Available online at www.sciencedirect.com
Proteomics and diagnostics: Let
’s Get Specific, againDom Zichi1, Bruce Eaton1,2, Britta Singer1 and Larry Gold1,3DNA array technology has changed all discussions about
proteomics. Whole genome arrays allow unbiased
experimentation, and the surprises that flow from those
approaches. ‘Whole proteome’ proteomics is not possible
today, and might never be possible unless experiments are
guided by careful evaluation of reagent specificity. In this paper
we explore some possible ways to increase the content of
proteomic analysis.
Addresses1 SomaLogic, 1775 38th Street, Boulder, CO 80301, USA2 Department of Chemistry and Biochemistry, University of Colorado,
Boulder, CO 80309, USA3 Department of Molecular, Cellular and Developmental Biology,
University of Colorado, Boulder, CO 80309, USA
Corresponding author: Zichi, Dom ([email protected]), Eaton,
Bruce ([email protected]), Singer, Britta
([email protected]), and Gold, Larry ([email protected]),
Current Opinion in Chemical Biology 2008, 12:78–85
This review comes from a themed issue on
Proteomics and Genomics
Edited by Natalie Ahn and Andrew H.-J. Wang
Available online 7th March 2008
1367-5931/$ – see front matter
# 2008 Elsevier Ltd. All rights reserved.
DOI 10.1016/j.cbpa.2008.01.016
IntroductionSome years ago we published a short paper called ‘Let’s
Get Specific’ [1��] in which we tried to understand what we
thought to be the extraordinary binding specificity of
aptamers [2,3]. We thought that elements of aptamer
affinity and binding specificity were derived in part the
large libraries used to find aptamers [often as many as 1015
molecules for a SELEX experiment [4,5], and recent
developments [6��]], as well as the structural possibilities
explored by single-stranded oligonucleotides. More
recently we have focused our attention on the more general
nature of biochemical specificity, and we have wondered
about the connection between reagent specificity and
proteomics. When one considers proteomics, the reagents
of historical interest are antibodies; we shall try in this short
review to compare binding specificity for both antibodies
and aptamers and the ways in which proteomics might be
scaled to deliver on the promise of high analyte density and
extreme sensitivity and specificity.
A beautifully written (and referenced) recent review by
Borrebaeck and Wingren [7��] is aimed at a substantially
Current Opinion in Chemical Biology 2008, 12:78–85
different question than we address in this article. Borre-
baeck and Wingren have focused attention on the
impressive (and growing) list of improvements to various
components of an antibody-based proteomic array. Our
focus is on what specificity is possible with various
reagents, and to raise the possibility that array formats
must solve any intrinsic limitations of those reagents.
High analyte density is often abbreviated as ‘content’ – in
the case of nucleic acid arrays, ‘content’ eventually
included probes for entire genomes of viruses, bacteria,
yeast, model organisms (flies, worms, and the mouse), and
humans. Using large arrays, scientists have utilized
mRNA and SNP profiling, along with epigenetic DNA
methylation, as genome-wide biomarker-discovery tech-
nologies. All three platforms are possible because gen-
ome-wide specific hybridization is possible – that is, by
judicious use of proper probe lengths and appropriate
base composition along with the right temperature, buffer
conditions, and hybridization time, specific sequences
can be recognized in the face of an entire genome.
DNA ‘chip’ technologies are remarkable as engineering
marvels, but their discovery power flows from the high
specificity of nucleic acid hybridization. In fact, not
surprisingly, this same high specificity is the hallmark
of how nucleic acids perform their functions in biochem-
istry. Remarkably, however, the array manufacturers have
reached elegant solutions even as they built their content
on to compromised platforms. Probes bound to surfaces,
be they beads (Luminex), slides (Agilent, Affymetrix,
NimbleGen), or things in the middle (Illumina) present
less than perfect kinetics and slow approaches to equi-
librium. These approaches are quite unlike hybridization
in solution as it was originally developed [8,9]. Only rarely
did platform builders include approaches that overcame
the slow kinetic approach to equilibrium (NanoGen,
MetriGenix, PamGene, etc.), and those sensible plat-
forms seem to have lost in the market place.
