Research Article
Pseudorandom Vs. Random Polymers - How to Improve the Efficiency of Lithography-based Synthesis
Author Affiliations
VP Bioinformatics, Caris Life Sciences, Phoenix, Arizona, United States
Corresponding Author
Phillip Stafford, VP Bioinformatics, Caris Life Sciences, Arizona, United States
Citation
Stafford P. Pseudorandom Vs.
Random Peptides-How to Improve the
Efficiency of Lithography-based Synthesis of
Peptide Microarrays. Syst Comput Biol J. 2018
Dec;1(1):102
© 2018 Stafford P. This is an openaccess article distributed under the terms of the
Creative Commons Attribution 4.0 international
License, which permits unrestricted use,
distribution and reproduction in any medium,
provided the original author and source are
credited.
Abstract
Computer lithography has been used for decades in the electronics industry. Biochemists
have used light-directed lithography and photoacids and photobases to enable BOC- and
FMOC-based peptide synthesis, enabling the creation of high-density peptide microarrays.
By reducing the size of the features, more peptides can be produced on industry-standard
microscope slides. Phage display has set the standard for the utility of probing a high density
peptide library. However, a sufficiently high-density peptide microarray can produce similar
results in a few hours vs. weeks of panning and selection.
For the specific case of random peptide microarrays, we examined a method to reduce
manufacturing costs by exchanging perfectly random peptides to pseudorandom ones. The
simplest mathematical formula for predicting the number of synthesis steps is M (steps)=AA
(# of different amino acids) * L (peptide length). However, when M< AA*L, peptides can no
longer be perfectly random, and a sequence bias will result. M affects synthesis time, AA and
L affect performance given that performance is directly related to the lack of sequence bias.
We isolated M as a variable and tested how different values of M affect the performance
of immunosignatures. Immunosignatures require random peptides to detect patterns of
antibody binding. As M decreases, the peptides gain sequence bias. The immunosignature
assay is used to quantify the impact of M. We first measured the predicted properties of
peptide libraries for M=35, 70, 140 and 272, then we synthesized actual peptides and tested
how 51 different monoclonal antibodies bound to the non-random peptides. We then tested
whether Coccidiomycoses possadasii (Valley Fever) infections could be distinguished – a
very simple task for a random peptide library. At low M the disorder in the peptide library
decreases, and one sees a near-monotonic signal with low information content. There is a
value of M, however, where synthesis steps can be reduced without unduly compromising
the information content of the resulting immunosignature. This tunable parameter should
be examined as a way to reduce synthesis time and manufacturing costs for random peptide
microarrays.
Keywords
in situ peptide synthesis; Photolithography; Random peptide; Immunosignature
Introduction
Many peptide microarrays are in the form of tiled epitope arrays [1]. They are commonly
used to map antibodies to a segment of a protein. However, there are other uses for peptide
microarrays-cell binding [2], small-molecule binding [3], protein binding [4], kinase activity
[5], even enzyme modulation [6] and a process that we use called ‘immunosignaturing’ [7].
Epitope arrays rely on antibodies recognizing and binding the exact sequence against which
they were raised. Immunosignatures, on the other hand, use random sequence peptides
and do not rely on perfect antibody-peptide matches. On an immunosignature peptide
microarray, antibodies can bind to their cognate peptide, but they can also bind apparently
unrelated peptides with equal or higher affinity [8,9]. This non-cognate binding may be
to mimotopes of the eliciting antigen [10]. Another explanation is that off-target binding
is an intrinsic property of an antibody, theoretically as important as on-target binding.
Immunosignatures detect both properties. There is a great deal of information that can be
extracted from an immunosignature assay, even to the point of diagnosing disease [11-14].
Epitope arrays can be very expensive to manufacture, limiting their use to smaller
experiments. Immunosignature arrays can be produced at very high numbers, since they use
manufacturing processes taken from the semiconductor industry. Increasing the number of
arrays, manufactured would actually reduce the cost per unit. We will describe a method to
reduce manufacturing complexity and describe how simplifying manufacturing affects the
randomness of a peptide library and how randomness affects immunosignatures.
Immunosignatures use random sequence peptides to
gauge specificity of antibodies
Immunosignatures measure the binding between random
peptides and antibodies. Immunosignatures measure both on-target
and off-target binding of antibodies simultaneously [8]. As a reference,
phage display or phage panning probe extremely large libraries. The
selection process leaves only strong binders. Those survivors are
usually closely related to the actual epitope. In contrast, peptide
microarrays do not utilize a selection process. All of the binding data
is retained, and is scalable. Phage libraries can be extremely large
[15] which is to their benefit, while microarrays typically contain a
few million peptides at best [16]. However, each peptide interaction
can be more informative than the final pool of phage display peptides.
