Antibody biomarker discovery through in vitro directed evolution of consensus recognition epitopes

Significance

The diagnosis of many diseases is dependent upon accurate detection of particular antibodies present in blood. However, the development of biochemical reagents that can reliably detect these antibodies has proved remarkably challenging. This study describes a process to create biochemical reagents that can accurately and reliably detect disease-associated antibodies, without requiring knowledge of the cause or mechanisms of disease. Simultaneously, this process enabled identification of a critical environmental agent involved in celiac disease. Thus, the process presented here may enable the development of effective diagnostic tests for other medical conditions where such tests are lacking and the identification of environmental factors involved in disease.

Abstract

To enable discovery of serum antibodies indicative of disease and simultaneously develop reagents suitable for diagnosis, in vitro directed evolution was applied to identify consensus peptides recognized by patients’ serum antibodies. Bacterial cell-displayed peptide libraries were quantitatively screened for binders to serum antibodies from patients with celiac disease (CD), using cell-sorting instrumentation to identify two distinct consensus epitope families specific to CD patients (PEQ and ^E/_DxFV^Y/_FQ). Evolution of the ^E/_DxFV^Y/_FQ consensus epitope identified a celiac-specific epitope, distinct from the two CD hallmark antigens tissue transglutaminase-2 and deamidated gliadin, exhibiting 71% sensitivity and 99% specificity (n = 231). Expansion of the first-generation PEQ consensus epitope via in vitro evolution yielded octapeptides QPEQAFPE and PFPEQxFP that identified ω- and γ-gliadins, and their deamidated forms, as immunodominant B-cell epitopes in wheat and related cereal proteins. The evolved octapeptides, but not first-generation peptides, discriminated one-way blinded CD and non-CD sera (n = 78) with exceptional accuracy, yielding 100% sensitivity and 98% specificity. Because this method, termed antibody diagnostics via evolution of peptides, does not require prior knowledge of pathobiology, it may be broadly useful for de novo discovery of antibody biomarkers and reagents for their detection.

The diagnosis of many diseases relies heavily upon the accuracy of antibody detection. Assays to detect antibodies using known antigens are used extensively to diagnose infectious and autoimmune diseases. And antibodies exhibiting unique antigen-binding patterns have been shown to occur in diverse human diseases, including oncological (1), inflammatory (2), and neurological and psychiatric disorders (3). The utility of antibodies in diagnostics derives from their intrinsic affinity and specificity, biochemical stability, and abundance in blood. Nevertheless, the identification of rare antibody specificities indicative of disease and the development of reagents for their accurate detection have proved exceptionally difficult (4). Intersubject variability of antibody specificities is a major challenge to the development of accurate tests. Specifically, individual genetic and stochastic variations that shape the antibody repertoire introduce heterogeneity in disease antibody subpopulations (polyclonal variation, specificity, affinity, and titer) that hinders uniform antibody detection (5, 6).

Random peptide libraries (RPLs) have been proposed as a potential source of diagnostic reagents capable of mimicking diverse biological antigens in the environment (7–9). Individual peptides identified from RPLs using patient sera have been capable of identifying patients with disease with modest accuracy (9, 10). Diagnostic accuracy can be improved in some cases, using panels of library-isolated peptides coupled with statistical classification algorithms (11), with the drawback of requiring multiple independent measurements. Despite these advances, peptides identified from random libraries have exhibited insufficient diagnostic efficacy (sensitivity and specificity) to foster their clinical development (11–13). Although approved antibody-based diagnostic assays often exhibit sensitivity and/or specificity values in excess of 95% (14, 15), library isolated peptides that mimic antigens (mimotopes), used alone or in combination, rarely meet these stringent requirements. For example, peptides from RPLs selected against serum antibodies from patients with Crohn’s disease (16), multiple sclerosis (12, 17, 18), celiac disease (11, 13), rheumatoid arthritis (19), or type-1 diabetes (20–22) have exhibited insufficient diagnostic accuracy. Although these studies have provided support for continued investigation of antibodies as candidate biomarkers, they have not yielded clinically efficacious diagnostic reagents. Consequently, there remains a need for discovery processes to produce antibody detection reagents exhibiting accuracies desired for clinical development.