Human diagnostics is better served by protein measure-
ments than by nucleic acid measurements, and served
best by protein measurements in blood samples (a matrix
with a vast number of proteins – see below). Human blood
is an integrator of much of what happens in the body, and
an effecter of much of biology, often because of secreted
proteins (such as growth factors) whose job is to move
through the blood to a nearby or distant site. In addition, a
variety of proteins make their way into blood when some
pathology causes localized cell death and the uninten-
tional release of proteins. Since blood equilibrates quickly
with all human tissues, including brain, panels of protein
biomarkers should become the earliest warnings one has
www.sciencedirect.com
Proteomics and diagnostics: Let’s Get Specific, again Zichi et al. 79
for the early stages of disease, even when a person is
asymptomatic. Of course these same biomarkers can be
present at vanishingly low concentrations and measuring
these low abundant proteins is the major hurdle proteo-
mics must overcome.
Of the approximately 23,000 human genes and their
>100,000 encoded proteins (comprising splice variants,
post-translationally modified proteins, and even more rare
events [10�], we do not know how many proteins are found
in blood. Probably every human protein is present in blood
at a very low level (if only from cell death), and perhaps
several thousand are present between the concentrations of
the most abundant blood proteins (albumin at just under
1 mM) down to protein concentrations at about 1 fM – a
dynamic range of 12 logs or more! The problems in human
proteomics and diagnostics are to scale proteomics to high
content to discover useful biomarkers and to make avail-
able diagnostic products that utilize (smaller) panels of
proteins for specific medical purposes. That is, in the same
spirit as made possible by nucleic acid array technologies,
one must survey large fractions of the proteome content inan unbiased manner for novel biomarker discovery.
One might think this is a simple task, given the stunning
success with nucleic acid arrays. The problem of course is
that typing nucleic acid complement sequences (probes) is
a lot easier than understanding the biophysics of protein
recognition biochemistry. Indeed, the allure of typing
enthralled the antisense, ribozymes, and siRNA thera-peutic researchers with the hope that typing would be a
great way to identify new drugs. For proteomics the idea
has been to replace typing of nucleic acid complements
with typing orders to the antibody suppliers. However,
when commercial antibodies are printed (as though they
were analogues of nucleic acid probes) and then tested
with various protein mixtures, performance (meaning
specificity) probably is not adequate. The purpose of this
review is to discuss these attempts and the aptamer-based
alternatives, and to note the intrinsic kinetic problems
that must be solved by proteomics.
Reagent-free proteomicsWe mention briefly that reagent-free proteomics would
be a wonderful development, although it appears diffi-
cult. Since 1975, with the publication of Pat O’Farrell’s
extraordinary work on 2D gels [11], such reagent-free
proteomics has been possible. The quality of the first 2D
gels was amazingly high (something like 1100 proteins
were visible), major and minor proteins seemed to differ
quantitatively by three logs or so, and differently charged
species of the same molecular weight were not a major
source of additional spots that cluttered the patterns. In
Pat’s thesis seminar he showed wild-type Escherichia coligels versus ‘lactose-operon deletion’ gels, and the missing
spots were a powerful demonstration of what might be
done, at least qualitatively.
www.sciencedirect.com
Blasting through many samples (tissue extracts or plasma
or serum, or urine, or whatever) with 2D gels had a
moment (‘back in the day’ as our children say . . .).However, the problems with reproducibility and quanti-
fication were serious, and the cost per analysis was high –
eventually people added MS to the methodology, and it is
now common for an entire 2D gel to be extracted (feature
by feature) for MS analysis. Furthermore, the limitations
in protein number (content again) have never been solved
– Pat O’Farrell’s number from E. coli really has remained,
after a lot of work, about the number of protein spots one
can visualize, and the proteins observed are thus inevi-
tably the most abundant proteins in the sample.
