 1% HAc. After desalting, the eluate was vacuum centrifuged to dryness and re-dissolved in 60 μL 0.1% trifluoroacetic acid prior to nano-LC-MS/MS.
The nanoLC-MS/MS experiments were performed using a 7 T hybrid LTQ FT mass spectrometer (ThermoFisher Scientific, Bremen, Germany) fitted with a nano-electrospray ionization (ESI) ion source. On-line nanoLC separations were performed using an Agilent 1100 nanoflow system (Agilent Technologies, Waldbronn, Germany). Each sample was analyzed by RP-nanoLC-MS/MS in duplicates (technical replicates) in the order according to S2 Table. The peptide separations were performed on in-house packed 15-cm fused silica emitters (75-μm inner diameter, 375-μm outer diameter). The emitters were packed with a methanol slurry of reversed-phase, fully end-capped Reprosil-Pur C18-AQ 3 μm resin (Dr. Maisch GmbH, Ammerbuch-Entringen, Germany) using a PC77 pressure injection cell (Next Advance, Averill Park, NY, USA). The injection volumes were 5 μL and corresponded to 2 μg of tryptic peptides. The separations were performed at a flow of 200 nL/min with mobile phases A (water with 0.5% acetic acid) and B (89.5% acetonitrile, 10% water, and 0.5% acetic acid). A 100-min gradient from 2% B to 50% B followed by a washing step with 98% B for 5 min was used. Mass spectrometric analyses were performed using unattended data-dependent acquisition mode, in which the mass spectrometer automatically switches between acquiring a high resolution survey mass spectrum in the FTMS (resolving power 50000 FWHM) and consecutive low-resolution, collision-induced dissociation fragmentation of up to five of the most abundant ions in the ion trap.
Antibody suspension bead arrays
A bead-based microarray platform was used as an orthogonal method for analysis of proteins selected based on the mass spectrometry results. All the 20 samples and a technical triplicate represented by a sample pool were processed and analyzed as previously described . In brief, crude CSF was diluted 1:2 and the protein content was labeled with biotin. In parallel, Human Protein Atlas antibodies (http://www.proteinatlas.org) generated towards the selected proteins were immobilized onto color-coded magnetic beads (Luminex corp.), each antibody assigned to a specific bead ID, and later combined into an array in suspension. The labeled samples were then further diluted 1:8 and heat treated at 56°C for 30 min followed by cooling to RT for 15 min before overnight incubation with the array. For readout, unbound proteins were washed off using a liquid handler (Biotek EL406) and detection mediated through a streptavidin-conjugated fluorophore (Invitrogen). At least 50 beads per identity were measured in a FM3D instrument (Luminex corp.) and the median fluorescence intensity (MFI) per bead and sample used for further analysis.
Statistical analysis on CSF protein amount.
The percentage of protein in the depleted CSF fraction was calculated as the ratio between protein amount in depleted fraction divided by the sum of amounts in both fractions, multiplied by hundred. A two-sample t-test was performed on percentage and amount of protein in depleted fraction as well as on the sum of total protein amount. The basic assumption of normally distributed residuals was not violated.
Mass spectrometry identification and quantification.
The following five mass spectrometry data processing programs were used to perform identification and quantification: Sieve v 2.1 (Thermo), DecyderMS v2.0 (GE healthcare), Maxquant , PEAKS (Bioinformatics Solutions Inc.) and OpenMS . The raw data was imported into the programs and retention time was cropped to the range from 1500 to 5400 seconds. The quantification was performed using the following parameters (default settings were used for unmentioned parameters): Sieve: low charge: 1, high charge: 4, retention time alignment window: 2 min, maximum number of frames: 7000; DecyderMS: ion peaks were automatically detected using a typical peak width of 0.4 min, signal to background threshold of 3, and uniform background subtraction. The resulting intensity maps were aligned using DeCyder MS 2.0, allowing a time tolerance of 2 min and m/z tolerance of 0.01 Da; Maxquant: Type: Standard, Multiplicity: 1, Match time: 1 min, alignment time window: 2 min; PEAKS: Retention time shift tolerance: 2 min, Mass error tolerance: 10 ppm; For OpenMS, we used an automated label free pipeline introduced by  using the following parameters: FeatureFinderCentroided: Mz tolerance: 0.07 Da, min spectra length: 6, max missing peaks: 3, slope bound of mass trace: 1, low charge: 1, high charge: 4, isotope Mz tolerance: 0.05; IDMapper: RT tolerance: 40 seconds, Mz tolerance: 10 ppm; MapAlignerIdentification: Mz: 10 ppm, RT: 120 seconds; FeatureLinkerUnlabeledQT: 10 ppm, RT: 60 seconds. The following software search engines were used to perform the identification: Mascot for Sieve and DecyderMS; Andromeda  for Maxquant; combination of SPIDER , PEAKS , and PEAKS DB  for PEAKS; and a combination of Xtandem  and omssa  for openMS (combined using “ConsensusID”  tool in openMS) using the default settings. For protein identification the UniProt/Swiss-Pro human database (release 2014_03, containing 20272 entries) with ovalbumin chicken protein sequence added to data database and combined with a decoy database (the sequences were reversed) was used; for identifying peptides the following settings were used for all the search engines: Enzyme: Trypsin, missed cleavages: 2 precursor mass tolerance: 10 ppm, fragment mass tolerance: 0.7 Da, minimum charge: 2, maximum charge: 3, fixed modifications: Carbamidomethyl (C), variable modifications: Oxidation (M) and Deamidated (N and Q). False discovery rate (FDR) was calculated based on the target/decoy database and the peptides with FDR lower than 0.05 were chosen as true positive hits (considering the risk of having one false positive in 20 observation). Peptides with FDR lower than 0.05 and log2 transformed data was used in all the subsequent analysis.
The five data analysis tools were compared in terms of number of identified and mapped peptides (peptides that were assigned to quantified features in MS intensity map), number of identified proteins, reproducibility between technical replicates (samples were analyzed in duplicate), unbiased separation of AD and healthy control using principal component analysis (PCA), and peptide level correlation to levels of the targeted proteins using antibody-based profiling.
To compare the number of identified peptides and proteins between the programs, only the peptides mapped to quantified features, with FDR lower than 0.05, and found in more than 90 percent of the replicates were included.