In addition, even a relatively small random peptide library contains
many unique epitopes [17-19] if longer peptides are used [1,14],
while phage display often relies on 6-10mers [20]. Phage libraries
may be diverse but microarrays can extract low binders as well
as high binders. By analyzing binding intensities one can identify
residues at key positions that play a role in enhancing binding vs.
those that prevent binding. When sera or an antibody is applied to a
random peptide microarray, the resulting ‘immunosignature’ [7] can
provide information about how many different peptides the antibody
binds, how diverse are the binding sequences, where the critical
residues lie in the epitope, and how amino acids bind to the paratope.
The immunosignature can be so informative and reproducible that
diseases can be diagnosed with high accuracy [11,13,14,21,22].
This diagnostic feature does not require information about peptide
sequences but even so, the peptide sequences in a signature can still
predict linear epitopes from infectious agents [17] or cancer antigens
[19].
In situ peptide synthesis
There are many ways to manufacture peptide microarrays. For a
review see Gao et al. [27,28]. Light-directed chemical synthesis has
been quite successful, dating back to 1995 (Fodor et al. [29], US Patent
5405783A). Some synthesis methods, direct light on a photoacid [16]
or photobase [30] generator which removes amino acid protecting
groups. Other methods use photoactivatable amino acids [29]. Some
methods use shadow masks [16], some use digital micromirrors [31],
some utilize a peptide ‘toner’ and a 20 ‘color’ laser printer [32], while
still others use electrical circuits to alter the pH in a microwell [33].
Each of these synthesis methods relies on the systematic addition of
amino acids to a growing peptide. For random peptide libraries, all of
these methods can take advantage of reduced synthetic steps using
pseudorandom peptides.
One of the highest throughput methods for the production of
peptide microarrays is shadow-mask lithography [34]. Shadowmasks are pre-made so they are not as flexible as programmable
light-directed arrays, and they require substantial up-front costs.
However, by implementing high-resolution steppers, automated
aligners, synthesizers and large wafers, mask-based peptide
microarrays can be very inexpensive, high density, high precision
and high throughput. These characteristics are important for high
volume/low costproduction.
Random vs. pseudorandom libraries
Many immunological experiments make use of random sequence
peptides [35]. Phage display uses random sequence libraries [36]
and has been used to identify therapeutics [37], epitopes [38], even
enzyme modulators [39]. Panning large libraries can identify very
specific biomarkers, but these methods are neither cheap nor high
throughput. Immunosignatures use a fixed library of a given size
for any sample. A diverse surface of 3D shapes produced by random
sequence peptides has been very successful in many applications
[11,13,21,40-42]. Surprisingly, 10,000 peptides were sufficient to
distinguish over 15 different diseases simultaneously, with >95%
accuracy [14]. However, 10,000 peptides likely had a finite level of
specificity and sensitivity. It was decided to create at least 300,000
peptides for the next immunosignature iteration. When M=AA*L,
synthesis of 300,000 17mers would take over 2 weeks of continuous
synthesis. Altering the library from random to pseudorandom could
drastically reduce manufacturing time.
How in-situ peptide synthesis is affected by amino acid
selection
The simplest algorithm for the number of synthesis steps (or
photomasks) required to produce a peptide of length L is M (number
of masks)=AA (number of amino acids used) * L (length). For a 17mer
using 20 amino acids, 20*17=340 synthetic steps are required. In
mask-based peptide photolithography, 340 steps would take over
2 weeks to complete. For immunosignature peptides, M can be less
than AA*L, but the resulting peptides become pseudorandom. Each
mask costs several thousand dollars. Each synthesis step takes
approximately 15 minutes. It is beneficial to reduce M. Figure 1
illustrates the bias imposed by forcing M to be less than AA*L.
Experimental design
We asked two questions about reducing synthesis steps. First,
how do the chemical characteristics of the peptide library change
as mask numbers are reduced? Second, how does reduced diversity
alter the outcome of a biochemical assay? The first question was
addressed using predicted characteristics of hypothetical peptides.
For the second question, an immunosignature of monoclonal
antibodies and an immunosignature of human sera are used to
measure performance. We first established the starting parameters.
Kuznetsov [43] noted that some amino acids are preferentially used
in immunosignatures [43]. To maintain consistency with the actual
immunosignature arrays that will be manufactured, we eliminated
Methionine, Threonine and Isoleucine because they appear rarely in
peptides that make up diagnostic immunosignatures [43]. Cysteine can form disulfide bonds under aqueous conditions, and was
removed for that reason. Thus our virtual 17mer peptide library used
16 amino acids (no C, T, M, or I). We created two control sets: M=272
uses 16 amino acids and no other mask restrictions. M=323 leaves
out only Cysteine. The M=323 case matches exactly the existing
10,000 peptide spotted peptide library that has been used in several
early publications (NCBI GEO accession GPL17600). We also created
intermediate test sets where M=35, M=70, and M=140.
