Identification of the Feline Humoral Immune Response to Bartonella henselae Infection by Protein Microarray

Background Bartonella henselae is the zoonotic agent of cat scratch disease and causes potentially fatal infections in immunocompromised patients. Understanding the complex interactions between the host's immune system and bacterial pathogens is central to the field of infectious diseases and to the development of effective diagnostics and vaccines. Methodology We report the development of a microarray comprised of proteins expressed from 96% (1433/1493) of the predicted ORFs encoded by the genome of the zoonotic pathogen Bartonella henselae. The array was probed with a collection of 62 uninfected, 62 infected, and 8 “specific-pathogen free” naïve cat sera, to profile the antibody repertoire elicited during natural Bartonella henselae infection. Conclusions We found that 7.3% of the B. henselae proteins on the microarray were seroreactive and that seroreactivity was not evenly distributed between predicted protein function or subcellular localization. Membrane proteins were significantly most likely to be seroreactive, although only 23% of the membrane proteins were reactive. Conversely, we found that proteins involved in amino acid transport and metabolism were significantly underrepresented and did not contain any seroreactive antigens. Of all seroreactive antigens, 52 were differentially reactive with sera from infected cats, and 53 were equally reactive with sera from infected and uninfected cats. Thirteen of the seroreactive antigens were found to be differentially seroreactive between B. henselae type I and type II. Based on these results, we developed a classifier algorithm that was capable of accurately discerning 93% of the infected animals using the microarray platform. The seroreactivity and diagnostic potential of these antigens was then validated on an immunostrip platform, which correctly identified 98% of the infected cats. Our protein microarray platform provides a high-throughput, comprehensive analysis of the feline humoral immune response to natural infection with the alpha-proteobacterium B. henselae at an antigen-specific, sera-specific, and genome-wide level. Furthermore, these results provide novel insight and utility in diagnostics, vaccine development, and understanding of host-pathogen interaction.


Introduction
Controlling Bartonella infection in its cat reservoir is integral to preventing cat scratch disease (CSD) in humans. B. henselae infection is mainly asymptomatic in cats, but has been associated with kidney disease and urinary tract infections, stomatitis, and lymphadenopathy [1]. The prevalence of Bartonella infection in cats ranges from 25% to as high as 41% throughout the world [2]. Infected cats can have bacterial titers of .10 6 colony forming units (CFU)/ml of blood and can remain bacteremic for several months to several years. Cats that are bacteremic, especially with high titers, are more likely to infect humans by scratches or bites. Although antibiotic treatment of infected cats has been associated with reduction of bacteremia levels, treatment does not appear to be sufficient to completely eradicate B. henselae from the blood stream [3]. Indeed, treatment can result in increased transmission of B. henselae to humans during attempts to administer antibiotics pills to uncooperative, infected cats.
Preventing initial infection of cats by vaccination is a potential strategy for limiting B. henselae infections in humans. With an estimated 90 million pet cats in the US and a predicted 8-20 million cats with chronic bacteremia, prevention and reduction of morbidity in humans from CSD could be achieved through extensive cat vaccination programs [4]. Profiling the feline host antibody response to B. henselae infection is central to diagnostics development and the identification of potential subunit vaccine candidates. Importantly, there are two major genotypes of B. henselae that can cause CSD in humans: types I (Houston); and II (Marseille) [5]. Cats are most often infected with one or the other type, but some cats are co-infected with both types, and both types can be transmitted to humans from pets [5]. Thus, establishing and comparing the host immune profile to infection with both types may be necessary for optimizing candidate antigen selection to prevent feline infection with type I and type II.
We previously developed a protein microarray technology that allows construction of the complete predicted proteome of a microorganism [6,7,8,9,10]. Utilization of arrays constructed from in vitro transcription reactions can identify the repertoire of seroreactive antibodies to proteins encoded by an infectious agent. These arrays are limited to detection of antibodies against recombinant proteins and would not detect post-translational modifications and non-protein antigens [11]. However, these arrays can be utilized to address basic questions about the pattern of the host humoral immune response to infectious agents [12,13,14], and to identify individual antigens that could be used as diagnostic reagents or for inclusion in vaccines [6,15]. The data derived from these studies can also be used to evaluate and improve the accuracy of in silico predictions of seroreactive antigens, and can provide a more detailed understanding of the adaptive immune response to infection. In this study, we developed a B. henselae genome-wide protein array and used the arrays to profile the antibody response in naturally infected cats and uninfected cats.