Reproducibility was measured as the coefficient of determination (R2) and variance ratio between the technical replicates calculated based on the peptides (non-normalized data) found in all the biological and technical replicates (full coverage) and across all software tested. The variance ratio for each sample (with two technical replicates, t1 and t2) was given by:
Where Xt1 and Xt2 are the vectors of peptide intensities, and are the mean of peptide intensities, and Nt1 and Nt2 are the number of peptides quantified in the first and the second technical replicate, respectively. A variance ratio close to one was regarded as a low ratio. To evaluate technical reproducibility and measure distance between AD and control samples, the peptide intensities were transformed using PCA. Mahalanobis distance between the technical replicates as well as between AD and control samples was calculated based on the first two components of the PCA result. Hypothesis testing was performed using Mahalanobis distance and p-values were derived using F distribution, showing cluster similarity between the technical replicates as well as between AD and control samples (A higher p-value indicates more cluster similarity) .
Disease related proteins.
To find proteins present with altered levels in AD patients compared to healthy controls, the data was first analyzed using reference normalization . The correlation between the technical replicates was estimated using “duplicateCorrelation”  and the “lmFit” function applied to fit multiple linear models using the “limma”  package in R . Finally, the “ebayes” function  was used to compute moderated t-statistics for comparison of AD versus control (assuming normal distribution of intensities). Using a liberal approach, peptides with p-value lower than 0.05 and with at least three observations (in 10 ADs and 10 controls) were selected for calculating proteins p-value and fold change (multiple isoforms of a proteins were regarded as different proteins). The significantly altered peptides (p-value <0.05) were manually curated based on the quality of quantified features, for OpenMS a specific cutoff of 0.2 for the feature quality score was used as quality cutoff and for remaining software all quantified features were visually inspected and incorrectly quantified or linked features were removed. For estimation of protein expression, the protein p-value was calculated as median of peptide p-values (as well as fisher's combined probability test ) and the protein fold change was calculated as median of peptide fold changes.
Antibody suspension bead arrays.
The Wilcoxon rank sum test was applied using raw data for group wise comparisons based on the antibody suspension bead array data and p-values lower than 0.05 were regarded as statistically significant. The log2 fold change was calculated from the ratio of medians in the AD group over controls and used for comparisons to the mass spectrometry data. For proteins with multiple antibodies, the one with the lowest p-value was selected for the comparison.
Correlation between mass spectrometry and antibody-based profiling.
In order to calculate the correlation between antibody-based profiling and MS the average of technical replicates was used. Pearson correlation coefficient between MS and antibody-based profiling was calculated between all the peptides (allowing no missing values) of a protein in MS and all the antibodies used for the corresponding protein in the antibody-based profiling method. The correlation was calculated for the raw and the three types of normalized MS data:
Where is the median of jth sample.
Where (sj − sref)median is the median of differences between sample sj and a reference sample (sref) which was selected as the sample with the lowest number of missing features.
Spiked-in normalization : peptides of chicken ovalbumin protein (which was spiked in the samples as an internal standard) were used to compute normalization factor for each sample separately for each of the five programs. The chicken ovalbumin peptides were first mean-centered across the samples:
Where cpij is intensity of ith chicken ovalbumin peptide in the jth sample and is mean of the ith peptide across all the samples. The peptide intensities in each sample were then subtracted by mean of the chicken ovalbumin intensities in the corresponding sample:
Where pij is the intensity of ith peptide in the jth sample and is mean of chicken ovalbumin peptides in the jth sample.
For each antibody, the highest correlating peptide in mass spectrometry was selected to examine the effect of normalization on the correlations.
CSF protein amount
The total amount of protein in the two pools obtained after the MARS Hu-7 column affinity purification of 0.5 mL CSF was 187±61 μg (mean±SD) for nine non-demented controls and 178±81 μg for ten AD patients, showing no significant difference. The result for one of the control samples was omitted, as the value of 587 μg was an outlier compared to the range 90–330 μg for the other 19 samples. Estimation of protein amount in crude CSF verified that this sample was aberrantly high compared to the other samples (we suspected that this sample was contaminated with blood and therefore, it was omitted only for statistical analysis performed on the protein amount but it was included in MS and antibody-based analysis).
The protein amount in the depleted CSF fraction (the flow-through fraction) was 46.2±17 μg and 30.8±14 μg for the control (n = 9) and AD patient (n = 10) group, respectively, showing a statistically significant difference (p-value = 0.046). Calculation of the fraction between unbound and bound proteins showed that 24.8±5.5% and 17.5±2.4% was obtained in the depleted CSF fraction for the control and AD patient groups, respectively. The difference between the groups was statistically significant (p-value = 0.003). No significant difference was found for the amount of proteins captured and released by the column (Fig 1).
The highest number of unique peptides was identified and mapped using PEAKS (Fig 2A). The number of identified and mapped peptides in PEAKS was almost twice as many compared to the other programs. Similar identification performance was observed using Maxquant and OpenMS. The lowest number of identified peptides was found using Sieve and DecyderMS. The highest number of proteins was identified using OpenMS, but PEAKS identified the highest number of proteins characterized with more than one peptide. The lowest number of proteins was mapped and identified using Sieve and DecyderMS. Only 173 proteins out of 894 proteins was identified and mapped across all five programs whereof the majority of the proteins were identified and mapped by only one program (Fig 2B and 2C).
The highest correlation between the technical replicates was found using Sieve (median R2 = 0.981). OpenMS (median R2 = 0.971), PEAKS (median R2 = 0.971), and DecyderMS (median R2 = 0.971) performed the second best. Maxquant (median R2 = 0.911) showed substantially lower correlations between the technical replicates compared to other programs. The variance ratio comparison also showed that Sieve (median ratio = 1.001) had the lowest ratio of variation between the technical replicates, followed by OpenMS (median ratio = 1.003), DecyderMS (median ratio = 1.004), PEAKS (median ratio = 1.033), and Maxquant (median ratio = 0.976) (S1A and S1B Fig).
In general, we observed high reproducibility between replicates as indicated by clear clustering of technical replicates using PCA (S2 Fig). However, in terms of Mahalanobis distance between the first two component of PCA, the highest reproducibility (higher p-value) was found using PEAKS and the lowest reproducibility was found using Maxquant (Fig 3). Furthermore, assuming a difference between AD and controls, OpenMS and Sieve showed slightly better separation between AD and controls compared to the other programs.
Using the combined results of all the five programs and after manual curation, we found 162 statistically significantly altered proteins (p-value<0.05) between Alzheimer’s disease and non-demented controls (S3 Table) where 31 proteins was identified and found to be statistically significantly differentially altered by all the five programs.
Effect of normalization on quantification.