The first analysis was completed using 100,000 virtual peptides.
For each library, an algorithm (see Supplemental Information) creates
M masks with 100,000 features and ‘synthesizes’ peptides using
these virtual masks. The resulting library is analyzed for chemical
properties using ProtParam ((http://web.expasy.org/protparam/).
The second analysis involved synthesizing a library of actual
peptides corresponding to the same mask restrictions. 384 peptides
were selected at random from each 100,000 virtual peptide libraries
and sent to Sigma Genosys (St. Louis, MO) for synthesis. We already
had several M=323 10,000 peptide microarrays made, so the human
sera experiments for M=323 were done by randomly picking 384
peptides from the 10,000 existing peptides to simulate a 384-peptide
library. Each of these 384-peptide sets (except M=323) were
tested using 51 different commercial monoclonals (Supplemental
Information Table 1). We previously published data on monoclonals
using the 10,000 peptide microarray [7]. Human sera from 8 controls
(healthy) humans and 8 patients diagnosed with Coccidiomycoses
[11] were used for all libraries including M=323.
Materials and Methods
In silico peptide generation
We created an R script called “Peptide Generator” in order
to simulate how peptides would be created when the number of synthetic steps is less than AA*L (Supplemental Information). The
program uses the following variables:
P is the library size, number of peptides
M is the number of virtual masks
N is peptide length
DD is the number of amino acids used
aa is the one-letter amino acid code
peptide is the output of P peptides
The script can use one-letter code for D- or L-amino acids. The
algorithm currently uses loops to cycle through the synthesis steps to
allow non-coders to clearly see the steps, but for efficiency the loops
should be replaced by ‘apply’ statements.To compare the predicted chemical characteristics of peptides
made by Peptide Generator, we created 100,000 virtual 17mer
peptides using 16 amino different acids for each mask restriction
scenario (M=35, 70, 140 and 272) as well as 100,000 completely
random peptides using 19 amino acids (M=323). The method for
creating the peptide libraries is by the greedy algorithm, where masks
are made available as amino acids are cycled through one by one. The
amino acids are added one step at a time to the growing peptide chain.
The program cycles through each amino acid alphabetically and stops
when every peptide is exactly 17 amino acids long and every mask is
used once. The algorithm as written does not attempt to make the
distribution of amino acids even during synthesis. As a result, some
peptides tend to catch up in length as the masks run out by quickly
adding any available amino acid(s) to the N-term of the synthesis.
This creates a bias in sequence,driven by the order of the amino
acid addition. This bias was deemed acceptable since each peptide
library has the same bias, and performance comparisons were
relative. Methods to artificially compensate for this bias introduced
an unpredictable distribution that was more complex to decipher.
In order to examine the chemical characteristics of each library
of 100,000 peptides, we used ProtParam (http://web.expasy.org/
protparam/). A library of 100,000 peptides was deemed sufficient to
summarize the trends for each mask set. ProtParam takes as input
one-letter amino acid sequences and then predicts the chemical
properties of each peptide. We examined molecular weight and
theoretical pI to gauge how our generated peptides differed from
completely random peptides, and we measured the number of
repeated 2mer, 3mer, 4mer, 5mer and 6mer sequences to estimate
library diversity. Each mask set is summarized by the average
molecular characteristics of 100,000 peptides.
Peptide synthesis and array printing
The second part of the project is a biochemical test with actual
peptides. Since 100,000 synthesized peptides would cost millions
of dollars and many months to synthesize, we chose to use a much
smaller library and increased the number of antibodies tested on
these arrays to 51 very disparate monoclonals. Small-scale peptide
synthesis yielded enough peptide to print thousands of microarrays.
384 peptides per mask set were synthesized at 80% purity, 5 mM
scale by Sigma Genosys (St. Louis, MO). Failed syntheses were
repeated until Sigma deemed the synthesis successful using their
quality metrics. MALDI spectrograms indicated that each peptide had
a major peak at the expected mass corresponding to an average purity
of >92% according to Sigma Genosys. S1a and S1b Figures show an
example of the MALDI chromatograms of a partial truncation that
was resynthesized (S1a Figure) and a properly synthesized peptide
(S1b Figure). Peptides were synthesized C-term to N-term, with a
terminal Glycine-Serine-Cysteine linker that enabled attachment of
the free N-term cysteine to a sulfo-SMCC linker (sulfosuccinimidyl
4-N-maleimidomethyl) cyclohexane-1-carboxylate, catalog number
22322, Pierce/Thermo Fisher Inc.) to a Schott A+ Nexterion
aminosilane slide (catalog number 1064875, Schott, Jena, Germany).
