A list of candidate cancer biomarkers for targeted proteomics.

1The study of cancer biomarker proteins began in 1847 with the discovery by Henry Bence-Jones of what turned out, more than 100 years later, to be a tumor-produced free antibody light chain “Bence Jones protein” in the urine of a multiple myeloma patient where it was present in large quantities and could be revealed by simple heat denaturation. One hundred and 40 years later this protein was demonstrated to be present also in the serum, and in 1998 a routine immunodiagnostic test was approved by the FDA. Hormones produced by tumors were also detected early on: adrenocorticotropic hormone (ACTH), calcitonin, and chorionic gonadotropin (hCG), for example, are elevated in specific cancer types, though not with the tumor specificity of Bence-Jones proteins.

2Unfortunately, the paradigm in which an overproduced tumor-specific protein can be easily detected as a marker of cancer has turned out to be the exception rather than the rule: in the nearly 160 years since Bence-Jones’ discovery, less than 10 proteins have progressed to the level of FDA-approved cancer diagnostic tests, and most of these lack ideal sensitivity and specificity for cancer.

3In recent years “… the emerging science of genomics and proteomics have generated a plethora of candidate cancer biomarkers”. Unfortunately few of these markers immediately stand out as superior prognostic or diagnostic tools, and even fewer have been validated and approved. Several factors might account for the slow pace of advance in cancer biomarkers. On the one hand, available proteomics technology has limited power to detect low-abundance cancer biomarkers against the background of high-abundance plasma proteins, and many of the best markers may thus be missed until discovery technology improves. On the other hand, the capacity to verify and validate existing candidate markers (through rigorous testing in large sample sets from many diseases) is limited, and it is therefore possible that the required biomarkers have already been “discovered” but not yet validated. In this paper, we are concerned with the latter possibility, and specifically with the problem of selecting among the existing candidates those that are most promising for systematic validation.

4This line of enquiry immediately raises the question: where is the list of known candidate cancer biomarkers? While a number of useful reviews and books discuss specific cancer markers with clinical promise, these generally concentrate on proven, or at least well-developed, markers or specific disease states. We were unable to find a list that draws together a large population of candidates at all stages of development from multiple discovery sources, and thus our first step has been to create one through a combination of literature search and other methods.

5The value of a list of existing candidates could be limited by the general lack of sensitivity and specificity exhibited by most of the cancer markers found to date, a factor that may have discouraged others from undertaking this task previously. Most candidates that have been followed up in larger studies have shown poor diagnostic value (Table 1), and even those that have been approved for clinical use exhibit lower sensitivity and specificity than the well-known markers of, eg, acute cardiovascular events (ie, troponin in myocardial infarction or B-type natriuretic peptide in congestive heart failure, Table 1).

6On the other hand, there seems to be a growing consensus that panels of markers may be able to supply the specificity and sensitivity that individual markers lack. For example a panel combining four known biomarkers (leptin, prolactin, osteopontin, insulin-like growth factor II), none of which used alone could distinguish patients from the controls, achieved a sensitivity and specificity of 95% for the diagnosis of ovarian cancer. In this case a combination of known proteins in a novel panel provided a significant advance. Xiao et al. identified 299 proteins in tissue culture by 1-D page and nano-ESI-MS/MS but then used ELISA to test 13 of the most interesting in serum. They reported that CD98, fascin, the secreted chain of the polymeric immunoglobulin receptor and 14-3-3 eta provide greater sensitivity when used together as a panel than any of the markers used alone.

7If, as we and others believe, panels of proteins provide the most promising avenue towards early and accurate cancer detection, then a re-examination of known candidates provides a logical approach to panel generation, with the expectation that a stream of new markers can be added as they are identified by marker discovery studies. This candidate-based, or targeted, approach will require a comprehensive list of prioritized candidates coupled with a technology able to assay these in large sets of plasma and serum samples from clinical and epidemiological studies (together a “biomarker pipeline”).

