The field of diagnostics is undergoing revolutionary changes. Historically, searching for new diagnostic analytes – biomarkers – has targeted single molecules with sufficient information content to answer a clinical question definitively. This search has been frustrating; aside from genetic testing for a short list of diseases, today there are few analytes that provide the yes/no diagnostic answers provided by a test such as human chorionic gonadotropin for pregnancy. Even screening for a disease as prevalent as prostate cancer (using prostate-specific antigen [PSA], a comparatively good marker) is controversial, due to the limitations of the assay. The lack of results in this search for biomarkers reflects both the framing of the problem and the tools being employed.
The notion that single biomarkers can provide sufficient information about complex disease states is challenging. Diseases involve multiple, often overlapping, physiologic and pathophysiologic processes. Individual analytes that reflect these processes are thus particularly likely to suffer from a lack of specificity. In the context of prostate cancer screening, PSA provides an example: elevated PSA levels can reflect prostate cancer or, more often, benign prostatic hyperplasia (BPH). Sensitivity can also be an issue due to both the inherent trade-off between sensitivity and the natural variability in individuals with a given disease (some men with apparently normal PSA levels do have prostate cancer). Patterns of biomarkers, or disease signatures, that reflect the true complexity of the clinical situation are much more likely to provide the accuracy required for routine clinical use.
Numerous 2D polyacrylamide gel electrophoresis (PAGE) analyses of the most abundant 1500 or so human serum proteins show that 1–4% of these proteins change abundance in systemic diseases. Protein signatures comprising of five to ten independent markers for each clinical question (if the markers were of the same usefulness as PSA for prostate cancer screening) would yield 99–99.99% accuracy for that diagnostic decision. For a relatively rare but lethal disease such as ovarian cancer, an accuracy of 99.99% would be adequate for screening for that disease even in a normal-risk population.
Biomarker discovery and development strategies must focus on specific clinical questions. These are the problems faced by clinicians where initial information is required in order to determine how to most appropriately manage the patient. The most important categories of clinical questions include:
| • | Screening: is prostate cancer present in this asymptomatic patient? | ||||
| • | Differential diagnosis: does a patient with an enlarged prostate, which was detected by digital rectal examination, have prostate cancer, BPH or some other condition? | ||||
| • | Therapy selection: will this prostate cancer patient respond well to androgen ablation? | ||||
| • | Monitoring: has this prostate cancer patient suffered a recurrence? | ||||
Aside from the issue of how the biomarker discovery challenge has been framed, currently available tools have also hampered the search for novel biomarkers. The number of potential biomarkers is quite large; however, their discovery remains a huge challenge. Although there are only 25,000–40,000 human genes, they give rise (through alternative splicing, protein processing, glycosylation and phospory-lation) to 300,000–1,000,000 human proteins. This suggests that there are potentially 3000–40,000 biomarkers for severe systemic diseases, based on the 1–4% prevalence of biomarkers cited above. Researchers seeking protein biomarkers (the focus of this edi-torial) have faced an unacceptable trade-off across key performance parameters. Techniques such as 2D-PAGE, often coupled to mass spectroscopic analysis, allow researchers to look at a large number of proteins simultaneously. Unfortunately, 2D-PAGE analysis is limited to proteins that are sufficiently abundant to be visualized with a variety of stains. Quantification and reproducibility can be challenging, and routine clinical use of 2D-PAGE is difficult to imagine. The lack of quantification can be particularly problematic, since information about more subtle changes in protein levels is lost.