Although antibody profiling methods using RPLs, including phage and bacterial display, lend themselves to various in vitro directed evolution protocols, this capability has not been exploited using blood specimens from patients. Given this, we applied bacterial display peptide libraries to first screen for disease-specific antibody binding peptides and subsequently to evolve peptides to achieve diagnostically useful levels of sensitivity and specificity. We selected celiac disease (CD) as a model disease because two distinct antibody specificities, transglutaminase 2 (TG2) and deamidated gliadin, have been characterized extensively (23) and serve as clinically important antibody biomarkers. Our results demonstrate that in vitro directed evolution can be applied for de novo generation of reagents that exhibit requisite levels of diagnostic sensitivity and specificity for clinical translation. Finally, our results raise the intriguing possibility that in vitro evolution of such diagnostic reagents may provide a route to identify previously unknown environmental antigens involved in disease and thereby elucidate pathobiology mechanisms.

Results

Discovery of Celiac Disease-Specific Peptide Epitopes.

Bacterial display random peptide libraries of the form X₁₅, X₁₂CX₃, and X₄CX₇CX₄ were screened using fluorescence-activated cell sorting (FACS). For screening, individual patient sera were pooled into three groups of CD cases and three groups of non-CD sera [i.e., healthy and gastrointestinal (GI)-illness control subjects], with each group composed of sera pooled from eight subjects. Alternating rounds of library enrichment were performed with CD sera using FACS and subtraction with non-CD sera using magnetic cell sorting (MACS) (Fig. 1). To determine whether enriched library members were specific for sera from CD groups and thereby guide screening, flow cytometry was applied to quantitatively measure reactivity levels after each cycle of sorting (Fig. S1). Libraries were sorted independently based on isotype-specific reactivity, using anti-IgG, anti-IgA, and anti-IgM secondary reporters. Alternating cycles of enrichment/subtraction resulted in large reactivity differences between pooled CD and non-CD sera for IgA and IgG, but not IgM, binding peptides (Fig. S2 A and B). Peptide sequences from IgG and IgA isotype-specific library screening revealed two prevalent epitopes among 195 clones: PEQ and DxFV^F/_YQ (Fig. 2A and Table S1). Peptides with the PEQ tripeptide emerged from both linear and constrained libraries, whereas those with DxFV^F/_YQ were identified almost exclusively from the constrained library pool.

Fig. 1.

Library screening algorithm to identify and evolve antibody-detecting peptides. A bacterial display random peptide library is subjected to repeated cycles of enrichment and subtraction with a sequence of pooled sera from CD groups or non-CD groups. Using consensus information from the primary library, a second-generation library is constructed and similarly screened with a new set of CD and non-CD sera.

Fig. 2.

Directed evolution of antibody-detecting peptides increases their sensitivity and specificity. (A) Sequences of individual peptides from the three most abundant consensus groups in each cycle of epitope evolution. See SI Materials and Methods for a complete list. (B) Bacterial clones expressing the PEQ-related peptides in A, Upper were pooled and assessed for IgG reactivity to five CD and five non-CD sera groups. Shown is a box-and-whiskers plot of the reactivity (fluorescence intensity) of each CD and non-CD sera group. The median value is plotted as a line with each box displaying the distribution of the inner quartiles, with whiskers showing the upper and lower quartiles (all differences are statistically significant, P < 0.0001). (C and D) Evolved consensus epitopes for (C) PEQ motif and (D) CSE generated using WebLOGO3.0.

In Vitro Evolution of CD-Specific Peptides.

