Immunosignature system for diagnosis of cancerPhillip Stafford1, Zbigniew Cichacz, Neal W. Woodbury, and Stephen Albert Johnston

Center for Innovations in Medicine, The Biodesign Institute, Arizona State University, Tempe, AZ 85287-5901

Edited by Philippa Marrack, Howard Hughes Medical Institute, National Jewish Health, Denver, CO, and approved June 23, 2014 (received for reviewJune 19, 2014)

Although the search for disease biomarkers continues, the clinicalreturn has thus far been disappointing. The complexity of thebody’s response to disease makes it difficult to represent this re-sponse with only a few biomarkers, particularly when many arepresent at low levels. An alternative to the typical reductionistbiomarker paradigm is an assay we call an “immunosignature.”This approach leverages the response of antibodies to disease-related changes, as well as the inherent signal amplification associ-ated with antigen-stimulated B-cell proliferation. To perform animmunosignature assay, the antibodies in diluted blood are incu-bated with a microarray of thousands of random sequence peptides.The pattern of binding to these peptides is the immunosignature.Because the peptide sequences are completely random, the assayis effectively disease-agnostic, potentially providing a comprehen-sive diagnostic on multiple diseases simultaneously. To explorethe ability of an immunosignature to detect and identify multiplediseases simultaneously, 20 samples from each of five cancercohorts collected from multiple sites and 20 noncancer samples(120 total) were used as a training set to develop a referenceimmunosignature. A blinded evaluation of 120 blinded samplescovering the same diseases gave 95% classification accuracy. Toinvestigate the breadth of the approach and test sensitivity tobiological diversity further, immunosignatures of >1,500 histori-cal samples comprising 14 different diseases were examined bytraining with 75% of the samples and testing the remaining 25%.The average accuracy was >98%. These results demonstrate thepotential power of the immunosignature approach in the accu-rate, simultaneous classification of disease.

cancer diagnostic | immunodiagnostic | antibody biomarker |peptide microarray

Cancer is the most likely disease for which an early diagnosticwould be immediately beneficial. Unfortunately, finding

specific biomarkers, especially for cancer, has been complicatedby the fact that biological molecules (RNA, DNA, proteins, orpeptides) that are uniquely released by a small tumor into thebloodstream are extremely dilute. Classical biomarker assays arebased on one-to-one molecular recognition events to detect oneor a few specific analytes that are often measured by antibody–protein interactions. There are three fundamental limitationswith this approach, all of which are confounded by the dilutionproblem alluded to above. The first is that the cross-reactivity ofsuch interactions poses a formidable problem in distinguishingdiseases. Biology’s promiscuous use of a limited number of ho-mologous sequences, folds, and domains makes specificity diffi-cult. The second is that diseases such as cancer are themselvesheterogeneous, and individual response to disease, at a molecu-lar level, can vary considerably. It is unlikely that this level ofcomplexity can be quantitatively assessed by one or a few specificproteins or metabolites in a way that supports robust diagnosis.Third, many of the biomarkers that have been proposed are oflow stability or require substantial preassay purification orpreparation; these aspects introduce substantial variation intothe measured values (1, 2). As a result, although considerableeffort has been put into the development of biomarkers, onlya small fraction of candidates make it to clinical practice, and theutility of those that are used is sometimes only modest (3–5).Here, we explore the ability of the immunosignature technology

to address the ideal of a simple, comprehensive diagnostic formultiple cancers.An “immunosignature” is the pattern obtained when circu-

lating antibodies in blood are allowed to bind to a large micro-array of randomized-sequence peptides affixed to a solid surface(6). Cancers generate neoantigens by virtue of their mutagenicnature, and they tend to release native proteins and bio-molecules not normally encountered by the immune system (7–9).These behaviors can elicit an immune response (6, 10, 11). Byvirtue of the tremendous amplification afforded by B-cell repli-cation (12), the signal elicited by the disease-specific antigens ismassively amplified. In fact, a key aspect of the immunosignatureassay is that the blood is greatly diluted before application to thearray, such that only the antibodies that have been sufficientlyamplified give distinct signals (13).Another somewhat counterintuitive aspect of the method is

that the peptide sequences used on the microarray are pur-posefully not chosen to represent the natural antigens of theantibodies produced in response to disease. In fact, in the arraysof 10,000 peptides used in this study, the peptide sequences weregenerated with a random number generator. This enables thesame microarray to be used for diagnosis of any disease. Despiteusing random-sequence peptides, monoclonal antibodies gener-ated from a wide variety of antigens show specific patterns ofbinding on these arrays, to both cognate and noncognatesequences (14, 15). Many of the peptides bound by a monoclonalantibody against a known linear epitope have no obvious se-quence similarity to that epitope. Most of the peptides thusidentified have demonstrated low affinity in solution for theantibody but are retained on the arrays due to avidity created byclose spacing of individual peptides (15).

Significance

Over much of the world, healthcare systems are facing an un-precedented challenge to meet the medical needs of an agingpopulation while controlling costs. The early detection and treat-ment of diseases that are prevalent in older people is likely to bea key aspect of economically efficient, high-quality healthcare. Inthe case of cancer, the resection of a stage I or stage II tumor isoften effectively a cure. An ideal diagnostic would allow earlydetection of disease on a single platform that could be used forany disease. Here, we demonstrate that the immunosignaturediagnosis platform could potentially meet the universal plat-form requirement. Ongoing work will address the early de-tection requirement separately.

Author contributions: P.S. and S.A.J. designed research; P.S. and Z.C. performed research;P.S., Z.C., and N.W.W. analyzed data; and P.S., Z.C., N.W.W., and S.A.J. wrote the paper.

Conflict of interest statement: S.A.J. and N.W.W. cofounded HealthTell, Inc., which usesimmunosignaturing to diagnose cancer.

This article is a PNAS Direct Submission.

Data deposition: The data reported in this paper have been deposited in the Gene Ex-pression Omnibus (GEO) database, www.ncbi.nlm.nih.gov/geo (superseries accession no.GSE52582).1To whom correspondence should be addressed. Email: phillip.stafford@asu.edu.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.1073/pnas.1409432111/-/DCSupplemental.

E3072–E3080 | PNAS | Published online July 14, 2014 www.pnas.org/cgi/doi/10.1073/pnas.1409432111