In contrast to 2D-PAGE and mass spectroscopic analysis, antibody-based sandwich assays, such as enzyme-linked immunosorbent assays (ELISAs), provide the highest quality information (e.g., limits of detection and reproducibility) about individual proteins. Low limits of detection and superior reproducibility are critical for biomarker discovery directed toward serum or plasma. These two matrices are the preferred methods for routine use since they are readily accessible. They also provide, in a sense, a molecular survey of the whole patient as they are in equilibrium with the physiologic and pathologic processes in tissues. Urine is another attractive matrix, though an edited one, since its make-up also reflects filtration by the kidneys. Despite their advantages, antibody-based sandwich assays suffer from a lack of scalability, which is critical for research use. To date, sandwich assay arrays have been limited to approximately 30–40 analytes; larger multiplexes result in unacceptable performance. Therefore, they are more useful for assaying analytes of interest that have already been identified than for seeking new biomarkers.
The preferred tools for seeking protein patterns, or signatures, will break through this compromise and allow the simultaneous assay of large numbers of proteins with the quality of individual measurements approaching the standards set by antibody-based sandwich assays. Discovering protein signatures for clinical questions will require a technology that permits analysis of a few hundred to a thousand (or more) simultaneous assays of high quality. The technology must also allow clinical labor-a-tories to translate signature discoveries to routine clinical use with minimal technological hurdles, such as fundamental switching of assay formats. This is a particular challenge for approaches that rely upon mass spectrometry, either alone or in combination with other techniques.
Aptamer arrays may offer a solution to the problem of developing new diagnostics. By combining a high degree of scalability with high performance for individual assays, they will enable the rational search for patterns of biomarkers. SomaLogic, Inc. is developing multiplexed protein arrays based on its proprietary aptamer technology. Aptamers are single-stranded nucleic acids derived from an in vitro evolutionary selection-amplification scheme. This selection scheme works since, unlike double-stranded nucleic acids, single-stranded nucleic acids fold up into unique 3D structures in a similar manner to proteins. Each unique structure is dictated by the sequence of the nucleic acid. By starting with 1015 random DNA sequences (thus, to a first approximation, 1015 specific shapes), it is possible to select (through 10–15 rounds of selection and amplification) specific binding reagents for each targeted human protein. Although aptamers are useful for many of the applications for which antibodies are employed, single aptamers, such as single antibodies, are generally not good enough for clinical diagnostics. The reason for this is the need for distinguishing specific protein signaling in a vast excess (serum or plasma) of hundreds to thousands of other proteins. More specificity is required.
For antibodies, the solution is the sandwich assay. In this format, two antibodies are used for two different epitopes of the same protein. Therefore, the specificity of antibody one (defined here as the probability of a nontarget protein being bound by the antibody) is multiplied by the specificity of antibody two to improve the signal-to-noise ratio and increase the specificity of the entire system. Aptamer-based assays can use the same approach of selecting two aptamers to different epitopes, but an alternative strategy allows aptamers to incorporate two multiplicative dimensions of specificity within a single aptamer per protein. Photoaptamers are single-stranded DNA (or RNA) molecules that include bromodeoxyuridine (BrdU) in place of T residues at specific locations in their sequences. Each photoaptamer is selected and screened to ensure that when it has folded up into the structure that binds tightly to a protein through affinity, a BrdU residue is sufficiently closely apposed to an electron-rich amino acid (usually one with an aromatic side chain). Short pulses of ultraviolet light at 308 nm then induce a chemical crosslink between the BrdU residue and the amino acid, yielding a covalently bound aptamer–protein complex. This gives rise to multiplicative specificity since two pseudoindependent parameters are used: affinity binding and crosslinking. This ability to achieve covalent attachment of proteins to aptamers using a single analytic reagent has tremendous consequences for multiplexed protein measurements in serum and plasma.