To improve the reactivity of and consensus between first-generation peptides, a focused library of the form X₆PEQX₆ was screened as above. Pooled sera groups (n = 3 subjects per group) were used only once for library enrichment to favor peptides cross-reactive with antibodies from many patients with CD. The X₆PEQX₆ library was enriched for IgG- and IgA-specific binders, but IgG binders were more rapidly enriched and cross-reactive to multiple CD groups in comparison with IgA binders; thus, our subsequent analysis focused on IgG isotype reactivity. From the enriched library population, three highly represented consensus motifs were observed: PEQxFP, PEQPL, and ^A/_VFPEQ (Fig. 2A). To assess the diagnostic sensitivity and specificity of individual peptides, the reactivity of one representative clone from each motif group was measured using CD case (n = 18) and non-CD control sera (n = 5) not used for screening. The PEQxFP motif derived peptide VWDRGVPEQMFPRKG reacted with 18/18 CD sera, whereas VAWTMGPEQPLVRAL reacted with 11/18, and GQGQAFPEQGSVPIN reacted with 14/18. None of the peptides were reactive with control sera. To increase the information content and diagnostic performance of the most reactive consensus motif, a second cycle of epitope expansion was performed. Thus, a library of the form X₅PEQXFPX₄ composed of 10⁸ members was screened as above, using sera dilutions of 1:500 and 1:1,000. Epitopes identified from the final screening cycle exhibited an evolved consensus of PFPEQxFP, AFPEQxFP, or QPEQ^A/_SFPE (Fig. 2A). Collectively, the entire set of peptides obtained from the second focused library exhibited the evolved consensus dodecamer sequence PxE^P/_A^Q/_FPEQxFP^E/_D (Fig. 2C), after adjusting the final position for the overrepresentation of arginine that results from random-codon–generated RPLs. To assess whether epitope evolution improved the sensitivity and specificity of the identified peptide epitopes, four to five clones from each PEQ motif group (Fig. 2A) were pooled and assayed for reactivity with pooled sera from five patients with CD or non-CD subjects. Pooled clones from each expansion cycle exhibited increased reactivity (P < 0.0001) with sera from patients with CD and decreased reactivity with non-CD sera (P < 0.0001) (Fig. 2B), demonstrating that epitope expansion increased the diagnostic sensitivity and specificity of the identified peptides. Thus, in vitro directed evolution yielded peptide epitopes specifically recognized by IgG antibodies of patients with CD.

To evolve the DxFV^F/_YQ epitope, a second-generation library of the form X₆^D/_ExFV^Y/_FQCX₄ was screened. This library was more readily enriched for IgA, rather than IgG binders. Additional consensus residues emerged within the randomized region and cysteine-constrained epitope variants were preferred, including CRD^S/_TFV^F/_YQC, RCxD^S/_TFV^F/_YQC, and DCFV^F/_YQC (Fig. 2A and Table S2). Similarly, screening of a linear third-generation library of the form X₆D^S/_T/^AFV^F/_YQX₄ identified a preference for cyclic peptides having the consensus CEDSFV^F/_YQC (Fig. 2D) and nonconstrained linear epitopes with the consensus ΩD^S/_TFV^F/_YQ, where Ω = [L/I/M/F/E] (Table S2). Importantly, the unique celiac-specific epitope (CSE) was not a mimic of TG2 or deamidated gliadin (DGP) because antibody titers against these CD antigens were unaffected by depletion of antibodies binding to the unique epitope (Fig. S3 A–C). Given the weak consensus at the Ω position, the degenerate search motif D^S/_T^/CFV^F/_YQ was used along with ScanProsite to identify a panel of candidate antigens (Table S3).

Evolved Peptide Epitopes Exhibit High Diagnostic Sensitivity and Specificity.

To evaluate the diagnostic utility of expanded peptide epitopes from one cohort of cases and controls (Tables S4 and S5), sera from a second cohort of CD cases and controls (n = 78) were assayed in a one-way blinded test. Cases (35/38) were positive for TG2 and/or endomysial antigen serology with partial or total villous atrophy. Of the remaining 3 cases, 2 had total villous atrophy with negative or unavailable serology (Tables S7–S9). All control sera were from healthy donors negative for TG2 IgA. Two peptides (DGP3, RGRAQPEQAFPESVG; and DGP6, GPQPFPEQLFPDPFR) exhibiting high sensitivity and specificity in a preliminary set of 10 CD and 10 non-CD sera were assayed for IgG reactivity, and a diagnostic cutoff was established using the individual patient reactivity dataset. Epitope DGP3 correctly identified 100% of CD cases (38/38) and 97.5% (39/40) of non-CD controls; epitope DGP6 correctly identified 92.1% of CD cases (35/38) and 97.5% (39/40) of non-CD controls (Fig. 3A). For comparison, a commercially available Quanta Lite DGP IgG assay, using a cutoff value of 10 units, achieved 98% sensitivity and 100% specificity (Fig. 3B). Furthermore, assay results with epitope DGP3 correlated with those obtained using Quanta Lite (Fig. 3C). Thus, a single peptide generated using sequential epitope expansion performed equivalently to a proprietary, Food and Drug Administration-approved diagnostic assay.