In order to examine the effect of normalization methods on the number of statistically significantly altered proteins between AD and healthy controls, we compared the number of proteins with positive and negative fold changes in each program after applying each normalization method. Using the median or reference normalization, nearly the same proportion of the proteins with increased and decreased level was statistically significantly altered across all the programs. However, when the normalization was performed locally (spiked in normalization), the number of statistically significantly altered proteins with increased level was decreased and the number of proteins with decreased level was increased in all the programs (Fig 4). The overlap of significantly altered proteins between the programs was low using non normalized data (S3 Fig). Higher overlap was found using spiked-in and median normalization (S3B and S3C Fig) compared to reference normalization (S3D Fig).
Verification using antibody-based profiling.
At the time of analysis and to the best of our knowledge, 70 out of the 162 proteins have previously not been reported to be statistically significantly altered in levels between AD and healthy controls. Based on antibody availability in the Human Protein Atlas, 52 proteins (represented by in total 112 antibodies) were selected for analysis using the suspension bead array technology. After initial data quality control, one of the AD samples was excluded from further analysis.
The antibody-based analysis revealed 11 of the 52 proteins (S4 Table) as statistically significantly different (p-value<0.05) between AD and healthy controls in two repeated experiments with a median technical variation of 5%. For the majority of the proteins we found decreased level in AD samples compared to healthy controls. Fold change comparison of mass spectrometry (reference normalized data) and antibody-based profiling measurements revealed inconsistencies between the two techniques in the direction of the fold changes of 30 out of 52 targeted proteins (Fig 5A). However, when the data was normalized using spiked-in normalization method and the same peptides (as used for reference normalization) were used to compute the fold changes and p-values, we found that the fold changes for a majority of the proteins with inconsistencies were reduced to 20 proteins (Fig 5B). Furthermore, with the spiked-in normalization the number of statistically significant proteins was decreased from 52 to 17 proteins. Overall, with this method there were consistent fold changes between mass spectrometry and antibody-based profiling for 22 out of 24 proteins (Fig 5C). Taking both MS and antibody-based profiling into account, eight proteins were statistically significant (Table 2). Among the eight proteins, four proteins were found to be significantly differentially altered by at least two programs and the remaining proteins were only found by one program (Table 2). These proteins included leucine-rich α2 glycoprotein (LRG), apolipoprotein M (ApoM), complement C1q subcomponent (subunit B, C) (C1QB and C1QC), сomplement C1S (C1S), EGF-containing fibulin-like extracellular matrix protein 1 (fibulin-3, FIBL3), receptor-type tyrosine-protein phosphatase zeta (PTPRZ) and seizure protein 6 homolog (SEZ6), all displaying lower levels in AD compared to controls (Fig 6A–6H). Additionally, when the level of eight proteins in antibody-based analysis was transformed using PCA, we observed that the control sample with the abnormal concentration (labeled as C4) was clearly deviating from the other samples as well as the AD sample (AD9) with the longer duration of the disease (Fig 6I). Moreover, nearly the same pattern of deviation was found in the clustering of significantly differentially abundant peptides/peaks in the MS data (Fig 6J).
Correlation to antibody-based profiling.
Out of 52 targeted proteins, 27 proteins were identified and quantified by all the programs (S5 Table, bold and underlined entities represent the overlapping proteins across the programs). Comparing correlation distribution of the overlapping proteins from the three normalization methods revealed that performing local normalization (spiked-in) resulted in a considerable improvement of correlations between the mass spectrometry and antibody-based profiling measurements compared to raw data and global normalization methods (median correlation for each normalization: spiked-in: 0.826, raw data: 0.794, reference: 0.542, median: 0.553) irrespective of the software used for quantification (Fig 7; an example for correlation improvement is illustrated in S4 Fig). Furthermore, quantification using raw data led to higher correlations compared to the global normalization methods (reference and median normalization). The reference and median normalization methods resulted in similar correlations between mass spectrometry and antibody-based profiling measurements. Moreover, we observed similar correlation patterns in all the programs (median correlation using spiked-in normalization: PEAKS: 0.870, OpenMS: 0.836, Maxquant: 0.820, Sieve: 0.802, DecyderMS: 0.793). The same overall pattern of correlation was found when the programs were compared using all the proteins (S5 Fig).
Lower fraction of low-abundance proteins in Alzheimer’s disease
After depleting of the seven most abundant proteins from the CSF samples, the protein amount and the fraction of proteins in the unbound depleted fraction in the subject with AD was statistically significantly lower compared to the healthy elderly controls. The lower protein amount might reflect loss of several proteins due to the depletion procedures  but since the samples were randomized before depletion the possibility of protein loss only in the AD group is low, but it can not be ruled out. The difference in protein amount might also represent the age difference between the study groups , where the elderly healthy controls are on average nine years older compared to AD. However, previous studies are suggesting novel CSF AD biomarkers with lower levels in AD CSF compared to controls [47–51] that might reflect more general differences in protein amount between AD and non-AD.
Effects of initial data analysis on peptide identification and reproducibility
There are several programs developed for initial mass spectrometry data processing and protein identification. However, due to the different algorithms implemented in these programs, they might produce different and even contradictory results . Multiple studies have evaluated and compared different data processing and identification algorithms used in these programs [52–57]. We found that the PEAKS program identified and mapped considerably higher number of peptides compared to the other programs. PEAKS uses a combination of three identification search engines through the PEAKS database (a de novo sequencing database) . It has previously been reported that combining multiple search engines substantially increases the number of identified peptides and proteins , a finding in line with our observation that both PEAKS and OpenMS identified more peptides than Sieve and DecyderMS (which both use the Mascot search engine). Despite identifying substantially higher number of peptides using PEAKS, the number of identified proteins was similar to the other programs. We found that PEAKS search engine is capable of characterizing each unique protein by multiple peptides. In addition, we found that Maxquant identified higher number peptides compared to Sieve and DecyderMS which is in line with other studies by Cox, J., et al  and Colaert, N., et al using SILAC . However, we found that many of the peptides identified only in Maxquant (not in OpenMS) were in fact non-proteotypic peptides (the peptides that were assigned to more than one protein). After removing these peptides, the peptide identification performance was similar between the two programs. Finally, observed that the majority of the proteins were identified by one program. This can be explained by multiple factors such as different pre-processing methods, scoring functions as well as failure to build a mass feature for the peptides (unmapped peptides). This implies that to gain appropriate peptide/protein coverage, multiple programs can be employed not only for the identification but also for the quantification. As for identification there are several methods to score combined identification from multiple search engines . However, to the best of our knowledge there are no approaches to score or combine quantification results from several programs.
