| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Oak Ridge Conference |
1 PerkinElmer Life and Analytical Sciences, Boston, MA.
2 PerkinElmer Life and Analytical Sciences, Shelton, CT.
3 Rush Alzheimer’s Disease Center, Rush University Medical Center, Chicago, IL.
4 PerkinElmer Life and Analytical Sciences, Beaconsfield, Bucks, United Kingdom.
5 Clinical Proteomics Reference Laboratory, SAIC-Frederick, Inc., NCI Frederick, Gaithersburg, MD.
6 Predictive Diagnostics, Inc., Vacaville, CA.
aAddress correspondence to this author at: PerkinElmer Life and Analytical Sciences, 249 Albany St., Boston, MA 02118. Fax 617-350-9658; e-mail mary.lopez{at}perkinelmer.com.
Abstract
Background: Researchers typically search for disease markers using a "targeted" approach in which a hypothesis about the disease mechanism is tested and experimental results either confirm or disprove the involvement of a particular gene or protein in the disease. Recently, there has been interest in developing disease diagnostics based on unbiased quantification of differences in global patterns of protein and peptide masses, typically in blood from individuals with and without disease. We combined a suite of methods and technologies, including novel sample preparation based on carrier-protein capture and biomarker enrichment, high-resolution mass spectrometry, a unique cohort of well-characterized persons with and without Alzheimer disease (AD), and powerful bioinformatic analysis, that add statistical and procedural robustness to biomarker discovery from blood.
Methods: Carrier-protein–bound peptides were isolated from serum samples by affinity chromatography, and peptide mass spectra were acquired by a matrix-assisted laser desorption/ionization (MALDI) orthogonal time-of-flight (O-TOF) mass spectrometer capable of collecting data over a broad mass range (100 to >300 000 Da) in a single acquisition. Discriminatory analysis of mass spectra was used to process and analyze the raw mass spectral data.
Results: Coupled with the biomarker enrichment protocol, the high-resolution MALDI O-TOF mass spectra provided informative, reproducible peptide signatures. The raw mass spectra were analyzed and used to build discriminant disease models that were challenged with blinded samples for classification.
Conclusions: Carrier-protein enrichment of disease biomarkers coupled with high-resolution mass spectrometry and discriminant pattern analysis is a powerful technology for diagnostics and population screening. The mass fingerprint model successfully classified blinded AD patient and control samples with high sensitivity and specificity.
The discovery of accurate diagnostic markers for disease states is one of the most pressing problems in biology and medicine. Many of the most devastating diseases, such as cancers, have far better prognoses if detected early (1)(2)(3)(4)(5). The availability of noninvasive, quick diagnostic tests also facilitates the development and monitoring of treatments and drugs. Unfortunately, the biological complexity of many diseases has precluded the identification of single diagnostic markers with a high degree of confidence. Often cited as cases in point are, CA125 and prostate-specific antigen, the currently preferred tests for ovarian and prostate cancers, respectively (6)(7)(8). These tests have limited utility because of high false-positive and -negative rates, but to date, there are no better alternatives (6)(7)(8).
Typically, most researchers have searched for disease markers using a "targeted" approach in which a hypothesis about the disease mechanism is tested and experimental results either confirm or disprove the involvement of a particular gene or protein in the disease. This approach can be very fruitful, particularly in delivering mechanistic information on disease processes. However, the targeted approach also is inherently biased in that it necessarily starts from a model generated by the researcher based on the currently available information. This limits the complexity of the models that can be proposed and produces a reductionist view of disease that may underestimate the global interaction of several systems or biochemical pathways.
