TECHNOLOGY R EPORT
Mass spectrometry-based clinical proteomics Wolfgang Pusch1†, Mark T Flocco2, Sau-Mei Leung2, Herbert Thiele1 & Markus Kostrzewa3 †Author
for correspondence Daltonik GmbH, Fahrenheitstrasse 4, D-28359 Bremen, Germany Tel: +49 421 2205 270 Fax: +49 421 2205 103 E-mail:
[email protected] 2Bruker Daltonics, 40 Manning Park, Billerica, MA 01821, USA 3Bruker Daltonik GmbH, Permoserstrasse 15, D-04318 Leipzig, Germany 1Bruker
In recent years, mass spectrometry (MS) has been recognized as a ‘Gold Standard’ tool for the identification and analysis of individual proteins in expression proteomics studies. Moreover, MS has proven useful for the analysis of nucleic acids for single nucleotide polymorphism (SNP) genotyping purposes. With the increased usage of MS as a standard tool for life science applications and the advancement of MS instrumentation, sample preparation and bioinformatics, MS technology has entered novel screening and discovery application areas that are beyond the traditional protein identification and characterization applications. The areas of clinical diagnostics and predictive medicine are just two prime examples of these fields. Predictive markers or biomarkers for early diagnosis of diseases are of growing importance for the human healthcare community. The goal of using MS in clinical proteomics is to generate protein profiles (mass to charge [m/z] ratio versus signal intensity) from readily available body fluids like serum, saliva and urine to detect changes in protein levels that reflect changes in the disease states. Whereas the results originating from individual protein markers may be intriguing, data resulting from the analysis of complex, multiple biomarker patterns may be unequivocal. These biomarker patterns are hidden in complex mass spectra and are not always obvious to the human eye. Sophisticated bioinformatics algorithms have to be applied to determine these unique biomarker patterns. Here, we review the latest developments concerning the use of MS for the discovery of biomarker patterns and for the identification of individual biomarkers in the field of clinical proteomics applications.
Introduction
Keywords: biomarker, MALDI, mass spectrometry, pattern matching Ashley Publications Ltd www.ashley-pub.com
The ability to screen an individual’s serum for specific biomarkers is an important tool in the early diagnosis of cancer. This screening procedure enables a non-invasive examination of patients who are predisposed to cancer. Commonly used biomarkers include prostate-specific antigen (PSA) for prostate cancer [1], CA125 for ovarian cancer [2] and estrogen receptor for estrogen-dependent mammary cancer [3]. However, most of these biomarkers do not produce an unequivocal answer as they lack specificity and sensitivity [4,5]. Sensitivity problems arise if the biomarker protein is not present in all patients or if its level is too low to be detected in early stages of the disease. Specificity problems arise due to the expression of the respective protein in tissues other than the cancer of interest. The approach of using serum biomarkers has been successful in principle, although the use of individual biomarkers in many cases is insufficient for a valid clinical diagnosis. The next logical step is to take serum screening one step further by discovering and utilizing multiple biomarkers, consisting of a pattern of
2003 © Ashley Publications Ltd ISSN 1462-2416
upregulated and/or downregulated proteins. These biomarker patterns may give a more valid clinical diagnosis and may appear before the onset of symptoms. Traditionally, two-dimensional gel electrophoresis has been used in the research community to discover protein changes between normal and disease states. However, its use in clinical proteomics has been limited because it is labor intensive and time consuming. The ability to screen and discover multiple biomarkers simultaneously in clinical proteomics has been advanced by the recent success of mass spectrometry (MS), especially matrix assisted laser desorption/ionization-time-of-flight (MALDITOF) MS. MALDI-TOF MS has enabled investigators to analyze, in parallel, a large number of serum proteins and peptides from patients. In contrast to the widely used enzyme-linked immunosorbent assays (ELISA), which use specific antibodies for an indirect detection of biomarkers, MALDI-TOF MS detects the analyte directly by an intrinsic property – its molecular weight. A MS protein profile spectrum can be used to detect multiple biomarkers simultaneously. Initially, Pharmacogenomics (2003) 4(4), 463–476
463
TECHNOLOGY REPORT
these biomarkers do not need to be identified nor is there a need for specific antibodies for capture or detection. By utilizing the advancements of computer technology and bioinformatics software, the door has been opened for a completely new approach. Complex MS biomarker patterns hidden in the ‘bar code’ of serum proteins and peptides of patients are now a major focus for clinicians. These diagnostic biomarkers are discovered by the use of sophisticated pattern determination and class prediction algorithms. Analytical problems in clinical proteomics
MS can be used for a diversity of analytical problems in the field of clinical proteomics. Possible applications include screening, initial diagnosis, monitoring of disease progression, and prediction of response to therapy. In contrast to many well-established antibody-based assays, when starting a study using MS, no information about individual biomarkers indicating a specific disease state is necessary. When a tumor is perfused by the blood, a portion of its specific proteins may be actively secreted into the circulatory system (Figure 1). Additionally, proteins may also be released by necrosis and apoptosis of tumor cells. These processes are accompanied by extensive morphological changes in the cell nucleus [6,7]. This is why the nuclear matrix proteins are relatively well-characterized individual biomarkers for tumors [8]. The MS-based pattern recognition using protein profiles starts with an interactive inspection of two test data sets originating from different patient cohorts (e.g., tumor versus control) with appropriate visualization tools, in order to find obvious differences in the spectra indicating possible biomarkers. If the differences between the classes are not obvious to the human eye, sophisticated mathematical algorithms have to be applied to discover biomarker patterns suitable to distinguish both sample groups and to detect samples that do not fit into either group (Figure 2), for example, due to benign transformations, inflammation processes or other unrelated diseases. The patterns may indicate prominent differences between the groups that justify the efforts of identification of the potential biomarker peak by a MS/MS analysis. Since serum contains a complex mixture of many proteins and peptides, it is possible that an identified biomarker peak is not due to a single protein or peptide but consists of multiple proteins or peptides of very similar molecular weight. Accordingly, a 464
high quality (multi-dimensional) fractionation and purification of the samples will be preferable for MS/MS-based protein identification. Additionally, profiling the potential biomarkers in a higher resolution MS, for example, using an instrument which is equipped with a reflectron, will result in a better peak resolution compared to the linear mode. But even if a pattern model, which was automatically generated by bioinformatics software, shows only gradual differences between the two sample classes, it may nevertheless prove useful for screening approaches. As the level of peptides and proteins may be up- or downregulated in the course of a disease process, this leads to a complex pattern of minor peptide and protein signal intensity changes, which may be used as an indicator for the disease. The capability to observe complex pattern changes by mass spectrometric multivariate analysis has the potential for significantly higher specificity and sensitivity compared to the conventional univariate assays. The aim of a screening approach is to match unknown samples against the existing pattern model to support a clinical diagnosis. In this case, the pattern itself is the diagnostic biomarker. In addition, identified individual proteins may also be used for screening applications. An example may be certain nuclear matrix proteins [8] that are used to characterize, for example, breast cancer [101]. Another analytical problem using MS for clinical proteomics is the detection of larger proteins [9]. In general, larger proteins have limited peak resolution and are more difficult to ionize in the MALDI procedure. This requires different optimization methods for data acquisition and the subsequent data interpretation. For all applications, it is a prerequisite to start with well-characterized samples with known clinical pedigrees to develop a biomarker pattern model or to identify individual biomarkers. A training data set of samples is used to establish the pattern, which is subsequently applied to a test data set to indicate the sensitivity and specificity of the model. The samples utilized for the model generation must contain populations from both affected and unaffected individuals so that the software can develop an accurate model. Sample collection and data preparation
Body fluid samples, which can be obtained by minimal invasive methods with low patient health risks, are preferred for diagnostic approaches and, therefore, may be suitable for early screening programs. Analyses of protein Pharmacogenomics (2003) 4(4)
TECHNOLOGY REPORT
Figure 1. Secretion of specific biomarkers into the blood circulation by tumors.