8Here we have begun to compile and prioritize a database of candidate biomarkers reported to be differentially expressed in studies of human cancer. We have included changes observed either at the protein (plasma or tissue) or nucleic acid (tissue DNA or RNA) level for any cancer, and excluded results restricted to animal, cell culture systems, or single case report studies in hopes of focusing on the most promising clinical biomarker candidates. We hypothesize that the protein version of most, if not all of these markers should be detectable in blood plasma at some level, irrespective of the tissue source, ultimately allowing for their use in patient screening, diagnosis or follow-up.

9The principal strategy for creation of our list involved compilation of designated cancer related proteins from: our previously published work, PubMed literature searches, cancer microarrays (868 proteins from 111 human cancer Superarrays (http://www.superarry.com and in supplemental material), Circulating Tumor Markers of the New Millennium, American Association for Clinical Chemistry abstracts and general literature perusal. PubMed searches included de novo PubMed literature searches: [plasma (Title/Abstract) NOT membrane (Title/Abstract) NOT stimulation (Title/Abstract) NOT drug (Title/Abstract) NOT dose (Title/Abstract) AND protein (Title/Abstract) AND cancer] and [“cancer antigen” AND human], as well as the PubMed literature search used for proteins from other sources [“protein name” AND cancer AND human AND (where necessary) diagnostic AND (where necessary) expression] and PubMed “related article” searches. Only proteins for which we found at least one published study on human cancer utilizing primary samples were retained (639 of the array proteins). Each biomarker reference was then manually tabulated and curated as to disease and tissue (including plasma). Single case studies were excluded.

10FDA approval dates for tests were obtained from the FDA Center for Devices and Radiological Health database. Proteins designated here as clinical markers are those offered commercially by ARUP or by Mayo Medical Laboratories, or else offered for internal use by either NIH or the Fred Hutchinson Cancer Research Center.

11Each documented protein on the resulting list was searched against the literature (via PubMed) using the query [“protein name” AND human AND cancer AND diagnostic]. This is admittedly a crude metric of research interest in a biomarker, but provides a useful method of relative prioritization among markers. In tabulating citation frequencies we did not exclude those categories ruled out in compiling the list initially: studies of animal systems, single clinical cases, or cell lines. If the “protein name” was not found by this search strategy it was counted as zero. It must be noted that PubMed is not a static archive but rather constantly changing both by additions, subtractions, and redefinition of MeSH headings. Still this exercise allowed some relative ranking of interest and therefore importance. Total cancer citations per year were determined using the query [human AND cancer AND diagnostic] limited to a specific publication year.

12Swiss-Prot/Uniprot accession numbers were obtained where possible. Most of the TrEMBL annotations were done prior to the addition of species information to the annotation number and so this form of the annotation was maintained. Candidate cancer biomarkers were annotated with GO numbers and IDs from EBI’s human GOA 30.0 (gene_association.goa_human, ftp://ftp.ebi.ac.uk/pub/databases) and the Gene Ontology’s GO.def version 1.213 (http://www.geneontology.org/ontology/GO.defs) respectively. Similar ID groupings were then combined. The entire Human GO file was treated in an identical fashion for comparison with the candidates.

13Where possible, normal or control values for the plasma concentration of each protein were obtained by literature search. Unless specifically noted, protein concentrations are for the intact protein not individual subunits.

14A search strategy combining literature search, extraction from microarray data, and a review of existing clinical tests, followed by manual curation, provided a list of 1261 candidate protein biomarkers (supplemental material) for which we found evidence of a quantitative change in some human cancer. As shown in Table 2, the candidates included proteins known to occur in plasma (274), proteins detected in tissue samples (542), and proteins whose corresponding mRNA or DNA levels were differentially expressed between cancerous and normal samples (656). These categories are non-exclusive in that a significant number of the candidates were found in more than one type of study. Proteins detected in the plasma represent 22% of the total proteins documented to date.

15Citation frequency analysis was used as one method of prioritizing the biomarkers, on the assumption that proteins most widely studied in the context of cancer had more promise as biomarkers. Citation frequency was determined using a PubMed query intended to count citations in which the authors considered the proteins to have diagnostic value (Figure 1, Table 3). When this is done, 29% of the 1,261 biomarkers have no such citations, 67% have fewer than 10, and 74% fewer than 20. Likewise only a very limited number of biomarkers have extensive citations, 62 proteins or only 5% of the total number of biomarkers were found to have greater than 500 citations.