Materials and Methods
Bacterial strains DNA extracted from B. henselae wild type strain JK33R was used for template DNA from which all ORFs were amplified prior to cloning. This B. henselae strain was isolated from the blood of an AIDS patient with bacillary angiomatosis and was cryopreserved after only several passages on agar. JK33R retains the rough colony phenotype characteristic of primary B. henselae isolates obtained from human and feline blood.

Cat serum samples
All procedures involving animals followed NIH protocols and were approved by and performed according to guidelines of the Institutional Animal Care and Use Committee of University of California, Davis. Serum samples were collected in 2008 from 124 cats housed in two shelters, one in California (Sacramento) and one in Michigan (Muskegon). These sera were tested for the presence of Bartonella antibodies at two serial dilutions 1:32 and 1:64 by two of the authors (BBC, RWK). A cat was reported serologically positive when the scoring on a scale from 0 to 4 was $2 at the 1:64 dilution. Uninfected cats had an IFA score range of 0.360.5 for B. henselae and 0.560.6 for B. clarridgeiae at the 1:64 dilution and the IFA score range at the 1:64 dilution for infected cats was 2.061.3 for B. henselae and 2.860.8 for B. clarridgeiae. Of these cats, 62 were identified as Bartonella blood culture negative and seronegative for Bartonella IgG antibodies by an IFA test [16]. Another 62 cats were determined to be either seropositive (IFA titer $1:64) or both IFA and culture positive (24 of the 62 seropositive cats were also culture positive). Similarly, serum samples from 8 Bartonella-uninfected (''specific pathogen free [SPF]'') cats were submitted for testing and were both culture and IFA negative. Cats ranged in age from ,3 months to adult (average uninfected cat age was approximately 19.9 months old with a standard deviation of 19.7 months; average infected cat age was approximately 19.4 months old with a standard deviation of 20.3).

PCR amplification of linear acceptor vector
Each predicted open reading frame (ORF) from the B. henselae genome sequence was amplified from B. henselae JK33 strain DNA, and was cloned into the pXT7 vector using a high-throughput PCR cloning method previously described [6]. The pXT7 plasmid (3.2 kb, KanR) encodes an N-terminal 10 x histidine (HIS) tag and a C-terminal hemagglutinin (HA) tag. pXT7 (10 mg) was linearized with BamHI (0.1 mg/ml DNA, 0.1 mg/mL BSA, 0.2 U/ml BamHI, Invitrogen) overnight at 37uC. The digest was purified using a PCR purification kit (Qiagen, Valencia, CA), quantified using a NanoDrop (Thermo Scientific), and verified by agarose gel electrophoresis. PCR was used to generate the linear acceptor vector in 50 ml PCR reactions with 0.5 mM of each primer (CTACCCATACGATGTTCCGGATTAC and CTCG-AGCATATGCTTGTCGTCGTCG). PCR was performed using 0.02 U/ml AccuPrime Taq DNA polymerase (Invitrogen), 0.2 mM of each dNTP, and 1 ng pXT7 diluted in AccuPrime Buffer II. The following PCR conditions were used: 95uC for 5 min; 30 cycles of 95uC for 0.5 min, 50uC for 0.5 min, 72uC for 3.5 min; and a final extension at 72uC for 10 min.

Open reading frame cloning
Primers were designed to all 1,493 ORFs that did not contain an internal stop codon (1493/1612), as predicted from the genome sequence of the B. henselae Houston-1 strain (BX897699.1) [17]. PCR primers were designed for the 59 and 39 ends of each ORF, with the addition of a 20 bp homologous recombination ''adapter'' sequence (ACGACAAGCATATGCTCGAG and TCCGGAA-CATCGTATGGGTA respectively). The adapter sequences, which become incorporated into the termini flanking the amplified gene, are homologous to the cloning sites of the linearized T7 expression vector pXT7. ORFs that are larger than 3000 bp were split into smaller fragments with 150 bp overlap for efficient amplification. PCR reactions were prepared using 0.02 U/ml AccuPrime Taq DNA polymerase (Invitrogen), 0.2 mM of each dNTP, diluted in Buffer II, with 2.5 ng of B. henselae template DNA, using the following conditions: 95uC for 2 min; 30 cycles of 95uC for 0.33 min, 55uC for 0.25 min, 50uC for 0.25 min, 68uC for 3 min; and a final extension of 68uC for 10 min. All B. henselae ORF-PCR reactions were confirmed by gel electrophoresis for correct insert size prior to cloning into pXT7.