Artery
Tumor Serum protein profile before tumor passage
*
*
*
Vein Serum protein profile after tumor passage
Tumor-specific proteins may be actively secreted by tumor cells or released into the circulatory system by necrosis and apoptosis of these cells. Either of these conditions leads to an alteration of the serum protein profile. This may result in detectable differences based either on relative or unique signal intensities when comparing sera from normal and disease samples.
profiles from serum [10], urine [11] or nipple aspirate [12] have been reported for cancer diagnosis. For studies in clinical proteomics, the case and the control cohorts have to be selected very carefully in order to find meaningful biomarker patterns. Samples from the same ethnic groups should be used in both cohorts. Moreover, it may be necessary to use only samples of the same sex in both groups. For female samples, the time point in the estrous cycle may also have an influence. If urine samples are investigated, the food intake will have considerable influence on resulting biomarker patterns, especially in the low mass range, due to secondary metabolites. If these constraints are not considered during the www.pharmaco-genomics.com
sample collection, the resulting differences may be meaningless, as they may be not due to the disease of interest. Moreover, inflammation reactions may result in similar biomarker patterns to certain diseases. This must also be considered during the sample collection and clinical characterization, as well as during subsequent analysis. Moreover, to generate comparable raw data from different patients, collected by different people, in different laboratories is a major requirement to minimize sample to sample variation. Variations may be due to the sample collection, sample storage, or sample processing and handling techniques used by the individual facilities. These variations, which are not caused by the disease, have to be adjusted by a normalization procedure. For example, day-to-day variation may lead to altered absolute signal intensities of the spectra, but the overall pattern of the signal intensities may not be affected. Accordingly, with normalized signals, the bias between independent spectra may be better controlled. One of the strategies for correcting these values currently pursued by the National Institute of Health (NIH) and Food & Drug Administration (FDA) Clinical Proteomics Programme [102] is to normalize each data point in a series of spectra according to the following equation (where I = intensity): CurrentI – MinimumI NormalizedI = -----------------------------------------------------------MaximumI – MinimumI
As the ‘Current Intensity’ of a data point is smaller than or equal to the ‘Maximum Intensity,’ this procedure leads to a normalization of the intensities ranging between 0 and 1, and 0 and 100%, respectively. Sample preparation
Direct MS analysis of biological fluids, such as serum, without prefractionation is a challenging assignment because serum is a very complex mixture of proteins and peptides. Instead, a selective enrichment of specific protein and peptide classes according to their physical, chemical or biological properties combined with a fractionation prior to MS analysis could be performed to enhance signal quality. Prefractionation of a complex serum sample may decrease signal suppression from major proteins (e.g., albumin) so that minor proteins can be detected. It can also separate similar mass to charge (m/z) ratio peptides and proteins into subfractions so that they can be detected separately. The overall goal of the prefractionation step is to maximize the 465
TECHNOLOGY REPORT
Figure 2. Principle of mass spectrometry based serum profiling.
Binding
Patients Serum samples
Fractionation
Washing
Normal
Normal
Cancer
Elution *
*
* Bioinformatics
MALDI-TOF MS
Clinical result
Serum samples from both normal and diseased subjects are fractionated on functionalized surfaces according to their chemical, physical or biological properties. Serum protein profiles are generated using a mass spectrometer. Patterns of up- and downregulated proteins that are suitable for use as biomarkers to separate disease from normal samples are discovered by combining appropriate bioinformatics software with an independently determined clinical diagnosis. The resulting biomarker pattern can then be used for the class prediction of unknown samples or for the identification of new individual biomarkers. MALDI-TOF: Matrix assisted laser desorption/ionization time-of-flight.
number of proteins and peptides that can be detected by MS. One of the approaches for sample preparation prior to MS analysis is using MALDI targets with functionalized surfaces (so-called ‘chips’) for specific enrichment of peptides and proteins. Different surface functionalities such as cation exchange, anion exchange, hydrophobic, hydrophilic, immobilized metal or antibody affinity capture can be used for capturing of different peptides/protein. After processing, these chips can be introduced directly into the mass spectrometer for measurement of bound analytes. This procedure is known as the surfaceenhanced laser desorption/ionization (SELDI)TOF MS approach [103]. A single chip may have eight or sixteen sample spots. This allows for convenient selective protein or peptide
466
enrichment directly on the sample target but with considerable costs for the chips. A possible alternative may be off-line enrichment and fractionation procedures prior to conventional MS analysis [13]. For example, magnetic particles with different functionalized surfaces offer a simple, easy to use method for sample preparation. Magnetic bead handling is well-known as an automation-friendly and scalable method for sample preparation in the life sciences. The bead surface can be functionalized with an anion exchange, cation exchange, reversed phase or immobilized metal affinity purification [104]. Two- or three-step protocols using combinations of different surface functionalities may allow for capturing more potential biomarkers. Scalability enables a detailed analysis of a larger amount of sample and
Pharmacogenomics (2003) 4(4)
TECHNOLOGY REPORT
possibly an identification of a detected biomarker molecule. Moreover, the surface functionality may be customized with specific antibodies to allow epitope mapping experiments [14-16].