Truncated peptides were capped by the manufacturer and would not
have a terminal cysteine ensuring that they do not attach to the slide.
Peptide stocks were made by diluting 2 mg of peptide in distilled
water + 20% acetonitrile at a final concentration of 2 mg/ml, based
on measurement of peptide quantity by Agilent BioAnalyzer 2100 in
protein mode (Agilent Inc., Santa Clara, CA). Microarray spotting plates
were prepared using 384-deepwell plates by diluting the stock 2 mg/
ml peptides into HEPES pH 7.3 with 2% TCEP (tris(2-carboxyethyl)
phosphine) to a final concentration of 0.5 mg/ml. Printing was done
at Applied Microarrays (Tempe, AZ) by piezo non-contact printing of
approximately 150-200 pL per feature. Post-dispense, slides were
stored at RT in 80% humidity to allow the maleimide conjugation
reaction to occur. Prior to use, the arrays were washed with 30%
ethanol, 500 mM EDTA pH 7.8, 40% acetonitrile to remove unbound
peptides (Immunosignature effects paper). Every peptide was
printed twice per assay. By duplicating the spots we ensured that any
discrepancy in printing could be measured. The duplicate peptides
were averaged for analysis. We used a 2-up assay system from Tecan
(HS 4800 Hybridization Station, Tecan, Månnedorf, Switzerland) to
perform the antibody incubations. The automation of the HS 4800.
reduced operator and day-to-day variability. For every monoclonal,
two technical replicates were run on different slides. Averaging across
replicates accounts for a possible discrepancy in the assays. The
technical replicates were run on physically different slides to reduce
systematic bias in the assay. The technical replicates must exceed
0.85 correlation coefficient in order to achieve our quality threshold.
Assays that failed would have been re-processed. However, during the
experiment no replicate failed to meet our quality threshold. Postassay measurements yielded peptide-to-peptide correlation average
>0.99 suggesting that peptide printing was an insignificant source of
variation. The average correlation across technical replicates was 0.96
suggesting that day-to-day and operator variance were insignificant
sources of variation. Each microarray was incubated with one of the
100 different commercially sourced monoclonals listed in Table 1.
Immunosignature assays-monoclonals
For the monoclonal assay, only M=35, M=70, M=140 and M=272
peptide libraries were used. M=323 was only tested against human
serum. Each of the 51 commercial monoclonals was added to 800
uL incubation buffer (3% BSA, 1x PBS pH 7.2, 0.05% Tween 20) to
a final concentration of 5 nM. Incubation buffer + monoclonal was
added to each microarray of 384 peptides. The Tecan HS4800 was
programmed to incubate each slide with shaking for 1 hr at 37°C. The
wash program included 3 washes of 5’ each with 1x Tris-buffered
saline pH 7.3 followed by 3 washes of 5’ each using distilled water.
When complete, the HS 4800 dried each slide with compressed
nitrogen. Detection was by goat anti-Mouse IgG (H+L) directly
conjugated to Alexafluor-555 (Life Technologies/Thermo Fisher)
incubated at 4 nM for 1 hr followed by identical wash steps. Images
were recorded by an Agilent ‘C’ scanner. Slides were scanned at
100% PMT, 70% laser power at 10 μm resolution using the 545 nm
laser. Arrays were aligned using GenePix 6.0. Data was analyzed in R
(CRAN.org) and GeneSpring 7.3.1 (Agilent, Santa Clara, CA).
Immunosignature assays–human sera
Eight patients who were diagnosed with Coccidiomycoses [11]
with a measured titer of 1:256 against coccidioidin were tested
against eight otherwise healthy age- and gender-matched volunteers
in a standard immunosignature assay [11]. We measured the
minimum p-value obtained between case and control for each mask
set. All immunosignature assays were run in duplicate and averaged.
T-tests were performed across all 384 peptides for each mask set and
recorded. For M=323, a 10,000 random peptide microarray was used
(NCBI GEO platform accession GPL17600), and 384 peptides were
selected at random for analysis.
Analysis methods
In order to compare how effective different peptide cohorts
work for immunosignatures, we first made the assumption that
signatures for different monoclonal antibodies must be different
[7]. We therefore calculated the Pearson’s Correlation Coefficient for
each set of 384 peptides across each of the 51 different monoclonal
antibodies. If every monoclonal was mathematically distinguishable using immunosignatures, it follows that the correlation coefficients
across the different antibodies would be low. Conversely, if the
monoclonals could not be distinguished, or the information content
per peptide library is low, it is likely that the correlation coefficients
would be high. We did this for the 4 different mask sets from M35 to
M272. This data is listed in Table 1.
