In Situ-Synthesized Virulence and Marker Gene Biochip for Detection of Bacterial Pathogens in Water

Pathogen detection tools with high reliability are needed forvarious applications, including food and water safety and clinicaldiagnostics. In this study, we designed and validated an insitu-synthesized biochip for detection of 12 microbial pathogens,including a suite of pathogens relevant to water safety. Toenhance the reliability of presence/absence calls, probes weredesigned for multiple virulence and marker genes (VMGs) of eachpathogen, and each VMG was targeted by an average of 17 probes.Hybridization of the biochip with amplicon mixtures demonstratedthat 95% of the initially designed probes behaved as predictedin terms of positive/negative signals. The probes were furthervalidated using DNA obtained from three different types of watersamples and spiked with pathogen genomic DNA at decreasing relativeabundance. Excellent specificity for making presence/absencecalls was observed by using a cutoff of 0.5 for the positivefraction (i.e., the fraction of probes yielding a positive signalfor a given VMG). A split multiplex PCR design for simultaneousamplification of the VMGs resulted in a detection limit of between0.1 and 0.01% relative abundance, depending on the type of pathogenand the VMG. Thermodynamic analysis of the hybridization patternsobtained with DNA from the different water samples demonstratedthat probes with a hybridization Gibbs free energy of approximately–19.3 kcal/mol provided the best trade-off between sensitivityand specificity. The developed biochip may be used to detectthe described bacterial pathogens in water samples when paralleland specific detection is required.

INTRODUCTION

Top

INTRODUCTION

Pathogen detection and identification using virulence and markergenes (VMGs) as genetic targets are receiving considerable interestdue to the rapidly increasing availability of gene and whole-genomesequence information for most major pathogens. VMGs are expectedto be more specific than the 16S rRNA gene due to the limitedability of the 16S rRNA gene to differentiate among closelyrelated microorganisms. Information about the presence or absenceof selected VMGs also yields information about the pathogenicpotential of detected microorganisms. Screening for multipleVMGs for each pathogen is necessary to reduce the potentialof erroneous calls due to differences in the occurrence of variousVMGs among pathogenic strains belonging to the same species(59). Use of a single VMG or a limited number of VMGs may leadto an increased rate of false-negative calls if the selectedVMGs are unevenly distributed among different strains.

One of the first microarrays for the detection of microbialpathogens using multiple VMGs was described by Wilson et al.(66), who used the Affymetrix platform. This microarray wasvalidated for 11 bacterial, five viral, and two eukaryotic pathogensand employed genomic DNA (gDNA) of pathogens spiked in totalDNA extracted from filtered air. This study also highlightedthe advantage of using highly redundant probe sets to inferpresence/absence calls based on the positive fraction (PF).The PF was defined as the number of positive probes for a giventarget divided by the total number of probes used to detectthat target. Setting a sufficiently high threshold for the PFto define a target as present substantially reduces the rateof false-positive calls due to cross-hybridization. When redundantprobe sets are used, the reliability of the presence/absencecalls is expected to be somewhere between the reliability achievedby PCR-based detection of the VMGs and the reliability achievedby sequencing of the corresponding VMG amplicons. Utilizationof redundant probe sets may also result in more efficient validationof newly designed microarrays. This is because the success levelsfor probe design are generally high and presence/absence callsbased on redundant probe sets are not affected by failure ofa small number probes within a set.

Among the challenges in the development of pathogen detectionmicroarrays is the limited ability to detect low-abundance targetsequences within complex mixtures. Microarrays without targetgene amplification only allow detection of microorganisms ata relative abundance of approximately 1 to 5% (4, 11, 26, 50,67). Obviously, pathogens may be present at levels well belowthis limit in various matrices. Hence, microarray-based pathogendetection generally relies upon target gene amplification strategies,typically using PCR (6, 7, 35). For conserved genes (e.g., the16S rRNA gene), amplification using universal or group-specificprimer sets in multitemplate PCR has been extensively used toenrich target genes prior to hybridization (12, 13, 16, 17,24, 36-38, 43, 44, 51, 52, 55, 63, 65). Methods for simultaneousamplification of VMGs, however, are less well developed. Frequently,multiple PCRs or multiplex PCR has been used to amplify multipleVMGs to enhance probe signals and to improve detection limitsfor genotyping of isolates and pathogen detection in variousmatrices (1, 2, 8, 10, 20, 25, 31, 32, 45, 54, 61, 62, 66).However, development of robust and sensitive multiplex PCR assaysfor multiple VMGs still requires careful primer design and significantoptimization of the reaction parameters (41). If a large numberof VMGs could be simultaneously amplified in a robust mannerin complex matrices, VMG-based microarrays could provide therequired specificity, sensitivity, and target throughput, allin the same assay, for diagnostic purposes.

The sensitivity of microarrays is also influenced by the strengthof probe signals obtained after hybridization. Probes with hightarget binding affinities may be preferred for detecting low-abundancetargets. However, such probes are also more prone to cross-hybridization(23) and may decrease specificity. This trade-off between specificityand sensitivity needs to be assessed experimentally to defineoptimal probe selection criteria.