with: s t
Equation III: ν = -
where: Instrumentation MALDI-TOF MS
The breakthrough of MS in biotechnology was introduced by the development of new gentle ionization methods. The ionization of the analyte is necessary for all MS techniques. These new technologies prevent the immediate decay of large biomolecules like proteins or DNA during the ionization/desorption process. Nowadays, two ionization techniques are commonly used for biomolecules – electrospray ionization (ESI) [17] and MALDI [18-20]. ESI can be used for liquid samples, for example, after HPLC separation. It is predominantly used for the analysis of peptides, proteins or larger DNA fragments (Figure 3). For MALDI-TOF MS, the analyte is first co-crystallized in a surplus of matrix molecules, commonly an organic acid on a solid target support. Then the target is introduced into the vacuum of the mass spectrometer’s ion source (Figure 4). A pulsed UV-laser is used to evaporate the matrix and the analyte into the gas phase. In addition to desorption support, the matrix ionizes the analyte molecules by proton transfer. In this process, predominantly single-charged ions are generated. These ions are accelerated by an electrical field that is switched on a few 100 nanoseconds (ns) after the ionization and then enter the field-free flight tube of the mass spectrometer. Entering the flight tube the kinetic energy of the analyte is identical with the work applied to accelerate the ions: 1 2
2
= z×e×U
where: e = elementary charge m = mass of the analyte U = acceleration voltage v = speed of the analyte z = number of charges in the analyte ion Equation I can be transformed to: 2×e×U
Equation II: ---- = ----------------------
www.pharmaco-genomics.com
Insertion in Equation I results in Equation IV: m z
2×e×U×t
2
Equation IV: ---- = ---------------------------------
s
2
The only remaining unknown variable to calculate the m/z ratio is t [21]. Accordingly, by simple measurement of the TOF of the analyte ions after acceleration in an electric field, their m/z ratio can be determined. For single charged ions, this value is equivalent to their molecular weight. This technique can be applied to the analysis of very large biomolecules like proteins [22] and nucleic acids [23,24]. SELDI-TOF MS
SELDI-TOF MS (Ciphergen, Fremont, CA, USA) does not differ generally from MALDI-TOF MS. The approach uses a linear TOF mass spectrometer and the analyte molecules are embedded into a surplus of matrix molecules that enable the gentle ionization process. Like in conventional MALDI analysis, the analyte molecules are desorbed and ionized by a laser irradiation. The difference is the functionalization of the MALDI target (‘chip’) to apply selective fractionation and enrichment of proteins and peptides together with an online detection in the MS. MS/MS analysis
Equation I: --- × m × ν
m z
s = length of the flight tube t = time of flight of the analyte
ν
2
MALDI-TOF MS is a suitable tool to discover specific protein biomarker patterns. This applies if the only purpose of a study is the classification of new samples with unknown diagnosis against an established biomarker pattern without the identification of the individual biomarkers. However, if the pattern displays prominent differences between the disease and normal samples, it may be desirable to allow an in-depth analysis of individual peaks. This facilitates the potential identification and classification of new biomarkers. In addition to these advanced classification opportunities, the knowledge about the biology of new biomarkers also increases the understanding of the disease mechanism. These new biomarkers could also serve as new targets 467
TECHNOLOGY REPORT
Figure 3. Use of different mass spectrometry technologies for different analytical problems.
Sample throughput
Very high
High
MALDI-TOF MS
MALDI-TOF MS/MS
Low
High
Low Nano-ESI MS/MS
Very high
Information content
ESI: Electrospray ionization; MALDI-TOF: Matrix assisted laser desorption/ionization time-of-flight; MS: Mass Spectrometry.
for drug development or could potentially be used as drug candidates themselves. Usually, a combination of multiple MS analyzers into one MS/MS instrument is used to identify individual biomarkers from the protein profile. The first analyzer serves as a mass filter to separate the mass of interest from the rest of the ions. Configurations currently used for this purpose are so-called quadrupole TOF (QqTOF) and MALDI TOF/TOF mass spectrometers. In the case of the QqTOF, a quadrupole analyzer is used as a mass filter to select a defined precursor ion mass out of a mixture of ions, for example, the mass of a potential biomarker can be selected out of a serum profile. The ions of this mass are then directed into a second quadrupole analyzer, which serves as a collision cell to generate smaller fragments of the precursor ion. The masses of these fragment ions are subsequently analyzed in an orthogonal TOF [25]. Similarly, in a MALDI TOF/TOF instrument (Figure 5) the first TOF analyzer is used as a mass filter. Metastable fragments as well as collision induced decay (CID) products are then accelerated into a second TOF analyzer for mass analysis. The MS/MS approach offers the opportunity to identify the respective protein(s) underlying prominent peak differences between spectra from disease or control serum. Pattern recognition
A central part of a complete clinical proteomics solution is the algorithm that is used to distiguish the protein profiles of disease and normal samples. Since the analytical problem is very complex, simple peak comparisons from
468
different spectra are not sufficient for high quality results. Since no information is available regarding at which distinct m/z ratios the significant differences between normal versus disease samples are hidden, nor the number of peaks that have to be included in the multivariate analysis, a supervised approach with a training data set is used by bioinformatics algorithms to learn the differences between the two sample classes. Then a biomarker pattern is established with intensities at different m/z ratios to differentiate the two sample classes. The clinical diagnoses of all training and test samples have to be determined by an independent procedure in advance to ensure the quality of the results. Subsequently, a class prediction determines to which group a new unclassified sample belongs (Figure 6). In the literature and in commercially available products, different approaches have been followed. So far, genetic algorithms [105], decision trees (Ciphergen, Fremont, CA, USA) and the unified maximum separability algorithm [28,31] have the best visibility. However, artificial neuronal networks [26] are promising as well. Moreover, the growing interest in the field of clinical proteomics may very well lead to the development of further suitable algorithms in the near future. The concept of genetic algorithms is similar to the principle of biological evolution. It optimizes the results by applying multiple cycles of mutation and survival of the fittest. Each spectrum consists of several thousands of data points of signal intensities at distinct m/z ratios. Groups of 5–20 individual data points of a
Pharmacogenomics (2003) 4(4)
TECHNOLOGY REPORT
Figure 4. The principle of MALDI-TOF MS.