In label free shotgun proteomics, it is common to make technical replicates for each sample to increase the reliability of the downstream analysis. In general, all the programs produced acceptable correlation and low variation between the technical replicates also indicating high reproducibility of the MS analysis. Sieve produced the highest reproducibility and in agreement with previous studies, we also found that OpenMS produced higher reproducibility across technical replicates compared to Maxquant  as well as DecyderMS and PEAKS. Considering the variation ratio, PEAKS program was deviating from the other programs whereas it resulted in similar correlation pattern to the other programs, achieving higher correlation than DecyderMS and Maxquant. We found that these high correlations were caused by a number of peptides with large intensities and high correlation between the technical replicates that masked the low correlation of peptides with small abundance. Therefore, despite having high correlation between the technical replicates, variance ratio comparison resulted in large difference in variance in some of the technical replicate pairs in PEAKS program.
Data normalization affects biological conclusions
Normalization is performed to reduce the technical variation and remove bias caused by differences in protein concentration and other technical aspects of the MS analysis [40, 61]. Using relative protein levels generated by antibody-based profiling, we compared the mass spectrometry results using three different normalization methods (median, reference, and spiked-in normalization) based on data from five different data processing programs. We found that global normalization strategies resulted in low correlation to the affinity data. Most of the global normalization methods are based on the assumption that the distribution of proteins with increased and decreased abundance between groups of interest is nearly symmetrical . The violation of this assumption in the investigated CSF samples was not clear until removal of the seven most abundant proteins prior to MS analysis, which is contributing with more than 75% on the total protein mass. By performing local normalization based on a spiked-in protein, thus only correcting for experimental bias, the correlations to the antibody-based profiling results were substantially increased. It is important to note that the antibody-based profiling was performed on the crude and not depleted CSF. Spiked-in normalization resulted in a considerable reduction of fold changes comparing AD to non-AD, a more consistent result to that of antibody-based technique and this was found irrespective of the program used for the MS quantification. In addition, spiked-in normalization resulted in equal or higher overlap of result between the programs (despite finding lower number of altered proteins). This suggests that global normalization should be used with care when analyzing CSF in case of systematic differences between groups of interest can not be ruled out. We suggest adding one or several recombinant proteins from a different species than the investigated, which can be used for correcting for experimental bias and to investigate if the assumption for using global normalization is valid.
Disease associated proteins
Eight proteins were found to display concordant results using the two technologies, all displaying lower levels in the AD patients compared to the controls. Several of the reported proteins, including ApoM, LRG, FBLN3 and PTPRZ, have functions related to cell adhesion, migration, and morphology [62–71] and have been reported as important in the development of various cancer types [72–79] as well as diabetes [80, 81]. Also related to immune system, (C1QB and C1QC), complement C1S, and SEZ6 have been reported to be involved in synapse development [82–84]. However, the relation of these proteins to the neurological diseases is not understood. We are conducting a study with a large cohort to further verify the presence of these proteins in CSF.
S1 DataFile. Antibody based technique quantification data.
S2 DataFile. DecyderMS median normalized data quantification data.
S3 DataFile. DecyderMS raw data quantification data.
S4 DataFile. DecyderMS reference normalized data quantification data.
S5 DataFile. DecyderMS spiked in normalized data quantification data.
S6 DataFile. Maxquant median normalized data quantification data.
S7 DataFile. Maxquant raw data quantification data.
S8 DataFile. Maxquant reference normalized data quantification data.
S9 DataFile. Maxquant spiked in normalized data quantification data.
S10 DataFile. OpenMS median normalized data quantification data.
S11 DataFile. OpenMS raw data quantification data.
S12 DataFile. OpenMS reference normalized data quantification data.
S13 DataFile. OpenMS spiked in normalized data quantification data.
S14 DataFile. PEAKS median normalized data quantification data.
S15 DataFile. PEAKS raw data quantification data.
S16 DataFile. PEAKS reference normalized data quantification data.
S17 DataFile. PEAKS spiked in normalized data quantification data.
S18 DataFile. Sieve median normalized data quantification data.
S19 DataFile. Sieve raw data quantification data.
S20 DataFile. Sieve reference normalized data quantification data.
S21 DataFile. Sieve spiked in normalized data quantification data.
S22 DataFile. Identified and mapped (to a feature) proteins and peptides using DecyderMS software.
Charge state and mass/charge are not available for DecyderMS since the program combines the charge states.
S23 DataFile. Identified and mapped (to a feature) proteins and peptides using Maxquant software.
S24 DataFile. Identified and mapped (to a feature) proteins and peptides using OpenMS software.
S25 DataFile. Identified and mapped (to a feature) proteins and peptides using PEAKS software.
S26 DataFile. Identified and mapped (to a feature) proteins and peptides using Sieve software.
S1 Fig. Reproducibility comparison between five data processing programs.
(A) Distribution of coefficient of determination between the technical replicates in five mass spectrometry data processing programs. The higher the correlation the closer the replicates quantification. (B) Distribution of variation ratios between the technical replicates in each tool. The closer the values to 1 the lower the variation between the technical replicates.
S2 Fig. Reproducibility of mass spectrometry experiment.
PCA of peptide intensities showing how study groups (AD: Alzheimer’s disease; C: healthy control) and the technical replicates (the number after underline) are clustered. (A) DecyderMS. (B) Maxquant. (C) OpenMS. (D) PEAKS. (E) Sieve.
S3 Fig. Overlap of significantly altered proteins between different programs using raw data and three normalization methods.
(A) Raw data. (B) Spiked-in normalization. (C) Median normalization. (D) Reference normalization.
S4 Fig. An example of correlation improvement using three normalization method and raw data.
Scatter plot of the highest correlated peptide between mass spectrometry and luminex (Protein P3IP1). Protein names are shown as Uniprot ID. (A) Reference normalization. (B) Median normalization. (C) Raw data. (D) Spiked-in normalization. Abb. ABA: antibody-based analysis.
S5 Fig. Distribution of all the highest correlated peptide-antibody pairs between mass spectrometry and antibody-based analysis.
The results obtained for the five programs and three normalization methods used were correlated to the antibody-based analysis.
S1 Table. Depletion setup for CSF samples.
S2 Table. LC-MS/MS run order CSF samples.
S3 Table. Differentially altered proteins.
S4 Table. The proteins with statistically significantly altered levels in the antibody-based profiling.
S5 Table. Correlation of the highest correlated peptide-antibody pair between Mass spectrometry and antibody-based technique.