High-throughput recombination cloning
All ORFs amplified from B. henselae strain JK33R DNA were cloned into the plasmid expression vector pXT7 using a highthroughput PCR recombination cloning method previously developed in our laboratory [6]. Linearized pXT7 was diluted to 10 ng/ml, mixed with 1 ml of B. henselae ORF PCR reaction mixture at a volume ratio of 4:1, and incubated on ice for 2 min, followed by addition of 10 ml of competent E. coli DH5a cells (MCLabs). Reactions were mixed, incubated on ice for 30 min, heat shocked at 42uC for 1 min, and chilled on ice for 2 min. 250 ml of SOC media was added and cells were incubated for 1 hr at 37uC. The entire reaction mixture was added to 1.5 ml of LB medium supplemented with 50 mg/mL of kanamycin, and incubated overnight at 37uC with shaking. Plasmids were isolated using QIAprep 96 Turbo kits (Qiagen, Valencia, CA) without colony selection. Minipreps of all attempted clones were analyzed by agarose gel electrophoresis to confirm insert size. 25% of all clones were confirmed for insert size by PCR using ORF sequence-specific primers. An additional 25% of all clones were selected at random and sequenced in both directions. Sequences were analyzed for fidelity, orientation, and for mutation in the overlapping region of the homologous recombination sites. The cloning efficiency for all amplified ORFs was 98.4%, resulting in 1520 plasmids encoding proteins from 1464 ORFs.

Protein microarray chip printing
The expression of cloned ORFs was carried out for five hours in in vitro transcription-translation (IVTT) reactions (RTS 100 kits, Roche) according to the manufacturer's instructions. Protein microarrays were printed onto nitrocellulose-coated glass FAST slides (Whatman) using an Omni Grid 100 microarray printer (Genomic Solutions). 3.3 ml of 0.2% Tween-20 was mixed with 10 ml of IVTT and transferred to 384-well plates. Plates were centrifuged at 16006 g to pellet any precipitate and remove air bubbles prior to printing. Supernatants were printed immediately without purification, and all ORFs were spotted in duplicate. Data values reported herein represent an average of the pair, unless otherwise mentioned. In addition, each chip was printed with control spots consisting of IVTT reactions without plasmid, purified IgGs, and purified EBNA1 proteins. Protein expression was confirmed using monoclonal anti-polyhistidine (clone His-1, Sigma) and anti-hemagglutinin (clone 3F10, Roche).

Microarray probing
Feline sera were preabsorbed with E. coli lysate prior to array staining, to remove background reactivity to E. coli proteins in the IVTT reactions. The sera were diluted to 1:200 in Protein Array Blocking Buffer (Whatman) containing 15 mg/ml reconstituted E. coli lysate (McLabs) and incubated at room temperature for 30 minutes with constant mixing. The arrays were rehydrated in blocking buffer for 30 min and probed with the preabsorbed sera overnight at 4uC with constant agitation. The slides were then washed five times in 10 mM Tris (hydroxymethyl) aminomethane buffer (pH 8.0) containing 0.05% (v/v) Tween-20 (TTBS), and incubated in biotin-conjugated, goat anti-cat immunoglobulin (anti-IgGfcc, Jackson Immuno Research) diluted 1/200 in blocking buffer. After washing the slides three times in TTBS, bound antibodies were detected by incubation with streptavidinconjugated SureLightH P-3 (Columbia Biosciences). The slides were then washed three times in TTBS and three times in Tris buffer without Tween-20 followed by a final water wash. The slides were air dried after brief centrifugation and analyzed using a Perkin Elmer ScanArray Express HT microarray scanner.

Immunostrip assay
Fifteen sequence-confirmed plasmids were expressed in fivehour IVTT reactions, according to the manufacturer's instructions. Proteins were printed on Optitran BA-S 85 0.45 mm Nitrocellulose membrane (Whatman) using a BioJet dispenser (BioDot) at 1 ml/cm, and cut into 3 mm strips. Individual strips were then blocked for 30 minutes in 10% non fat dry milk dissolved in TTBS. Prior to immunostrip probing, cat sera were diluted 1:250 in 10% nonfat dry milk solution containing 15 mg/ ml E. coli lysate, and incubated for 30 min with constant mixing at room temperature. Preabsorbed sera were then applied to each strip and incubated overnight at 4uC with gentle mixing. Strips were washed five times in TTBS, and then incubated for 11hour at room temperature in alkaline phosphatase conjugated goat anticat immunoglobulin (anti-IgG, Fcc fragment-specific, Jackson ImmunoResearch), that was diluted to 1:5000 in TTBS. The strips were then washed three times in TTBS, followed by another three washes in Tris buffer without Tween-20. Reactive bands were visualized by incubating with 1-step Nitro-Blue Tetrazolium Chloride/5-Bromo-4-Chloro-39-Indolyphosphate p-Toluidine Salt (NBT/BCIP) developing buffer (Thermo Fisher Scientific) for 2.5 minutes at room temperature. The enzymatic reaction was stopped by washing the strips with tap water. Strips were air dried and scanned at 2,400 dpi (Hewlett-Packard scanner). Images were converted to gray scale format by Photoshop and unaltered images are shown. Band intensities were quantified using ImageJ software [18] (found at http://rsbweb.nih.gov/ij/).