A. Measurement • Matrix-embedded analyte on microtiter plate target
Acceleration Drift
+
+ + +
+
+
Detector
Electrodes Laser desorption/ionization
Time-of-flight B. Mass spectrum • Analyte ions separated according to their mass/charge ratio
Intensity m/z
The samples are prepared on a microtiter plate target embedded in a surplus of matrix molecules. A pulsed UV-laser is used to evaporate the matrix and the analyte into the gas phase. In addition to desorption support, the matrix ionizes the analyte molecules by proton transfer. The ions are then accelerated by an electrical field and enter the field-free flight tube of the mass spectrometer. By simple measurement of the TOF of the analyte ions, their mass to charge ratio can be determined. Modified from [24]. m/z: mass to charge; TOF: Time-of-flight.
spectrum are combined to so-called logical chromosomes. These logical chromosomes are then altered by mutation and recombined by a single point crossover. Afterwards, a test procedure determines if these alterations lead to a model that better fits the data. If not, this alteration and all possible combinations derived thereof are discarded. Otherwise, the new combination will be further used for the next generation of mutation and selection. After several hundreds of iterations (= generations) of mutation and selection, a biomarker pattern is established. This biomarker pattern consists of an ndimensional space, and each dimension is the signal intensity of one data point, which is included in the final pattern. In this n-dimensional space, classical cluster analysis techniques can then be used to determine the class of a new sample from its mass spectrum. This principle was successfully used in a retrospective study to detect ovarian cancer [27]. Multiple runs using genetic algorithms will probably generate different biomarker patterns. This effect is due to the danger of optimizing on local maxima. Accordingly, different runs and different parameter
www.pharmaco-genomics.com
settings should be used to generate the best fitting model by a genetic algorithm. A second mathematical model used for this purpose is decision trees [28,29]. The Biomarker Patterns Software package (Ciphergen, Fremont, CA, USA) is based on on the Classification and Regression Tree (CART) analysis [106]. Once the decision tree model is established, the user can easily track the individual decisions leading to a certain classification of a new sample. Specific diagnostic peaks are used as information nodes in a decision process. Based on the absence or presence, and on the intensity of such a peak, one branch or the other is followed in the decision tree. At the next node, a further decision is made and so on. Finally, a result is determined according to the decisions from multiple nodes. This was successfully used to differentiate prostate cancer from benign disease and healthy men [30]. A third published algorithm is the Unified Maximum Separability Algorithm (UMSA) [31], which is a modified support vector machine [32]. Support vector machines avoid the problem of overfitting [33,34]. Typically, the quality of a new 469
TECHNOLOGY REPORT
Figure 5. The principle of MALDI-TOF/TOF MS.
Reflectron
Selector
LIFT
Detector Target
TOF 1
TOF 2
Ions are desorbed and ionized from the microtiter plate target and accelerated into the first TOF analyzer. Here, metastable and collision-induced fragments continue to travel with the velocity of the intact molecular ions they are derived from. This means that they stay close together in space. TOF 1 ends with a selector. Only analytes of the selected mass and the derived fragments can pass to the ion source of the second TOF analyzer (called LIFT™ here). All other molecular ions are rejected. In the ion source of TOF 2 the selected parent and fragment ions are accelerated. Subsequently, these analyte ions are separated in the conventional TOF 2 analyzer. This allows the mass analysis of the fragments, by measuring their arrival time on the detector. The reflector in the second TOF analyzer is used to enhance the resolution of the mass spectrum of the fragments. TOF: Time-of-flight.
pattern model is evaluated by measuring how well it fits the data. However, good fits can be misleading as they may result from properties of the model that have nothing to do with the cognitive process of interest (overfitting) [35]. In contrast to a Principal Component Analysis (PCA) [36,37], UMSA [38] does not concentrate on the signals with the highest variation but on those signal patterns that collectively best separate different groups of patients. USMA has been used to detect serum biomarkers in breast cancer [31]. Currently, multiple algorithms have been applied successfully in the field of pattern recognition for MS-based clinical proteomics. Additional algorithms will certainly be published in the near future. Currently, there does not seem to be a single solution for all possible analytical problems. Different diseases may require different algorithms and spectra originating from different body fluids may require different approaches as well. Therefore, in order to get optimal results for a specific disease, one will probably require the use of multiple mathematical algorithms in parallel. Seamless bioinformatics integration
A high-quality pattern matching algorithm is needed but not totally sufficient for the application of MALDI-TOF MS in the field of clinical proteomics. 470
In addition to all necessary requirements concerning sample preparation, a convincing bioinformatics concept has to lead from the control of robotic elements via automated data acquisition with the mass spectrometer to automated data interpretation. The software should, as far as possible, also support the user working under FDA or Good Laboratory Practice (GLP) regulated environments, if necessary. Internal quality control issues may make it necessary to offer more than one algorithm in parallel for the pattern determination and/or for the class prediction. The quality of an algorithm can be scored by means of the sensitivity and specificity, as well as the positive and negative predictive values of the resulting models (Table 1). Furthermore, the classification algorithms should also support the interactive interpretation of the results. If the classification is a ‘black box’, where the user cannot easily understand how and why this result was generated, then he will only gain a limited understanding of the results. Therefore, the acceptance of black box algorithms by clinical users may be lower than for those algorithms that allow for an easy understanding and explanation of the results, like decision trees. Suitable visualization tools have to support the interactive interpretation of results. In particular, sophisticated visualization of results is necessary to support the
Pharmacogenomics (2003) 4(4)
TECHNOLOGY REPORT
Table 1. Outcome of a fictive retrospective whole population study with high sensitivity and high specificity. Sensitivity (percentage of true positives)
98%
Specificity (percentage of true negatives)
98%
Samples
100,000
Normal
99,900
Cancer
100
Result: True positives (detected cancer cases)
98
False negatives (missed cancer cases)
2
True negatives
97902
False positives
1998
Positive predictive value
98/[1998 + 98] = 4.6%
Negative predictive value
97902/[97902 + 2] = 99.9%
identification of biomarker candidates, for example, by MS/MS analysis. This also creates the need to link clinical proteomics analysis tools and their results with those from the expression proteomics identification studies. An appropriate database solution should be the melting pot that combines expression proteomics and clinical proteomics data, thus leading from data to knowledge. Quality of resulting biomarker profiles
The quality of a diagnostic cancer assay is addressed by its sensitivity and specificity. The sensitivity is the percentage of detected true positives from the total number of true positive samples. The specificity is the percentage of detected true negatives from the total number of true negative samples. Accordingly, the optimal assay would have a sensitivity of 100% and a specificity of 100%; then, no cancer case would be missed and no sample would be incorrectly classified as cancer. These ideal values will not be reached in practice. To be broadly accepted, the software algorithms for mass spectrometric analysis based on multiple biomarkers have to generate better results than the existing individual cancer biomarkers that are analyzed by other methodologies. Even if a high specificity is achieved, the large numbers of patient samples to be screened combined with a low occurrence of the disease lead to a considerable number of false positives compared to true positives (Table 1). This may become a serious problem if a screening study is not restricted to specific patient groups at higher risk to acquire the disease.