Dow

nloa

ded

by g

uest

on

Janu

ary

17, 2

021

An immunosignature of an individual consists of an overlay ofthe patterns from the binding signals of many of the mostprominent circulating antibodies. Some of the binding signals arepresent in most individuals (whether sick or healthy), and someare unique to an individual, but if the individual has a diseasesuch as a cancer, a subset of the binding signals will be due todisease-associated antigens that are common to most individualswith the disease (16). An important aspect of this approach isthat it senses essentially all antibodies raised to the disease anddetects each of the antibodies as separable binding patternscomposed of unique molecular recognition elements. This differsfrom, for example, an ELISA, which might sum the contributionsof many different antibodies using a single protein, cell, or viruscapsid. Again, from a statistical perspective, the high dimensionalityof this readout affords much more specificity than could beobtained from a set of cognate sequences or from an array ofthe native antigens themselves.Not only does the use of highly dilute blood and random

peptide sequences in the immunosignature assay paradoxicallygive rise to improved sensitivity and specificity but these aspectsof the assay also result in several other unique benefits of theimmunosignature approach. Because of the dilution (1:500 inthese studies), blood proteins other than antibodies do not sig-nificantly bind to the arrays, meaning that there is no samplepreparation involved other than dilution (17). The dilutionensures the assay is sample-sparing. Finally, the assay is disease-agnostic. The arrays can be used for the simultaneous detectionand identification of multiple diseases.It is simultaneous detection and identification of multiple

diseases with a single assay that underlies the true potential ofthis approach as a disruptive force in healthcare. This, combinedwith the fact that serum antibodies are robust to handling (17,18) such that a drop of blood can be sent dried on filter paperthrough the mail (17), should enable frequent, inexpensivemonitoring for many different diseases. The goal of the currentwork is to test the multidisease aspect of immunosignaturesrigorously. Although the approach has previously been used todiscriminate various subtypes of brain cancer (19), it has not yetdemonstrated multiplexed cancer diagnosis. Here, we performa blinded train/test validation study wherein a group of 120individuals with five different cancers from various geographicregions was used as a training set to define a multicancer sig-nature. The signature predicted the disease status of a test cohortof equal size and composition. To explore the ability of the ap-proach to discriminate between an even larger set of diseases,1,516 different individuals spanning 14 different disease cohortsplus a diverse cohort of healthy controls were assayed and theability to distinguish between these diseases was evaluated.

Materials and MethodsImmunosignaturing Microarrays. The immunosignature peptide microarrayused in this work has been described previously (15). Two different librariesof 10,000 random-sequence, 20-residue peptides were used (each peptidecontains 17 variable amino acids and a common 3-aa linker). The two li-braries of random-sequence peptides are of completely different sequences.Library version 1 contains 10,420 peptides and was used for trial 1. Libraryversion 2 contains 10,286 peptides and was used for trial 2. Library 1 wasprinted such that two completely isolated assays are available on one slidebut only a single replicate of each sequence is available per assay. Library 2was printed with duplicate peptides but with only one assay per slide.Peptides for library 1 were synthesized by Sigma Genosys, and those for li-brary 2 were synthesized by Alta Biosciences. A common linker sequence,GSC, was synthesized on the amine terminus of all peptides in library 1 andon the carboxyl terminus in library 2. The variable 17 residues for eachpeptide in both libraries were determined by a random number generator.Peptides were printed onto aminosilane-coated glass slides (Schott) by Ap-plied Microarrays using noncontact piezo printing.

Assays. Microarrays are preincubated with blocking buffer [BB = 10 mM PBS(pH 7.3), 0.5% BSA, and 0.5% Tween] for 1 h before addition of a 1:500dilution of serum into sample buffer (SB = BB less 0.5% Tween) for 1 h at 25 °C.The primary antibody is washed off with BB, and the peptide-boundantibodies are detected by addition of 5 nM AlexaFluor 647-conjugatedanti-human secondary antibody (Rockland Antibodies) for 1 h in SB at 25 °Cand are then washed three times in SB and five times in 18 MΩ of water,followed by centrifugation at 1,800 × g for 5 min to dry. Arrays are scannedat 10-μm resolution at 647-nm wavelength by an Agilent C scanner usinghigh laser power and 70% gain for the photomultiplier tube. The resultingTagged Image File Format images are aligned using the correspondingGenePix Array List file which assigns a measured fluorescence intensity to itspeptide feature. All data are publicly available in the Gene ExpressionOmnibus (GEO) in superseries GSE52582, which contains data from trial 1(GSE52580) and trial 2 (GSE52581).

Samples. Serum samples were received at Arizona State University throughInstitutional Review Board Protocol no. 0912004625, “Profiling BiologicalSera for Unique Antibody Signatures,” which was renewed in March 2013 bythe Western Institutional Review Board (Olympia, WA). All patient sampleswere obtained under informed consent and deidentified by the donatingclinic. All disease states were assessed by a trained pathologist in consulta-tion with an oncologist at each clinic. Details of a patient’s age, sex, out-come, date of diagnosis, or disease substratification are restricted by theagreement with the donating clinics. However, every effort was made toensure no patient was undergoing therapeutic antibody treatment. Nopatients were censored due to age, sex, or subsequent outcome. Table 1describes the samples for trial 1. Table 2 describes the samples for trial 2.Other than the class designated as “BC second tumor” in trial 2, which onlyincluded women who were diagnosed with a new, spontaneous tumor fol-lowing resection of a primary breast tumor, patients followed the samerestrictions for inclusion as used in trial 1. No patients were censored due toage, sex, or subsequent outcome. Collaborators are listed by name inAcknowledgments, and the abbreviations used for their institutions are lis-ted here: Arizona State University collection, Tempe, AZ (ASU); BarrowNeurological Institute, St. Joseph’s Hospital and Medical Center, Phoenix,AZ (BNI); Cleveland Clinic, Cleveland, OH (CC); Fred Hutchison Cancer ResearchCenter, Seattle, WA (FHCRC); Memorial Sloan–Kettering Cancer Center, NewYork, NY (MSKCC); Multiple Myeloma Research Foundation, Norwalk, CT(MMRF); Mt. Sinai Hospital, New York, NY (MS); Pancreatic Cancer ResearchTeam, Phoenix, AZ (PCRT); University of Texas Southwestern Medical Center,Dallas, TX (UTSW); University of California, Irvine, CA (UCI); University ofPittsburgh Department of Immunology, Pittsburgh, PA (UPitt); and Univer-sity of Washington Medical Center, Seattle, WA (UW).

Table 1. Description of trial 1 samples

Disease state Training Test Collection site

Healthycontrols

20 (6, 2, 8, 4) 20 (5, 3, 4, 8) ASU, PCRT, FHCRC, UTSW

GBM 20 (20) 20 (20) BNIPC 20 (11, 6, 3) 20 (8, 5, 7) CC, PCRT, UWLung cancer 20 (20) 20 (20) FHCRCMM 20 (20) 20 (20) MMRCBC 20 (11, 5, 4) 20 (8, 6, 6) ASU, FHCRC, UTSW

In total, 240 serum samples were randomly selected frommultiple sites. Onehundred twenty unblinded samples were used for training, and 120 blindedsamples were tested. Column 1 lists the disease state reported for the trainingsamples, as noted at the time of diagnosis. Any comorbidities were ignored forthe purpose of classification. Blinded test samples were kept unknown untilcompletion of the classification process. Column 2 (Training) and column 3(Test) refer to the total number of unique serum samples. The numbers inparentheses refer to the number of samples from each clinic, respectively, aslisted in column 4. Column 4 lists the clinic(s) providing the samples; definitionsof abbreviations are provided in Materials and Methods. Brain cancer, lungcancer, and MM samples were available from a single clinic only. The raw datafrom trial 1 are available for public download in the GEO’s public repository(www.ncbi.nlm.nih.gov/geo/) under accession no. GSE52580. The download-able file for trial 1 lists each serum sample as belonging to one of the six classeslisted in column 1, and designated as either “Training” or “Test.” MMRC,Multiple Myeloma Research Consortium.