Data and statistical analysis
The protein microarrays used here do not meet the criteria for required deposition under MIAME guidelines [19], and alternatives to standardize protein microarray results are in development to insure that all information can be easily interpreted (description of minimum information about a proteomics experiment [MIAPE] can be found here [20]). Intensities were quantified using QuantArray software utilizing automatic background subtraction for each spot. Proteins were considered to be expressed if either tag's signal intensity was greater than the average signal intensity of the IVTT reaction without plasmid, plus 2.5-times the standard deviation. ''No DNA'' controls consisting of IVTT reactions without addition of plasmid were averaged and used to subtract background reactivity from the unmanipulated raw data. All results presented are expressed as signal intensity. As previously reported [21], the ''vsn'' package in the Bioconductor suite (http://Bioconductor.org/) in the R statistical environment (http://www.R-project.org) was used to calculate seroreactivity. In addition to the variance correction, this method calculates maximum likelihood shifting and scaling calibration parameters for different arrays, using known non-differentially expressed spots. This calibration has been shown to minimize experimental effects [22]. We used raw values for the positive and negative controls to calibrate, and then normalize, the entire data set using the vsn package. Differential analysis of the normalized signals was then performed using a Bayes-regularized t-test adapted from Cyber-T for protein arrays [23,24,25,26]. Benjamini-Hochberg p-value adjustments were applied to account for multiple test conditions [27]. All p-values shown are Benjamini-Hochberg corrected for false discovery, unless otherwise noted. Multiple antigen classifiers were built using Support Vector Machines (SVMs). The ''e1071'' and ''ROCR'' packages in R were utilized to train the SVMs and to produce receiver operating characteristic curves, respectively.

Results
Construction and probing of a Bartonella henselae protein microarray with sera from infected and uninfected cats A protein microarray comprised of 4032 spots from 1433 ORFs spotted in duplicate, with positive and negative controls was fabricated, as described in Methods. IVTT expression efficiency was determined by probing against the amino-terminal HIS and carboxy-terminal HA tags for each spot (Fig. 1a and 1b). 95.1% of the B. henselae proteins spotted onto the microarray had signals greater than the average of ''No-DNA'' control reactions plus 2.5 times the standard deviation, and were considered positively expressed. Sera from cats housed in two animal shelters were used to probe the B. henselae protein microarray in order to map the feline host anti-B. henselae antibody profile after naturally acquired infection. The B. henselae protein microarray was probed with a collection of 132 cat sera. Sixty-two cats identified as seropositive or seropositive and culture-positive were compared to 62 cats identified as seronegative and culture negative. The seronegative and culture negative cats could have been exposed earlier in their lives to Bartonella species and cleared the infection, with either lack of antibody titers or titers below the positive threshold (score of 2 or more at 1:64 dilution). Additionally, the array was probed with 8 Bartonella-free cat sera (from cats never exposed to Bartonella) to establish non-specific reactivity. In total, microarray data from 132 cat sera were used to generate a profile of host reactivity to B. henselae infection. Representative images of protein microarrays probed with sera from B. henselae-infected and naïve cats samples are shown in Figure 1c and 1d. Signal intensities of duplicate spots were recorded and averaged for each antigen and for each cat serum, individually. Antigens were considered seroreactive if the average signal intensity exceeded the average signal intensity of the IVTT reaction without plasmid (No-DNA controls) plus 2.5-times the standard deviation. As expected, duplicate spots were highly correlative with a total R squared of 0.99, similar to our previous published results from a Chlamydia trachomatis protein microarray [32].