www.pharmaco-genomics.com
Moreover the algorithm should not only distinguish diseased from normal patients; it is especially desirable that the cancer is detected at a very early stage before the onset of symptoms. If possible, it should also indicate the stage of the disease. Furthermore, it should distinguish benign from malignant transformations, an effect that is of special interest for prostate cancer diagnosis, where the prevalence of benign hyperplasia is common [39]. The algorithm ideally should distinguish samples that do not fit in any of the groups. These samples may belong to a new cluster. Unclassified spectra may indicate an unhealthy state that is due to a different disease but they also may be the result of poor analysis data. Automatically generated biomarker patterns are destined to cluster samples in one of two or several groups. While an interactive inspection of the data may propose a certain number of potential biomarkers, a subset may be sufficient for an unequivocal class prediction by a mathematical algorithm. Accordingly, automated pattern generation should always be accompanied by an interactive inspection by an expert to get an overview of all potential biomarkers. Good laboratory practice and FDA compliance
Major requirements concerning software issues arise when clinical diagnosis is performed in environments that are regulated according to ‘GLP’ or according to the requirements of the FDA. The FDA rules are defined in Part 11 of the Code of Federal Regulation 21 (21 CFR Part 11) [40-42,107].
471
TECHNOLOGY REPORT
Figure 6. Example for the usage of complex biomarker patterns for serum profiling.
Step 1: Discovery
Step 2: Evaluation
Step 3: Class prediction
Training data set
Test data set
Unknown data set
Disease
Normal
Disease
Normal
Profile 1
Profile 2
Profile 1
Profile 2
*
*
*
Pattern discovery x
*
Cluster analysis
*
Cluster analysis y
x
x Use biomarker pattern for step 2.
Profile 2
*
y
y
Profile 1
Determination of: • Sensitivity • Specificity • Positive predictive value • Negative predictive value
Disease
Normal
In the first step, a training data set containing both disease and normal samples is used to discover a biomarker pattern. The clinical diagnoses of all training data samples have been previously determined. Bioinformatic tools then use various mathematical algorithms to determine a biomarker pattern of up- and down-regulated – or possibly even unique – signals. In the second step, the biomarker pattern is evaluated by the classification of a test data set. Similar to the training data, it too contains diseased and normal samples with known clinical diagnoses. Typically cluster analysis is performed to predict the class (normal or disease) of a sample. In the figure, the cluster analysis is shown for a pattern of two biomarkers, x and y. Accordingly, only the two dimensions x and y have to be considered. Therefore, for a pattern of n biomarkers the cluster analysis would be performed in an n-dimensional space. The resulting assay is then evaluated by calculating the sensitivity and the specificity as well as the positive and negative predictive values. If the quality of the resulting multivariate biomarker pattern is reliable, then unknown samples can be matched against this pattern. This class prediction/independent validation leads to a better estimate of sensitivity, specificity and positive/negative predictive values.
21 CFR Part 11 strictly applies only to the US but it serves as a benchmark standard that defines how to handle computer-generated raw, meta data and analysis results. These requirements have also been adopted by laboratories that are regulated according to the ‘predicate rules’ of GLP and the Good Manufacturing Practice (GMP), which were defined by the Organisation for Economic Cooperation and Development [108,109]. Main topics of the 21 CFR Part 11 are electronic records, electronic signatures, audit trails, long-term archiving and data security. 21 CFR Part 11 challenges the user as well as the vendor who provides an analysis system. The user has to make sure that his/her laboratory practice as defined in standard operation procedures (SOPs) is compliant with the FDA regulations. The
472
system vendor has to support the work in FDA regulated environments as far as possible by according software functionality. Since the rules and especially all possible exceptions have far reaching consequences concerning software architecture, we can only present a short outline of the software requirements here. As a central point, whenever computer systems are used for the data acquisition or analysis, a hard copy of the resulting record is no longer sufficient. Once the data are saved on a storage medium, they have to be handled as an electronic record. Electronic records have to be supplied with electronic signatures that are typically comprised of a user identifier and an individual password. This electronic signature is legally binding and replaces a manual signature. The electronic record contains the information concerning who did what,
Pharmacogenomics (2003) 4(4)
TECHNOLOGY REPORT
when and possibly why. All changes to the record have to be tracked in an audit trail, which has to be stored at least as long as the respective electronic record – typically 10–15 years. New results must not replace older results, unless the changes are tracked in the audit trail. When starting the work with an analysis system, the user has to be identified by an appropriate user management. This ensures that each operation on the system and each possible manipulation of the data can be unequivocally assigned to a specific user. Different users may have different rights on the system. For example, only a user with administrator rights may generate and release a new analysis method. For the field of clinical proteomics that may mean that only an administrator has the right to confirm a newly detected biomarker pattern as an accepted and released method for the diagnosis of a certain cancer type [43-45]. Advantages and limitations of MS
Previously, clinical proteomics was mainly performed by investigating a single biomarker protein, for example, the CA125 serum biomarker for ovarian cancer [46-48]. However, judged by sensitivity and specificity, the use of single biomarkers in clinical diagnosis is often limited. For ovarian cancer, this means that the disease remains a ‘silent killer’ as the biomarker is usually not detected until an advanced stage of the disease [49]. Moreover, in order to develop an ELISA, respective biomarker proteins have to be purified and specific antibodies must be (commercially) available, with high purity and high specificity. As a result, only a limited number of markers are currently applied for cancer detection [50]. In contrast to this univariate approach, MS allows multivariate analysis of multiple biomarkers simultaneously. Here, differences in biomarker patterns between disease and control data complement an individual biomarker. This approach increases the sensitivity and specificity of the assay and promises a diagnosis at an earlier stage of a disease, possibly before the onset of symptoms. Besides MS, a further methodological approach may be offered by antibody arrays. However, issues such as steric hindrance, competition, conformation and specificity make the development of protein microarrays both costly and provocative. Cross-talk between related analytes or detector antibodies on a multiplexed protein microarray is the primary cause of non-specific binding or background signal. This will ultimately limit the number of www.pharmaco-genomics.com