Mass spectrometry (MS) has had a similar fate, so far.
Even though the sensitivity of a great mass spectrometer
might enable analysis of samples quite a bit lower than
nmolar in a few microliters, when complex matrices are
explored the noise obscures all but the most abundant
proteins. In fact, it appears that 2D gels and MS query
more or less the same most abundant proteins from within
a complex sample. Clearly, the resolving power of 2D gels
coupled to MS in proteomics has yet to match what is
capable in DNA hybridization micro arrays.
Reagent-dependent proteomics: antibodiesNo one really knows what fraction of the human pro-
teome has been used to generate high quality antibodies
[see the Human Protein Atlas (http://www.proteinatla-
s.org/), [12��]] or protein reagents with alternative frame-
works [13�]. The assumption has been that antibody
production could be scaled up to meet the need, and
that people would be able to print antibodies the same
way that people print nucleic acid probes.
Just how specific are antibody reagents, and (thus) would
an array of printed antibodies allow quantitative proteo-
mics? Sadly we do not have a literature quite like the
definitive literature for nucleic acid hybridization. The
‘Turner’ rules [14] and hundreds of other careful papers
(e.g. [15,16,17] and many, many others) allow one to
calculate the likely interactions between two single-
stranded oligonucleotides. Proteins are likely to interact
with other proteins with low affinity (comprising diffusion
limited association rates along with fast dissociation
rates), but we have no rules. However, three independent
lines of evidence suggest that proteins are sloppy in their
intermolecular interactions and that one ought to expect
the equivalent of ‘error-prone DNA hybridization.’
First, phage display has been used to identify short
peptides (many experiments were done with peptides
10 amino acids in length) that will bind to a target protein
[18,19]. Usually binding peptides are found, with Kds
between 1 mM and about 10 nM. The classic phage
display libraries in the early days contained roughly
107 peptides, about the number present in the human
Current Opinion in Chemical Biology 2008, 12:78–85
80 Proteomics and Genomics
proteome. Many peptides are found to bind weakly to
almost any target protein. Binding interactions of low
affinity should be expected as ‘noise’ in human biology.
Second, intracellular protein-interaction maps have been
made (and continue to be made) using some variant of the
yeast two-hybrid system [20,21]. Again one fights noise –
whatever bait protein is used, many candidate proteins
emerge that interact weakly with the bait and are either
noise or subtle reflections of meaningful biology – it is
very difficult to tell which. Probably the ways that the
two-hybrid system is tuned matters greatly – if both
proteins are expressed at high intracellular levels, weak
interactions (probably reflecting mM Kds or even greater)
are sufficient to activate transcription.
Third, antibodies found in healthy people (these obser-
vations are not about people who have known auto-
immune disorders) react with roughly 28% of the
human proteins [13�]. This is an astonishing observation,
and almost certainly most often reflects weak and unin-
tended binding. That is, in the format employed, anti-
bodies against other targets react with human proteins.
But how could this be? Did not we learn that antibodies
are magic bullets? In fact, this platform is shown
(Figure 1B – below) and represents a particularly com-
promised platform because of very slow effective dis-
sociation rate constants for unintended non-target
Figure 1
Specificity: A function of rebinding rates. In the binding reactions shown, the c
green or blue. Intrinsic on and off rates are identical in the examples. (A) An
mixture of antigens. (B) Reverse arrays with homogenous antigen spots and m
example tissue homogenate, probed with a single antibody. Intrinsic on and
limited on rates, with solution off rates). Rebinding rates (for cognate and inc
very slow in (C).
Current Opinion in Chemical Biology 2008, 12:78–85
proteins. The slow effective dissociation rate results from
multivalent binding of antibodies to the ‘wrong’ human
proteins.