Profile of humoral immune response to the Bartonella henselae proteome
The antibody response profile to B. henselae in its natural cat host is shown as a heatmap for both infected and uninfected cats henselae ORFs. (C and D) Comparison of arrays probed with naïve serum and positive serum. The arrays were read in a laser confocal scanner, analyzed, and the data normalized as described in Materials and Methods. The signal intensity for each antigen is represented by a rainbow palette of blue, green, red and white, corresponding to increasing signal. doi:10.1371/journal.pone.0011447.g001 (Fig. 2.). Of the total 1433 ORFs, 105 or 7.3% were found to be seroreactive. Cellular localization prediction of seroreactive ORFs revealed that 34 of 105 contained a signal peptide. Twenty-three were predicted to be localized to the cytoplasm, 21 to the cytoplasmic membrane, 9 to the outer membrane, 3 were periplasmic, and 49 were unable to be predicted by PSORTb. The mean reactivity for each protein on the array was compared between the infected and uninfected groups and plotted as a histogram (Fig. 3). Fifty-two antigens were differentially reactive (p,0.05), and 53 antigens were cross-reactive (p.0.05) between infected and uninfected cats. All of the sera reacted similarly to the cross-reactive antigens whether from infected or uninfected cats. Mean reactivity of seroreactive antigens was correlated to IFA score, and the Pearson's R was determined to be 0.70 for differentially reactive antigens and 0.17 for cross-reactive antigens, indicating a strong association with current IFA diagnostic assays and the identified differentially reactive antigens. A complete list of all seroreactive antigens is shown in Table S1. Twenty-four of the 62 seropositive cats were culture positive. The average seroreactivity from cats that were seropositive and culture positive was not significantly different from the average seroreactivity of cats that were seropositive and culture negative for either differentially reactive or cross-reactive antigens (Student's t-test p-value = 0.75 and 0.36, respectively).

Bartonella henselae type-specific immune response
Differential antibody profiles were investigated for type I and type II B. henselae (Fig. 4). Of the cat sera profiled by microarray, 10 were B. henselae type I and 14 were B. henselae type II. Thirteen antigens were statistically differentially reactive between these two groups for all 105 seroreactive antigens (Student's t-test p-value ,0.05). The other 92 were equally reactive, including the two most reactive antigens (BH13530 (MopA) and BH07870 (LemA)). B. henselae type I cats were most significantly differentially reactive to BH01250 (CyoD) and BH13260 (VirB2, p = 5.5610 23 and 5.0610 23 ). BH01250 is the 6 th most reactive antigen, and BH13260 is the 23 rd most reactive antigen of the infected cat group. BH12700 (VceA) is the 5 th most reactive antigen of the infected cat group and had a mean signal intensity of 25k. The mean reactivity for B. henselae types I and II was 29 k and 21 k, respectively, suggesting that the distribution of type I and type II infected cats was evenly represented among the 62 infected cats. The same was true for the second most type-specific differentially reactive antigen (BH01250). BH01250 had a mean signal intensity of 24 k, and B. henselae types I and II had signal intensities of 37 k and 18 k, respectively. One antigen (BH11090) was found to be significantly more reactive in the type II infected cat group than in the type I infected group (p-value = 1.7610 22 ).

Validation of seroreactivity with immunostrips
In order to validate the seroreactivity of the protein microarrays and to test the feasibility of using serodiagnostic antigens in an alternative analytical diagnostic assay, 15 antigens were printed onto nitrocellulose membranes, referred to as immunostrips. These 15 antigens were chosen for being highly significant and highly seroreactive by protein microarray. Individual immunostrips were probed with 30 infected and 28 uninfected cat sera chosen from the collection at random (Fig. 5). Bartonella-infected cat sera reacted strongly with the differentially reactive antigens, although the intensity pattern varied depending on the individual cat. Uninfected cat sera had low reactivity with these antigens, and produced a different pattern of reactivity than infected cats. Quantitative analysis of the immunostrips was used to directly compare seroreactivity of infected and uninfected cats on the immunostrips platform. The three most differentially reactive antigens in the protein microararay BH13530, BH07870, BH12700 (p-value = 7.0610 214 , 2.3610 213 , and 3.3610 28 , respectively) were also the most significantly different in the immunostrip, showing validation of the microarray and transference to a separate diagnostic platform (p-value = 1.39610 214 , Figure 2. Probing a collection of B. henselae infected, uninfected, and SPF control cat sera. Arrays containing 1433 B. henselae proteins were probed with cat sera organized into 3 groups as described in the text. Heatmap showing normalized intensity with red strongest, bright green weakest, and black in between. Cat samples are in columns and sorted left to right by increasing average intensity to differentially reactive antigens, and antigens are listed in rows sorted by decreasing average seroreactivity of infected cats. Only seroreactive antigens are displayed (n = 105). Seroreactivity is higher in the infected cats against differentially reactive antigens, but equally reactive in the cross-reactive antigen set. doi:10.1371/journal.pone.0011447.g002 1.73610 213 , and 4.0610 210 respectively). Furthermore, the immunostrips were probed with SPF naïve sera to confirm and evaluate background reactivity in uninfected cats. As expected, SPF naïve cats displayed minimal reactivity to all bands, similar to most of the uninfected group.