Stafford et al. PNAS | Published online July 14, 2014 | E3073

IMMUNOLO

GYAND

INFLAMMATION

STATIST

ICS

PNASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

17, 2

021

Microarray Data Analysis. For trial 1, three technical replicates for each samplewere processed and averaged. Technical replicates with a Pearson’s corre-lation coefficient <0.85 were reprocessed. On average, 10% of the arrayshad to be reprocessed due to high background, uneven aminosilane coating,or other image anomalies. Data were median-normalized per array. Initialfeature selection in the training set of samples used multiple test-correctedANOVA. Further filtering was done by pattern matching using “ExpressionProfile” in GeneSpring 7.3.1 (Agilent) with Euclidean distance/average link-age as the similarity measure. For this filter, each disease group (disease) wascompared with all other disease groups (cumulatively referred to as non-disease). Peptides with consistently high signal in disease and consistentlylow signal in nondisease were chosen as being the most selective for thatgiven disease. This was repeated for every disease until equal numbers ofpeptides were selected for every disease. For trial 1, 120 peptides from thetraining set were chosen to classify the five cancers and one control cohort(Table S1). For trial 2, 280 peptides from the training set were chosen toclassify the 14 diseases and one control cohort (Table S2). Classification ofthe test samples was done in R, version 2.6.2, using support vector machine(SVM) as the main classifier, with default parameters. All code used togenerate the classification is listed in Dataset S1. To avoid biased results fromthe choice of classifier, four other classifiers were used (Table S3), with codeincluded in the Dataset S1. Table 3 contains the SVM results for trial 1 [usingthe “confusionMatrix” function in the ROCR package (20)], with the corre-sponding receiver operating characteristic (ROC) plot shown in Fig. S1 asproduced by pROC in S+ (Insightful) (21). Confidence intervals were reportedfor each of the values (22). False-positive (FP) results are reported such thatthere is a corresponding false-negative (FN) result in the missed class.

For trial 2, the size of the dataset dictated a simple and conservativeholdout method for training and testing. Leave-one-out cross-validationoften underestimates error and leads to overfitting. Therefore, we randomlychose 75 ± 7% of the samples for training, leaving the remaining ∼25% fortesting. The 7% variation was designed to simulate differences in naturalcohort sizes. This was repeated 100 times, and an average ± 95% confidenceinterval was reported for accuracy, specificity, sensitivity, positive predictivevalue, and negative predictive value for each of three different classifiers:linear discriminant analysis (LDA), naive Bayes (NB) and SVM. Table 4 lists the100-fold holdout results for each disease and each classifier separately. The95% confidence interval is an indicator of how well each disease is predictedby the chosen peptides. Smaller cohorts demonstrated higher variance, be-cause each sample had greater relative impact on the classification perfor-mance. Exactly 280 total features were selected per iteration. There were197 features in common across all 100 training iterations.

Study Design. Trial 1. A blinded test–train trial was created using three tech-nical replicates of 120 unblinded training samples representing five differentcancers plus controls. An equivalently sized test cohort was created by usingsamples that remained blinded by the collaborator. Collection site, collectiondate, age, and sex were randomized. Samples were serum or plasma (17)from venous draws of 2–10 mL each, stored at −20 °C. Initial patient consentwas obtained by each clinic independently. Due to differences in patientprotocols across clinics, this report can only provide the disease class asreported to ASU by the pathologist and oncologist for each clinic.Trial 2. To reveal any underlying sensitivity to extraneous factors, such ascollection site, microarray manufacturing variance, and sample processing,2,118 samples from 10 different collaborators were processed betweenSeptember 2007 and January 2011. This large and diverse sera bank containssamples from more geographical locations than trial 1 and contains uneven

Table 2. Description of trial 2 samples

Disease state Disease cohort Test size Collaborator(s)

Second BC 61 15 ± 1 UCIAstrocytoma 166 42 ± 3 BNIBC stage II, III 141 35 ± 3 ASU, FHCRC, UTSW, UCIBC stage IV 42 11 ± 1 UTSWGBM 27 7 ± 1 BNIHealthy control 249 62 ± 4 ASU, BNI, CC, FHCRC, MSKCC, PCRT,

UTSW, UCI, UPitt, UWLung cancer 107 25 ± 2 FHCRCMM 112 28 ± 2 MMRCOligodendroglioma 48 12 ± 1 BNIMixed oligoastrocytoma 97 25 ± 2 BNIOvarian cancer 86 22 ± 2 MS, MSKCCPancreatitis 82 20 ± 1 CC, UWPC 136 34 ± 3 CC, UWEwing sarcoma 20 5 ± 0 ASUValley fever 142 36 ± 3 UA

In total, 1,516 serum samples were used in trial 2. For each class listed in column 1 (Disease state), the totalnumber of unique samples for that disease is listed in column 2 (Disease cohort). A 100-fold resampling methodselected 75 ± 7% of the samples from each disease class for training, with the remaining held back for testing.This reselection process was done 100 times. The average and SD of the test size are listed in column 3 (Testsize). Collaborators who donated the samples are listed in column 4; definitions of abbreviations are providedin Materials and Methods. The raw data from trial 2 are available for public download in the GEO’s publicrepository (www.ncbi.nlm.nih.gov/geo/) under accession no. GSE52581. The downloadable file for trial 2 listseach serum sample as belonging to one of the 15 classes in column 1.

Table 3. Classification scores for trial 1 using SVM

True calls and miscalls are listed in the gray area of the chart. All perfor-mance statistics are calculated from these calls and are listed in the whitearea. Average accuracy is 0.95, with a 95th percentile confidence interval of0.8943, 0.9981(kappa = 0.94). Note that any missed call (FN) for a given classwould yield a simultaneous FP call in another class, because all samples mustbe either a true call or a miscall. BrC, brain cancer, ND, nondisease; NPV,negative predictive value; PPV, positive predictive value. Statistics in thelower part of the graph are calculated using confusionMatrix (R 2.6.2, pack-age “carat”). The code for analysis is provided in Dataset S1.

E3074 | www.pnas.org/cgi/doi/10.1073/pnas.1409432111 Stafford et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

17, 2

021

numbers of patients per disease with unequal distributions of age, sex,ethnicity, and reported comorbidities. Microarrays are subject to a number oftechnical issues, including image artifacts that can affect the resulting dataquality. We examined the 2,118 microarrays and reduced the total number ofsamples to 1,922 by eliminating those samples whose technical replicates hada Pearson’s correlation coefficient <0.85. Most failures resulted from grossimage artifacts, such as high background, uneven aminosilane deposition,and scratches or other surface defects. No sample was used without a tech-nical replicate to estimate reproducibility; therefore, 21 samples were sub-sequently excluded because their replicate was removed. We then examinedthe remaining microarrays for signs of microarray batch bias using ComBat

(23, 24). Because each microarray was printed in batches, it was possible thata printing failure could result in low-quality arrays or arrays that did notmatch previously manufactured arrays. In total, 406 arrays were removeddue to manufacturing batch-bias that exceeded the disease-related signal.No disease class lost significantly more samples than another. The remaining1,516 samples were used for holdout testing.