The authors wish to thank the entire staff of the Human Protein Atlas for their efforts to produce the antibodies included in the study.
Conceived and designed the experiments: PEK AH LL JB MI PN KK GS. Performed the experiments: PEK AH ML MM LK KK GS. Analyzed the data: PEK AH MM KK. Contributed reagents/materials/analysis tools: PEK ML LL JB MI PN KK. Wrote the paper: PEK AH ML MM MI PN KK GS.
- 1. Terry RD. Alzheimer's disease and the aging brain. Journal of geriatric psychiatry and neurology. 2006;19(3):125–8. Epub 2006/08/02. doi: 10.1177/0891988706291079 pmid:16880353.
- 2. Banerjee S. The Macroeconomics of Dementia-Will the World Economy Get Alzheimer's Disease? Arch Med Res. 2012;43(8):705–9. pmid:WOS:000314378000017. doi: 10.1016/j.arcmed.2012.10.006.
- 3. Ferri CP, Prince M, Brayne C, Brodaty H, Fratiglioni L, Ganguli M, et al. Global prevalence of dementia: a Delphi consensus study. Lancet. 2005;366(9503):2112–7. pmid:WOS:000234146000026.
- 4. Blennow K, de Leon MJ, Zetterberg H. Alzheimer's disease. Lancet. 2006;368(9533):387–403. pmid:WOS:000239441900033.
- 5. Blennow K, Hampel H, Weiner M, Zetterberg H. Cerebrospinal fluid and plasma biomarkers in Alzheimer disease. Nat Rev Neurol. 2010;6(3):131–44. pmid:WOS:000275290200004. doi: 10.1038/nrneurol.2010.4.
- 6. Kang JH, Korecka M, Toledo JB, Trojanowski JQ, Shaw LM. Clinical Utility and Analytical Challenges in Measurement of Cerebrospinal Fluid Amyloid-beta(1–42) and tau Proteins as Alzheimer Disease Biomarkers. Clin Chem. 2013;59(6):903–16. pmid:WOS:000321548500009. doi: 10.1373/clinchem.2013.202937.
- 7. Mattsson N, Portelius E, Rolstad S, Gustavsson M, Andreasson U, Stridsberg M, et al. Longitudinal cerebrospinal fluid biomarkers over four years in mild cognitive impairment. Journal of Alzheimer's disease: JAD. 30. Netherlands2012. p. 767–78. doi: 10.3233/JAD-2012-120019. pmid:22475796
- 8. Hansson O, Buchhave P, Zetterberg H, Blennow K, Minthon L, Warkentin S. Combined rCBF and CSF biomarkers predict progression from mild cognitive impairment to Alzheimer's disease. Neurobiol Aging. 30. United States2009. p. 165–73. pmid:17646035
- 9. Tumani H, Teunissen C, Sossmuth S, Otto M, Ludolph AC, Brettschneider J. Cerebrospinal fluid biomarkers of neurodegeneration in chronic neurological diseases. Expert Rev Mol Diagn. 2008;8(4):479–94. pmid:WOS:000257765700013. doi: 10.1586/14737126.96.36.1999.
- 10. Garton MJ, Keir G, Lakshmi MV, Thompson EJ. Age-related changes in cerebrospinal fluid protein concentrations. Journal of the neurological sciences. 104. Netherlands1991. p. 74–80. pmid:1717663
- 11. Yin GN, Lee HW, Cho JY, Suk K. Neuronal pentraxin receptor in cerebrospinal fluid as a potential biomarker for neurodegenerative diseases. Brain Res. 2009;1265:158–70. pmid:WOS:000265673000017. doi: 10.1016/j.brainres.2009.01.058.
- 12. Henkel AW, Muller K, Lewczuk P, Muller T, Marcus K, Kornhuber J, et al. Multidimensional plasma protein separation technique for identification of potential Alzheimer's disease plasma biomarkers: a pilot study. J Neural Transm. 2012;119(7):779–88. pmid:WOS:000305525800007. doi: 10.1007/s00702-012-0781-3.
- 13. Song F, Poljak A, Smythe GA, Sachdev P. Plasma biomarkers for mild cognitive impairment and Alzheimer's disease. Brain Res Rev. 2009;61(2):69–80. pmid:WOS:000271169000003. doi: 10.1016/j.brainresrev.2009.05.003.
- 14. Beranova-Giorgianni S. Proteome analysis by two-dimensional gel electrophoresis and mass spectrometry: strengths and limitations. 2003;22(5):273–81. doi: 10.1016/S0165-9936(03)00508-9.
- 15. Abdi F, Quinn JF, Jankovic J, McIntosh M, Leverenz JB, Peskind E, et al. Detection of biomarkers with a multiplex quantitative proteomic platform in cerebrospinal fluid of patients with neurodegenerative disorders. J Alzheimers Dis. 2006;9(3):293–348. pmid:WOS:000241016700009.
- 16. Musunuri S, Wetterhall M, Ingelsson M, Lannfelt L, Artemenko K, Bergquist J, et al. Quantification of the Brain Proteome in Alzheimer's Disease Using Multiplexed Mass Spectrometry. J Proteome Res. 2014;13(4):2056–68. pmid:WOS:000334016400025. doi: 10.1021/pr401202d.
- 17. Donovan LE, Higginbotham L, Dammer EB, Gearing M, Rees HD, Xia QW, et al. Analysis of a membrane-enriched proteome from postmortem human brain tissue in Alzheimer's disease. Proteom Clin Appl. 2012;6(3–4):201–11. pmid:WOS:000303194800007.
- 18. Andreev VP, Petyuk VA, Brewer HM, Karpievitch YV, Xie F, Clarke J, et al. Label-Free Quantitative LC-MS Proteomics of Alzheimer's Disease and Normally Aged Human Brains. J Proteome Res. 2012;11(6):3053–67. pmid:WOS:000304682500001. doi: 10.1021/pr3001546.
- 19. Shevchenko G, Wetterhall M, Bergquist J, Hoglund K, Andersson LI, Kultima K. Longitudinal Characterization of the Brain Proteomes for the Tg2576 Amyloid Mouse Model Using Shotgun Based Mass Spectrometry. J Proteome Res. 2012;11(12):6159–74. pmid:WOS:000311925900051. doi: 10.1021/pr300808h.
- 20. Chen S, Lu FF, Seeman P, Liu F. Quantitative Proteomic Analysis of Human Substantia Nigra in Alzheimer's Disease, Huntington's Disease and Multiple Sclerosis. Neurochem Res. 2012;37(12):2805–13. pmid:WOS:000311537300018. doi: 10.1007/s11064-012-0874-2.