Serodiagnostic accuracy
In order to determine the diagnostic ability of the differentially reactive antigens, we used kernel methods and support vector machines [33,34] to build linear and nonlinear classifier from microarray and immunostrips data. As such, we utilized the discriminatory power of multiple ORFs in order to assess their ability to separate uninfected from infected sera. Serodiagnostic antigens were ranked according to p-value with the top 3 antigens having p-values less than 7.0610 214 (Table S1). We input the most significantly different antigens in sets of 1, 2, 3, 4, 5, 10, 25, and 52 antigens on the basis of p-value. The boxplots for these predictions were plotted and the results show that increasing the number of antigens from 1 to 2, and 2 to 3 produces an improvement in the classifier (Fig. 6). The classifier was able to accurately predict 93% of infected cats using 3 antigens. Increasing the diagnostic set from 3 to 4 antigens produced no increase in accuracy, and increasing past 4 antigens results in a reduction in accuracy due to over-fitting.
A subset of sera that was tested on the microarray platform was selected at random and demonstrated seroreactivity on the immunostrip platform. In order to determine the potential diagnostic utility of these differentially reactive antigens, quantified  immunostrip results were used to build a classifier and accuracy was determined, as shown in Figure 6. As in the microarray classifier, increasing the number of antigens produced a more accurate diagnostic classifier until a peak was reached followed by a reduction in accuracy due to over-fitting. We found that the most accurate diagnostic ability utilized 4 antigens and produced a diagnostic accuracy capable of identifying 98% of the infected cats. Differences in accuracy and the number of antigens that increase the diagnostic accuracy between immunostrips and microarrays are an expected result from utilizing different platforms for diagnostics.  Functional classification of the reactive antigens We next classified the serodiagnostic and cross-reactive antigens according to their annotated and computationally predicted features. The summary of this analysis can be found in Table 1. The NCBI database of Clusters of Orthologous Groups of proteins (COGs) was used for annotation. Each COG consists of individual proteins or groups of paralogs from at least 3 lineages, and is comprised of 25 categories of functional definitions. Each protein in the database is assigned to one or more COGs, with a total of 1569 COGs assigned to the 1433 ORFs on the array. The results in Table 1 illustrate that reactivity is not evenly distributed between various COGs and some are much more likely to contain seroreactive ORFs. For example COG V, containing proteins involved in defense mechanisms, was found to be enriched for seroreactive ORFs (4.20 fold enrichment, p-value = 2.9610 22 ). In contrast, COG E (amino acid transport and metabolism) contained no seroreactive proteins, despite the presence of 100 COG E ORFs in our array. Underrepresentation was also found in proteins involved in transcription (1/78 or 1.3%, p-value = 3.9610 22 ) and translation (2/138 or 1.4%, p-value = 2.9610 23 ). Analysis of individual COGs showed that no category was entirely reactive. For instance, the most predictive COG (COG U -intracellular trafficking and secretion) contained only 23.5% seroreactive proteins (16/68, 3.30 fold enrichment, pvalue = 1.0610 25 ). The next most predictive COG (COG M -cell wall/membrane biogenesis) contained 17.7% seroreactive proteins (16/90, 2.49 fold enrichment, p-value = 3.8610 24 ). Utilization of the COG database allowed characterization of functional group enrichment and under-representation; however, 292 B. henselae proteins were not defined by the COG database. Interestingly, this undefined category contained a significant number of seroreactive proteins (32/292, 1.54 fold enrichment, p-value = 7.7610 23 ) compared to ORFs that were grouped into various COGs (80/ 1277, 0.88 fold enrichment, p-value = 7.7610 23 ).
In addition to looking for COG enrichment, we also analyzed enrichment based on subcellular localization of proteins. Locali- zation was predicted from ORF sequence with pSORTb software. From our entire ORF collection, pSORTb predicted 28 outer membrane proteins, of which 9 were seroreactive (32%). This was the most enriching feature based on localization (4.39 fold enrichment, p-value = 9.2610 25 ). Conversely and as expected, we found pSORTb predicted cytoplasmic proteins were significantly underrepresented (0.63 fold enrichment, p-value = 2.9610 23 ). Interestingly, the pSORTb predicted periplasmic proteins, as well as proteins in COG categories C, M, and O, were significantly enriched in only differentially reactive antigens, but not in cross-reactive antigens. These categories may prove useful for in silico prediction of potentially protective antigens and diagnostics. Proteins predicted to have isoelectric points between 5 and 7 are significantly underrepresented in the seroreactive group (0.69 fold enrichment, p-value = 1.5610 22 ). This group is comprised of differentially reactive and cross-reactive proteins, in which only the cross-reactive proteins are significantly underrepresented (0.54 fold enrichment, p-value = 7.3610 23 ), and would be poor targets for in silico predictors of diagnostic antigens. We found that more acidic proteins are less likely to be seroreactive and are more likely to be located in the cytoplasm, which may explain their underrepresentation. Proteins predicted to have isoelectric points between 5-7 were also significantly underrepresented in seroreactive and cross-reactive groups. Some molecules can be expressed at high levels in vivo making them more likely targets for immune recognition, independent of their functional category. In order to determine the validity of this assumption, we looked at frequency distribution of codon adaptation index (CAI). Generally, higher CAI-values reflect potentially higher expression levels of an ORF. The CAI-values of B. henselae range from 0.27 to 0.71. While we found significance in ORFs that score in the middle on the CAI index range (from 0.4 to 0.5, fold enrichment 1.28, p-value = 3.4610 22 ), we were surprised to find that antigens with high CAI values (which tend to be over-expressed) were not significantly enriched ( Table 2). The bias for identifying highly expressed proteins discovered using 2D gels could be an inherent artifact not found using in vitro expression based platforms, like protein microarrays.