16Of the 34 biomarkers with more than 1000 citations (Figure 2, Table 3) 79% are found in the plasma and 56% are presently used clinically (89% of which are reported in plasma). Of the 28 markers with between 500 and 1000 citations (Figure 3) 57% are plasma proteins but only 7% are used clinically. Both of the markers used clinically are plasma proteins. Some proteins with high citation frequency (eg, albumin) are somewhat surprising to see in the context of cancer biomarkers; these have been retained nevertheless because they appear to have reasonable relevance (low serum albumin levels are prognostic of poor survival as noted in the table contents).

17In an effort to include more recently discovered biomarkers we also looked at the proteins that had greater than 100 citations in 2004 or greater than 50% of their citations in 2004. Of the proteins with more than 100 citations in 2004, all but COX2 are represented in figures 2 and 3. Of those with a majority of total citations occurring in 2004, most have a low number (<10) of absolute citations (Figure 4, Table 3), 32% are detected in plasma and none are presently being used clinically.

18We tracked the number of citations per year for selected cancer biomarkers over the last 35 years (Figure 5). The number of times a protein was cited in a given year (“protein name” AND cancer AND human AND diagnostic) was divided by the total number of cancer citations for that year (cancer AND human AND diagnostic) to give a rough index of the prominence of the biomarker in cancer research. Although frequently cited in the 1970’s and 1980’s, interest in CEA has dropped dramatically. The most cited marker in this group, PSA, has well-documented limitations as a diagnostic yet it continues to be cited either as the only option or as the biomarker upon which to improve. Interest in most of these biomarkers evolves in a fairly similar way: each appears to take a few years to be recognized, followed by gradually increasing interest over the following 15 to 20 years. Of these markers the FDA has approved only three as diagnostic cancer antigens: alpha-fetoprotein, CEA, and PSA (approved May 31, 1988, October 15, 1980 and February 25, 1986 respectively; Figure 5). To date only 6 additional markers have been approved by the FDA under the category of tumor associated antigens: CA 19-9 in May of 2002, Her2/Neu in September of 2000, CA 15.3 in February of 1981, bladder tumor marker in April of 1997, thyroglobulin in March of 1999 and CA 125 in July of 1987 (Table 4). None of these markers, used singly, has over 90% sensitivity and specificity. Although these numbers are for specific assays, they are representative of the general lack of specificity and sensitivity of the individual cancer markers currently available.

19We attempted to collect normal plasma concentrations for candidate cancer biomarkers reported in the literature. The resulting 211 values were histogramed (Figure 6) for comparison with the distributions of concentrations of either unselected plasma proteins from PPI’s plasma protein database, or a set of candidate cardiovascular biomarkers. The cancer candidates cover a >10-log concentration range with proteins such as immune modulating interleukins (1α andβ, 2, 5, 6, 9, 10, IFN-γ and GM-CSF) being present in normal plasma or serum in the pg/mL range while classical plasma proteins (albumin, transferrin, fibrinogen, and α-2-macroglobulin) are present at mg/mL levels. When the cancer candidate distribution is compared to the concentrations for all plasma proteins (unpublished results) and plasma markers of cardiac disease, a greater proportion of the cancer candidates appear in the lower concentration ranges than general plasma proteins or cardiac markers. Thus normal values for 185 (88%) of the markers for which we know the plasma concentration fall below 10 microgram/mL and 103 (49%) fall below 10 ng/mL. Tabulated concentrations are those found in controls not patients. Thus in many cases these may increase in cancer, thereby aiding in their detection.

20We compared the distribution of GO annotations for the cancer candidates with the distribution for all annotated human proteins over a series of summary categories, with the aim of finding any large biases in the cancer group. In comparing “Biological Process” GO annotation, the cancer biomarkers show an increased representation of apoptosis, cell cycle and proliferation annotations; processes blocked or increased in tumors (Figure 7); while metabolism, catabolism and transport proteins are decreased. When the two sets are compared by “Cellular Component” GO terms (Figure 8), the extracellular category is over represented in the cancer biomarker database in comparison with the whole human database (20% versus 6% respectively). This is true even if the proteins found experimentally in plasma are excluded (12% extracellular). The other Cellular Component categories show only small differences between the proteins sets. Comparing “Molecular Function” GO terms, only small differences are apparent between the cancer candidates and the whole annotated human proteome.