Antibody arrays: printed antibodies, reversearrays, and sandwich formatsThree kinds of protein-based proteomic arrays have been
tried. The easy approach was to simply print antibodies
and see what would happen, and then to make even
better antibodies to see if performance was improved.
These arrays are called ‘antibody capture’ arrays and are
commercially available. When the difficulties of sensi-
tivity and specificity with these arrays became clear,
people tried ‘reverse arrays’ in which, for example, plasma
or tissue homogenates were spotted directly on to a
surface and probed with antibodies. For reasons visible
in Figure 1 reverse arrays may in fact be more specific
than antibody-capture arrays. Finally, people have tried
very hard to deploy antibody sandwiches onto an array
format.
First, arrays of single antibodies are now available com-
mercially (reviewed in Borrebaeck and Wingren, [7��]).The content is impressive – hundreds of antibodies are
probed (often) with a standard mixture of proteins and an
unknown but similar mixture, with one mixture labeled
with one color dye and the other with another. While the
Brown lab at Stanford was an early proponent of such
orrect interaction is blue with green. The wrong interactions are pink with
tibody capture arrays, with single antibody in each spot, probed with
ixture of antibodies, for example plasma. (C) Mixed antigens spotted, for
off rates are identical in the examples (A), (B), and (C) (usually diffusion
orrect interactions) are extremely fast in (B), somewhat slower in (A), and
www.sciencedirect.com
Proteomics and diagnostics: Let’s Get Specific, again Zichi et al. 81
experiments, using commercial antibodies [22], the stron-
gest work thus far comes from Borrebaeck’s lab in Sweden
[7��]. Over several years that lab has developed surfaces
for printing antibodies, discovered which antibody frame-
work was best at obtaining high quality antibodies with
stable structures and good analyte recognition, and so on.
The work is lovely. The use of such arrays becomes
straightforward. If, for example, a large number of plasma
samples from healthy women were mixed together and
another set of plasma samples from similar women with
stage II ovarian cancer were compared on a large antibody
array, any differences in the average protein concen-
tration in the two sample sets would show up as a change
in the ratio of dye #1 to dye #2 on a specific antibody
feature. This now-standard experiment (patterned after
mRNA profiling arrays) has been reported several times
[23,24�,25], and soon will become a cottage industry. We
hope that data from antibody arrays will be confirmed by
high quality ELISAs, used to probe the same plasma
mixtures for that analyte (or, even better, each individual
sample from the plasmas that made up each mixture [26]).
The formalism is obvious – oligonucleotide array (hybrid-
ization) data are confirmed by something precise (QPCR)
because even hybridization arrays (as specific as they are)
do not guarantee that the measured oligonucleotide is
the intended oligonucleotide. The proteomics com-
munity will learn about this problem, although slowly,
exactly as the DNA-array community learned about the
problem.
Figure 1A pictures the problem and suggests a key metric
for evaluating even very good antibody arrays. Put
simply, the burden of proof lies with the manufacturers,
distributors, and academic scientists who claim high
specificity in array formats. The question is at least
partially answered by spiking plasma samples with
non-human proteins at low concentrations to determine
the limit of detection and to measure ‘background’ in
unspiked plasma. We imagine that a set of (ten or so?)
non-human proteins will become a standard ‘specificity
metric’ so that various platforms and reagents may be
compared. The severity of the problem goes beyond
limits of detection, of course. If an abundant protein in
plasma binds inappropriately to an antibody selected for
a non-abundant protein, the measured protein will not be
the intended analyte. This is a real issue – if several
plasma proteins are 106 to 1012 times more abundant than
the analyte of interest and those abundant proteins have
even submicromolar Kds for the capture antibodies aimed
at the intended analyte, a significant fraction of the signal
on that capture antibody will result from the wrong
proteins, due only to equilibrium binding. Moreover,
avidity/rebinding components can increase the noise
on an antibody array. Equilibrium binding for target
analytes and all other proteins in a mixture is quite a
hurdle to overcome in printed antibody arrays (and see
next paragraph).