ResultsTrial 1. The primary goal of this effort was to determine the ca-pability of immunosignatures to classify multiple diseases. Thisissue was explored in two separate trials. In trial 1, equal numbers

Table 4. List of classification scores for trial 2Disease Accuracy Sensitivity Specifi city PPV NPVLDA Second BC 97.80 ± 0.14 69.10 ± 2.82 99.21 ± 0.10 81.05 ± 3.46 98.48 ± 0.11 Astro 96.93 ± 0.17 90.10 ± 1.3 97.82 ± 0.17 83.79 ± 1.11 98.73 ± 0.18 BC 99.51 ± 0.05 99.71 ± 0.2 99.49 ± 0.08 95.45 ± 0.68 99.97 ± 0.02 BC stage IVa 99.62 ± 0.06 89.85 ± 1.49 100 ± 0 100 ± 0 99.6 ± 0.06 GBM 99.18 ± 0.10 94.33 ± 2.00 99.25 ± 0.09 62.10 ± 4.24 99.92 ± 0.03 Lung cancer 99.02 ± 0.12 92.37 ± 0.58 99.59 ± 0.09 94.79 ± 1.27 99.35 ± 0.05 MM 98.72 ± 0.11 100 ± 0 98.62 ± 0.12 85.13 ± 1.13 100 ± 0 ND 96.62 ± 0.17 85.45 ± 0.77 99.31 ± 0.10 96.66 ± 0.47 96.60 ± 0.23 Oligodendroglioma 99.65 ± 0.07 92.57 ± 1.95 99.86 ± 0.03 95.21 ± 1.19 99.78 ± 0.06 Oligoastrocytoma 98.94 ± 0.15 98.45 ± 0.82 98.95 ± 0.12 86.41 ± 1.78 99.91 ± 0.04 Ovarian cancer 99.92 ± 0.03 100 ± 0 99.91 ± 0.03 98.67 ± 0.47 100 ± 0 Pancreatitis 99.67 ± 0.05 95.42 ± 1.00 99.91 ± 0.03 98.50 ± 0.54 99.74 ± 0.05 PC 97.69 ± 0.11 86.61 ± 1.39 98.79 ± 0.08 87.22 ± 1.19 98.67 ± 0.12 Sarcoma 98.81 ± 0.11 54.15 ± 5.48 99.67 ± 0.07 71.55 ± 5.65 99.12 ± 0.12 Valley fever 99.67 ± 0.08 100 ± 0 99.64 ± 0.09 96.87 ± 0.74 100 ± 0 Total 98.77 ± 0.04 89.87 ± 1.32 99.33 ± 0.08 88.89 ± 1.59 99.33 ± 0.07NB Second BC 96.00 ± 0.16 56.07 ± 1.46 99.46 ± 0.07 90.37 ± 11.68 96.31 ± 0.15 Astro 91.92 ± 0.23 91.69 ± 1.25 91.91 ± 0.25 31.39 ± 10.61 99.66 ± 0.06 BC 98.78 ± 0.07 97.75 ± 0.46 98.91 ± 0.12 90.55 ± 9.81 99.73 ± 0.06 BC stage IVa 99.40 ± 0.09 84.48 ± 2.05 100 ± 0 100 ± 0 99.38 ± 0.09 GBM 96.85 ± 0.17 43.19 ± 2.17 99.72 ± 0.05 88.81 ± 16.61 97.04 ± 0.19 Lung cancer 99.08 ± 0.10 92.40 ± 0.89 99.74 ± 0.06 97.32 ± 6.18 99.25 ± 0.08 MM 96.45 ± 0.15 81.51 ± 2.07 97.76 ± 0.14 75.72 ± 11.16 98.38 ± 0.2 ND 95.84 ± 0.17 93.18 ± 0.62 96.41 ± 0.18 83.88 ± 7.21 98.54 ± 0.15 Oligodendroglioma 98.54 ± 0.14 74.38 ± 2.24 99.94 ± 0.03 98.56 ± 5.95 98.54 ± 0.14 Oligoastrocytoma 97.75 ± 0.15 86.11 ± 0.86 98.72 ± 0.13 84.75 ± 13.01 98.85 ± 0.09 Ovarian cancer 99.79 ± 0.05 98.48 ± 0.43 99.90 ± 0.03 98.81 ± 3.75 99.87 ± 0.04 Pancreatitis 99.30 ± 0.11 92.27 ± 1.49 99.80 ± 0.05 97.40 ± 5.82 99.45 ± 0.11 PC 95.91 ± 0.20 78.67 ± 0.96 98.26 ± 0.09 85.62 ± 7.73 97.13 ± 0.17 Sarcoma 96.73 ± 0.20 25.21 ± 1.44 100 ± 0 100 ± 0 96.69 ± 0.21 Valley fever 97.96 ± 0.22 97.48 ± 0.60 97.99 ± 0.22 84.63 ± 12.45 99.73 ± 0.07 Total 97.35 ± 0.15 79.52 ± 1.27 98.57 ± 0.10 87.19 ± 8.13 98.57 ± 0.12SVM Second BC 98.89 ± 0.03 91.04 ± 0.59 99.19 ± 0.04 81.16 ± 8.55 99.65 ± 0.03 Astro 97.12 ± 0.06 84.11 ± 0.31 98.93 ± 0.03 91.56 ± 2.18 97.82 ± 0.06 BC 99.78 ± 0.02 99.39 ± 0.13 99.82 ± 0.02 98.40 ± 1.34 99.93 ± 0.01 BC stage IVa 99.89 ± 0.02 96.26 ± 0.75 100 ± 0 100 ± 0 99.88 ± 0.02 GBM 99.08 ± 0.03 100 ± 0 99.07 ± 0.03 46.42 ± 21.1 100 ± 0 Lung cancer 99.73 ± 0.02 96.82 ± 0.18 99.97 ± 0.01 99.65 ± 1.12 99.73 ± 0.02 MM 99.58 ± 0.01 99.89 ± 0.08 99.55 ± 0.01 94.7 ± 1.19 99.99 ± 0.01 ND 98.13 ± 0.07 91.33 ± 0.35 99.70 ± 0.02 98.60 ± 0.81 98.03 ± 0.09 Oligodendroglioma 99.82 ± 0.01 94.76 ± 0.30 99.96 ± 0.01 98.67 ± 3.38 99.85 ± 0.01 Oligoastrocytoma 99.29 ± 0.03 100 ± 0 99.24 ± 0.03 89.66 ± 4.09 100 ± 0 Ovarian cancer 99.92 ± 0.01 98.7 ± 0.10 100 ± 0 100 ± 0 99.92 ± 0.01 Pancreatitis 99.73 ± 0.02 96.27 ± 0.27 99.94 ± 0.01 99.07 ± 1.58 99.77 ± 0.02 PC 98.62 ± 0.03 90.98 ± 0.21 99.45 ± 0.02 94.74 ± 2.16 99.02 ± 0.02 Sarcoma 99.19 ± 0.04 100 ± 0 99.18 ± 0.03 38.81 ± 31.06 100 ± 0 Valley fever 99.82 ± 0.01 99.66 ± 0.10 99.83 ± 0.01 98.53 ± 0.80 99.96 ± 0.00 Total 99.24 ± 0.03 95.95 ± 0.22 99.59 ± 0.02 88.66 ± 5.29 99.57 ± 0.02

Approximately 75% of each class was removed for training, with 25% held out for testing using these differentclassifiers. This was done 100 times, and the results from LDA (yellow), NB (red), and SVM (green) classificationsare shown as the average from the 100 different tests, with the 95th percentile confidence interval displayed.