21Given the size of the list of candidates resulting from our assembly procedure, we attempted to select a smaller subset of higher priority candidates as a starting point for consideration of assay development and clinical validation. This subset comprising 260 proteins (Table 3) was compiled from the most highly cited proteins, the “recent” markers, plasma proteins of known concentration (indicating existence of an assay) and any marker presently in any type of clinical use. Many of these markers fall into expected categories such as immune modulation molecules (acute phase proteins, coagulation factors, immune modulators); and mediators of classical cancer pathways (oncoproteins, angiogenic or apoptosis factors, tumor suppressors or antigens, cellular homing or proliferation molecules). Somewhat less expected perhaps is that almost 22 (8%) of these top 262 proteins are involved in hormonal action.

22For each of the 260 high priority candidates, we performed web searches, primarily through the Exact Antigen website (www.exactantigen.com), to determine whether an antibody with potential utility in a plasma assay is commercially available. Relevant antibodies were found for 186 (72%) of the 260 high priority candidates.

23According to the Centers for Disease Control, 1 in every 4 deaths in the United States is due to cancer. Many of these deaths could be averted by improved early cancer detection, since existing therapies, especially surgery, are much more effective in early cancer stages as compared to later stages. Billions of dollars have been spent on basic research looking for molecular differences related to cancer-work that has been at least partly motivated by the need for improved in vitro diagnostic tests to detect or monitor progression of cancer. Yet to our knowledge no centralized database of known candidate cancer biomarkers exists. Such a list could serve to confirm new results, eg, from proteomic comparisons of cancer and control sera, by placing them in a context of earlier work. Additionally it could serve as a reservoir of current and future candidates to be tested in large sample sets by candidate-based (“targeted” or “directed”) proteomics methods. The latter use is important, since candidate-based methods, consisting of specific assays for defined targets, are likely to be much more sensitive than proteome profiling methods, and hence could cover a much broader universe of protein candidates and potentially detect disease states earlier.

24The present catalog of 1261 human candidate cancer biomarkers is a first attempt at such a database. We did not select specific cancer types or specific detection methods, choosing instead to cast a broad net. In the resulting list, it will be apparent that the strength of evidence and likelihood of ultimate usefulness of the candidates varies widely. Even candidates that have been tested and found to have poor diagnostic specificity and sensitivity were retained, as they may nevertheless contribute to useful panels as in the work of Mor and Xiao. Looking at the list, one might question why the most abundant plasma protein (serum albumin) is included – though perhaps counter-intuitive, albumin does meet the search criteria used, and is in fact a useful negative acute phase indicator likely to be altered in cancer along with many inflammation-related proteins. Other well-known proteins not usually considered as cancer-specific are also included (eg, protein and peptide hormones overproduced by endocrine tumors or through ectopic synthesis). Overall, the list is not easily recognizable by inspection as a list of cancer markers.

25Of the 1261 proteins, 22% are reported to occur in plasma. This is an appreciable fraction considering that many of the large array studies, capable of finding many markers per experiment, have looked for differential protein or DNA expression in tissues. For bona fide cell-associated cancer markers such as Her-2, there is persuasive evidence that at least a fragment of the protein molecule is released into the plasma and can be detected as a cancer biomarker, and other proteins documented here in the tissues of cancer patients have been demonstrated to be found in plasma in other disease indications. These cases provide some support for the hypothesis that most if not all of the 1261 proteins should be detectable at some level in plasma, the diagnostic sample of choice, given a sensitive enough assay. Whether current assay technologies will be sensitive enough to see a large fraction of the candidates in plasma is a major question at this point, and one that will require vigorous efforts to resolve.

26As might be expected, there is a smooth distribution in the number of literature citations per candidate, ranging from almost 8,000 (for PSA) to zero (for candidates not mentioned as diagnostic by the publication’s authors). This result suggests that our literature analysis did not identify a crisply defined set of cancer markers, but rather part of a continuum extending from a few established markers through plausible candidates into more speculative possibilities. Given the complexity of cancer, such an outcome is not surprising.

