Abstract

The application of mass spectrometry to discover new cancer biomarkers is in its infancy. Many of these new markers are low-abundance proteins that exist as fragments associated with carrier proteins. Although reproducibility is key to the use of mass spectrometry for ion fingerprint analysis, the scientific community has yet to establish a common platform or standardized operating procedures that are necessary for intra- and inter-laboratory comparison. In an effort to assist others who are perfecting mass spectrometry platforms for profiling, ongoing experimental data were posted for public consumption. An unintended consequence of unrestricted access to experimental data is the risk of inappropriate conclusions drawn and publicly disseminated that could have been avoided by communication between the producers and consumers of the data. Such disputes, however, should not divert us from the validation of this promising new approach.

Proteomic Profiling for Ovarian Cancer: Dangers of Analysis in a Vacuum

In this issue, Baggerly et al. ( 1 ) report on their analysis of conclusions made recently by Zhu et al. ( 2 ) concerning our publicly posted ovarian cancer-related mass spectral data study sets ( http://home.ccr.cancer.gov/ncifdaproteomics/ ). Baggerly et al. focus on the apparent ability by Zhu et al. to identify diagnostic signatures in one ovarian cancer experimental data set, which was posted on our website in April 2002. That set was able to accurately predict phenotypes in a second separate and independent data set, which was produced and posted in August 2002. On the basis of their own analysis of these two data sets, Baggerly et al. ( 1 ) claim that the diagnostic signatures reported by Zhu et al. are not accurate and conclude that reproducibility within the two posted data sets used in these studies has not been demonstrated. Baggerly et al. further generalize that between-set mass spectral reproducibility itself remains an open question and justify their conclusions as follows: 1) The case and control samples in the second data set could be separated by randomly selected values. 2) Some peaks that go up in one set go down in the second set. 3) The m/z features found by Zhu et al. appear to change in amplitude across the two experimental sets posted.

We are not surprised that Baggerly et al. found differences between our posted ovarian cancer studies data sets. Each data set that we have posted is generated from an experimental setting in which one or more of the parameters have been purposefully changed. Under classic experimental design, we held some parameters constant as we deliberately changed other parameters to study the effect of the changes on the characteristics of the output spectra. Thus, we expected that many differences between the research data sets would exist due to the differences in the experimental design used to generate each data set.

Each web posting represents an experiment designed with specific goals and questions. Investigators familiar with matrix-assisted laser desorption ionization time-of-flight (MALDI-TOF) mass spectrometry understand that the output spectra are highly dependent on the chemical characteristics of the samples analyzed, on the reagents used, and on the exact details of the protocol. The mass spectral portraits for the specific study sets in question were derived from the subset of molecules that were retained on the specific chip retentate surface chromatography and specific type of matrix used. The spectral fingerprint can be altered, among other things, by the laser energy and intensity, pH, matrix quality, ion signal suppression, binding time and incubation conditions (i.e., temperature, pH, and so forth), the manufacturing of the chip surfaces themselves, and the distribution of the functional groups on the chips. The ion spectra peak mass/charge ( m/z ) values and their relative intensities are the result of the many physical and chemical events that generate the spectra. If the protocols change between two experiments that use the same sample sets, then results of these experiments (i.e., the specific ion amplitude and m/z ) are likely to be different.

Comparisons of mass spectral data reproducibility across time can be applied only to study sets in which all process variables, except time, are constant. The ovarian cancer study sets analyzed by Baggerly et al. did not fall into this category, and we are puzzled as to why they were used to judge reproducibility. Data set 4–3–02 was produced with a low-resolution first-generation surface-enhanced laser desorption ionization time-of-flight (SELDI-TOF) mass spectrometer, manual sample application and washing, and a weak cation-exchange binding surface. Although data sets 8–7–02 and 4–3–02 were produced with the same type of chip surface, data set 8–7–02 was produced with a different and roboticized sample application method, with different washing and binding conditions, and with an entirely different, upgraded, low-resolution mass spectrometer. We would be surprised if the experimentally designed process changes between these two studies did not result in altered spectra. In fact, a goal of these experiments was to study the spectral alterations produced by changing the process!