www.sciencedirect.com
Second, reverse-phase protein microarrays (RPAs) have
become popular. A key reference is from the Petricoin
and Liotta labs [27��]. As shown in Figure 1C, multivalent
recognition in an RPA is unlikely – a spotted tissue extract
or plasma sample will have thousands of proteins, and
statistically it is unlikely that identical proteins will fall
near each other, and thus provide opportunities for multi-
valent binding. However, in the review cited above, the
authors write that ‘Currently the two biggest technical
challenges facing this techniques are the need for specific
antibodies and . . .’ We agree, but in fact this is just the
equilibrium binding problem mentioned above, but noth-
ing worse than that. Again, as previously mentioned,
antibody specificity almost never is good enough to yield
binding only to the intended analyte in the presence of
many other vastly more abundant proteins. We believe
this is inherent in the biochemistry of antibody CDRs
(complementarity-determining regions) and cannot be
solved easily even with recombinant antibody technol-
ogy.
There is, of course, an additional issue facing RPAs. The
spotted analytes are of uncertain protein state, from
native to denatured and everything in between. Anti-
bodies often are not well characterized with respect to the
preferred analyte structure, which could be an important
attribute in biology, and this must confound the use of
RPAs. The good news (and the bad) is that RPAs are easy
to use, and thus another cottage industry has been born,
but getting the right assay for the serious study of human
health is unlikely to be easy.
Third, an enormous amount of time and money has been
spent trying to build arrays of antibody sandwiches. We
show (Figure 2a) why sandwich assays won the battle for
single analyte measurements for diagnostics – specificity is
the product of the two specificities of the two antibody
reagents in a sandwich. If a capture antibody is bound by
proteins other than the intended analyte, the second
antibody (which recognizes a different analyte epitope)
will NOT also bind to the inappropriately captured
unwanted protein. One problem with adapting this sand-
wich format for protein arrays is that any nonspecifically
bound protein with a secondary antibody in the assay will
now signal for the wrong analyte (Figure 2b). As the
sandwich array content grows, these nonspecific inter-
actions will grow, limiting the practical size of this format
for arrays of antibody sandwiches. The content limita-
tions to sandwich arrays are depicted (Figure 2b) and
argue powerfully that sandwich arrays will not be used for
novel biomarker discovery because they cannot scale to
high content.
So formats matter here (Figures 1 and 2). Forcing protein–
protein interactions as required to drive high content
proteomics arrays is in conflict with the rather delicate
structural integrity of these biopolymers. When proteins
Current Opinion in Chemical Biology 2008, 12:78–85
82 Proteomics and Genomics
Figure 2
Antibody sandwich assays. (a) As shown, antibody sandwiches help with specificity. The correct interaction is blue with green. The wrong interactions
are pink with green. The second antibody (dark green) provides additional specificity and the incorrect interaction with the capture antibody is not
recognized by the second antibody. (b) The specificity is lost when many proteins are probed in the same assay. When antibodies to the pink antigen
(red is the capture antibody, and brown is the secondary antibody) are added to the array, the wrong interaction in (a) will now generate signal and
specificity is lost.
within a printed feature are crowded they may denature
and further lose specificity. In addition, when the soluble
partner is bivalent, weak binding might be sufficient for
kinetic entrapment, which really is nothing more than
avidity (which is nothing more than artificially increased
rebinding rates or even multivalent binding caused by
packed proteins attached to the surface with close pack-
ing). So these features– substantial weak interactions
between many abundant proteins, along with bivalent
probes, must lead to measurements that are obscured by
noise.
The take home message is clear: antibodies will bind
more specifically to protein analytes in solution or even in
Current Opinion in Chemical Biology 2008, 12:78–85
cells or in blood than they will if they are used as capture
antibodies or as antibody probes against printed human
proteins. Surprisingly, reverse arrays offer some advan-
tages over antibody capture arrays, which we did not
appreciate when we started writing this review.