For the coccidiomycoses set, we assayed eight different case
and eight different control serum samples according to previous
protocols [11]. We report the minimum p-value for each mask
set, M=35, 70, 140, 272 and 323. For M=323, we performed the
immunosignatures on the CIM 10,000 random peptide microarrays,
and selected 384 peptides at random for the analysis. This prevented
having to re-synthesize completely random (M=323) peptides, since
completely random peptides already exist on the 10,000 peptide
printed microarray. All peptide lists and data are publicly available at
GEO (series accession numbers GSE50044, GSE50045).
In order to assess how peptide disorder affected immunosignature
performance, we calculated Shannon’s Entropy for each of the
peptide libraries, since this method is a direct measure of disorder
[44]. We used Support Vector Machines as the classifier for predictor
of the identity of each monoclonal using a train/test paradigm. We
mathematically removed 38 peptides from the 384 peptide libraries,
and trained the SVM classifier. We then gave the classifier the leftout 38 peptides and asked it to predict the identity of each of the 51
different monoclonals. Thus, the classifier is completely naïve to the
predictor peptides. This provided a way to directly assess peptide
entropy vs. immunosignature performance. We used Shannon’s
Entropy formula:
Entropy for each peptide library was calculated per peptide and
averaged.
Table 1
Table 1: Test of immunosignaturing performance-This table lists values that were used to characterize immunosignature performance. From
left to right: M represents the number of masks; Avg. CC Ab represents the average correlation coefficient calculated across every possible
combination of the 51 monoclonal antibodies; Ave CC serum is the average correlation coefficient calculated across every possible combination
of control and infected sample; ANOVA p -val is the p -value from the F-test using every monoclonal as a class, and the technical replicates as
the measurements; Avg. raw value Ab is the average of all 51 monoclonals for each of the tested mask libraries; Median raw value is the median
of all 51 monoclonals for each of the tested mask libraries; CV is the Coefficient of Variation (standard deviation corrected by mean) for every
peptide across every monoclonal for each mask set. Note that the M=323 library was only used for the serum experiments
Figure 1: Analysis of amino acids prevalence across each peptide position using mask restrictions.
100,000 peptides were contructed in silico using a randomization method that takes into account the total number of shadow masks used
in synthesis of 17mer peptides. Each graph shows the number of times each of the 16 amino acids is used at each of the 17 positions for a
given number of masks. The X-axis for each panel represents the position in the peptide from C-terminus (position 1, facing the surface of
the microarray) to the N-terminus (position 17, facing away from the surface). The Y-axis is the number of times each amino acid appears at
that position. The legend associates graph colors with an amino acid. The lower middle graph is M323, and is used as a baseline comparison.
M323 is unique as it uses 19 amino acids (excluding cysteine) and mimics a library of random peptides if synthesized on beads and cleaved/
purified, with no mask number restrictions. M323 should have no positional bias. M=35 (upper left) shows the highest positional bias,
suggesting that diversity of sequence would be compromised using only 35 masks, even though the number of steps is nearly 10-fold
less than fully random. M=70 (middle top), M=140 (right top), and M=272 (bottom left) show progressively decreasing bias trends, and
smoother amino acid distributions. M=323 has the lowest variation in positional bias across the length of the peptide since each position
was determined by a random number generator, independently
Figure 2: Left 5 panels: distribution of duplicated 5-mers for M=35,
M=70, M=272 and M=323.
As the number of masks decreases, diversity in the resulting
sequences is reduced, creating redundant sequences. The 5 panels
above illustrate the number of duplicate 5mers. The X-axis is the total
counts of duplicated 5mers in the 100,000 peptides, the Y-axis is the
proportion of 5mers. The average for M323 (far right, top) is 1. The
average number of duplicated 5mers for M=35 is 2 suggesting that
redundancy is a hallmark of reduced mask numbers. Right 5 panels:
Analysis of predicted peptide characteristics as a function of mask
number. As in the left panels, each mask set created a distribution
of data. Here the data is Molecular Weight. The trend seen here is
similar for Extinction coefficient, beta sheet tendency, Alpha helix
tendency, Beta turn tendency, Antigenicity, Isoelectric point, Aliphatic
character, Average residue volume, Hydrophilicity, Flexibility,
Sequence complexity, and Accessibility (all taken from ProtParam).
X-axis is MW, Y-axis is the proportion of the total library for each
MW value. The lower mask numbers tend to create sets of peptides
with very specific and extreme molecular weights. For M35, weights
ranged from 1856 to nearly 2600 while for M=323, weights ranged
from 2380 to 2418. High mask numbers tend toward a more Gaussian
distribution of molecular weights.
Figure 3: Display of unique 2mer, 3mer, 4mer, 5mer and 6mers
contained within 100,000 peptides using the M=35, M=70, M=140,
M=272 and M=323 mask restriction sets.