The experimental goals of the two study sets and the way the samples were processed were different. The goal of data set 4–3–02 was to evaluate the ability of a weak cation-exchange chip surface to generate mass spectral information that could be used to segregate case and control ovarian cancer samples. In contrast, the goal of data set 8–7–02 was to determine whether the between-day total process variance was more or less than the variance between the case and control groups. For data set 8–7–02, case and control samples were run in separate batches on separate days and not commingled. We used protocols recommended by the Clinical and Laboratory Standards Institute (formerly NCCLS) to explore the individual sources of between-day spectral variation generated within a single batch or run, independent of differences in the phenotype of the samples themselves. Important and ongoing questions have been whether differences in sample collection between case patients and control subjects could contribute to analytical bias and whether qualitative differences could contribute to the source of variance within each phenotype. As with any unprocessed electronic signal, scalar changes in the relative intensity values, reported by mass spectrometry, can change dramatically from day to day. Consequently, investigators often use normalization procedures (e.g., dividing by total ion current or by maximum ion signal) that may correct for scalar differences. If the spectra are not normalized (and it appears that Baggerly et al. did not normalize the spectra for their analysis), it is likely that randomly chosen values would be able to separate the two phenotypes. The design of the experiment required that the cancer samples and control samples be run on separate days, thus generating scalar intensity differences across the two groups. The conclusion of Baggerly et al. that “systematic differences in spectra are more likely associated with procedural bias, such as incomplete randomization, that confounds our ability to recognize potentially reproducible biological factors” was expected, and so it inappropriately casts a negative view on a planned component of the experimental design of this study that investigated relative sources of process variance.

Need for Optimization and Protocol Standardization Before Analysis of Reproducibility

Certainly, the issue of reproducibility is of major importance to investigators evaluating and researching mass spectral profiles for disease detection. Nevertheless, a meaningful analysis of reproducibility requires communication between those producing the data (in this case, us) and those analyzing, or consuming, the data. Why is this communication necessary? Communication helps to ensure that data are being evaluated within the context of the experimental conditions used and the questions that were being asked when the data were produced ( 3 , 4 ) . Without such communication, data can be misinterpreted, unwarranted and overextended conclusions can be drawn, and misinformation can be spread. Unfortunately, we believe that this may be exactly what has happened here.

The discussion of experimental reproducibility is a key question to those attempting to use mass spectrometry for discovery or profiling approaches. Although our own laboratory efforts are now emphasizing sequencing and characterization of the low-molecular-weight biomarker candidates, we are also continuing to seek a better understanding of the sources of analytical and physical variability along the entire process (from sample collection to data generation). We and others in the field have begun to develop appropriate quality control and quality assurance procedures to increase process stability. During this process, we have posted examples of our raw mass spectral data in the public domain, much of it experimental and not published in peer-reviewed journals. In fact, the two ovarian cancer data sets in question were not part of any publication. These postings are meant to provide the community an unbiased and transparent snapshot into our ongoing experimental research work and to further provide data sets for their own work, whatever that may be. As a consequence of the unrestricted Web access to multiple data sets, many investigators have downloaded the data and analyzed them.

In hopes of enabling the clearest understanding of our data, we have annotated our Web-posted data in the hope that they will be analyzed appropriately and will be judged only in the context of the data collection experiment. We are hopeful that processes and platforms will evolve to a point where mass spectral pattern analysis can be used for routine diagnostics and commercialized after rigorous validation and scientific publication after peer review. Academic consortium groups, such as the National Cancer Institute's Early Detection Research Network, have reported very promising initial results for the systematic evaluation of reproducibility of mass spectral profiling for prostate cancer detection applications ( 5 ) , which indicates that reproducibility can be attained when this is the goal of the study. This effort involves six centers where the same experimental conditions, methodology, and types of equipment are being used to generate the type of data that is representative of how reproducible a given process is. Although bias and reproducibility will continue to be a major issue when using any diagnostic platform in which the identities of the peaks are unknown, we anticipate that the underlying identities will eventually be elucidated—in fact, some peaks have been identified ( 6 ) . The transition from an information profile derived from unknown identities to a profile derived from identified and characterized molecules should minimize bias and increase reproducibility because the measurement of the markers should then be independent of the platform.