Stafford et al. PNAS | Published online July 14, 2014 | E3075

IMMUNOLO

GYAND

INFLAMMATION

STATIST

ICS

PNASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

17, 2

021

of training and test samples were collected from multiple sites(Table 1) and used in a train/blinded test format. In the trainingphase, 20 sera samples from each of six cohorts were used: (i)patients with advanced pancreatic cancer (PC), (ii) therapy-naiveglioblastoma multiforme (GBM; an aggressive form of astrocy-toma), (iii) esophageal adenocarcinoma (EC), (iv) multiple my-eloma (MM), (v) stage IV breast cancer (BC), and (vi) mixed“nondisease” controls. The nondisease controls were obtainedfrom collaborating clinics, as well as healthy samples collectedlocally. These sample cohorts were assayed and used to definethe signatures for each disease.Assays were performed as described in Materials and Methods.

Each assay was performed in duplicate, and the average Pear-son’s correlation coefficient between replicates for all 120 sam-ples in the training set was 0.92 ± 0.05. BC demonstrated thelowest average replicate correlation (0.87), and EC demon-strated the highest (0.96). We first performed a t test betweeneach of the n = 20 cancer cohorts and the n = 20 control cohort,one by one. The number of peptides with P < 9.6 × 10−5 is listedin Table 5, along with the minimum P value obtained. In each case,there were at least 600 and typically >1,000 peptide features withP < 9.6 × 10−5. The minimum P value in each case was more the sixorders of magnitude smaller than random chance would predict,implying that the separation between each disease and healthycontrols was statistically sound. The sensitivity in distinguishingeach sample ranged from 80–100%, with PC having the lowestsensitivity. The specificity was greater than 98% for each di-agnosis. In a pairwise test against control patients, MM displayedthe most significantly different peptides by t test, at 3.25 × 10−34.Of the top 100 peptides selected in this way, only BC showed nooverlap with any other disease.The analysis described above indicates that a signature dis-

tinguishing each cancer from noncancer controls can be estab-lished. Clinically, it would also be relevant to be able to distinguisheach cancer from the other types. In the analysis described above,there was overlap in the signatures distinguishing each cancerfrom noncancer, as shown in the rightmost column of Table 5. Inthe case of BC, the top 100 peptides that distinguished it fromhealthy controls via t test were completely unique (i.e., none ofthose peptides appeared in the top 100 peptides of any otherdisease); however, for GBM, 26 peptides appeared at least oncein another list. This implies that greater stringency is required toobtain sufficiently high specificity in a multiclass analysis than canbe obtained by t test.To assess the performance of multiple classifications, multi-

class peptide feature selection was performed as described in

Materials and Methods. Twenty-four of the most distinguishingpeptides per disease were selected for a total of 120 peptides inthe final feature set. PC and BC had relatively low overall sig-nals, whereas EC and brain cancer had much higher signals.Therefore, it was important to perform the feature selection insuch a way that the classifier is not overwhelmed by diseases withstronger average signal strength (Materials and Methods). Aleave-one-out cross-validation of the training set produced onlytwo miscalls of 120 calls when using SVM as a classifier.The blinded test set consisted of 20 samples from each of the

six cohorts. These samples were held blinded by collaboratorsuntil the analysis was complete. Nondisease controls were se-lected at random from a mixture of blinded controls from col-laborators, as well as internally blinded, locally collected healthycontrols (distribution of collection sites is shown in Table 1). Thetest dataset was classified using the 120 peptides obtained fromthe training described above with SVM as the classifier. Theresults are shown in Table 3, with the associated ROC curveshown in Fig. S1. Visualization of the relative group-wise sepa-ration is shown in Fig. 1 for each of the tested classifiers (SVM iscoded in orange and is always listed first). To calculate FP andFN results, any sample that was incorrectly called was counted asan FP result for the called disease and counted as an FN resultfor the cohort to which it actually belonged. Even using thisstringent scoring approach, with the exception of PC, the sensi-tivity was greater than or equal to 95% (80% for PC) using SVM.The specificity for all cancers was at least 98% using SVM. Thelow sensitivity for PC may be due to general immune suppressionin later stages of this disease (16, 25, 26).To test the analysis for dependence on a particular type of

classifier, we examined four other classifiers in addition to SVM(Table S3 and Figs. S2–S5 for the ROC charts and Fig. 1, Upperand Lower, for each classifier’s group-wise separation). Noclassifier or feature set was optimized to increase accuracy. Ofthe classifiers tested, SVM gave the best results and principalcomponent analysis (PCA) gave the worst results, particularly interms of sensitivity. This is not surprising, because the hyper-planes associated with the SVM algorithm allow for a more ac-curate description of multidimensional (many disease cohorts)data space than a methodology like PCA that attempts to de-scribe the dataset as a minimum number of components (addi-tional details of classifier performance and limitations are providedin SI Text).To give a visual indication of the intensity data and disease

differences resulting from trial 1, Fig. 2 shows a heat map of the120 feature-selected peptides (y axis) and 120 patients (x axis)ordered by divisive hierarchical clustering using Euclidean dis-tance with average linkage to estimate node separation. Thishierarchy is explicitly depicted in the colored dendrogram (Fig.2, Left). In Fig. 2, the results from a k-means clustering of thepeptides, where k = 5 classes (shown as I to V), are listed to theright of each heat map. The noncancer controls were not used toselect nondisease peptides; thus, there were five groups of pep-tides and six groups of patients. One heat map (Fig. 2, Left)shows the training dataset using the 120 selected features, andthe other heat map (Fig. 2, Right) shows the unblinded test dataclustered using the same 120 peptides.In trial 1, the fact that signals from 10,000 peptide features

were used to select a small subset of features that discriminatedthe five diseases and healthy controls might give rise to over-fitting if the analysis was done incorrectly. The blinded testanalysis verified that there was no pure statistical overfitting ofthe cohorts, as one might get by selecting features from datadominated by random noise. However, one might still be con-cerned that the discrimination seen was not for the disease butfor other aspects of the sample cohorts. It was therefore im-portant to exclude the possibility of fitting to a particular samplecollection, whether the protocol, the geographical location, the

Table 5. Statistical analysis of trial 1 peptides using t test

Disease P < 9.6 × 10−5 Minimum P valueCommon peptides

per 100

BC 608 1.54 × 10−14 0ECl 3,103 4.8 × 10−25 14GBM 3,596 9.05 × 10−30 26Healthy NA NA NAMM 4,478 3.52 × 10−34 19PC 1,126 3.67 × 10−11 12

A standard t test using multiple testing correction (familywise error rate =5%) was used to compare the 20 training samples from trial 1 for eachdisease against the single set of 20 noncancer controls, as a binary compar-ison. Column 1 lists the disease cohort. Column 2 lists the number of peptideswith a P value <9.6 × 10−5 (corresponding to one FP result per 10,480 pep-tides). Column 3 is the minimum P value calculated for each pairwise com-parison. Column 4 is the number of peptides of the top 100 most significantthat overlap with peptides from at least one other disease, used as a test forspecificity. BC had no overlap with any other disease, whereas GBM over-lapped with peptides from three different diseases. NA, not applicable.