With the advent of advanced mass spectrometry (MS) 1 technologies offering the ability to investigate proteins and peptides over a broad range of molecular weights, there has been interest in developing disease diagnostics based on unbiased quantification of the differences in global patterns of protein and peptide masses, typically in blood from individuals with and without disease (9)(10)(11)(12)(13). Advantages to this approach include that it does not presume any mechanism and should therefore be unbiased with respect to the identification of disease-related proteins or peptides. Thus, it offers the possibility of including proteins and peptides in the diagnostic that are not known to be involved in the disease process. Disadvantages include that any error in data collection, processing, or analysis may lead to the identification of artifacts instead of true distinguishing features. The method also assumes accurate and unbiased identification of phenotype. Finally, the approach is data-derived such that given a sufficiently large pool of potential proteins and peptides, one is almost certain to identify a pattern that discriminates between persons with and without disease within any given data set. This can be mitigated by (a) proper sample collection and handling procedures and a robust, reproducible processing and analysis protocol that can identify and reject poor-quality data; (b) detailed evaluation of individuals to be included in the analyses; and (c) adequate attention to statistical techniques. The true power in this approach lies in its ability to potentially uncover previously unexpected metabolic connections in a particular disease. In addition, the collection of features that provide a fingerprint for a disease state have the potential to be far more accurate than a single marker, which may vary considerably because of different genetic and biological backgrounds. Several reports have focused on the discriminant mass pattern analysis approach mainly for the diagnosis of various cancers with promising, and sometimes controversial, results (14). Most of the controversy has centered on the availability, resolution, and analysis of the raw data, not the validity of the concept (15)(16). This is, in many ways, similar to the controversy that has dogged genomic array research with respect to statistical analysis and validity. As with any disruptive technology that leads to a paradigm shift, there are clear technical hurdles to be overcome before discriminant mass pattern analysis is useful in a diagnostic setting.
The current study brings together a suite of methods and technologies, including novel sample preparation based on carrier-protein capture and enrichment of biomarkers (17), high-resolution mass spectrometric detection, and bioinformatics analysis that adds statistical and procedural robustness to biomarker discovery from blood, and a unique cohort of well-characterized persons with and without Alzheimer disease (AD). This technologic platform was applied to the discovery of discriminant features in AD, for which no definitive early diagnostic test is available.
Clinically, AD is characterized by progressive loss of memory and other cognitive abilities. Loss of cognition is often accompanied by changes in motor function and behavioral disturbances. The disease is the leading cause of dementia in older persons (18). An estimated 5 million people in the United States suffer from the disease, and that number is expected to double or triple by the middle of this century (18). AD is a progressive disease. The lifespan of an AD victim is generally reduced, although a person may live 10 or more years after onset (19). Pathologically, AD is characterized by the accumulation of neuritic plaques and neurofibrillary tangles. In neuritic plaques, a central core of ß-amyloid peptide fibrils is associated with swollen and tortuous dystrophic neurites, reactive microglia and astrocytes, and glial filaments. In neurofibrillary tangles, abnormally phosphorylated tau proteins wound into paired helical filaments accumulate intracellularly. These changes are thought to lead to neuronal degeneration and loss of synapses and dendrites. AD is a complex disease with multiple genetic and environmental risk factors (20). Currently, the disease is diagnosed clinically (21); however, the identification of a biological marker for the disease is an urgent issue that reliably studied would be extremely beneficial for the acceleration of drug and therapeutic development (22).
Materials and Methods
samples
Serum samples were obtained from participants in the Religious Orders Study, a longitudinal, clinicopathologic study of aging and AD of older Catholic clergy. Each participant agreed to an annual clinical evaluation, blood donation at baseline, and signed an informed consent and an Anatomical Gift Act donating his/her brain to Rush investigators at the time of death. The study was approved by the Human Investigations Committee of Rush University Medical Center. Since January 1994, more than 1000 persons have enrolled in the study and completed the baseline evaluation. The clinical evaluation documented the presence of mild cognitive impairment (MCI) and dementia and its major causes, including AD, stroke, and Parkinson disease. Details of the clinical evaluation have been reported previously (23). Because data suggest that many persons with MCI have the pathology of AD (24), we combined the groups for some analyses. The demographics of the samples used in this study are shown in Table 1
.