Refocusing Attention on the Biological Content

We hope that the claims by Baggerly et al. ( 1 ) do not divert attention away from the tremendous promise that proteomics holds for biomarker discovery and the accelerating scientific progress in this infant field. In the late 1990s, our laboratory used mass spectrometry profiling applications to determine whether tissue cell lysates obtained with laser capture microdissection contained a proteomic portrait that could discriminate different tumor types, differences between primary and metastatic disease tissue, or differences between early-stage premalignant lesions and invasive cancers ( 7 ) . The results of these and other later studies ( 8 ) indicated that there were important differences in the proteomic fingerprints of the tumor cells themselves, especially in the low-molecular-weight range (mass-to-charge ratio of less than 40 000). The next question at that time was, “How much of this new information can be captured from a blood sample?”

We and our colleagues set out to test the hypothesis that the low-molecular-weight range of the circulatory proteome contained diagnostic information. Previous work by Goodacre et al. ( 9 ) and Lay and colleagues ( 10 ) indicated that it was possible to combine mass spectral data with pattern recognition methodology to identify fingerprints that could discriminate among bacterial species without prior knowledge of the molecules themselves. Using a variety of different pattern recognition methods, high-throughput mass spectrometry, and serum/plasma from patients with a variety of disease states, we and others ( 5 , 6 , 1121 ) have generated data that appear to indicate that discriminatory information can be found within the specific study sets tested. Results of these independent studies appear to support the initial hypothesis that the circulatory proteome is indeed a rich source of biomarker information, even though the identity of the underpinning molecules was not known ahead of time.

This new information archive (i.e., the circulatory proteome) is being explored actively by essentially two complementary but independent approaches. One approach uses the fingerprints, or patterns, of mass spectral information as the diagnostic tool and does not require or use the identity or sequence of the molecules contributing to the mass spectra. These investigators are examining collections of mass spectral amplitude values and then evaluating whether the combined relative intensities of the m/z values can be used to classify disease states accurately. The alternative approach is to sequence the proteins in the new set of candidate biomarkers directly. After a candidate biomarker is identified, the capture reagents (e.g., antibodies) are developed to measure multiplexed analyte panels of subsets of the candidate biomarkers. Both of these approaches have substantial advantages, disadvantages, and obstacles ahead. It is not clear which approach will have the earliest clinical impact. Our opinion is that both approaches should be applied concomitantly because any method that might provide a clinical benefit warrants rigorous and serious investigation ( 4 ) .

What are the impediments facing investigators choosing between these approaches? The impediment facing an approach based on patterns of unidentified molecules is reproducibility across platforms, time, and laboratories. Because mass spectrometry platforms are in a constant state of technologic evolution, with new, improved systems coming online every year, the scientific community has not yet adopted a common platform or standard operating procedure. Lack of agreement on the utilization and type of reference standards further complicates this issue. Because the molecules that underpin the pattern are not known, another impediment is the difficulty of assuring that experimental bias is not a contributor to the discrimination. Experimental bias can result from differences in how the case and control specimens are collected and processed or from the procedure and process of generating the mass spectra itself. In addition to the problems of experimental bias, investigators must recognize the further challenges presented by high-dimensional data analysis. Rigorous validation with blinded study sets is absolutely required to guard against overfitting. Nevertheless, mass spectrometry profiling remains an attractive and very rapid analytical approach that is well suited for commercialization. This discovery-based approach does not require the lengthy development and validation of antibody reagents and immunoassay-based systems.