Discussion
Our results represent a large-scale analysis of B. henselae proteins that are immunogenic in the context of naturally acquired feline infection. We have constructed the first protein microarray for B. henselae, which allowed us to assess the humoral immune response to B. henselae from 62 naturally infected cats. The seroreactive antibody profile revealed many unknown seroreactive antigens, and confirmed previously identified ones. In this research, we sought to identify potentially protective and diagnostically relevant B. henselae antigens, and to understand the repertoire of the humoral immune response to infection in the natural feline host reservoir. In a genetically diverse population, the antibody repertoire is expected to vary among individual cats. Despite this diversity, the subset of differentially reactive antigens we identified provides a predictive accuracy rate of 93% for diagnosis of B. henselae exposed cats using the microarray (and 98% with the immunostrips), and it is likely that this set of antigens will form the basis of new and more accurate serodiagnostic assays for Bartonella exposure. Additionally, utilization of these antigens in alternative platforms (e.g., ELISA format) may provide an easy, universal, and rapid diagnostic assay. Importantly, some of the antigens we have discovered could be ideal candidates for generating subunit vaccines that protect cats from infection, thus limiting the morbidity and mortality of incidental human infections. All differentially reactive antigens, as well as the level of differential seroreactivity are not expected to correlate with protection, and empirical evaluation of selected, differentially reactive antigens will have to be determined. Of note, vaccination with facultatively intracellular pathogens, like B. henselae, could also depend on effective T-cell memory for mediating host defense and production of specific antibodies. Our study investigated the immune response of cats at a single time point, and we were not able to document previous infection of these cats by other Bartonella species or genotypes, including B. clarridgeiae or B. koehlerae, for which some cross-reactivity may occur [1]. Therefore, in some cats, the antibody profile could represent undetected coinfection or superinfection. In fact, protection of cats from infection with both B. henselae type I and type II would be ideal, because both types cause human zoonotic infections.
The most differentially reactive antigens on the immunostrips were able to accurately predict 98% of infected cats using 3 antigens (MopA, VceA, and LemA). Inclusion of the next highest ranked antigen, VirB8 (a type IV secretion system [T4SS] molecule), did not significantly improve the predictive accuracy. Antigens that are involved in T4SS are key factors in mediating Bartonella-host cell interactions, and five T4SS components are found in the serodiagnostic set (TrwE, TrwG, VirB2, VirB8, and VirB10). Interestingly, no T4SS molecules were found in the crossreactive antigen set. T4SS is comprised of a protein complex that is used to transport effector molecules directly into the host target cell. T4SS molecules are attractive targets for the development of new therapeutic agents, and investigation of monoclonal antibody therapy or small molecule inhibitors to block the secretion of virulence factors could provide protection or limit disease. The cross-reactive protein ParA is a plasmid partitioning protein, and the gene encoding this protein was previously identified as the only unique gene absent in B. quintana, although it is present in the genomes of both B. koehlerae and B. henselae [35]. Lindroos, et al., believe it is a pseudogene and not involved in host specificity, because the gene is shorter than normal in B. henselae and it is not flanked by a homolog to parB, as are most other parA genes. The role of this protein in association with the feline pathogen remains to be investigated, as well as its role in establishing acute and chronic infection. BH12700 (VceA, a multidrug resistance gene), and BH13260 (VirB2, a T4SS molecule), are both differentially reactive between genotypes I and II, and could facilitate an understanding of the differences between the two types. Experimental infection with B. henselae type I or type II in SPF cats provides complete protection against a challenge with the same genotype [3,36]. However B. henselae type I provides partial protection from a challenge with B. henselae type II. On the contrary cats infected with B. henselae type II were not protected when challenged with the heterologous strain, B. henselae type I [37]. The two most differentially reactive antigens are equally reactive in both types I and II, and subunit vaccines including these antigens should be ideal candidates for protection against both types I and II of B. henselae. Further investigation of the antibody profiles of protected and non-protected SPF cats could provide insight into individual antigens or profiles of antigens responsible for mediating cross-protection against both types, as our study was a cross-sectional study and some of the bacteremic cats may have had a previous infection with the other B. henselae genotype or even with other Bartonella species leading to production of cross-reactive antibodies [37].