|
Serum was drawn in a tiger-top serum separator tube (BD) at participating sites. After 30 min, the serum was centrifuged for 10 min at 3900g after which 0.5-mL aliquots were placed in separate 2-mL vials (Nalge Nunc International) placed on dry ice and returned to the laboratory, where they were transferred to an ultralow-temperature (–80 °C) freezer. Sera from 302 individuals were used in this study. Frozen human serum was purchased from Sigma-Aldrich and was used as reference material to monitor method performance.
sample processing
Serum samples were processed by use of prototype ProXPRESSIONTM Biomarker Enrichment Kits (PerkinElmer). This Cibachron blue (CB) dye affinity-chromatography–based technology is designed to capture high-abundance carrier proteins (such as albumin) in blood and then enrich for the peptide and protein fragments bound to the carrier proteins. CB filtration plates and spin columns were either provided by VIVASCIENCE or were prepared by use of MULTISCREEN-DV plates and CB resin (Sigma-Aldrich). Approximately 400 µL of CB resin was layered into each well. ZipPlatesTM and the vacuum manifold were purchased from Millipore. Millipore also provided custom-fitting adapters for direct spotting of samples on single-use MALDIchipTM Target plates (PerkinElmer). 
-Cyano-4-hydroxycinnamic acid was purchased from Sigma-Aldrich and recrystallized from ethanol before use, or alternatively premixed matrix was provided by Laser BioLabs. The matrix was prepared in house by addition of 1.25 mL of 500 mL/L acetonitrile containing 1 mL/L trifluoroacetic acid (TFA) to 5 mg of dry matrix and then vortex-mixing until the matrix was well dissolved. This gave a matrix concentration of 4 µg/µL of solution. The vials of matrix solution were kept protected from light at room temperature until ready for use.
Serum samples were processed on 96-well plates by the following procedure: Plates were placed on the vacuum manifold and prewashed with 180 µL of Sample Binding Buffer (SBB; PerkinElmer). The wash step was repeated 3 times. Low vacuum (<25 mmHg) was applied to remove any remaining buffer from the wells. Serum samples were diluted 1:10 in SBB; 100 µL of each diluted serum sample was then applied to each well of the CB plate. The proteins were allowed to bind to the CB medium for 15 min. Low vacuum (<25 mmHg) was applied to pass the sample through the CB medium. The wells were then washed 11 times with 180 µL of SBB. A clean collection plate was then placed under the CB plate, and 150 µL of Sample Elution Buffer (SEB; PerkinElmer) was applied to each well and the biomarkers were eluted with low vacuum (25 mmHg).
Alternatively, serum samples were processed with spin columns by the following procedure: CB spin columns were washed 3 times with 400 µL of SBB. A 25-µL portion of each serum sample was diluted with 300 µL of SBB. Each diluted sample was then loaded on a column and centrifuged at 12 000g for 5 min. The flow-through from the collection tube was recovered and reloaded on the column and centrifuged at 12 000g for 5 min. The column was then washed with 400 µL of SBB, and the biomarkers were eluted with 200 µL of SEB and centrifuged at 12 000g for 5 min. The elution step was repeated once more, and the 2 elution volumes were combined.
Eluted biomarkers were concentrated and desalted on ZipPlate and applied directly to MALDIchip targets by vacuum elution as follows: A C18 ZipPlate was placed on the vacuum manifold. The resin was prewashed 3 times with 180 µL of pure acetonitrile, followed by two 180-µL washes with 1 mL/L TFA, which was followed by two 180-µL washes with pure acetonitrile. Vacuum (375 mmHg) was applied to clear the wells during the prewash steps. The C18 resin was prewetted by applying 3 µL of pure acetonitrile directly to each well of the ZipPlate. Pipetting was timed to allow 
4 min for prewetting of the C18 matrix before sample application. The entire elution volume from each CB column or CB well was loaded into 1 well of a ZipPlate, and the sample was bound by applying vacuum (125 mmHg). Vacuum was gradually increased after 5 min to 375 mmHg. All samples were allowed to pass completely through the resin before proceeding to the wash steps. The ZipPlate was washed 3 times with 300 µL per well of 1 mL/L TFA. Six additional wash steps were performed with 180 µL of 1 mL/L TFA per well per step. The vacuum (375 mmHg) was applied for up to 10 min to remove any remaining liquid.