In contrast to direct mass spectrometry profiling of blood or tissue, sequencing and characterization of the underlying constituents are a very laborious process. In fact, the cycle time for protein sequencing, characterization, antibody (or analyte-specific ligand) development, validation in clinical research study sets, and immunoassay development is the biggest impediment to the direct characterization approaches. The advantage of this approach is that, for characterized biomarkers, reproducibility of analyte measurements with well-tested and validated immunoassay platforms is not an issue. In addition, when the molecules are identified, bias and overfitting can be assessed directly.

Because of the need to identify the biomarkers, our research efforts have centered on identifying and sequencing the underlying discriminatory information that exists in the mass range profiled by direct mass spectrometry profiling. Recently, we have discovered that many of the low-molecular-weight fragments with molecular information that exist in circulation are bound to high-abundance carrier proteins, such as albumin and immunoglobulins ( 22 , 23 ) . By harvesting these carrier proteins with fragments bound, the concentration of these fragments is amplified to a level that is detectable by mass spectrometry and assessable by traditional immunoassays ( 24 ) . We can now take advantage of the carrier protein binding as an in vivo analytical fractionation step and selectively procure and directly sequence the bound biomarker fragments. When we use this method with serum obtained from women with ovarian cancer at either stage I or stages III-IV or from women with a high risk of ovarian cancer, who were monitored for 5 years after serum collection, and then compare the differential sequenced protein fragments found in the three sets, a complex and rich source of candidate biomarkers is revealed ( Table 1 and Fig. 1 ). To increase confidence in the findings, each of the example candidates was found iteratively with multiple peptide hits for each entity. This same information archive, in a complex fashion, appears to underpin serum mass spectral profiles. Thus, a list of sequence-identified proteins or peptides that reside in the mass range of a mass spectral profile contributes information between profile-based approaches and multiplexed immunoassay-based systems. Such associations, we believe, should expedite the clinical translation of this knowledge, independent of the analytical method used. Moreover, as soon as previously unidentified peaks in a mass spectral pattern have been identified, investigators will be better prepared to evaluate the influence of experimental bias.

Table 1.

Examples of stage I ovarian cancer-specific albumin-bound candidate serum biomarkers

Phosphodiesterase γ Adenomatous polyposis coli protein Aldican tyrosine kinase receptor Anaplastic lymphoma tyrosine kinase receptor precursor α tectorin BCL2-associated athanogene 1 protein Baculoviral inhibitor of apoptosis protein 1 BCL2 interacting killer Breast cancer antigen NY-BR-1 Caspase recruitment protein 12 Centrosomal protein 2 cGMP-Gated cation channel Cysteine- and histidine-rich domain-containing, zinc binding protein 1 Disks large-associated protein 2 Elastin microfibril interface-located protein Insulin-like growth factor complex Interleukin-17 precursor Junctophilin 1 Kinase suppressor of ras-1 Mouse double minute 4 protein Melanoma antigen B4 Pannexin 2 Proto-oncogene tyrosine protein kinase Retinoblastoma assoc. factor 600 Rho GDP-dissociation inhibitor 1 Ryanodine receptor 3 Arginine/serine-rich nuclear protein 25 Sex-determining region Y box 3 transcription factor Transmembrane fibronectin-like domain containing leucine-rich transmembrane protein 1 precursor Netrin transmembrane receptor unc 5 Tumor-associated calcium signal transducer 2 Tuftelin-interacting protein 11 c-Src protein tyrosine kinase Ubiquitin-activating enzyme E1 Vascular noninflammatory molecule 2 precursor Zinc finger homeobox protein 
Phosphodiesterase γ Adenomatous polyposis coli protein Aldican tyrosine kinase receptor Anaplastic lymphoma tyrosine kinase receptor precursor α tectorin BCL2-associated athanogene 1 protein Baculoviral inhibitor of apoptosis protein 1 BCL2 interacting killer Breast cancer antigen NY-BR-1 Caspase recruitment protein 12 Centrosomal protein 2 cGMP-Gated cation channel Cysteine- and histidine-rich domain-containing, zinc binding protein 1 Disks large-associated protein 2 Elastin microfibril interface-located protein Insulin-like growth factor complex Interleukin-17 precursor Junctophilin 1 Kinase suppressor of ras-1 Mouse double minute 4 protein Melanoma antigen B4 Pannexin 2 Proto-oncogene tyrosine protein kinase Retinoblastoma assoc. factor 600 Rho GDP-dissociation inhibitor 1 Ryanodine receptor 3 Arginine/serine-rich nuclear protein 25 Sex-determining region Y box 3 transcription factor Transmembrane fibronectin-like domain containing leucine-rich transmembrane protein 1 precursor Netrin transmembrane receptor unc 5 Tumor-associated calcium signal transducer 2 Tuftelin-interacting protein 11 c-Src protein tyrosine kinase Ubiquitin-activating enzyme E1 Vascular noninflammatory molecule 2 precursor Zinc finger homeobox protein 
Fig. 1.