The Y-axis is the proportion of unique nmers, and the X-axis is
increasing length of subsequences.
Figure 4: From left to right Top: Four images from antibodies AAT,
A10, 8A1 and 3A1, the same as those used in subsequent figures.
Bottom: Same monoclonals on the M=272 library. These images
show a promiscuous monoclonal that binds many peptides (AAT),
two different moderate-specificity antibodies (A10 and 8A1) and a
high specificity monoclonal that binds very few peptides but with
high intensity (3A1). These four antibodies produce very different
binding profiles on M=272 while M=35 produced very similar
binding profiles.
Figure 5: Bar graphs of each mask set (M=35, red; M=70, yellow; M=140, cyan; M=272, blue) for all of the 51 tested monoclonals.
Each dataset was median normalized, ensuring that the median signal is 1 for every monoclonal. The solid box is the first quartile, centered
on the median. The dashed lines are the 95% percentile, outliers are show as colored dots
Figure 6: Heat map four of the 51 total monoclonals across four of
the mask libraries: 3A1, 8A1, A10 and AAT, all processed on M=35,
M=70, M=140, and M=272.
Each peptide was converted from a sequence to a number, from
1 to 384, and sorted from lowest to highest. The rows represent
the rank-order of the intensity values for each of these four
monoclonals. The vertical colored bars (red, yellow, cyan, blue, and
purple) represent the k-means clustering of the values, with the
purple group containing the highest intensities and the red group
containing the lowest intensities.
Figure 7: Principal Components Map (PCA) of all 51 monoclonals (colored dots) for each of the four mask sets (colored by mask set).
The first two principal components are plotted on the X and Y-axes, respectively. The least dispersed is M=35, the most M=272 (as it is
spread across both component 1 and component 2).
Figure 8: Shannon’s Entropy for each of four peptide libraries vs. cross-validation accuracy of prediction of monoclonals Shannon’s
Entropy, a standard measure of disorder, was calculated for each of four different peptide libraries (Y axis).
Each of 50 different monoclonals was predicted by Support Vector Machines using a randomly selected 1/10 of the total 384 peptides
per library. This was done 10 times, and the average accuracy for each peptide library is plotted on the X axis. R2
for a linear fit is 0.7029.
Conclusions
The data provided here can be applied to maskless peptide
syntheses as well, by reducing the number of steps. In cases where
amino acids are added simultaneously to growing chains of peptides
in multiplexed reaction vessels, the number of cycles can be reduced.
As the number of peptides simultaneously synthesized increases,
the value of restricting the number of steps increases. Perhaps most
importantly, it is not necessary to use the output of Peptide Generator
as-is. A more useful method is to produce every possible peptide
for a given mask restriction. Then, filter this large collection of
peptides. The user can decide what filter is most important, whether
reducing redundancy, reducing certain physical characteristics
such as hydrophobicity, or picking random peptides with the most
life-space motifs. Many of the biases described in this manuscript
can be moderated by this post-hoc filter. Depending on the assay,
performance of a restriction-set of peptides, if properly filtered,
might be competitive with a fully random set. The assay drives the
balance between reducing synthesis steps and retaining peptide
diversity and disorder. One of the quickest methods for improving the
utility of a random peptide library is reducing 3mer, 4mer and 5mer
redundancy. Even without restriction, a random library groomed for
minimal 5mer repeats is more likely to contain more unique linear
targets than an unfiltered set. That single change could enhance
epitope identification by an order of magnitude or more. Judicious
selection of a library of mask-restricted peptides would not only improve the performance of the resulting random peptide library, but
the impact to synthesis time and cost would be substantial.
Although this manuscript focused on shadow-mask lithography,
our results can be applied to many peptide synthesis methods. Random
peptide libraries are quite useful for a number of immunological and
biochemical techniques. As peptide microarrays become denser, the
advantages of using random microarrays over phage display grows.
Although there are many ways of making peptide microarrays, costs
are still very high and production volume remains relatively low. In
order to use these arrays for any sort of diagnosis, the costs must
drop orders of magnitude and production volume must scale up as
well. Shadow mask synthesis can facilitate extremely high throughput
production, low cost synthesis equipment, and scalable costs that
trend the same way semiconductor manufacturing does. This enables
commodity-level high quality peptide microarrays, facilitating new
applications in medical care, diagnostics, and even therapeutic
antibody quality control and assessment.
Data are provided at the Gene Expression Omnibus at NCBI.
Platform: GPL17490, GPL17679 and series Series GSE49217,
GSE50044 and GSE50045, “In situ synthesis of peptide microarrays
- shadow mask design”.
Acknowledgement
There were no funding agencies for this work. Data are
freely available at the Gene Expression Omnibus at NCBI. Platforms:
GPL17600, GPL17490, GPL17679, Experimental Series: GSE49217,
GSE50044, GSE50045.