E3076 | www.pnas.org/cgi/doi/10.1073/pnas.1409432111 Stafford et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

17, 2

021

patient population, or the manner in which the sample wasstored. To control for this, training and test samples for two ofthe diseases were randomly selected from three sites (BC andPC; Table 1). More importantly, the healthy controls were se-lected from four sites. Thus, if sample-specific problems were anissue, one would have expected that some of the 20 healthycontrols would have been miscalled. In fact, all 20 of the healthycontrols were called correctly (Table 3), making it unlikely thatnonbiological factors dominated the classification performance.

Trial 2. One of the more important results from trial 1 is thata signature can be defined that accurately detects and identifiesa complex and heterogeneous disease, such as stage IV BC rel-ative to healthy controls and four other cancers. We wanted toextend this idea by asking how many different diseases could bedistinguished on this type of array. Trial 2 was designed to ex-plore this question. The number of diseases was increased fromfive to 14, and the number of total samples was increased from120 to over 1,500. Also included separately are several differenttypes of brain cancers, two cohorts of early BC by stage, anda cohort of BC that involves a new occurrence of a tumor (nota metastasis). In addition, PC and pancreatitis are included sothat two distinct diseases of the same organ are present. Thequestion was whether disease heterogeneity limited the numberof diseases that could be distinguished, at least with arrays of10,000 peptides.Table 2 describes the samples used in the 1,516-sample cohort.

From each disease sample set, 75% of test samples were ran-domly selected as the training set as described in Materials andMethods, and feature selection and training were performedusing these samples independent of the test samples. Theremaining 25% of test samples were then called based on theresulting signature. This random resampling, training, and test-ing were repeated 100 times.

Fig. 3 is a heat map depicting the binding intensity associatedwith 280 classifier peptides from one of the feature sets arisingfrom resampling (Materials and Methods) across the entire 1,516-patient sample set, with cohort size listed in parenthesis. Thecolors in Fig. 3 distinguish high (red) from low (blue) intensities,and the patterns that remain after hierarchical clustering of bothpeptides (y axis) and patients (x axis) help to visualize the relativedifference within and across disease cohorts. Fig. 4 illustrates theways by which individual peptides contribute to the overalldisease classification performance.Table 4 displays the average results of the resampling con-

ducted 100 times using an SVM classifier (Materials and Meth-ods). The average accuracy of assigning the test cohort was 97%or greater for each disease and healthy controls, in support of thethesis that distinct signatures can be determined for each of thedisease cohorts even in the background of such a varied andcomplex set of samples. Table 4 includes results for two otherclassifiers (LDA and NB), which were similar. We conclude that10,000 peptides used to develop immunosignatures may be ableto distinguish up to 15 different disease conditions.

DiscussionWe first established signatures for five different cancers relativeto noncancer controls, using 20 known training samples for eachdisease. This was done using a case vs. control method for eachdisease vs. nondisease controls. The signatures thus created wereable to classify the blinded test set with less than perfect speci-ficity. Using this type of pairwise feature selection method led tooverlap in the peptides designed to classify the different cancertypes. To distinguish each cancer optimally, peptides were se-lected to optimize multiclass separation. Peptides were selectedfrom the training set using a more stringent process, reducing theoverlap in peptides to zero. Greater than 95% accuracy wasobtained when testing the 120 blinded samples. The rigor of thistype of analysis was extended in a second trial involving 14 dis-ease classes and a nondisease control group using 1,516 differentsamples. Seventy-five percent of the samples for each diseaseclass were used for training, following the same selection processas the first trial, with the remaining 25% used as the test set. Theprocess was repeated 100 times to arrive at overall accuracy. Thesensitivity for each disease group was >95%, except for PC(80%), and the overall specificity was 95%.With 109 different circulating antibodies in the blood (27), it

may seem remarkable that the signature of antitumor antibodies isdetectable. It is clear that B cells respond to tumors at early stages(26, 28, 29). Antibodies to both self-antigens and neoantigens are

Fig. 1. Classification visualizations for trial 1 using five classifiers: SVM, PCA,LDA, k-nearest neighbors (k-NN), and NB. The five charts shown here com-prise a visual representation of the relative inter- and intragroup differencesas determined by the clustering or classification methods used. The plots arelimited by the attempt to unify the visual representation for very differentclassifiers. As such, we do not display axis values but suggest that the readernotes the general trend. Classifier limitations and properties are discussed inSI Text. Classification performance values are listed in Table 3 for SVM and inTable S3 for PCA, LDA, k-NN, and NB. (Upper Left) Support vectors for SVMare extrapolated as a relative distance and then summed for display on an xyplot. BC is black for all figures, EC is red, normal donors are magenta, PC isbrown, MM is blue, and GBM brain cancer is cyan. (Upper Center) PCA dis-plays the first two principal components plotted on the x and y axes, re-spectively. (Upper Right) LDA graph displays the relative difference betweenthe discriminants on an xy plot. (Lower Left) k-NN as a classifier yields within-and across-group relationships with groups assigned by ellipses. The grouprelationships with each other are shown, but the group identities are notshown to keep the display legible. (Lower Left) Estimated probability den-sities from two estimator variables are plotted for each disease group.

Fig. 2. Heat map of test samples from trial 1. Diseases use the same colorsand abbreviations as in Fig. 1. The values for each of the 120 peptides and120 patient samples are plotted with blue, indicating low binding, and withred, indicating high binding. Hierarchical clustering using Euclidean distanceas the measure of similarity was used to cluster the peptides (y axis) andpatients (x axis). The hierarchy to the far left is based on this clustering re-sult. The colored bars to the right were created by k-means clustering usingfive clusters.

Stafford et al. PNAS | Published online July 14, 2014 | E3077

IMMUNOLO

GYAND

INFLAMMATION

STATIST

ICS

PNASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

17, 2

021

induced. Both nucleic acid-encoded and nonencoded antigens(e.g., glycosylated) can elicit a B-cell response. The results wereport indicate that antibodies unique to each type of tumorcan bind to the array and be discerned from other antibodies.Presumably, neoantigens would elicit higher affinity antibodiesthan self-antigens or natural antibodies (30), which is an im-portant presumption, given that we have demonstrated a high-affinity monoclonal antibody can be diluted 100-fold into healthyserum without diminishing the signal of the monoclonal anti-body (15). This implies that high-affinity antibodies (i.e., thoseelicited by foreign antigens and by antibody maturation) havehigher affinity to the random peptides than low-affinity anti-bodies. This may, in part, explain the ability to discern the tumor-specific signature. Even though multiple individuals recognizethe same tumor-specific antigen, individuals respond to theseantigens slightly differently. Immunosignaturing likely detectsthe unique antibodies that individuals raise against tumor anti-gens, but the feature selection process excludes peptides that arenot common across cohorts with the same disease. Thistraining process is critical to establishing a trustworthy signaturethat works in a broader population.Given that the sequencing of thousands of tumors has led to

the conclusion that tumor mutations are very specific, it mayseem remarkable that a common signature can be discovered foreach tumor. The definition of the signature is based on the re-quired distinction between samples. The immunological impli-cations of these common signatures are that each class of tumorspresents at least some common antigens to the immune systemand each individual makes a similar antibody to that antigen.This antigen can be misproduction or unusual posttranslationalmodification of a native protein, such as Her2 or MUC1. If thetumors present common protein variants (e.g., frameshift, mu-tation, posttranscriptional or posttranslational variant), they wouldalso create a common immunosignature. An important implicationis that common tumor-specific antigens exist that are not beingdetected by genomic or transcriptome sequencing.An unusual feature of the immunosignature procedure is that

because it involves a true signature, the training phase alsobecomes the de facto discovery phase. The implication is that thesignature peptides are determined by the distinction required,whether a single disease vs. control or multidisease vs. control. Thiswas evident in trial 1, where the peptides chosen to distinguish each

cancer from the control cohort using a pairwise case vs. controlmethod were different from those discerning multiple cancers si-multaneously. There may be peptides that react to more than one

Fig. 3. Heat map of samples from trial 2. In total, 1,516 samples (x axis) are shown with the values for each of the 255 predictor peptides (y axis). Each diseaseis listed, with the total number of patients indicated in parentheses. The heat map is generated as in Fig. 2. Oligo/Astro, oligoastrocytoma.