Reagent-dependent proteomics: aptamersOur colleagues and we have been working on aptamer
arrays for almost two decades. We collectively come to
proteomics from backgrounds in genetics, molecular
biology, biochemistry, and physical chemistry. For at least
ten years we worked to format arrays of aptamers to
measure proteins – we even published a paper on a mixed
sandwich protocol (using an aptamer and an antibody –
www.sciencedirect.com
Proteomics and diagnostics: Let’s Get Specific, again Zichi et al. 83
[28]). Equilibrium binding assays led to comparable per-
formance (and limitations) as those achieved by anti-
bodies in any of the formats discussed above. Recently,
we have been using kinetic manipulations so that mere
equilibrium binding will not need to do the impossible –
to distinguish between the intended analyte and to not
bind significantly to any abundant and inappropriate
protein.
We have tried to solve the problems identified in this
review article without using sandwich arrays (which prob-
ably would not scale with aptamers any better than they
scale with antibodies, as in Figure 2). The principle
obstacle we needed to overcome was to identify a second
specificity element (beyond equilibrium binding) that
could be built into the assay; at one point we worked
very hard on photo-crosslinking as the second specificity
element [29�].
We have had some success (Figure 3) at quantifying many
human proteins with aptamers that have both low Kds for
their cognate proteins and higher Kds for the abundant
proteins in plasma. In addition, we have been able to
Figure 3
Multiplexed proteomics with aptamers work well. (a) Multiplexed readout of
(blue) and 5% plasma (red) and measured on arrays.
www.sciencedirect.com
select aptamers with very slow dissociation rates, some-
thing we had tried to do unsuccessfully many times in the
past. These slow dissociation rates allow us to remove the
(abundant) non-target proteins that would otherwise con-
tribute to noise. Most importantly, we have been able to
format assays so that binding discrimination occurs in
solution before array read-out.
In Figure 3 we show some proteomic data, both for an
aptamer-array with many proteins measured simul-
taneously (Figure 3a), and spike-and-recovery exper-
iments (Figure 3b) in 5% plasma (with an expected
neurotrophin-3 concentration of about 1–5 pM) and buf-
fer. The significance of these data is that broad and
quantitative measurements of (low) protein concen-
trations are now available.
The details of these experiments will be submitted
shortly (personal communication from Dan Schneider,
Sheri Wilcox, Jeff Carter, Marty Stanton, and many others
at SomaLogic). We continue to be instructed, intellec-
tually, by the kinetic descriptions of John Hopfield from
decades ago; his descriptions of the intrinsic problems of
serum proteins with aptamers. (b) Neurotrophin-3 was spiked into buffer
Current Opinion in Chemical Biology 2008, 12:78–85
84 Proteomics and Genomics
specific binding in a complex sample (cells or plasma or invitro) have much to say about sound experimentation
[30��,31].
ConclusionsThis short review was intended at first to highlight the
obvious: if one measures enough proteins in plasma in
people with and without a variety of diseases, one ought
to be able to identify novel biomarkers that could be used
in small panels to facilitate appropriate evidence-based
medical decisions for specific indications. Ultimately one
expects that small specific proteomic panels will be
aggregated into larger panels that enable more compre-
hensive medical decisions to be made. This appears to us
to be within reach [32�].
But as we wrote we came to the conclusion that novel
biomarker discovery is constrained by reagent limitations
(with respect to limits of detection and content/scale) –
that is, old-fashioned biochemistry remains important.
We find the work reviewed here (and also a lot we did
not review) to be stunning in its medical implications if
(and only if) the reagents and/or platforms are up to the
task. We have tried to make clear what these old or new
reagents must do for the dream to be realized. We are
optimists about the diagnostic and medical future through
proteomics.
Conflicts of interest statementDr Eaton consults for SomaLogic. Drs Singer, Zichi, and
Gold are employees of SomaLogic.