elements in an array and the overall sensitivity of individual assays. The necessary, highly specific monoclonal antibodies can be generated both in vivo and in vitro, and they have been used with some success coupled to solid supports. Although this approach is feasible, one must remember that 20,000–50,000 human genes may require > 200,000 different antibodies to effectively find and capture biomarkers due to processing events, post-translational modification and degradation. One must also remember that many biomarkers are simply degraded forms of other proteins or structures. Upregulation, downregulation and post-translational modification of these proteins and peptides may indicate disease but they may not contain the epitope needed to bind to an antibody of choice. A classic example is PSA, which when complexed with α-2 macroglobulin, has hidden the classic epitope used to identify the PSA molecule. This renders it invisible in an immunoassay, forcing a multiplex analysis of free and total, or complex and total forms to provide a useful diagnostic algorithm [51]. There are many companies attempting to develop protein microarrays, BD Clontech [110], Zyomyx [111], Randox [112], Biosite [113], Somalogic [114] and Affibody [115] to name a few. With proper research and development, this can be seen as an excellent orthogonal tool in clinical proteomics and biomarker discovery. In contrast, MS allows the multivariate analysis of complex patterns of new biomarkers without knowing their individual identities and without having specific monoclonal antibodies available. The proteins underlying individual peaks used for the differentiation do not have to be identified initially, although it may become necessary to determine the relation to the disease for a broad acceptance by the clinicians. If desirable, the use of MS/MS technology, for example, through modern TOF/TOF instruments, promises the identification of biomarkers. Accordingly, the use of MS in clinical settings may in turn boost the availability of identified biomarkers. The sample preparation and the MALDITOF measurement itself are suitable for automation and allow a throughput of several thousand samples per day. MS currently has a high distribution in research institutions whereas the use by clinicians is still limited. However, basic MS instruments are currently in a transition from expert systems to simple user-friendly platforms for everyday use by researchers and technicians; a 473
TECHNOLOGY REPORT
process that may also pave the way into the clinical laboratory. Although the investigation of complex biomarker patterns may promise better results than single protein biomarkers, MS will probably be employed in conjunction with the widely used ELISA. Outlook
Currently, predominantly MALDI- and SELDI-TOF instruments are being used in clinical proteomics. Many MALDI-TOF instruments are compatible with the microtiter plate format. This facilitates the robotic solutions for the necessary sample preparation. The complete MALDI-TOF measurement can also be automated and is suitable for high-throughput applications, ideally even for a 24/7 operation. However, in the future, LC-MS instruments may be also useful for clinical proteomics applications. Typically, LC-MS are smaller benchtop Highlights • MALDI-TOF MS based approaches allow the classification of patient serum samples without the identification of individual markers. It also only requires a small amount of sample for analysis. • The analysis of complex biomarker patterns enables better sensitivities and specificities compared to the univariate analysis of individual markers. Improved cancer diagnosis has to be shown for a variety of cancers to establish MS in the field of clinical diagnostics. • MALDI-TOF MS based clinical proteomics detects the biomarkers directly and is not dependent on an indirect detection via a specific antibody. • Combined use of MALDI-TOF MS and MS/MS instruments allows not only for sample classification but has also the potential for the identification of new biomarkers. • Identification of new biomarkers may lead to a better understanding of the disease mechanism. New biomarkers may also serve as drug target candidates or as potentially new drugs. • MALDI-TOF based clinical research will not replace biopsies or examinations by physicians. However, it can be a valuable tool for early routine screening approaches.
Bibliography Papers of special note have been highlighted as either of interest (•) or of considerable interest (••) to readers. 1. Balk SP, Ko YJ, Bubley GJ: Biology of prostate-specific antigen. J. Clin. Oncol. 21(2), 383-391 (2003). 2. Rai AJ, Zhang Z, Rosenzweig J et al.: Proteomic approaches to tumour marker discovery. Arch. Pathl. Lab. Med. 126, 1518-1526 (2002).
474
3.
4.
instruments and they allow quantification of the analytes. However, they are less suitable for a higher sample throughput and for the analysis of larger analytes, a deconvolution of multiply charged components is necessary (Figure 3). Possibly, such instruments may complement the MALDI-TOF screening ‘work-horses’ for the indepth analysis of individual proteins. In the future, establishing highly significant biomarker patterns specific for a large number of cancer types will be a major milestone towards a better diagnosis from non-invasive sources, like serum or other readily accessible body fluids. A challenge will be the detection of the biomarkers long before the patient is symptomatic, to improve the prospects of a medical treatment. Intelligent algorithms could possibly make suggestions concerning other diseases if the sample fits neither the disease, nor the control model. The next step could be the monitoring of disease progression [52]. Besides tumor screening, possible further fields of interest are the analysis of a patient’s disease predisposition or of the exposure to carcinogenic substances by the evaluation of biomarker patterns [53]. Although currently the main application of MS in the field of clinical proteomics is cancer research, the technique could also be applied for the investigation of other disorders. Current screening applications are mainly established by researchers in clinical institutions. However, the superior technology of MS based clinical proteomics also has the potential to enter the clinical diagnostics market and to be accepted by many clinicians as a new cutting edge technology for enhanced cancer diagnosis. Acknowledgments The authors would like to thank Dr Armin Holle for helpful advice concerning the MALDI-TOF/TOF technology.