Enrichment of low-molecular-weight (1000–12 000 m/z ) proteomic information via carrier protein binding and amplification. Mass spectral comparison of the same serum where the input was either the albumin-bound fraction ( upper ) or the native total sample ( lower ). The spectra are scaled equally so that a direct comparison can be made. MALDI-QqTOF = matrix-assisted laser desorption ionization hybrid quadrupole time-of-flight; SELDI-TOF = surface-enhanced laser desorption ionization time-of-flight; MS = mass spectrometry.

Fig. 1.

Enrichment of low-molecular-weight (1000–12 000 m/z ) proteomic information via carrier protein binding and amplification. Mass spectral comparison of the same serum where the input was either the albumin-bound fraction ( upper ) or the native total sample ( lower ). The spectra are scaled equally so that a direct comparison can be made. MALDI-QqTOF = matrix-assisted laser desorption ionization hybrid quadrupole time-of-flight; SELDI-TOF = surface-enhanced laser desorption ionization time-of-flight; MS = mass spectrometry.

Lessons Learned and a Plea for Better Communication

It is our opinion that it is incumbent upon those investigators who are analyzing Web-posted scientific data (e.g., gene microarray or mass spectral data) to evaluate the information fairly and to include the data producers in the evaluation. Close communication between the data analysts and data producers will be a necessary and integral nexus if optimal clinical benefit is to be achieved. Without such communication, frankly, we can see no reason why investigators would want to post preliminary and ongoing experimental data on a website, not accompanying a publication, simply to have it analyzed in ways that may be inappropriate. This problem could actually slow down research and thus delay implementation of any clinical benefit. This problem is reflected in the title chosen by Baggerly et al.: “Signal in Noise: Evaluating Reported Reproducibility of Serum Proteomic Tests for Ovarian Cancer.” Whose reported reproducibility? What test? To our knowledge, there is currently no clinical or even experimental test using our Web-posted data, or the features themselves as the basis of any “test” offered for ovarian cancer detection. The title, in our opinion, appears to be potentially misleading, since the two data sets used as the basis for the Baggerly paper are experimental research study sets that were never part of any clinical “test.”

To be completely clear, many of the experiments that we conducted were designed with the goal of specifically identifying spectral changes and tracking those changes that resulted from deliberate changes in the protocol. These questions centered on some of the following aspects: What effect does changing the incubation parameters have on the spectra? What effect does changing the laser energy or mass spectrometry platform have on the spectra? What are the sources of variability between a nested set of defined serum samples from patients with a cancer and plasma sera from unaffected control subjects? The field itself is still in its early infancy, with every laboratory using its own internal methods, mass spectrometry instruments, chemistries, and processes. Consequently, the field is not yet ready to address the clinical reproducibility of this approach under guidelines such as those from the Clinical and Laboratory Standards Institute. Investigators are still refining their methodology to evaluate the effect of process on spectral quality, complexity, and stability. To move the field forward productively, we urge that effort be directed toward the following general goals: determination of appropriate reference material, the identification of well-controlled and handled sample study sets or the prospective collection of defined clinical study sets, and the development of standardized mass spectrometry and multiplexed immunoassay platforms and standardized operating procedures. Only after standardization should the community develop appropriate methods to assess reproducibility. The proteomics community is well poised to work together in this fashion as it methodically optimizes and evaluates various biomarker measurement and detection systems.