27Only 5% of the 1261 candidates have been extensively studied (500 or greater total citations over the years). When examined as a function of time, the citation history of individual markers appears to show a slow evolution of interest that peaks 15 to 20 years after the initial papers. Only in the cases of CEA and PSA was discovery of a biomarker followed by a rapid increase in publications over a few years and in the case of PSA the steady increase was seen only 10 years after the first citations appeared. Thus in order to catch recently emerged candidates, we focused on candidates with a high proportion of citations occurring in 2004 but with fewer total citations (often 10 or less). Of the total 1261 proteins only 41 are used in some clinical sense and even fewer have FDA approved assays.

28While the observed slow pace is easily explained by the deliberate nature of clinical research and the progressive, rather than abrupt, nature of adoption in medical practice, it presents a stark reminder of the challenge involved in making any rapid advance in cancer diagnostics.

29These candidate cancer markers, taken as a group, appear to be present in plasma at lower concentrations than comparable groups of cardiac markers or unselected plasma proteins. Although systematic biases in selection of these groups could affect this result, it tends to support the contention that plasma cancer marker discovery is, and may continue to be, a challenge in terms of detection sensitivity. Present discovery proteomics platforms typically detect proteins with plasma concentrations in the mg/mL to microg/mL range. For the proteins in our list with known plasma concentrations, we estimate that 86% would be missed by most conventional proteomics platforms, while 48% would be missed by high-end proteomics platforms with extensive multi-dimensional fractionation. For the present, the only way that many of these proteins can be detected is by specific assays: ie, by targeted proteomics. Targeted proteomics thus represents a preferred path to validation and further study of the candidate markers listed here.

30The distribution of our cancer biomarker candidate proteins among GO annotation categories shows remarkable similarity to the distribution for all annotated human proteins. There is some enrichment for proteins annotated as related to apoptosis, cell cycle and proliferation (in the GO biological process category), as would be expected on account of the fundamental involvement of these processes in cancer. The extracellular group (in the GO cell component category) is also somewhat over-represented, a trend favorable to detection in plasma. Nevertheless the candidates seem to represent a very wide sampling of the human proteome.

31The full set of these 1,261 candidates is too large to submit for immediate verification and validation in large sample sets by any available means, and some method of prioritization is required to initiate their evaluation. As an initial approach, we have selected a subset of the candidates based on a set of criteria including number of total citations, number of recent citations, proportion of recent citations, known plasma concentration (implying existence of an assay) and clinical use in any context. This subset of 260 candidates (presented in Table 3) includes 186 candidates for which a relevant antibody is commercially available, opening the possibility of testing this group using an antibody array or other miniaturized immunoassay technology in the near future.

32While the list of candidate cancer biomarkers assembled here is clearly a simplistic and therefore somewhat crude initial catalog, we believe the result will prove to be of sufficient value to justify extending the effort to provide an ongoing summary of the progress of cancer diagnostics. In particular we believe that linking a database of marker candidates to the bioinformatics architecture used in biomarker discovery will help to connect the discovery and validation phases necessary for progression of biomarkers to the clinic. One can envision a steady accumulation of candidates, regular revision of candidate priorities as evidence emerges from multiple sources (literature, microarrays, systems models, etc), and finally feedback in the form of specific measurements from validation studies in large sample sets. Such a collection of data would provide an up-to-date snapshot of the workings of a cancer diagnostic marker pipeline.

33Finally, lists such as this prompt important, but infrequently-asked questions regarding the most productive tack for future discovery efforts. Is it reassuring to find confirmation of fresh observations through overlap with a pre-existing list? Perhaps so, and particularly if the candidates involved appear repeatedly in similar independent studies. However the sieve used here is crude and so our list cannot really “confirm” a candidate seen in a new study–overlap just improves the odds of relevance. Further, since there are certain to be good cancer markers not on this list, failure to appear here in no way disqualifies a novel marker. Hence our hope is to contribute a mechanism for marginally improving chances of recognizing a valid marker, and a systematic source for enriched candidates available for validation and panel assembly efforts.