Bast RC, Ravdin P, Hayes DF et al.: 2000 Update of recommendations for the use of tumour markers in breast and colorectal cancer: Clinical practice guidelines of the American Society of Clinical Oncology. J. Clin. Oncol. 19(6), 1865-1878 (2001). Baltaci S, Aksoy H, Turkolmez K, Elhan AH, Ozden E, Gogus O: Use of percent free prostate-specific antigen density to improve the specificity for detecting prostate cancer in patients with rectal examinations and
5.
6.
7.
intermediate prostate-specific antigen levels. Urol. Int. 70(1), 36-41 (2003). Brawer MK: Clinical usefulness of assays for complexed prostate-specific antigen. Urol. Clin. North Am. 29(1), 193-203 (2002). Mommers EC, Poulin N, Sangulin J, Meijer CJ, Baak JP, van Diest I : Nuclear cytometric changes in breast cancer carcinogenesis. J. Pathol. 193(1), 33-39 (2001). Castelli C, Losa GA: Ultrastructural complexity of nuclear components during
Pharmacogenomics (2003) 4(4)
TECHNOLOGY REPORT
8.
9.
•• 10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
early apoptotic phases in breast cancer cells. Anal. Cell. Pathol. 23(1), 1-9 (2001). Lüftner D and Possinger K: Nuclear matrix proteins as biomarkers for breast cancer. Expert Rev. Mol. Diagn. 2(1), 89-97 (2002). Hanash S: Disease proteomics. Nature. 422(6928), 226-232 (2003). Review article introducing the field of clinical proteomics. Petricoin EF, Ornstein DK, Paweletz CP et al.: Serum proteomic patterns for detection of prostate cancer. J. Natl. Cancer Inst. 94(20), 1576-1578 (2002). Hampel DJ, Sansome C, Sha M, Brodsky S, Lawson WE, Goligorsky S: Towards proteomics in uroscopy: urinary protein profiling after radiocontrast medium administration. J. Am. Soc. Nephrol. 12, 1026-1035 (2001). Paweletz CP, Trock B, Pennanen M et al.: Proteomic patterns of nipple aspirate fluids obtained by SELDI-TOF: Potential for new biomarkers to aid in the diagnosis of breast cancer. Disease Markers 17, 301-307 (2001). Valerio A, Basso D, Mazza S et al.: Serum protein profiling of patients with pancreatic cancer and chronic pancreatitis: searching for a diagnostic protein pattern. Rapid Commun. Mass Spectrom. 15, 2420-2425 (2001). Parker CE, Tomer KB: MALDI/MS-based epitope mapping of antigens bound to immobilized antibodies. Mol. Biotechnol. 20(1), 49-62 (2002). Kelly MA, McLellan TJ, Rosner PJ: Strategic use of affinity-based mass spectrometry techniques in the drug discovery process. Anal. Chem. 74(1), 1-9 (2002). Peter JF, Tomer KB: A general strategy for epitope-mapping by direct MALDI-TOF mass spectrometry using secondary antibodies and cross- linking. Anal. Chem. 73(16), 4012-4019 (2001). Fenn JB, Mann M, Meng CK, Wong SF, Whitehouse CM: Electrospray ionization for mass spectrometry of large biomolecules. Science 246(4926), 64-71 (1989). Karas M, Hillenkamp F: Laser desorption ionization of proteins with molecular masses exceeding 10,000 daltons. Anal. Chem. 60(20), 2299-2301 (1988) Hillenkamp F, Karas M: Mass spectrometry of peptides and proteins by matrix-assisted ultraviolet laser desorption/ionization. Methods Enzymol. 193, 280-295 (1990). Hillenkamp F, Karas M, Beavis RC, Chait BT: Matrix-assisted laser desorption/ionization mass spectrometry of
www.pharmaco-genomics.com
21.
22.
23.
24.
25.
26.
27.
••
28.
29.
30.
••
31.
bioploymers. Anal. Chem. 63(24), 1193A-1203A (1991). Eckerskorn C: Massenspektrometrie. In: Bioanalytik. Lottspeich F, Zorbas H (Eds), Spektrum Akademischer Verlag, Heidelberg, Berlin 323-368 (1998). Eckerskorn C, Strupat K, Karas M, Hillenkamp F, Lottspeich F: Mass spectrometric analysis of blotted proteins after gel electrophoretic separation by matrix-assisted laser desorption/ionization. Electrophoresis 13, 664-665 (1992). Nordhoff E, Ingendoh A, Cramer R et al.: Matrix-assisted laser desorption/ionization mass spectrometry of nucleic acids with wavelengths in the ultraviolet and infrared. Rapid Commun Mass Spectrom. 6(12), 771-776 (1992). Pusch W, Wurmbach JH, Thiele H, Kostrzewa M: MALDI-TOF mass spectrometry-based SNP genotyping. Pharmacogenomics 3(4), 537-548 (2002) Aebersold R, Mann M: Mass spectrometrybased proteomics. Nature 422(6928), 198-207 (2003). Ball G, Mian S, Holding F et al.: An integrated approach utilizing artificial neural networks and SELDI mass spectrometry for the classification of human tumours and rapid identification of potential biomarkers. Bioinformatics 18(3), 395-404 (2002). Petricoin EF, Ardekani AM, Hitt BA et al.: Use of proteomic patterns in serum to identify ovarian cancer. Lancet 359, 572-577 (2002). Proof-of-principle for the application of genetic algorithms for the pattern determination in a retrospective study of ovarian cancer samples. Fung ET, Enderwick C: ProteinChip clinical proteomics: computational challenges and solutions. Biotechniques Suppl. 34-8 40-1 (2002). Adam B, Qu Y, Davis JW et al.: Serum protein fingerprinting coupled with a pattern-matching algorithm distinguishes prostate cancer from benign prostate hyperplasia and healthy men. Cancer Res. 62, 3609-3614 (2002). Qu Y, Adam BL, Yasui Y et al.: Boosted decision tree analysis of surface-enhanced laser desorption/ionization mass spectral serum profiles discriminates prostate cancer from noncancer patients. Clin. Chem. 48(10), 1835-1843 (2002). Proof-of-principle for the application of decision tree models to differentiate prostate cancer from control samples. Li J, Zhang Z, Rosenzweig J, Wang YY, Chan D: Proteomics and bioinformatics
••
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
44.