The views expressed here are expressed solely by the authors and should not be construed as representative of those of the Department of Health and Human Services, the US Food and Drug Administration, or SAIC, Inc. Moreover, aspects of the topics discussed have been filed as US government–owned patent applications. Drs. Petricoin and Liotta are co-inventors on these applications and may receive royalties provided under US law.

References

(1)
Baggerly KA, Morris JS, Edmonson SR, Coombes KR. Signal in noise: evaluating reported reproducibility of serum proteomic tests for ovarian cancer.
J Natl Cancer Inst
 
2005
;
97
:
307
–9.
(2)
Zhu W, Wang X, Ma Y, Rao M, Glimm J, Kovach JS. Detection of cancer-specific markers amid massive mass spectral data.
Proc Natl Acad Sci U S A
 
2003
;
100
:
14666
–71.
(3)
Petricoin EF, Fishman' DA, Conrads TP, Veenstra TD, Liotta LA. Proteomic pattern diagnostics: producers and consumers in the era of correlative science. Available at: http://www.biomedcentral.com/1471-2105/4/24/comments . [Last accessed: January 15, 2005.]
(4)
Petricoin EF, Fishman DA, Conrads TP, Veenstra TD, Liotta LA. Lessons from Kitty Hawk: from feasibility to routine clinical use for the field of proteomic pattern diagnostics.
Proteomics
 
2004
;
4
:
2357
–60.
(5)
Semmes OJ, Feng Z, Adam BL, Banez LL, Bigbee WL, Campos D, Cazares LH, Chan DW, Grizzle WE, Izbicka E, Kagan J, Malik G, McLerran D, Moul JW, Partin A, Prasanna P, Rosenzweig J, Sokoll LJ, Srivastava S, Thompson I, Welsh MJ, White N, Winget M, Yasui Y, Zhang Z, Zhu L. Evaluation of Serum Protein Profiling by Surface-Enhanced Laser Desorption/Ionization Time-of-Flight Mass Spectrometry for the Detection of Prostate Cancer: I. Assessment of Platform Reproducibility.
Clin Chem.
 
2005
Jan;
51
(1):
102
–12.
(6)
Zhang Z, Bast RC Jr, Yu Y, Li J, Sokoll LJ, Rai AJ, et al. Three biomarkers identified from serum proteomic analysis for the detection of early stage ovarian cancer.
Cancer Res
 
2004
;
64
:
5882
–90.
(7)
Paweletz CP, Gillespie JW, Ornstein DK, Simone NL, Brown MR, Cole KA, et al. Biomarker profiling of stages of cancer progression directly from human tissue using a protein biochip.
Drug Dev Res
 
2000
;
49
:
34
–42.
(8)
Stoeckli M, Chaurand P, Hallahan DE, Caprioli RM. Imaging mass spectrometry: a new technology for the analysis of protein expression in mammalian tissues.
Nat Med
 
2001
;
7
:
493
–6.
(9)
Goodacre R, Neal MJ, Kell DB, Greenham LW, Noble WC, Harvey RG. Rapid identification using pyrolysis mass spectrometry and artificial neural networks of Propionibacterium acnes isolated from dogs.
J Appl Bacteriol
 