Fig. 3.

Diagnostic assay enabled by ADEPt. (A and B) Measurement of blinded patient sera (n = 78) for IgG reactivity using (A) DGP3 (Left) and DGP6 (Right) and (B) Quanta Lite. (C) Assay results using ADEPt DGP3 epitope correlate with those obtained using Quanta Lite (Spearman’s coefficient, ρ = 0.89). (D) Serum IgA antibody reactivity to DSFVYQ epitope in 231 patient samples. (E) Matched sera from patients with CD before and after 1 y of GFD exhibit decreased reactivity to DSFVYQ.

To determine prevalence of anti-CSE antibodies in patients with CD and control subjects, CD and non-CD sera (n = 231) were assessed for reactivity to the CSE peptide: MDVRCRDSFVYQCHVGT. Overall, the CSE peptide exhibited 71% (65/92) sensitivity and 99% (2/139) specificity (Fig. 3D). To determine whether the serum antibody titer against the CSE epitope dissipated after the introduction of a gluten-free diet (GFD), sera from 11 CD cases obtained at time of diagnosis or after 1 y on a GFD were assayed. Patients with active CD (10/11) were reactive and 8/11 of these patients exhibited reduced, but nonzero, levels of epitope reactivity after a GFD (Fig. 3E); all patients were seronegative for TG2 and DGP antibodies after a GFD. Together, these results suggest the CD-specific peptide is derived from an antigen distinct from TG2 and DGP epitopes.

Directed Evolution of Peptide Epitopes Facilitates Nonself Antigen Discovery.

Due to the substantially increased information content within the third-generation evolved consensus epitopes (QPEQAFPE, PFPEQxFP) compared with the first-generation epitope PEQ, we reasoned that evolved epitopes might enable unbiased antigen identification within the entire protein database. Unbiased BLASTp searches of the epitopes QPEQAFPE and PFPEQxFP directly identified cereal grain proteins from the genus Triticeae, including gliadins, hordeins, and secalins (Fig. 4). For comparison, an identical search using the first- and second-generation motifs PEQ and PEQxFP yielded an excessive number of unrelated hits and did not enable antigen discovery. The highest-scoring antigen, obtained using the epitope consensus QPEQAFPE, was ω-gliadin from wheat (Fig. 4A). Similarly, use of the aggregate (i.e., using all sequences) consensus epitope from third-generation peptides (PxEP^Q/_FPEQxFPE; Fig. 2C) identified exclusively ω-gliadins among the 25 highest-scoring sequences. Searches performed with the third-generation motif PFPEQxFP also identified a diverse group of prolamins from wheat, barley, and rye (Fig. 4B). The third-generation motifs were identical to the prolamin epitopes that, in CD, result from posttranslational deamidation of glutamine to glutamic acid (Q→E) by TG2. Collectively, these results demonstrate that the in vitro directed evolution of epitopes can facilitate discovery of nonself antigens.

Fig. 4.

Protein antigens containing evolved epitope PEQ motif. (A and B) Proteins and organisms identified by query of (A) QPEQAFPE and (B) PFPEQXFP against the nonredundant protein database, using BLASTp (PAM30 Matrix) and rank ordered by total score.

Discussion

The antibody diagnostics via evolution of peptides (ADEPt) method presented here provides an effective route to evolve diagnostically efficacious peptides for de novo biomarker discovery and detection without knowledge of disease pathobiology. Previous methods to discover peptides binding to disease antibodies, including antibody profiling and signature analysis using peptide libraries (7, 24), have demonstrated the existence of unique antibody specificities in a broad range of diseases (25). And although the peptides identified have demonstrated diagnostic potential, alone or in panel format (8, 25), their translation to the clinic has been hindered by inadequate diagnostic sensitivity and specificity values. By applying concepts from in vitro directed evolution to human patient samples, we were able to screen large libraries in an iterative fashion for molecular properties (affinity, cross-reactivity, and molecular specificity) that favor diagnostic sensitivity and specificity. In agreement with many prior studies, our results demonstrate that a RPL, in the absence of directed evolution, is insufficient to identify peptides with optimal diagnostic efficacy. Only when the peptide search space was expanded through directed evolution were we able to achieve accuracies comparable to gold-standard diagnostics for CD. Thus, it may be possible to improve the diagnostic utility of previously reported peptides arising from RPLs using ADEPt. Although we concluded the directed evolution process after screening the third-generation focused epitope library wherein sensitivity and specificity were maximized (100%, 98%), further cycles of directed evolution could enhance the dynamic range between CD and non-CD signals. In short, our results demonstrate the potentially broad utility of directed evolution in the context of biomarker discovery and diagnostics development.