approaches for identification of serum biomarkers to detect breast cancer. Clin. Chem. 48(8), 1296-1304 (2002). Proof-of-principle for the application of the UMSA algorithm for the discovery of serum biomarkers in breast cancer samples. Zhang Z, Chan DW: Unified maximum separability analysis a supervised approach for biological expression data analysis. 2002 AACC Annual Meeting. Orlando, Florida (Jul 28 – Aug 1, 2002). Cai YD, Liu XJ, Xu XB, Chou, KC: Support vector machines for the classification and prediction of beta-turn types. J. Pept. Sci. 8(7), 297-301. Hua S, Sun Z: A novel method of protein seciondary structure prediction with high segment overlap measure: support vector machine approach. J. Mol. Biol. 308(2), 397-402. Pitt MA, Myung IJ: When a good fit can be bad. Trends Cogn. Sci. 6(10), 421-425 (2002). Raychaudhuri S, Stuart JM, Altman RB: Principal component analysis to summarize microarray experiments: application to sporulation time series. Pac. Symp. Biocomp. 455-466 (2000). Landgrebe J, Wurst W, Welzl G: Permutation-validated principal components analysis of microarray data. Genome Biol. 3(4), research0019 (2002). Zhang Z, Page G, Zhang H: Applying classification separability analysis to microarray data. In: Methods of Microarray Analysis: Papers from CAMDA ‘00. Lin SM, Johnson KF (Eds), Kluwer Academic Publishers, Boston, 25-26 (2001). Medina JJ, Parra RO, Moore RG: Benign prostatic hyperplasia (the ageing prostate). Med. Clin. North Am. 83(5), 1213-1229 (1999). Code of Federal Regulations, Title 21, Food and Drugs, Part 11, ‘Electronic Records, Electronic Signatures: Final Rule’ Federal Register 62 (54), 13429-13466. Huber L: Implementing 21 CFR Part 11 in Analytical Laboratories, Part 1: Overview and Requirements. BioPharm. 12(11), 28-34 (1999). Winter W, Huber L: Implementing 21 CFR Part 11 in Analytical Laboratories: Part 2, security aspects for systems and applications. BioPharm. 13(1), 44-50 (2000). Winter W, Huber L: Implementing 21 CFR Part 11 in analytical laboratories: Part 3, ensuring data integraty in electronic records. BioPharm. 13(3), 45-49 (2000). Huber L, Winter W: Implementing 21 CFR Part 11 in analytical laboratories: Part 4,
475
TECHNOLOGY REPORT
45.
46.
47.
48.
49.
50.
51.
476
data migration and long-term archiving for ready retrieval. BioPharm. 13(6), 58-64 (2000). Winter W, Huber L: Implementing 21 CFR Part 11 in analytical laboratories: Part 5, the importance of instrument control and data acquisition. BioPharm. 13(9), 52-56 (2000). O’Rourke J, Mahoon SM: A comprehensive look at the early detection of ovarian cancer. Clin. J. Oncol. Nurs. 7(1), 41-47 (2003). CA125 response: can it replace the traditional response criteria in ovarian cancer? Oncologist 7(5), 437-443 (2002). Cherry C, Vacchiano SA: Ovarian cancer screening and prevention. Semin. Oncol. Nurs. 18(3), 167-173 (2002). Goff BA, Mandel L, Muntz HG, Melancon CH: Ovarian carcinoma diagnosis. Cancer 89(10), 2068-2075 (2000). Duffy MJ: Biochemical markers in breast cancer: which ones are clinical useful? Clin. Biochem. 34(5), 347-352 (2001). Ellison L, Cheli CD, Bright S, Veltri RW, Partin AW: Cost-benefit analysis of total, free/total and complexed prostate-specific antigen for prostate cancer screening. Urology 60, 42-46 (2002).
52.
53.
Von Eggeling F, Junker K, Fiedler W et al.: Mass spectrometry meets chip technology: a new proteomic tool in cancer research? Electrophoresis 22, 2898-2902 (2001). Pritzker KPH: Cancer biomarkers: easier said than done. Clin. Chem. 48(8), 1147-1150 (2002).
108. http://www.olis.oecd.org/olis/1998doc.nsf/
109.
Websites 101. http://www.matritech.com
Matritech, Newton, MA, USA 102. http://clinicalproteomics.steem.com/
103. 104. 105. 106. 107.
proteome-detail.php Food and Drug Administration Clinical Proteomics Programme. http://www.ciphergen.com Ciphergen, Fremont, CA, USA. http://www.bdal.de Bruker Daltonik, Leipzig, Germany. http://www.correlogic.com Correlogic Systems, Bethesda, MD, USA. http://www.salford-systems.com Salford Systems, San Diego, CA, USA. http://www.fda.gov/ora/compliance_ref/cpg/ cpggenl/cpg160-850.htm FDA: Compliance Policy Guide. 21 CFR Part 11; Electronic Records; Electronic Signatures (CPG 7153.17).
110. 111. 112. 113. 114. 115.
LinkTo/env-mc-chem(98)17 OECD Series on Principles of Good Laboratory Practice and Compliance Monitoring, Number 1, OECD Principles on Good Laboratory Practice (as revised in 1997). http://www.olis.oecd.org/olis/1995doc.nsf/ LinkTo/ocde-gd(95)115 OECD Series on Principles of Good Laboratory Practice and Compliance Monitoring, Number 10, GLP Consensus Document. The Application of the Principles of GLP to Computerized Systems, Environment Monograph No. 116 (1995). http://www.clontech.com BD Clontech, Palo Alto, CA, USA. http://www.zyomyx.com Zyomyx, Hayward, CA, USA. http://www.randox.com Randox, Crumlin, UK. http://www.biosite.com Biosite, San Diego, CA, USA. http://www.somalogic.com Somalogic, Boulder, Colorado, USA. http://www.affibody.com Affibody, Bromma, Sweden.
Pharmacogenomics (2003) 4(4)