References
- Buus S, Rockberg J, Forsström B, Nilsson P, Uhlen M, et al. Highresolution Mapping of Linear Antibody Epitopes Using Ultrahighdensity Peptide Microarrays. Mol Cell Proteomics. 2012
Dec;11(12):1790-800.
- Domenyuk V, Loskutov A, Johnston SA, Diehnelt CW. A Technology
for Developing Synbodies with Antibacterial Activity. PLoS ONE.
2013 Jan;8:e54162.
- Boltz KW, Gonzalez-Moa MJ, Stafford P, Johnston SA, Svarovsky
SA. Peptide microarrays for carbohydrate recognition. Analyst.
2009 Apr;134(4):650-652.
- Sinzinger MD, Chung YD, Adjobo-Hermans MJW, Brock R.
A microarray-based approach to evaluate the functional
significance of protein-binding motifs. Analytical Anal Bioanal
Chem. 2016 Feb;408:3177–3184.
- Houseman BT, Huh JH, Kron SJ, Mrksich M. Peptide chips for the
quantitative evaluation of protein kinase activity. Nat Biotech.
2002 Mar;20:270-274.
- Fu J. Microarray Selection of Cooperative Peptides for Modulating Enzyme Activities. Microarrays. 2017 Jun;6(2):8.
- Stafford P, Halperin R, Legutki JB, Magee DM, Galgiani J, et al. Physical characterization of the ‘Immunosignaturing Effect’. Mol Cell Proteomics. 2012 Apr;11(4):M111.011593.
- Halperin RF, Stafford P, Legutki JB, Johnston SA. Exploring
antibody recognition of sequence space through randomsequence peptide microarrays. Mol Cell Proteomics. 2011
Mar;10(3):M110.000786.
- Geysen HM, Rodda SJ, Mason TJ, Tribbick G, Schoofs. Strategies
for epitope analysis using peptide synthesis. J Immunol Methods.
1987 Sep;102(2):259-274.
- Huang J, Ru B, Zhu P, Nie F, Yang J, et al. MimoDB 2.0: a mimotope
database and beyond. Nucleic Acids Res. 2012 Jan;40:D271-D277.
- Navalkar K, Magee DM, Galgiani J, Cichacz Z, Johnston SA, et al.
Application of immunosignatures to diagnosis of Valley Fever.
Clin Vaccine Immunol. 2014 Aug;21(8):1169-1177.
- Restrepo L, Stafford P, Magee DM, Johnston SA. Application of
immunosignatures to the assessment of Alzheimer’s disease. Ann
Neurol. 2011 Aug;70(2):286-295.
- Singh S, Stafford P, Schlauch KA, Tillett RR, Gollery M, et al. Humoral
Immunity Profiling of Subjects with Myalgic Encephalomyelitis
Using a Random Peptide Microarray Differentiates Cases from
Controls with High Specificity and Sensitivity. Mol Neurobiol.
2018 Jan;55(1):633-641.
- Stafford P, Cichacz Z, Woodbury NW, Johnston SA.
Immunosignature system for diagnosis of cancer. PNAS. 2014
Jul;111(30):E3072-E3080.
- Rodi DJ, Soares AS, Makowski L. Quantitative Assessment of
Peptide Sequence Diversity in M13 Combinatorial Peptide Phage
Display Libraries. J Mol Biol. 2002 Oct;322(5):1039-1052.
- Legutki JB, Zhao ZG, Greving M, Woodbury N, Johnston SA, et al.
Scalable high-density peptide arrays for comprehensive health
monitoring. Nat Commun. 2014 Sep;5:4785.
- Richer J, Johnston SA, Stafford P. Epitope Identification from
Fixed-complexity Random-sequence Peptide Microarrays. Mol
Cell Proteomics. 2015 Jan;14(1):136-147.
- Halperin R, Stafford P, Johnston SA. GuiTope: an application for
mapping random-sequence peptides to protein sequences. BMC
Bioinformatics. 2012;13:1.
- Donnell B, Maurer A, Papandreou-Suppappola A, Stafford P. TimeFrequency Analysis of Peptide Microarray Data: Application
to Brain Cancer Immunosignatures. Cancer Inform. 2015
Jun;14(Suppl 2):219-233.
- Zwick MB, Shen J, Scott JK. Phage-displayed peptide libraries.
Curr Opin Biotechnol. 1998 Aug;9(4):427-436.
- Restrepo L, Stafford P, Johnston SA. Feasibility of an early Alzheimer’s disease immunosignature diagnostic test. J Neuroimmunol. 2013 Jan;254(1-2):154-160.
- Williams S, Stafford P, Hoffman S. Diagnosis and early detection
of CNS-SLE in MRL/lpr mice using peptide microarrays. BMC
Immunol. 2014 Jun;15:23.