1994
;
76
:
124
–34.
(10)
Holland RD, Wilkes JG, Rafii F, Sutherland JB, Persons CC, Voorhees KJ, Lay JO Jr. Rapid identification of intact whole bacteria based on spectral patterns using matrix-assisted laser desorption/ionization with time-of-flight mass spectrometry.
Rapid Commun Mass Spectrom
 
1996
;
10
:
1227
–32.
(11)
Petricoin EF III, Ornstein DK, Paweletz CP, Ardekani A, Hackett PS, Hitt BA, et al. Serum proteomic patterns for detection of prostate cancer.
J Natl Cancer Inst
 
2002
;
94
:
1576
–8.
(12)
Petricoin EF, Ardekani AM, Hitt BA, Levine PJ, Fusaro VA, Steinberg SM, et al. Use of proteomic patterns in serum to identify ovarian cancer.
Lancet
 
2002
;
359
:
572
–7.
(13)
Ornstein DK, Rayford W, Fusaro VA, Conrads TP, Ross SJ, Hitt BA, et al. Serum proteomic profiling can discriminate prostate cancer from benign prostates in men with total prostate specific antigen levels between 2.5 and 15.0 ng/ml.
J Urol
 
2004
;
172
:
1302
–5.
(14)
Wadsworth JT, Somers KD, Cazares LH, Malik G, Adam BL, Stack BC Jr, et al. Serum protein profiles to identify head and neck cancer.
Clin Cancer Res
 
2004
;
10
:
1625
–32.
(15)
Vlahou A, Laronga C, Wilson L, Gregory B, Fournier K, McGaughey D, et al. A novel approach toward development of a rapid blood test for breast cancer.
Clin Breast Cancer
 
2003
;
4
:
203
–9.
(16)
Villanueva J, Philip J, Entenberg D, Chaparro CA, Tanwar MK, Holland EC, Tempst P. Serum peptide profiling by magnetic particle-assisted, automated sample processing and MALDI-TOF mass spectrometry.
Anal Chem
 
2004
;
76
:
1560
–70.
(17)
Hingorani SR, Petricoin EF, Maitra A, Rajapakse V, King C, Jacobetz MA, et al. Preinvasive and invasive ductal pancreatic cancer and its early detection in the mouse.
Cancer Cell
 
2003
;
4
:
437
–50.
(18)
Conrads TP, Fusaro VA, Ross S, Johann D, Rajapakse V, Hitt BA, et al. High-resolution serum proteomic features for ovarian cancer detection.
Endocr Relat Cancer
 
2004
;
11
:
163
–78.
(19)
Malyarenko DI, Cooke WE, Adam BL, Malik G, Chen H, Tracy ER, et al. Enhancement of sensitivity and resolution of surface-enhanced laser desorption/ionization time-of-flight mass spectrometric records for serum peptides using time-series analysis techniques.
Clin Chem
 
2005
;
51
:
65
–74.
(20)
Koopmann J, Zhang Z, White N, Rosenzweig J, Fedarko N, Jagannath S,et al. Serum diagnosis of pancreatic adenocarcinoma using surface-enhanced laser desorption and ionization mass spectrometry.
Clin Cancer Res
 
2004
;
10
:
860
–8.
(21)
Li J, Zhang Z, Rosenzweig J, Wang YY, Chan DW. Proteomics and bioinformatics approaches for identification of serum biomarkers to detect breast cancer.
Clin Chem
 
2002
;
48
:
1296
–304.
(22)
Liotta LA, Ferrari M, Petricoin E. Clinical proteomics: written in blood.
Nature
 
2003
;
425
:
905
.
(23)
Mehta AI, Ross S, Lowenthal MS, Fusaro V, Fishman DA, Petricoin EF III, Liotta LA. Biomarker amplification by serum carrier protein binding.
Dis Markers
 
2003
;
19
:
1
–10.
(24)
Zhou M, Lucas DA, Chan K, Issaq HJ, Petricoin EF III, Liotta LA, et al. Investigation into the human serum interactome.
Electrophoresis
 
2004
;
25
:
1289
–98.