- 21. Echan LA, Tang HY, Ali-Khan N, Lee K, Speicher DW. Depletion of multiple high-abundance proteins improves protein profiling capacities of human serum and plasma. Proteomics. 2005;5(13):3292–303. Epub 2005/07/30. doi: 10.1002/pmic.200401228 pmid:16052620.
- 22. Open source libraries and frameworks for mass spectrometry based proteomics: A developer's perspective ☆. 2014;1844(1):63–76. doi: 10.1016/j.bbapap.2013.02.032.
- 23. McHugh L, Arthur JW. Computational Methods for Protein Identification from Mass Spectrometry Data. PLoS Comput Biol. 42008.
- 24. Cappadona S, Baker PR, Cutillas PR, Heck AJ, van Breukelen B. Current challenges in software solutions for mass spectrometry-based quantitative proteomics. Amino acids. 2012;43(3):1087–108. Epub 2012/07/24. doi: 10.1007/s00726-012-1289-8 pmid:22821268; PubMed Central PMCID: PMCPmc3418498.
- 25. Jahn H, Wittke S, Zurbig P, Raedler TJ, Arlt S, Kellmann M, et al. Peptide fingerprinting of Alzheimer's disease in cerebrospinal fluid: identification and prospective evaluation of new synaptic biomarkers. PloS one. 6. United States2011. p. e26540. doi: 10.1371/journal.pone.0026540. pmid:22046305
- 26. Calza S, Valentini D, Pawitan Y. Normalization of oligonucleotide arrays based on the least-variant set of genes. BMC Bioinformatics. 2008;9(1):140. doi: info:pmid/18318917.
- 27. Teunissen CE, Petzold A, Bennett JL, Berven FS, Brundin L, Comabella M, et al. A consensus protocol for the standardization of cerebrospinal fluid collection and biobanking. Neurology. 2009;73(22):1914–22. pmid:WOS:000272205200016. doi: 10.1212/WNL.0b013e3181c47cc2.
- 28. Haggmark A, Bystrom S, Ayoglu B, Qundos U, Uhlen M, Khademi M, et al. Antibody-based profiling of cerebrospinal fluid within multiple sclerosis. Proteomics. 2013;13(15):2256–67. Epub 2013/05/23. doi: 10.1002/pmic.201200580 pmid:23696371.
- 29. Cox J, Mann M. MaxQuant enables high peptide identification rates, individualized p.p.b.-range mass accuracies and proteome-wide protein quantification. Nature Biotechnology. 2008;26(12):1367–72. doi: 10.1038/nbt.1511. pmid:19029910
- 30. Sturm M, Bertsch A, Gröpl C, Hildebrandt A, Hussong R, Lange E, et al. OpenMS–An open-source software framework for mass spectrometry. BMC Bioinformatics: BioMed Central Ltd; 2008. p. 163.
- 31. Weisser H, Nahnsen S, Grossmann J, Nilse L, Quandt A, Brauer H, et al. An automated pipeline for high-throughput label-free quantitative proteomics. J Proteome Res. 2013;12(4):1628–44. Epub 2013/02/09. doi: 10.1021/pr300992u pmid:23391308.
- 32. Cox J, Neuhauser N, Michalski A, Scheltema RA, Olsen JV, Mann M. Andromeda: a peptide search engine integrated into the MaxQuant environment. J Proteome Res. 2011;10(4):1794–805. Epub 2011/01/25. doi: 10.1021/pr101065j pmid:21254760.
- 33. Han Y, Ma B, Zhang K. SPIDER: software for protein identification from sequence tags with de novo sequencing error. Journal of bioinformatics and computational biology. 2005;3(3):697–716. Epub 2005/08/19. pmid:16108090.
- 34. Ma B, Zhang K, Hendrie C, Liang C, Li M, Doherty-Kirby A, et al. PEAKS: powerful software for peptide de novo sequencing by tandem mass spectrometry. Rapid communications in mass spectrometry: RCM. 2003;17(20):2337–42. Epub 2003/10/15. doi: 10.1002/rcm.1196 pmid:14558135.
- 35. Zhang J, Xin L, Shan B, Chen W, Xie M, Yuen D, et al. PEAKS DB: De Novo Sequencing Assisted Database Search for Sensitive and Accurate Peptide Identification*. Molecular & cellular proteomics: MCP. 112012.
- 36. Craig R, Beavis RC. TANDEM: matching proteins with tandem mass spectra. Bioinformatics (Oxford, England). 2004;20(9):1466–7. Epub 2004/02/21. doi: 10.1093/bioinformatics/bth092 pmid:14976030.
- 37. Geer LY, Markey SP, Kowalak JA, Wagner L, Xu M, Maynard DM, et al. Open mass spectrometry search algorithm. J Proteome Res. 2004;3(5):958–64. Epub 2004/10/12. doi: 10.1021/pr0499491 pmid:15473683.
- 38. Nahnsen S, Bertsch A, Rahnenführer J, Nordheim A, Kohlbacher O. Probabilistic Consensus Scoring Improves Tandem Mass Spectrometry Peptide Identification. J Proteome Res. 2011. doi: 10.1021/pr2002879.
- 39. Worley B, Halouska S, Powers R. Utilities for quantifying separation in PCA/PLS-DA scores plots. Analytical biochemistry. 2013;433(2):102–4. Epub 2012/10/20. doi: 10.1016/j.ab.2012.10.011 pmid:23079505; PubMed Central PMCID: PMCPmc3534867.
- 40. Kultima K, Nilsson A, Scholz B, Rossbach UL, Falth M, Andren PE. Development and evaluation of normalization methods for label-free relative quantification of endogenous peptides. Molecular & cellular proteomics: MCP. 2009;8:2285–95. Epub 07/15. doi: 10.1074/mcp.M800514-MCP200 10.1074/mcp.M800514-MCP200. Epub 2009 Jul 12 pmid:19596695.
- 41. Smyth GK, Michaud J, Scott HS. Use of within-array replicate spots for assessing differential expression in microarray experiments. Bioinformatics (Oxford, England). 2005;21(9):2067–75. Epub 2005/01/20. doi: 10.1093/bioinformatics/bti270 pmid:15657102.
- 42. Smyth GK. limma: Linear Models for Microarray Data. 2014:397–420. doi: 10.1007/0-387-29362-0_23.