- Hughes A, Cichacz Z, Scheck AC, Coons SW, Johnston SA, et al.
Immunosignaturing can detect products from molecular markers
in brain cancer. PLoS ONE. 2012;7(7):e40201.
- Stafford P, Wrapp D, Johnston SA. General Assessment of
Humoral Activity in Healthy Humans. Mol Cell Proteomics. 2016
May;15(5):1610-1621.
- Wang L, Whittemore K, Johnston SA, Stafford P. Entropy is a Simple
Measure of the Antibody Profile and is an Indicator of Health
Status: A Proof of Concept. Scientific Reports. 2017 Dec;7:18060.
- Young JD, Huang AS, Ariel N, Bruins JB, Ng D, et al. Coupling
Efficiencies of Amino Acids in the Solid Phase Synthesis of
Peptides. Pept Res. 1990 Aug;3(4):194-200.
- Gao X, Pellois JP, Na Y, Kim Y, Gulari E, et al. High density peptide
microarrays. In situ synthesis and applications. Mol Divers.
2004;8(3):177-187.
- Gao X, Gulari E, Zhou X. In situ synthesis of oligonucleotide
microarrays. Biopolymers. 2004 Apr;73:579-596.
- Fodor S, Read J, Pirrung M, Stryer L, Lu A, et al. Light-directed,
spatially addressable parallel chemical synthesis. Science. 1991
Feb;251(4995):767-773.
- Choung RS, Marietta EV, Van Dyke CT, Brantner TL, Rajasekaran
J, et al. Determination of B-Cell Epitopes in Patients with
Celiac Disease: Peptide Microarrays. PLoS One. 2016
Jan;11(1):e0147777.
- Pellois JP, Zhou X, Srivannavit O, Zhou T, Gulari E, et al. Individually
addressable parallel peptide synthesis on microchips. Nat
Biotechnol. 2002 Sep;20(9):922-926.
- Beyer M, Nesterov A, Block I, König K, Felgenhauer T, et al.
Combinatorial synthesis of peptide arrays onto a microchip.
Science. 2007 Dec;318(5858):1888.
- Maurer K, McShea A, Strathmann M, Dill K. The removal of the t-BOC group by electrochemically generated acid and use of an addressable electrode array for peptide synthesis. J Comb Chem. 2005 Sep-Oct;7(5):637-640.
- Agarwal PB, Pawar S, Reddy SM, Mishra P, Agarwal A. Reusable
silicon shadow mask with sub-5 μm gap for low cost patterning.
Sensors and Actuators A: Physical. 2016 May;242:67-72.
- Kay BK, Adey NB, Yun-Sheng H, Manfredi JP, Mataragnon AH, et al.
An M13 phage library displaying random 38-amino-acid peptides
as a source of novel sequences with affinity to selected targets.
Gene. 1993 Jun;128(1):59-65.
- Noren KA, Noren CJ. Construction of High-Complexity
Combinatorial Phage Display Peptide Libraries. Methods.
2001;23(20:169-178.
- Cai C, Dai X, Zhu Y, Lian M, Xiao F, et al. A specific RAGE-binding peptide biopanning from phage display random peptide library that ameliorates symptoms in amyloid β peptidemediated neuronal disorder. Appl Microbiol Biotechnol. 2016 Jan;100(2):825-835.
- Ge M, Yan A, Luo W, Hu YF, Li R-C, et al. Epitope screening of the
PCV2 Cap protein by use of a random peptide-displayed library
and polyclonal antibody. Virus Res. 2013 Oct;177(1):103-107.
- Chin CF, Tan SJ, Gan CY, Lim TS. Identification of Peptide Based
Inhibitors for α-Amylase by Phage Display. Int J Pept Res Ther.
2015 Feb;21(3):237-242.
- Kukreja M, Johnston SA, Stafford P. Immunosignaturing
microarrays distinguish antibody profiles of related pancreatic
diseases. J Proteomics Bioinform. 2012 Jan;S6:001.
- Legutki JB, Johnston SA (2013) Immunosignatures can predict
vaccine efficacy. PNAS. 2013 Nov;110(46):18614-18619.
- Boltz KW, Nagaraj V, Svarovsky S, Lake D, Stafford P, et al. High
throughput technology for the identification and characterization
of glycan binding peptides. Glycobiology. 2006;16(3):1155-1156.
- Kuznetsov IB. Identification of non-random sequence
properties in groups of signature peptides obtained in random
sequence peptide microarray experiments. Biopolymers. 2016
May;106(3):318-29.
- Giancarlo R, Scaturro D, Utro F. Textual data compression
in computational biology: a synopsis. Bioinformatics.
2009;25(13):1575-1586.