Here, environmental (i.e., nonhuman) protein antigens recognized by CD-specific antibodies were unambiguously identified using ADEPt. Multiple methods have been developed to identify candidate autoantigens, including synthetic peptide and peptoid arrays (3), whole-protein antigen arrays (1), and human cDNA or peptidome libraries (26). In contrast, methods to identify nonhuman antigens mostly closely associated with disease have not been reported. The rapidly expanding protein database, currently composed of more than 31 million protein sequences, is simply too large to enable database searching using the limited consensus data arising from a first-generation RPL. Epitope expansion using ADEPt dramatically reduced the frequency of antigen candidates within the nonredundant protein database, enabling precise identification of immunodominant B-cell epitopes (ω-gliadin, γ-gliadin, and B-hordein). Interestingly, the immunodominant B-cell epitopes were highly similar to recently elucidated immunodominant T-cell epitopes (27). We did not observe linear B-cell epitopes derived from the CD-specific autoantigen TG2, which is consistent with the proposed existence of immundominant structural epitopes within TG2 (28). However, we cannot rule out the possibility that lower-abundance linear epitopes or structural mimotopes were enriched during library screening but outcompeted by DGP and D^S/_TFV^Y/_FQ peptides. Future efforts using next-generation sequencing and bioinformatic tools may permit identification and characterization of a greater number and variety of disease-associated peptide epitopes.

Application of ADEPt to sera from patients with CD identified a previously unreported CSE. Antibodies binding CSE peptides with the consensus motif CXD^S/_TFV^Y/_FQC were present in 71% of patients with CD from geographically distinct cohorts and exhibited equivalent specificity (∼99%) for CD compared with gold-standard antibody biomarkers of CD (anti-TG2 IgA, anti-endomysial antibodies, and anti-DGP IgG). The sensitivity and specificity values observed with CSE are significant because many distinct antibodies have been reported to be present in patients with CD but the same specificities have been observed in unrelated disorders (29, 30). In contrast, the anti-CSE antibody specificity occurred exclusively within subjects with CD (29). The observation that anti-CSE antibody titers significantly decrease in matched sera from patients pre- and 1 y post-GFD further supports the disease specificity of this antibody specificity. Although the precise identity of the antigen mimicked by CSE remains to be elucidated, the ability of the evolved consensus epitope to narrow our search to <40 candidate antigens suggests that antigen discovery will be possible. Although these data highlight the need for an unbiased strategy to down-select candidate antigens, it is interesting to note the presence of human commensals and pathogens (Prevotella, Roseburia, Lactobacillus, Bacteroides, Vibrio, Burkholderia, Giardia, and Bacillus) and the common wheat fungal pathogen (Puccinia) among the candidates. Systematic analysis of fragments containing the epitope from each candidate antigen for their sensitivity and specificity may provide a means to uncover the antigen that gave rise to this antibody specificity. Finally, the well-established significance of DGP antibodies and TG2 autoantibodies to the pathobiology of CD suggests that confirmation of the antigen corresponding to CSE may provide additional clues regarding the mechanisms of CD pathogenesis.

In summary, we present ADEPt as a method enabling the simultaneous discovery of antibody biomarkers of disease and reagents for their sensitive and specific detection. In principle, ADEPt could be applied to a variety of antibody-containing specimens (e.g., serum/plasma, cerebrospinal fluid, urine, and saliva). Given the ubiquitous nature of antibody repertoire changes observed in diverse diseases, ADEPt may be useful to create diagnostics for early disease detection, stratification, and therapeutic monitoring (31). And finally, because this method is not constrained to searches for autoantigens, ADEPt may be useful to reveal previously unknown environmental factors involved in disease.

Materials and Methods

Bacterial Display Peptide Library Screening.