- 43. Zhou A, Brewer KD, Rizo J. Analysis of SNARE complex/synaptotagmin-1 interactions by one-dimensional NMR spectroscopy. Biochemistry. 2013;52(20):3446–56. doi: 10.1021/bi400230u pmid:23617808; PubMed Central PMCID: PMC3812274.
- 44. Smyth GK. Linear models and empirical bayes methods for assessing differential expression in microarray experiments. Statistical applications in genetics and molecular biology. 2004;3:Article3. Epub 2006/05/02. doi: 10.2202/1544-6115.1027 pmid:16646809.
- 45. Fisher RA. Statistical Methods for Research Workers. 1925.
- 46. Günther R, Krause E, Schümann M, Blasig IE, Haseloff RF. Depletion of highly abundant proteins from human cerebrospinal fluid: a cautionary note. Molecular neurodegeneration. 102015.
- 47. Hansson SF, Andreasson U, Wall M, Skoog I, Andreasen N, Wallin A, et al. Reduced levels of amyloid-beta-binding proteins in cerebrospinal fluid from Alzheimer's disease patients. Journal of Alzheimer's disease: JAD. 2009;16(2):389–97. Epub 2009/02/18. doi: 10.3233/jad-2009-0966 pmid:19221428.
- 48. Toledo JB, Korff A, Shaw LM, Trojanowski JQ, Zhang J, Initiative tAsDN. Low levels of cerebrospinal fluid complement 3 and factor H predict faster cognitive decline in mild cognitive impairment. Alzheimer's Research & Therapy. 2014;6(3):36. doi: info:pmid/25478014.
- 49. Podlesniy P, Figueiro-Silva J, Llado A, Antonell A, Sanchez-Valle R, Alcolea D, et al. Low cerebrospinal fluid concentration of mitochondrial DNA in preclinical Alzheimer disease. Annals of neurology. 2013;74(5):655–68. Epub 2013/06/25. doi: 10.1002/ana.23955 pmid:23794434.
- 50. Puchades M, Hansson SF, Nilsson CL, Andreasen N, Blennow K, Davidsson P. Proteomic studies of potential cerebrospinal fluid protein markers for Alzheimer's disease. Brain research Molecular brain research. 2003;118(1–2):140–6. Epub 2003/10/16. pmid:14559363.
- 51. Simonsen AH, Hagnelius N-O, Waldemar G, Nilsson TK, McGuire J. Protein Markers for the Differential Diagnosis of Vascular Dementia and Alzheimer’s Disease. International Journal of Proteomics. 2012;2012. doi: 10.1155/2012/824024.
- 52. Coble JB, Fraga CG. Comparative evaluation of preprocessing freeware on chromatography/mass spectrometry data for signature discovery. Journal of chromatography A. 2014;1358:155–64. Epub 2014/07/27. doi: 10.1016/j.chroma.2014.06.100 pmid:25063004.
- 53. Moulder R, Filen JJ, Salmi J, Katajamaa M, Nevalainen OS, Oresic M, et al. A comparative evaluation of software for the analysis of liquid chromatography-tandem mass spectrometry data from isotope coded affinity tag experiments. Proteomics. 2005;5(11):2748–60. Epub 2005/06/14. doi: 10.1002/pmic.200401187 pmid:15952233.
- 54. America AH, Cordewener JH. Comparative LC-MS: a landscape of peaks and valleys. Proteomics. 2008;8(4):731–49. Epub 2008/02/26. doi: 10.1002/pmic.200700694 pmid:18297651.
- 55. Brosch M, Swamy S, Hubbard T, Choudhary J. Comparison of Mascot and X!Tandem performance for low and high accuracy mass spectrometry and the development of an adjusted Mascot threshold. Molecular & cellular proteomics: MCP. 2008;7(5):962–70. Epub 2008/01/25. doi: 10.1074/mcp.M700293-MCP200 pmid:18216375; PubMed Central PMCID: PMCPmc2656932.
- 56. Chamrad DC, Korting G, Stuhler K, Meyer HE, Klose J, Bluggel M. Evaluation of algorithms for protein identification from sequence databases using mass spectrometry data. Proteomics. 2004;4(3):619–28. Epub 2004/03/05. doi: 10.1002/pmic.200300612 pmid:14997485.
- 57. Johnsen LG, Skov T, Houlberg U, Bro R. An automated method for baseline correction, peak finding and peak grouping in chromatographic data. The Analyst. 2013;138(12):3502–11. Epub 2013/05/15. doi: 10.1039/c3an36276k pmid:23665697.
- 58. Shteynberg D, Nesvizhskii AI, Moritz RL, Deutsch EW. Combining Results of Multiple Search Engines in Proteomics*. Molecular & cellular proteomics: MCP. 122013. p. 2383–93.
- 59. Colaert N, Vandekerckhove J, Martens L, Gevaert K. A case study on the comparison of different software tools for automated quantification of peptides. Methods in molecular biology (Clifton, NJ). 2011;753:373–98. Epub 2011/05/24. doi: 10.1007/978-1-61779-148-2_25 pmid:21604136.
- 60. Sturm M, Bertsch A, Gropl C, Hildebrandt A, Hussong R, Lange E, et al. OpenMS—an open-source software framework for mass spectrometry. BMC Bioinformatics. 2008;9:163. doi: 10.1186/1471-2105-9-163 pmid:18366760; PubMed Central PMCID: PMC2311306.
- 61. Quackenbush J. Microarray data normalization and transformation. Nature genetics. 2002;32 Suppl:496–501. Epub 2002/11/28. doi: 10.1038/ng1032 pmid:12454644.
- 62. Ai J, Druhan LJ, Hunter MG, Loveland MJ, Avalos BR. LRG-accelerated differentiation defines unique G-CSFR signaling pathways downstream of PU.1 and C/EBPepsilon that modulate neutrophil activation. Journal of leukocyte biology. 2008;83(5):1277–85. Epub 2008/02/15. doi: 10.1189/jlb.1107751 pmid:18272588; PubMed Central PMCID: PMCPmc2376838.
- 63. Takahashi N, Takahashi Y, Putnam FW. Periodicity of leucine and tandem repetition of a 24-amino acid segment in the primary structure of leucine-rich alpha 2-glycoprotein of human serum. Proceedings of the National Academy of Sciences of the United States of America. 1985;82(7):1906–10. Epub 1985/04/01. pmid:3856868; PubMed Central PMCID: PMCPmc397442.
- 64. Kobe B, Kajava AV. The leucine-rich repeat as a protein recognition motif. Current opinion in structural biology. 2001;11(6):725–32. Epub 2001/12/26. pmid:11751054.