The biomarkers were eluted directly by vacuum onto disposable MALDIchip targets by the following protocol: A prewashed and dried MALDIchip Target plate was placed beneath the ZipPlate (containing the bound biomarkers) in the vacuum manifold (Note: this requires use of a custom PerkinElmer MALDIchip Plate adapter and Millipore ZipPlate manifold). Vacuum (50 mmHg) was applied, and 3 µL/well of matrix solution was pipetted directly into each well of the ZipPlate. After 5 min, the vacuum was released and the top cover of the vacuum manifold was removed. The ZipPlate was carefully lifted from the MALDIchip Target plate, leaving small droplets on the plate, indicating successful sample transfer. Samples were allowed to air dry at room temperature, which led to the formation of matrix crystals.
ms
Mass spectra were acquired on a prOTOFTM 2000 matrix-assisted laser desorption/ionization orthogonal time-of-flight (MALDI O-TOF) mass spectrometer interfaced with TOFWorksTM software (PerkinElmer/SCIEX). Because of the orthogonal design, a single external mass calibrant was used to achieve better than 10 ppm mass accuracy over an entire sample plate (up to 384 samples). In this study, a 2-point external calibration of the prOTOF instrument was performed before acquiring the spectra in a batch mode from 96 samples. MALDI O-TOF MS is capable of collecting data over a wide range of mass values (300 kDa) in a single acquisition. An example of a spectrum acquired for a mass range of 700 to >18 000 Da is shown in Fig. 1
. In addition, collisional cooling of ions provides very high sensitivity by capturing a large cross-section of ions emerging from the MALDI laser plume. Rapidly reducing the ion energy distribution of the emerging ion beam allows capture of divergent ions that would otherwise be lost. These ions are focused toward the axis of the beam, reducing ion loss and increasing sensitivity. As a result, typical resolution for peptides and proteins up to 10 kDa was >10 000 full width at half maximum.
|
Mass spectra were exported from the SQL database as text files, preprocessed to restore the zero-intensity values, and if needed, the file size was reduced by use of internally developed scripts for MATLAB 7 (The MathWorks). The prOTOF data files generated 700 000 data points per spectrum. The raw mass spectral data used in this study are accessible without restriction athttp://www.perkinelmer.com/biomarkerdata and can be downloaded in their entirety.
processing and analysis of spectral profiles by definition of disease fingerprints with BAMFTM technology
Biomarker Amplification Filter (BAMF) technology (Predictive Diagnostics) was used to process and analyze the raw mass spectral data. The BAMF technology generates biomarker fingerprints for diagnostic and drug discovery applications, [available on a fee-for-service contract basis from Predictive Diagnostics (www.predictivediagnostics.com)]. The program performs quality-control steps, generates and scores potential biomarkers, builds the optimal biomarker model, and classifies the disease state according to an optimized computational model. For this study, the raw mass spectral data were uploaded to the BAMF technology LINUX cluster processing center directly from the prOTOF acquisition computer via a 128-bit encrypted web interface. Subsequent to upload of the raw mass spectral data, the following processes were conducted: spectral outlier rejection, feature discovery, and model building.
spectral outlier rejection
Raw spectra were received in BAMF as vector data (m/z along with an intensity value). Outlier rejection comprised quality-control steps used to identify and remove spectra represented by systematic error to ensure acceptable data uniformity. The data were not modified by this procedure before proceeding to the next step. The principle goal of outlier rejection is essentially to detect irregularities in spectra, and the outlier rejection algorithms produce a cluster plot based on the spectral entropy and the mean error from the ensemble average. Spectral entropy is defined as a measure of the intensity distributions of the spectra and is computed by creating an intensity histogram for a spectrum. By treating the intensity histogram as a probability distribution, spectral entropy can be computed by the traditional formula for entropy: S = –[sum over histogram index i] p_i * log(p_i).