Bacterial display peptide libraries of the form X₁₅, X₄CX₇CX₄, or X₁₃CX₂ were screened using FACS and MACS to identify peptides binding to antibodies in sera from patients with CD but not to those in non-CD sera (Tables S4 and S5). The library was depleted of non-CD (i.e., healthy and GI-illness controls) antibody-binding peptides, using MACS. A frozen aliquot of each library containing 20 times the expected diversity was inoculated into 500 mL LB (10 g tryptone, 5 g yeast extract, and 10 g/L NaCl) supplemented with 34 μg/mL chloramphenicol (Cm) and grown to OD₆₀₀ = 0.5 at 37 °C with vigorous shaking (250 rpm). Protein expression was induced by addition of l(+)-arabinose to a final concentration of 0.02% wt/vol with shaking at 37 °C for 1 h. Cells (2.5 × 10¹⁰) were centrifuged (3,000 × g, 4 °C, 10 min) and resuspended in cold phosphate buffered saline with Tween 20 (PBST). To deplete the library of streptavidin- and protein A/G-binding clones, washed streptavidin-conjugated beads and protein A/G beads were added to a ratio of one bead per 50 cells, and the mixture was incubated 45 min at 4 °C on an inversion shaker. A magnet was then applied to the tube for 5 min and the unbound cells in the supernatant were recovered. To deplete the library of secondary antibody-binding peptides, a 1:500 dilution of secondary antibody was incubated with the cells followed by incubation with streptavidin (SA) beads and removal by magnet similar to SA-binding peptide removal. Subtractive MACS steps for removal of nonspecific serum antibody-binding peptides were performed in a similar manner to that of SA and protein A/G depletions except that before incubation with biotinylated secondary antibody or beads, the library was first incubated with 1:100 pooled non-CD sera (n = 8 subjects per pool) for 45 min at 4 °C, followed by washing two times with PBST. For positive selection, pooled CD sera 1:100–1:200 (n = 8 subjects per pool) were added to the library and incubated for 45 min. Magnetic separation was used to wash the beads three times with PBST, and the pellet was resuspended in LB with Cm and 0.2% glucose (wt/vol) for overnight growth. For flow cytometric analysis and sorting, induced cells corresponding to five times the estimated remaining clonal diversity were incubated with 1:100–1:200 dilution of pooled sera for 45 min at 4 °C followed by centrifugation and removal of unbound supernatant (washing three times). The pellet was then resuspended in the respective 1:500 biotinylated goat anti-human secondary antibody (IgA/IgG/IgM) in 1× PBST at 4 °C for 45 min followed by centrifugation and removal of unbound supernatant (washing two times). For tertiary labeling, the pellet was resuspended in 15 nM SA-phycoerythrin in 1× PBST at 4 °C for 45 min, followed by another wash via centrifugation and resuspension (washing three times) in ice-cold PBST at a volume between 10⁷ and 10⁸ cells/mL. Resuspended cells were analyzed using a FACSAria cell sorter (Becton Dickinson), using 488-nm excitation. After sorting, retained cells were amplified for further rounds of sorting by overnight growth and plated to isolate single clones.

Epitope Evolution by Cytometric Screening.

Second-generation libraries were constructed of the form X₆PEQX₆ and X₆[E/D]XFV[YF]QCX₄ on the N terminus of eCPX, using degenerate NNS oligonucleotides (32) (Table S6), resulting in an estimated library diversity of 2 × 10⁸ and 1 × 10⁸ members, respectively. A third-generation library of the form X₅PEQXFPX₄ and X₄D[STA]FV[YF]QX₅ was similarly constructed (Table S6). Directed library screening was performed as above except that unique nonrepeating pools of sera from patients with CD (n = 3 subjects per pool) were used for each round of enrichment such that no pool was used more than once and a nonrepeating non-CD control pool (n = 3–5 subjects per pool) was used for each round of subtraction. In an effort to expand upon the known antigenic sequence, third-generation libraries were screened using 1:500 and 1:1,000 pooled disease sera. PEQ focused libraries were screened by IgG isotype and DXFV^F/_YQ libraries by IgA isotype.

Additional Methods.

Additional descriptions of reagents and methods, including cohort information, sample handling and preparation, clone reactivity assays, antigen mimicry assays, and protein database queries for candidate antigen identification, are available in SI Materials and Methods.