In clinical diagnostics, the ability to generate high signal-to-noise ratios determines the sensitivity of the assay. As many, probably most, potential protein disease biomarkers will be present at serum or plasma concentrations under 1 nM, the ability to measure multiple protein markers depends critically on having not only high signals, but also very low noise. In this context, the greatest single advantage of photoaptamer arrays becomes obvious, namely that the entire system involves elements that are all covalently bound to each other. Consider a competing assay system, the multiplexed ELISA chip. Multiple capture antibodies are deposited on a chip surface, almost always by hydrophobic adsorption. Thus, the initial contact of capture agent to chip surface, although strong, is noncovalent. Then serum is incubated on the chip and antibodies capture their protein targets. This interaction is limited by the dissociation constant (Kd) of the pair (and the relative Kds of all competing proteins for the same antibodies), again a noncovalent interaction. After mild washing (mild enough to disrupt neither of these two noncovalent linkages), the secondary (detection) antibodies are added, further washing is employed and finally a staining scheme (usually involving a third antibody) is applied to the chip.
Contrast this method to the photoaptamer chip. Using standard phosphoramidite chemistry, it is possible to introduce almost any functional group at the 3´ or 5´ end, and this can then be covalently linked to the appropriate surface. After incubation with serum or plasma, the chip can be mildly washed, as with antibodies. Irradiation at 308 nm light then covalently attaches the target protein to its photoaptamer; nontarget proteins are not bound. The entire chip can then be washed with detergents, base, acid or any harsh treatment that will lower noise by washing off proteins bound to the surface or proteins bound but not crosslinked to the aptamers. Since the array now consists entirely of aptamers with or without their bound target proteins, simple staining approaches can be used that target the primary amines in lysines (on the protein) but do not crossreact with the photoaptamers themselves. As a result, the signal generated from each feature on the array is proportional to the amount of bound (target) protein. In fact, it has been demonstrated that harsh washes, when used with the universal protein stain strategy described above, considerably reduce backgrounds on multiplexed protein assay chips.
Perhaps most importantly for biomarker and signature discovery, photoaptamers avoid the limitations associated with secondary antibodies. By avoiding the potential for crosstalk among secondary antibodies, the inherent scalability limit of sandwich arrays (∼30–40 analytes) is completely avoided. As noted above, the minimal scale needed for biomarker discovery is likely to be in the order of hundreds of proteins, therefore addressing scalability limits is critical. Photoaptamer arrays also benefit from the high consistency of photoaptamers. As they are manufactured synthetically, lot-to-lot variability is minimized, enhancing the reproducibility of photoaptamer-based assays. To date, no inherent limits on scalability have been identified, making arrays with hundreds of analytes a real possibility. It is not critical that these analytes be random proteins. In fact, driving content based on existing medical knowledge is a critical advantage for array-based approaches.
As photoaptamer arrays are scaled up, they can be used to discover and develop clinically useful diagnostics. As suggested above, the approach is medically driven, since it begins with the identification of a specific clinical question and then identifies serum or plasma biomarkers that are useful for distinguishing clinical populations (such as drug responders and nonresponders). These biomarkers are then combined to create the signatures. By using this strategy, researchers can focus on the areas of greatest need rather than those applications in which one might find, by serendipity, a single interesting biomarker. Initial feasibility studies have yielded encouraging results across a range of diseases, with an early focus on oncology and auto-immune diseases. As development continues, larger clinical studies will be pursued to confirm array performance in their intended settings. A key advantage of the photoaptamer array approach is that there is no need to make a transition between technologies (e.g., from discovering a signature with mass spectro-metry to measuring it with sandwich assays). In addition, there is no need to force an inappropriate technology into routine clinical use. Photoaptamer arrays and associated instruments are being developed for routine clinical laboratory use, beginning with clinical reference laboratories but ultimately extending even to the patient bedside. In fact, the only significant difference between research photoaptamer arrays and those used in clinical practice may be the scope of the array, since clinical arrays may feature a more focused set of analytes identified through the development process.
Although still in the developmental stage, photoaptamer arrays have the potential to transform the way new clinical tests are discovered, developed and used in practice. They can be part of a new approach to diagnostics that begins with the identification of a clinical need, recognizes the complexity of disease and emphasizes the absolute importance of high-quality results.
Acknowledgements
The authors would like to thank Larry Gold and all their other colleagues at SomaLogic, Inc. for all their work on this subject and contributions to this article.