AcknowledgementsWe thank our colleagues at both SomaLogic and the University of Coloradofor hundreds of serious conversations.
References and recommended readingPapers of particular interest, published within the period of review,have been highlighted as:
� of special interest�� of outstanding interest
1.��
Eaton BE, Gold L, Zichi DA: Let’s get specific: the relationshipbetween specificity and affinity. Chem Biol 1995, 2:633-638.
Shows that the factors that contribute to high affinity are the same asthose that determine specificity.
2. Tuerk C, Gold L: Systematic evolution of ligands by exponentialenrichment: RNA ligands to bacteriophage T4 DNApolymerase. Science 1990, 249:505-510.
3. Ellington AD, Szostak JW: In vitro selection of RNA moleculesthat bind specific ligands. Nature 1990, 346:818-822.
4. Schneider DJ, Feigon J, Hostomsky Z, Gold L: High-affinityssDNA inhibitors of the reverse transcriptase of type 1 humanimmunodeficiency virus. Biochemistry 1995, 34:9599-9610.
5. Gold L: Oligonucleotides as research, diagnostic, andtherapeutic agents. J Biol Chem 1995, 270:13581-13584.
6.��
Klussmann S: The Aptamer Handbook: FunctionalOligonucleotides and their Applications Weinheim: Wiley-VCH;2006.
An important collection of articles about aptamers that spans the rangefrom practical to theoretical.
Current Opinion in Chemical Biology 2008, 12:78–85
7.��
Borrebaeck CA, Wingren C: High-throughput proteomics usingantibody microarrays: an update. Expert Rev Mol Diagn 2007,7:673-686.
A useful review of the state of the art of antibody microarrays.
8. Nygaard AP, Hall BD: A method for the detection of RNA–DNAcomplexes. Biochem Biophys Res Commun 1963, 12:98-104.
9. Gillespie D, Spiegelman S: A quantitative assay for DNA–RNAhybrids with DNA immobilized on a membrane. J Mol Biol 1965,12:829-842.
10.�
Birney E, Stamatoyannopoulos JA, Dutta A, Guigo R, Gingeras TR,Margulies EH, Weng Z, Snyder M, Dermitzakis ET, Thurman REet al.: Identification and analysis of functional elements in 1%of the human genome by the ENCODE pilot project.Nature 2007, 447:799-816.
This work foreshadows the complete elucidation of the humantranscriptome, and leads to the possibility that far more of the genomethan expected appears as RNA – one wonders if all such RNAs areimportant or represent noise from unavoidably ‘sloppy’ DNA-dependentRNA polymerases.
11. O’Farrell PH: High resolution two-dimensional electrophoresisof proteins. J Biol Chem 1975, 250:4007-4021.
12.��
Uhlen M: Mapping the human proteome using antibodies.Mol Cell Proteomics 2007, 6:1455-1456.
A short overview of the HUPO Antibody Initiative and the Human ProteinAtlas program (http://www.proteinatlas.org/).
13.�
Hudson ME, Pozdnyakova I, Haines K, Mor G, Snyder M:Identification of differentially expressed proteins in ovariancancer using high-density protein microarrays.Proc Natl Acad Sci U S A 2007, 104:17494-17499.
Uses protein microarrays to detect autoantibodies in sera to identifyproteins associated with ovarian cancer.
14. Turner DH, Sugimoto N: RNA structure prediction. Annu RevBiophys Biophys Chem 1988, 17:167-192.
15. SantaLucia J Jr: A unified view of polymer, dumbbell, andoligonucleotide DNA nearest-neighbor thermodynamics.Proc Natl Acad Sci U S A 1998, 95:1460-1465.
16. Mathews DH, Sabina J, Zuker M, Turner DH: Expanded sequencedependence of thermodynamic parameters improvesprediction of RNA secondary structure. J Mol Biol 1999,288:911-940.