Fig. 4. Line graph for two of the 255 classifier peptides from trial 2. (Upper)Graph (red) shows the intensity across all 1,516 patient samples for peptideFLKWWGHIRAPTDHSRWGSC. (Lower) Graph (blue) shows the intensity forpeptide FPEILSTTIDRVVVNRGGSC. The y axis is the normalized intensity foreach peptide, and the x axis is the patient sera sample with the diseaseclasses split. (Upper) Example shows a peptide with high intensity for threedifferent diseases. This peptide is an example of one that is not perfect foran individual disease but contributes partly to discriminating three of thediseases from the other 11 diseases. (Lower) Example is a peptide that is veryhigh for a single disease and very low for every other disease. These twoexamples represent how individual peptides contribute to the ability of theclassifier to distinguish multiple diseases simultaneously. Disease classes aregrouped together and arbitrarily represented by numbers 1–15 rather thanthe disease name. Group 14 is unlisted due to size constraints.

E3078 | www.pnas.org/cgi/doi/10.1073/pnas.1409432111 Stafford et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

17, 2

021

cancer (see Fig. 4), but that knowledge remains undiscovered untilthe other cancer(s) is tested.The robustness of the assay becomes apparent by examining the

nature of the samples used. All samples came from historicalcollections, some >10 y old. Additionally, samples were collectedfrom multiple geographical sites, and in trial 2, the size of thecohort could vary and no effort was made to match age, sex, orethnicity. Using this type of diverse sample set for trainingdemands a robust signature, but also one that is less likely to fail intesting due to overfitting. That the immunosignature technique isapparently robust as well as simple and inexpensive may enhancesuccess with large and diverse training sets. Combined with theability to use historical samples (17), the cost of clinical validationof immunosignature diagnostics could be very low by comparisonwith standard practice. Also, unlike standard biomarker de-velopment, the platform for discovery of an immunosignaturewould be the same as that used in clinical practice. Changes indiagnostics that make them more clinic-friendly often createnew performance characteristics that require new evaluations.A potential value of the immunosignature diagnostic is that

the same platform would be used for any type of diagnosis. Thisfeature would drive the cost down, allowing broad use, even forearly detection of disease. To meet this objective, independentsignatures for multiple diseases would be obtained from thesame array. We had previously demonstrated that the same arraycould signature infections (15), Alzheimer’s disease (31, 32),pancreatic diseases (16), and four types of brain cancer (19).However, in trial 2, we significantly expanded the demand forcross-disease distinction. The signatures were chosen to distin-guish 14 diseases from each other and from a broad range ofnoncancer controls. The cancers included three different stagesof BC, four different brain cancers, two diseases of the pancreas(only one of which was PC), ovarian cancer, and two differentblood cancers. The conclusion was that 10,000 peptides weresufficient to select signatures that could distinguish these serumsamples by disease with at least 95% overall accuracy. How manymore diseases and how many different types could still be dis-tinguished on 10,000 peptides remain to be seen, because somecancers, such as GBM, yield unique and robust signatures,whereas others, such as PC, are far less distinct. Although 95%accuracy should be sufficient for regular monitoring of majordiseases, regular monitoring of less prevalent diseases mightrequire greater performance. We note that the next version ofthe immunosignature peptide microarrays currently under de-velopment has >300,000 peptides.The immunological implication of the diagnostic distinctions

we report here is that each of the disease conditions produceseither different antigens and/or different B-cell responses tocommon antigens. Both would create distinctive antibodies, andtherefore distinctive signatures. In the case of cancer, it seemslikely that the antigens eliciting the signature B-cell response aremade by the tumor cells, but the antigens could also be elicited in

nontumor cells. For example, it has been shown that stromalcells around a tumor are genetically altered (33, 34). We notethat the different stages of BC had distinguishing signatures.Perhaps each stage produces a different dominant set of antigensor the B-cell response to the same antigen matures, creatinga new signature. The end result is a unique signature, but theunderlying process remains unknown.The peptides on the array are chosen from nonnatural se-

quence space. This allows the same array to be used for anydiagnostic in any species. However, the limitation of this strategyis that one cannot simply align the peptides in the signature tothe natural proteomic space to identify the causative antigen. Wehave demonstrated that under some circumstances, a relevantalignment can be established (14), and we have developed analgorithm to facilitate this effort (35), but higher density pep-tide arrays will be required for routine use of this approach inantigen discovery.Finally, relative to the ideal diagnostic system, we have shown

that immunosignatures offer the potential of a single, simpleplatform for diagnosing multiple diseases. Another importantspecification in this regard is the ability to detect disease early(36). In this regard, immunosignatures may have an advantageover conventional biomarkers because the activated B cell canamplify its signal (specific antibody) 1011-fold in 1 wk (12, 37).We have demonstrated that the immunosignature of Alzheimer’sdisease in the mouse model can be detected very early (32) andthat pancreatic intraepithelial lesions in humans (an early formof ductal pancreatic adenocarcinoma) have distinct signatures(16). The extent to which this platform can generally be used todetect disease early with adequate specificity requires furtherinvestigation.In summary, we have demonstrated that the immunosignature

technology using a 10,000-peptide array is capable of high ac-curacy in a standard training and blinded test assay, and can beused for the simultaneous classification of multiple cancers.The arrays used in this study are available for purchase (www.peptidearraycore.com). Raw data are available at the GEO(National Center for Biotechnology Information) under super-series accession no. GSE52582.

ACKNOWLEDGMENTS. We thank our collaborators who supplied thenumerous patient samples from their respective institutions. In many cases,samples for a single cancer type from one collaborator were collected atmultiple sites; however, those sites are not listed. We thank Dr. Hoda Anton-Culver (UCI: BC, controls), Sam M. Hanash (FHCRC: BC and lung cancer,controls), Adi Gazdar (UTSW: BC and lung cancer, controls), John Galgiani(UAC: Valley fever, controls), Amy Stoll (PCRT: PC, controls), Doug Levine(MSKCC, ovarian cancer, controls), John Martignetti (MS: ovarian cancer),Adrienne Scheck (BNI: brain cancers), Teri Brentnall (UW: PC, pancreatitis,controls), Mary Bonner (CC: PC, controls), and Robert Penny (MultipleMyeloma Research Consortium). This work was supported by an InnovatorAward from the Department of Defense Congressionally Directed MedicalResearch Programs (to S.A.J.) and by funds supplied by the Biodesign Insti-tute at ASU (P.S. and S.A.J.).