We found that reactivity was not evenly distributed across the proteome and no individual category was completely seroreactive (Table 1). Cross-reactive and differentially reactive antigens are selectively enriched or underrepresented from specific functional categories. As expected, outer membrane proteins are more likely to be seroreactive (4.39 fold enrichment, p-value = 9.2610 25 ) and account for 8.6% of reactive antigens, but represent only 2.0% of the proteins on the array. As expected, antigens predicted to be cytoplasmic are significantly underrepresented (comprising only 4.6% of total seroreactive antigens, but constitute 34.8% of the proteins on the array). While these results classifying antigenicity by protein localization are consistent with expectations, they provide a quantitative and more informative understanding than previously reported. Importantly, there is no category that is entirely reactive, although there are some categories that are entirely unreactive. For example, there are 100 ORFs that are involved in amino acid transport and metabolism, but none of these antigens were seroreactive. Moreover, we found ORFs involved in translation and transcription to be significantly less likely to be seroreactive (p-value = 2.9610 23 and 3.9610 22 , respectively). Antigens predicted to contain at least one transmembrane domain were highly significantly enriched on the array (2.18 fold enrichment, p-value = 6.6610 211 ). Conversely, antigens that were predicted to not contain a transmembrane domain were significantly underrepresented (0.063 fold enrichment, p-value 6.6610 25 ). Identification of categories that are non-reactive is important for understanding immune evasion and pathogenicity, and is also valuable for developing exclusion criteria of in silico prediction algorithms. Similarly, cross-reactive antibodies may target identical proteins (or common epitopes) of both related and unrelated bacteria, and are thought to exist as a consequence of an indiscriminate or broad antibody response against numerous bacteria, whether symbiotic, environmental, or other. Moreover, cross-reactivity does not need to occur between closely related organisms, but can also occur between phylogenetically distant species. In this report, we identify numerous cross-reactive antigens in a cat population, which may be important for improved diagnostics and vaccine development. These results are consistent with our previously published results detailing the seroreactive profile of Burkholderia pseudomallei [21], another proteobacteria member. We again found significant enrichment of surface, chaperones, and PSORTb-predicted extracellular and outer membrane proteins. Moreover, inclusion of a predicted signal peptide or transmembrane domain were both enrichment features. Proteins predicted by PSORTb to be cytoplasmic or of unknown localization, and proteins predicted to not contain transmembrane domains were also significantly underrepresented in our previous profiling of seroreactive antigens from Burkholderia infected patients [21].
A comprehensive profile of the antibody repertoire to an infectious agent can provide new insight into disease pathogenesis. The data reported here provide specific insight into the antigens that are recognized by the humoral immune response after natural infection(s) in a naturally infected cat population. We found that the humoral immune response is not stochastic, and targets a diverse collection of antigens. The results presented here confirm that proteomic features can help predict seroreactive antigens, but these predictions are imperfect because the majority of proteins predicted are nonreactive. The empirically determined antibody repertoire against the comprehensive B. henselae proteome highlights the need for an improved understanding of seroreactivity determinants. Finally, our characterization of the feline antibody repertoire generated during B. henselae infection provides novel insight and utility in diagnostics, vaccine development, and in understanding the host-pathogen interaction of a zoonotic infectious agent.