17. Do CB, Woods DA, Batzoglou S: CONTRAfold: RNA secondarystructure prediction without physics-based models.Bioinformatics 2006, 22:e90-e98.
18. Scott JK, Smith GP: Searching for peptide ligands with anepitope library. Science 1990, 249:386-390.
19. Smith GP, Scott JK: Libraries of peptides and proteinsdisplayed on filamentous phage. Methods Enzymol 1993,217:228-257.
20. Fields S, Song O: A novel genetic system to detectprotein–protein interactions. Nature 1989, 340:245-246.
21. Young KH: Yeast two-hybrid: so many interactions, (in) so littletime. Biol Reprod 1998, 58:302-311.
22. Marinelli RJ, Montgomery K, Liu CL, Shah NH, Prapong W,Nitzberg M, Zachariah ZK, Sherlock GJ, Natkunam Y, West RBet al.: The Stanford tissue microarray database. Nucleic AcidsRes 2007, 36:D871-D877.
23. Wingren C, Ingvarsson J, Dexlin L, Szul D, Borrebaeck CA: Designof recombinant antibody microarrays for complex proteomeanalysis: choice of sample labeling-tag and solid support.Proteomics 2007, 7:3055-3065.
24.�
Wingren C, Steinhauer C, Ingvarsson J, Persson E, Larsson K,Borrebaeck CA: Microarrays based on affinity-taggedsingle-chain Fv antibodies: sensitive detection ofanalyte in complex proteomes. Proteomics 2005,5:1281-1291.
A description of the state of the art in recombinant antibody microarrays.
25. Ellmark P, Ingvarsson J, Carlsson A, Lundin BS, Wingren C,Borrebaeck CA: Identification of protein expression signaturesassociated with Helicobacter pylori infection and gastric
www.sciencedirect.com
Proteomics and diagnostics: Let’s Get Specific, again Zichi et al. 85
adenocarcinoma using recombinant antibody microarrays.Mol Cell Proteomics 2006, 5:1638-1646.
26. Mor G, Visintin I, Lai Y, Zhao H, Schwartz P, Rutherford T, Yue L,Bray-Ward P, Ward DC: Serum protein markers for earlydetection of ovarian cancer. Proc Natl Acad Sci U S A 2005,102:7677-7682.
27.��
Gulmann C, Sheehan KM, Kay EW, Liotta LA, Petricoin EF 3rd:Array-based proteomics: mapping of protein circuitries fordiagnostics, prognostics, and therapy guidance in cancer.J Pathol 2006, 208:595-606.
A description of the drawbacks and advantages of several formats ofmicroarrays used in proteomics.
28. Drolet DW, Moon-McDermott L, Romig TS: An enzyme-linkedoligonucleotide assay. Nat Biotechnol 1996, 14:1021-1025.
29.�
Petach H, Gold L: Dimensionality is the issue: use ofphotoaptamers in protein microarrays. Curr Opin Biotechnol2002, 13:309-314.
www.sciencedirect.com
Yet another discussion of obstacles on the path to high-densityproteomics.
30.��
Hopfield JJ: Kinetic proofreading: a new mechanism forreducing errors in biosynthetic processes requiringhigh specificity. Proc Natl Acad Sci U S A 1974,71:4135-4139.
One of our favorite papers of all time – a clear exposition about thelimitations of equilibrium binding when searching for a needle in a hay-stack.
31. Hopfield JJ, Yamane T, Yue V, Coutts SM: Directexperimental evidence for kinetic proofreading in aminoacylation of tRNAIle. Proc Natl Acad Sci U S A 1976,73:1164-1168.
32.�
Kessler A: The End of Medicine: How Silicon Valley (and NakedMice) will Reboot your Doctor New York: Collins; 2006.
An imaginative description of a better future for healthcare that dependson proteomic arrays and in vivo imaging.
Current Opinion in Chemical Biology 2008, 12:78–85