The second calculation averages all of the spectra and calculates the Euclidean distance of each spectrum to this average spectrum. The spectral entropy is then plotted against the Euclidean distance to create the outlier rejection graph, an example of which is shown in Fig. 4
. This plot illustrates how the spectra distribute (cluster) with respect to each other and allows the identification of those spectra that are clearly outliers.
|
Spectral entropy is plotted against the Euclidian distance because each measure of uniformity distributes the spectral ensembles in a different manner; therefore, plotting the measures against one another allows the identification of regions of stability.
It is important to note that Euclidian distance is relatively insensitive to inherited heterogeneity of samples; therefore, minor differences between disease and normal spectra will be lost in this global measurement. It thus is primarily an illustration of qualitative differences between the spectra. In this study, outlier rejection established the uniform data set for the subsequent steps, feature selection, and model building.
feature selection and model building
In the feature selection step, algorithms were run to identify features in the qualified spectral data that were clearly different among groups. Typically, the feature selection process will identify numerous features; however, only some may be useful as diagnostic predictors. A model consists of a particular combination of features. Model-building and classification parameters are derived from modified, self-organizing maps, and these are used as classifiers. The model-building process was used to build and test all possible models of the identified features in the data set to determine which features, in what combinations, had the best diagnostic utility. In many cases, the various processes in the model-building step may produce markers that are not necessarily intuitive because they do not represent known elements or significant peaks in the spectra. However, they may act as "helper markers" to provide a more accurate picture of mass regions of interest. Classification algorithms in the model-building step undergo a consensus function for determination of the final selected class of categorizers. The consensus function operates by polling the best and worst categories to identify overlaps, using a stringent and majority rules approach.
Results
sample complexity reduction
Several techniques for the fractionation of complex protein and peptide samples have been optimized for proteomics applications, including sucrose-gradient fractionation, microscale solution-phase isoelectric focusing, ultrafiltration, and membrane chromatography–based fractionation. Of these approaches, chromatographic prefractionation offers certain clear performance advantages in terms of convenience, reproducibility, and ready implementation on automated liquid-handling platforms. The chromatographic step is a straightforward approach to applying a second dimension of separation to a sample that is independent of the molecular weight obtained by MS. Reproducible sample complexity reduction is an essential first step in biomarker discovery experiments because the large dynamic range of serum/plasma necessitates the effective enrichment or concentration of low-abundance analytes.
For the present study, we developed a streamlined, microscale sample preparation protocol that used either chromatography-based spin columns or 96-well filtration plates. The approach is based on affinity capture of albumin (and other carrier proteins) by a novel dye affinity membrane absorber matrix or beads, with subsequent elution and concentration of the carrier-protein–bound peptides.
Eluted peptides were concentrated on a reversed-phase, C18 ZipPlate device, followed by automated vacuum spotting on a single-use MALDIchip Target plate for direct mass spectrometric analysis.