- 65. Timpl R, Sasaki T, Kostka G, Chu ML. Fibulins: a versatile family of extracellular matrix proteins. Nature reviews Molecular cell biology. 2003;4(6):479–89. Epub 2003/06/05. doi: 10.1038/nrm1130 pmid:12778127.
- 66. de Vega S, Iwamoto T, Yamada Y. Fibulins: multiple roles in matrix structures and tissue functions. Cellular and molecular life sciences: CMLS. 2009;66(11–12):1890–902. Epub 2009/02/04. doi: 10.1007/s00018-009-8632-6 pmid:19189051.
- 67. Van Vactor D. Protein tyrosine phosphatases in the developing nervous system. Current opinion in cell biology. 1998;10(2):174–81. Epub 1998/04/30. pmid:9561841.
- 68. Stoker AW. Protein tyrosine phosphatases and signalling. The Journal of endocrinology. 2005;185(1):19–33. Epub 2005/04/09. doi: 10.1677/joe.1.06069 pmid:15817824.
- 69. Kwon SK, Woo J, Kim SY, Kim H, Kim E. Trans-synaptic adhesions between netrin-G ligand-3 (NGL-3) and receptor tyrosine phosphatases LAR, protein-tyrosine phosphatase delta (PTPdelta), and PTPsigma via specific domains regulate excitatory synapse formation. The Journal of biological chemistry. 2010;285(18):13966–78. Epub 2010/02/09. doi: 10.1074/jbc.M109.061127 pmid:20139422; PubMed Central PMCID: PMCPmc2859559.
- 70. Bixby JL. Receptor tyrosine phosphatases in axon growth and guidance. Neuroreport. 2000;11(10):R5–10. Epub 2000/08/03. pmid:10923644.
- 71. Arregui CO, Balsamo J, Lilien J. Regulation of signaling by protein-tyrosine phosphatases: potential roles in the nervous system. Neurochem Res. 2000;25(1):95–105. Epub 2000/02/24. pmid:10685609.
- 72. Andersen JD, Boylan KL, Jemmerson R, Geller MA, Misemer B, Harrington KM, et al. Leucine-rich alpha-2-glycoprotein-1 is upregulated in sera and tumors of ovarian cancer patients. Journal of ovarian research. 2010;3:21. Epub 2010/09/14. doi: 10.1186/1757-2215-3-21 pmid:20831812; PubMed Central PMCID: PMCPmc2949730.
- 73. Luo R, Zhang M, Liu L, Lu S, Zhang CZ, Yun J. Decrease of fibulin-3 in hepatocellular carcinoma indicates poor prognosis. PloS one. 2013;8(8):e70511. Epub 2013/08/13. doi: 10.1371/journal.pone.0070511 pmid:23936443; PubMed Central PMCID: PMCPmc3731361.
- 74. Hwang CF, Chien CY, Huang SC, Yin YF, Huang CC, Fang FM, et al. Fibulin-3 is associated with tumour progression and a poor prognosis in nasopharyngeal carcinomas and inhibits cell migration and invasion via suppressed AKT activity. The Journal of pathology. 2010;222(4):367–79. Epub 2010/10/12. doi: 10.1002/path.2776 pmid:20927779.
- 75. Xu S, Yang Y, Sun YB, Wang HY, Sun CB, Zhang X. Role of fibulin-3 in lung cancer: in vivo and in vitro analyses. Oncology reports. 2014;31(1):79–86. Epub 2013/10/22. doi: 10.3892/or.2013.2799 pmid:24142183.
- 76. Perez-Pinera P, Garcia-Suarez O, Menendez-Rodriguez P, Mortimer J, Chang Y, Astudillo A, et al. The receptor protein tyrosine phosphatase (RPTP)beta/zeta is expressed in different subtypes of human breast cancer. Biochemical and biophysical research communications. 2007;362(1):5–10. Epub 2007/08/21. doi: 10.1016/j.bbrc.2007.06.050 pmid:17706593; PubMed Central PMCID: PMCPmc2084077.
- 77. Kairouz R, Daly RJ. Tyrosine kinase signalling in breast cancer: modulation of tyrosine kinase signalling in human breast cancer through altered expression of signalling intermediates. Breast cancer research: BCR. 2000;2(3):197–202. Epub 2001/03/16. pmid:11250710; PubMed Central PMCID: PMCPmc138775.
- 78. Wu CW, Chi CW, Lin WC. Gastric cancer: prognostic and diagnostic advances. Expert reviews in molecular medicine. 2002;4(6):1–12. Epub 2004/02/28. doi: 10.1017/S1462399402004337 pmid:14987390.
- 79. Lin W, Kao HW, Robinson D, Kung HJ, Wu CW, Chen HC. Tyrosine kinases and gastric cancer. Oncogene. 2000;19(49):5680–9. Epub 2000/12/15. doi: 10.1038/sj.onc.1203924 pmid:11114748.
- 80. Plomgaard P, Dullaart RP, de Vries R, Groen AK, Dahlback B, Nielsen LB. Apolipoprotein M predicts pre-beta-HDL formation: studies in type 2 diabetic and nondiabetic subjects. Journal of internal medicine. 2009;266(3):258–67. Epub 2009/05/22. doi: 10.1111/j.1365-2796.2009.02095.x pmid:19457058.
- 81. Zheng L, Feng Y, Shi Y, Zhang J, Mu Q, Qin L, et al. Intralipid decreases apolipoprotein M levels and insulin sensitivity in rats. PloS one. 2014;9(8):e105681. Epub 2014/08/22. doi: 10.1371/journal.pone.0105681 pmid:25144649; PubMed Central PMCID: PMCPmc4140822.
- 82. Stevens B, Allen NJ, Vazquez LE, Howell GR, Christopherson KS, Nouri N, et al. The classical complement cascade mediates CNS synapse elimination. Cell. 2007;131(6):1164–78. Epub 2007/12/18. doi: 10.1016/j.cell.2007.10.036 pmid:18083105.
- 83. Masliah E, Mallory M, Alford M, DeTeresa R, Hansen LA, McKeel DW Jr., et al. Altered expression of synaptic proteins occurs early during progression of Alzheimer's disease. Neurology. 2001;56(1):127–9. Epub 2001/01/10. pmid:11148253.
- 84. Boulanger LM. Immune proteins in brain development and synaptic plasticity. Neuron. 2009;64(1):93–109. Epub 2009/10/21. doi: 10.1016/j.neuron.2009.09.001 pmid:19840552.