1. De Cecco L, et al. (2009) Impact of biospecimens handling on biomarker research inbreast cancer. BMC Cancer 9(1):409.

2. Mischak H, et al. (2010) Recommendations for biomarker identification and qualifi-cation in clinical proteomics. Sci Transl Med 2(46):46ps42.

3. Poste G (2011) Bring on the biomarkers. Nature 469(7329):156–157.4. Ioannidis JPA (2013) Biomarker failures. Clin Chem 59(1):202–204.5. Cramer DW, et al. (2011) Ovarian cancer biomarker performance in prostate, lung, co-

lorectal, and ovarian cancer screening trial specimens. Cancer Prev Res (Phila) 4(3):365–374.6. Reiman JM, Kmieciak M, Manjili MH, Knutson KL (2007) Tumor immunoediting and

immunosculpting pathways to cancer progression. Semin Cancer Biol 17(4):275–287.7. Kotera Y, Fontenot JD, Pecher G, Metzgar RS, Finn OJ (1994) Humoral immunity

against a tandem repeat epitope of human mucin MUC-1 in sera from breast, pan-creatic, and colon cancer patients. Cancer Res 54(11):2856–2860.

8. Jäger E, et al. (1998) Simultaneous humoral and cellular immune response againstcancer-testis antigen NY-ESO-1: Definition of human histocompatibility leukocyteantigen (HLA)-A2-binding peptide epitopes. J Exp Med 187(2):265–270.

9. Dunn GP, Old LJ, Schreiber RD (2004) The immunobiology of cancer immunosurveillanceand immunoediting. Immunity 21(2):137–148.

10. Hudson ME, Pozdnyakova I, Haines K, Mor G, Snyder M (2007) Identification of dif-ferentially expressed proteins in ovarian cancer using high-density protein micro-arrays. Proc Natl Acad Sci USA 104(44):17494–17499.

11. Darnell RB, Posner JB (2003) Paraneoplastic syndromes involving the nervous system.N Engl J Med 349(16):1543–1554.

12. Cooperman J, Neely R, Teachey DT, Grupp S, Choi JK (2004) Cell division rates ofprimary human precursor B cells in culture reflect in vivo rates. Stem Cells 22(6):1111–1120.

13. Sykes KF, Legutki JB, Stafford P (2013) Immunosignaturing: A critical review. TrendsBiotechnol 31(1):45–51.

14. Halperin R, Stafford P, Legutki JB, Johnston SA (2010) Exploring antibody recognitionof sequence space through random-sequence peptide microarrays. Mol Cell Proteo-mics 28(1):e101230–e101236.

15. Stafford P, et al. (2012) Physical characterization of the “immunosignaturing effect”.Mol Cell Proteomics 11(4):011593.

16. Kukreja M, Johnston SA, Stafford P (2012) Immunosignaturing microarrays distin-guish antibody profiles of related pancreatic diseases. Proteomics and Bioinformatics,10.4172/jpb.S6-001.

Stafford et al. PNAS | Published online July 14, 2014 | E3079

IMMUNOLO

GYAND

INFLAMMATION

STATIST

ICS

PNASPL

US

Dow

nloa

ded

by g

uest

on

Janu

ary

17, 2

021

17. Chase BA, Johnston SA, Legutki JB (2012) Evaluation of biological sample preparationfor immunosignature-based diagnostics. Clin Vaccine Immunol 19(3):352–358.

18. Pauling L, Campbell DH (1942) The manufacture of antibodies in vitro. J Exp Med76(2):211–220.

19. Hughes AK, et al. (2012) Immunosignaturing can detect products from molecularmarkers in brain cancer. PLoS ONE 7(7):e40201.

20. Sing T, Sander O, Beerenwinkel N, Lengauer T (2005) ROCR: Visualizing classifierperformance in R. Bioinformatics 21(20):3940–3941.

21. Robin X, et al. (2011) pROC: An open-source package for R and S+ to analyze andcompare ROC curves. BMC Bioinformatics 12(1):77.

22. Fawcett T (2006) An introduction to ROC analysis. Pattern Recognit Lett 27(8):861–874.23. Chen C, et al. (2011) Removing batch effects in analysis of expression microarray data:

An evaluation of six batch adjustment methods. PLoS ONE 6(2):e17238.24. Johnson WE, Li C, Rabinovic A (2007) Adjusting batch effects in microarray expression

data using empirical Bayes methods. Biostatistics 8(1):118–127.25. Brand RE, et al. (2011) Serum biomarker panels for the detection of pancreatic cancer.

Clin Cancer Res 17(4):805–816.26. Bracci PM, Zhou M, Young S, Wiemels J (2012) Serum autoantibodies to pancreatic

cancer antigens as biomarkers of pancreatic cancer in a San Francisco Bay Area case-control study. Cancer 118(21):5384–5394.

27. Alberts B, et al. (2002)Molecular Biology of the Cell (Garland Science, New York), 4th Ed.28. Butts CA (2013) Anti-tumor immune response in early stage non small cell lung cancer

(NSCLC): Implications for adjuvant therapy. Transl Lung Cancer Res 2(5):415–422.

29. Savitskaya YA, et al. (2010) Circulating Natural IgM Antibodies Against Angiogenin inthe Peripheral Blood Sera of Patients with Osteosarcoma as Candidate Biomarkersand Reporters of Tumorigenesis. Biomark Cancer 2:65–78.

30. Notkins AL (2004) Polyreactivity of antibody molecules. Trends Immunol 25(4):174–179.

31. Restrepo L, Stafford P, Magee DM, Johnston SA (2011) Application of immunosignaturesto the assessment of Alzheimer’s disease. Ann Neurol 70(2):286–295.

32. Restrepo L, Stafford P, Johnston SA (2013) Feasibility of an early Alzheimer’s diseaseimmunosignature diagnostic test. J Neuroimmunol 254(1-2):154–160.

33. Gius D, et al. (2007) Profiling microdissected epithelium and stroma to model genomicsignatures for cervical carcinogenesis accommodating for covariates. Cancer Res67(15):7113–7123.

34. Vargas AC, et al. (2012) Gene expression profiling of tumour epithelial and stromalcompartments during breast cancer progression. Breast Cancer Res Treat 135(1):153–165.

35. Halperin RF, Stafford P, Emery JS, Navalkar KA, Johnston SA (2012) GuiTope: Anapplication for mapping random-sequence peptides to protein sequences. BMC Bio-informatics 13(1):1.

36. Hartwell L, Mankoff D, Paulovich A, Ramsey S, Swisher E (2006) Cancer biomarkers: Asystems approach. Nat Biotechnol 24(8):905–908.

37. Förster I, Rajewsky K (1990) The bulk of the peripheral B-cell pool in mice is stable andnot rapidly renewed from the bone marrow. Proc Natl Acad Sci USA 87(12):4781–4784.

E3080 | www.pnas.org/cgi/doi/10.1073/pnas.1409432111 Stafford et al.

Dow

nloa

ded

by g

uest

on

Janu

ary

17, 2

021