Shown in Fig. 2
are typical spectra obtained from raw human serum processed through a ZipPlate device only (spectra designated by A) vs serum processed with the biomarker enrichment protocol in addition to a ZipPlate device (spectra designated by B). The number of peptide species detected after MS was substantially increased when the biomarker enrichment protocol was used.
|
reproducibility of mass spectral profiles
Coupled with the biomarker enrichment protocol, the high-resolution O-TOF MALDI mass spectra produced information-rich, reproducible peptide signatures. Surface-based biomarker enrichment strategies have been criticized for their limited binding capacities and tendency to bind the more abundant serum proteins (25)(26). Although it did not quantitatively capture all of the albumin from the serum, this protocol routinely delivers highly reproducible peptide profiles with intensity CVs typically 5%–10% (Fig. 3
).
|
alzheimer study: experimental design
Serum samples were processed according to the following biomarker enrichment protocols: All (n = 302) samples were randomized before processing, and their clinical diagnoses were not revealed (blinded) until the study was completed. Each serum sample was fully processed in triplicate, effectively giving 3 separate data sets, (1050 individual spectra in total, including no-serum and normal-serum controls) 350 spectra in each replicate set. This was done initially to investigate the reproducibility of the individual data sets and to look for consistency in preliminary feature results (data not shown). Reference serum and no-serum controls were also processed in each experiment. For the final analysis, 70% of the total sample spectra were grouped into 3 training sets: Group 1 represented samples from persons with AD (n = 43) or MCI (n = 25); group 2 represented samples from those individuals with only AD; and group 3 represented samples from individuals without cognitive impairment (n = 143). The remaining spectra, 30% of the samples from each group (n = 91), were pooled in a single "blinded" testing group. The 3 training sets were used to build disease models with use of BAMF algorithms. These models were then challenged with the nonoverlapping, blinded testing set.
analysis and detection of distinguishing features
The raw spectral data derived in the study above were analyzed with BAMF technology in 3 general stages: (a) outlier rejection, (b) feature discovery and model search, and (c) model building, validation, and screening.
Outlier rejection.
Outlier rejection analysis reduced the total analyzed spectra from 694 (replicates included) to 544. This spectral pool contained 172 blind samples (groups 1, 2, and 3), 89 group 2, and 283 group 3 individual spectra. The distribution of spectra from which the final spectral pool was drawn is shown in Fig. 4
. Outlier rejection analysis is used to evaluate peak intensity consistency and in addition is used to calculate properties such as overall information content of the spectra. Specifically, BAMF performs a global analysis of all peaks across the entire spectrum. If the information content does not meet uniformity criteria, the spectrum is rejected. This quality-control process is specifically designed for the purpose of choosing the training set. In the Alzheimer study, 21% of the raw spectra were rejected based on the uniformity criteria chosen for this experiment. However, this has no direct correlation with the reproducibility of corresponding peptide peak intensities, which were calculated by use of multiple replicates of a control serum sample and yielded peak intensity CVs of 5%–10%. There are many factors that may have contributed to nonuniform individual spectra in the AD sample set. Possible causes include sample degradation and differences in handling or storage before processing. One of the useful features of BAMF is that it allows the establishment of criteria for the exclusion of artifactual sample spectra in each experiment.
Feature selection and model search.
Feature selection algorithms revealed several putative markers that were used to build the first model, and then subsequent models were iteratively derived from the first model. BAMF algorithms typically identify 150–200 putative markers that are used to construct the initial model, and subsequent models have 1 to 20 markers. Both high- and low-intensity peaks are chosen for model building. From the model searches in this study, 2 models were chosen for consideration, and these are shown in Table 2
. An example of one marker identified in model 1 (8692 Da) is shown in the graphical contour plots and histogram in Fig. 5
. Fig. 5A
is a stacked plot of AD patient samples in the y direction and m/z values in the x direction. Fig. 5B
is a stack plot of non-AD patient samples in the y direction and m/z values in the x direction. It is evident that the 8692 m/z peak is present in Fig. 5A
and relatively absent in Fig. 5B
. Likewise, peak 8707 m/z is absent in Fig. 5A
and present in Fig. 5B
. This infers that the protein corresponding to the 8692 m/z peak is being up-regulated in the disease state (AD) and the protein corresponding to the 8707 m/z peak is being down-regulated in the disease state. The intensity histogram in Fig. 5C
illustrates the marker 8692 in more detail. Markers with masses of 5855, 7969, 8707, 8692, and 9099 Da were clearly distinguished in these and similar plots (data not shown).
|
|
Screening and validation.
Training sets were chosen randomly for all models. The training set for models 0 and 1 consisted of 50 group 2 and 50 group 3 spectra chosen randomly. The remaining spectra (39 group 2 and 233 group 3 spectra) made up the initial validation set. The blinded study consisted of 172 spectra from the blinded group.
predictive results
The predictive values for models 0 and 1 are shown in Table 3
. There was some disparity between the blinded predictive values and the validation predictive values. This disparity may be speculated to arise from several factors, including a relatively limited sample set and the possibility that the AD sample set, in particular, is not entirely representative of the broad AD population. Nevertheless, the sensitivities and specificities were quite high and indicate that the set of proteins and peptides identified by this procedure may be involved in the disease process and may serve as biomarkers of AD. The exact role of each marker (mass) and how each individual marker contributed to the classification (with regard to sensitivity and specificity) were not specifically investigated. This is attributable to the fact that with the BAMF algorithm function and model calculation, the marker set is modeled in its entirety, and therefore, the set, not the individual markers in each model, delivers the classification.
|
Discussion
The low–molecular-weight portion of the human blood proteome comprises peptides and fragment of proteins. This "fragmentome" has attracted much interest because it is a relatively unmined region of the proteome (17). A percentage of the lower–molecular-weight species in blood are bound to high-abundance carrier proteins (such as albumin), extending their circulating half-lives in serum by preventing excretion by the kidneys. A blood-sample preparation approach that takes advantage of this inherent biological enrichment of disease-associated biomarkers, coupled with high-resolution MS and discriminant pattern analysis, can be an extremely powerful technology for diagnostics and population screening. This approach is also amenable to very high throughput and, unlike many other methods, is robust, reproducible, and scalable.
Several parameters are commonly used to assess the capabilities of a potential laboratory test. Diagnostic sensitivity may be defined as the probability of a test being positive in patients who actually are afflicted with the disease, whereas diagnostic specificity measures the ability of the test to distinguish patients with the disease from those who do not have the disease. The diagnostic efficiency (accuracy) of a test is the percentage of all patients tested who are correctly diagnosed by the test. None of the current immunologic or serologic marker assays based on conventional methodologies such as enzyme immunoassay, double immunodiffusion, counterimmunoelectrophoresis, particle agglutination, or immunoblotting, approach the 100% specificity and sensitivity of an ideal assay. The sensitivity and specificity values obtained in this study for AD, a disease that currently has no validated serologic or immunologic test, are certainly competitive with the cited tests used in the field of cancer diagnostics [Ref. (27) and Table 3
].
A key advantage of using a MS-based approach in mining the low-mass proteome is that the mass spectrometer can reproducibly record protein fragment peaks. If a biomarker for a given disease state is a fragment of a larger protein, it may be extremely difficult, or impossible, to produce effective antibodies for tests such as conventional ELISAs. Such antibodies would most likely demonstrate high cross-reactivity to one another (nested peptide fragments) and the larger parent protein as well.
In this report, we demonstrate the applicability of a novel MS-based approach to AD biomarker discovery and patient classification. Serum samples from individuals with AD, MCI, and no cognitive impairment were enriched for biomarkers and then processed and analyzed by high-resolution MS. The spectra were used to generate a fingerprint model that successfully classified blinded samples from AD patients and controls without cognitive impairment, with high sensitivity and specificity. Although the demographics of the sample set indicate an approximate average 10-year difference between the disease and nondisease groups, the total age distribution was similar between the 2 groups. In addition, the Mini Mental State Exam (MMSE) scores at baseline were all normal or above normal. The samples drawn for analyses in this report are from a cohort study with more than 1000 participants. To reduce introduction of bias based on case selection, we used the first 300 consecutive samples from patients without cognitive impairment or with MCI or AD without regard to other variables. The researchers that processed the samples for MS and analysis were blinded to all clinical data, including clinical diagnoses, and the blinding was not broken until the analyses were completed. It was at that time that we discovered the mean age differences. This limits the ability to draw inferences regarding the potential utility of this precise set of markers to discriminate solely between those with and without cognitive impairment. Future analyses will need to be performed to examine the potential effects of age on these particular markers. However, we do not feel that this limitation undermines the methodologic approach used in the present study to identify proteins that can potentially discriminate between two groups.
The fact that this type of analysis yields tenable predictive results for AD, which has multiple genetic and environmental risk factors, suggests great promise for this screening approach in early, noninvasive disease detection. The spectral pattern recognition approach has already demonstrated applicability and usefulness for various cancers (1)(7)(8)(9)(10)(11). This study was designed to discover a surrogate biomarker, an m/z mass pattern that would classify normal vs AD spectra, with high sensitivity and specificity. The characterization and identities of the individual markers featured in the diagnostic model will be the subject of future research and may facilitate the development of a blood-based diagnostic test.
Acknowledgments
We thank the participants in the Religious Orders Study and the staff of the Rush Alzheimer’s Disease Center. The study was supported in part by Grants P30 AG10161 (to D.A.B.) and R01 AG15819 (to D.A.B.) from the National Institute on Aging.
Footnotes
1 Nonstandard abbreviations: MS, mass spectrometry; AD, Alzheimer disease; MCI, mild cognitive impairment; CB, Cibachron blue; TFA, trifluoroacetic acid; SBB, Sample Binding Buffer; SEB, Sample Elution Buffer; MALDI, matrix-assisted laser desorption/ionization; O-TOF, orthogonal time of flight; and BAMF, Biomarker Amplification Filter. 
References
The following articles in journals at HighWire Press have cited this article:
|
|
 |
|
  E. P. Diamandis Oncopeptidomics: A Useful Approach for Cancer Diagnosis? Clin. Chem., June 1, 2007; 53(6): 1004 - 1006. [Full Text] [PDF]
|
|
|
|
|
|
M. F. Lopez, A. Mikulskis, S. Kuzdzal, E. Golenko, E. F. Petricoin III, L. A. Liotta, W. F. Patton, G. R. Whiteley, K. Rosenblatt, P. Gurnani, et al. A Novel, High-Throughput Workflow for Discovery and Identification of Serum Carrier Protein-Bound Peptide Biomarker Candidates in Ovarian Cancer Samples Clin. Chem., June 1, 2007; 53(6): 1067 - 1074. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. Albrethsen Reproducibility in Protein Profiling by MALDI-TOF Mass Spectrometry Clin. Chem., May 1, 2007; 53(5): 852 - 858. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 A. Hye, S. Lynham, M. Thambisetty, M. Causevic, J. Campbell, H. L. Byers, C. Hooper, F. Rijsdijk, S. J. Tabrizi, S. Banner, et al. Proteome-based plasma biomarkers for Alzheimer's disease. Brain, November 1, 2006; 129(Pt 11): 3042 - 3050. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
G. L. Hortin The MALDI-TOF Mass Spectrometric View of the Plasma Proteome and Peptidome Clin. Chem., July 1, 2006; 52(7): 1223 - 1237. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
G. L. Hortin, S. A. Jortani, J. C. Ritchie Jr, R. Valdes Jr, and D. W. Chan Proteomics: A New Diagnostic Frontier Clin. Chem., July 1, 2006; 52(7): 1218 - 1222. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
P. Kepper, R. Reinhardt, A. Dahl, H. Lehrach, and S. Sauer Matrix-Assisted Laser Desorption/Ionization Mass Spectrometric Analysis of DNA on Microarrays Clin. Chem., July 1, 2006; 52(7): 1303 - 1310. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