The main objective of this study was to design and validatean in situ-synthesized VMG biochip to detect 12 bacterial pathogens.Many of the pathogens included (Aeromonas hydrophila, Helicobacterpylori, Legionella pneumophila, Pseudomonas aeruginosa, Vibriocholerae, Vibrio parahaemolyticus, and Yersinia enterocolitica)were waterborne. Other pathogens (Clostridium perfringens, Salmonella,Staphylococcus aureus, Campylobacter jejuni, and Listeria monocytogenes)were relevant to clinical diagnostics and food safety. gDNAextracted from tap water, river water, and tertiary effluentfrom a municipal wastewater treatment plant served as the matrixused to assess sensitivity and specificity after samples werespiked with pathogen gDNA. Target gene amplification using asplit multiplex PCR assay allowed detection of pathogens ata relative abundance between 0.1 and 0.01%, depending on thepathogen and the VMG. Up to six VMG amplicons per pathogen andup to 35 probes per VMG amplicon were used to eliminate false-positivecalls. The effect of characteristics of probes on their hybridizationbehavior was also evaluated in order to derive probe designrules yielding the best trade-off between sensitivity and specificity.The described VMG biochip may have applications in diagnosticareas where parallel screening of multiple pathogens with highlevels of specificity is critical.

MATERIALS AND METHODS

Top

MATERIALS AND METHODS

Probe design and biochip synthesis.
Two sets of probes were included on the VMG biochip (see thesupplemental material for probe sequences). The first set ofprobes (n = 791) was designed to detect 35 VMGs covering 12pathogens. A total of 47 PCR primer sets were designed for thesegenes; each amplicon was targeted by 7 to 35 probes, and theaverage was 17 probes per amplicon (Table 1). The second setof probes (n = 2,034) targeted 67 VMGs for the 12 pathogensmentioned above, as well as VMGs for five additional pathogens.For these genes, no primers were designed, and the probes wereanalyzed as nontargeted probes to assess hybridization specificity.All VMGs selected in this study were used previously to detectthe selected pathogens in PCR-based assays. All probes weredesigned based on 671 sequences retrieved from the GenBank databaseand satisfied the following criteria: (i) the probes were complementaryto all retrieved sequences of a given gene with species-levelspecificity, and (ii) the probes contained at least two mismatchesto all other nontarget sequences within the database. The Gibbsfree energy of probe-target duplex formation assuming complementarytargets ( ${Delta}$ G°_duplex) ranged from –14.1 to –23.9kcal/mol; the mean was –18.4 kcal/mol, and the variability(standard deviation) was 2.1 kcal/mol.

View this table:
[in this window]
[in a new window]

TABLE 1. Overview of the validation results^a

The designed probes were synthesized in situ on microfluidicbiochips using a proprietary light-directed synthesis technologydeveloped by the University of Michigan (18, 19) and commercializedby Xeotron (Houston, TX; now part of Invitrogen, Carlsbad, CA).The microfluidic chip contained 8,000 microreactors with a diameterof 50 µm, which were interconnected with flow channels(Fig. 1a). The in situ synthesis technology employed conventionalphosphoramidite chemistry in conjunction with photogeneratedacid-triggered deprotection of the 5'-hydroxyl groups of nucleotidephosphoramidite monomers. Spatially resolved light patternsfor light-directed acid deprotection were generated using adigital micromirror device. The probes were attached to thesubstrate via a spacer consisting of Ts and C₁₈ with an effectivelength of 12 nucleotides. The biochip also contained 22 randomlyspaced control spots containing solely linker chemistry to assessbackground signal intensity (56, 64).

View larger version (20K):
[in this window]
[in a new window]

FIG. 1. Evaluation of 791 targeted and 2,034 nontargeted probes with a composite target mixture containing all 47 VMG amplicons (24 amplicons labeled with Cy3 and 23 amplicons labeled with Cy5). (a) Part of the Xeotron chip after hybridization and washing up to a temperature of 30°C. (b) Distribution of the SNRs for 791 targeted probes (filled bars) and 2,034 nontargeted probes (open bars). (c) Distribution of the PFs for 47 targeted probe sets (filled bars) and 67 nontargeted probe sets (open bars). (d) Success of probe selection for targeted and nontargeted probe sets (• and ${circ}$ , respectively). The indicated slopes of the dashed lines are numerically identical to the PFs.

Nucleic acid extraction.
Twelve reference bacteria were used to validate the VMG biochip.Cultures of A. hydrophila ATCC 7966, C. perfringens ATTC 12916,Salmonella sp. strain ATCC 13311, L. monocytogenes ATCC 15313,P. aeruginosa ATCC 10145, V. parahaemolyticus ATCC 43996, andY. enterocolitica ATCC 55075 were obtained from the AmericanType Culture Collection (ATCC) (Manassas, VA). Cells were grownovernight according to the instructions provided by ATCC forthe organisms. gDNA was extracted from 1 to 2 ml of the culturesusing a DNeasy tissue kit (Qiagen, Valencia, CA) according tothe manufacturer's instructions. For V. cholerae ATCC 39315,C. jejuni ATCC 700819, H. pylori ATCC 700392, S. aureus ATCC700699, and L. pneumophila ATCC 33152, purified gDNA was obtainedfrom the ATCC.

Tap water (40 liters), river water (10 liters; Red Cedar River,East Lansing, MI), and tertiary effluent (20 liters; wastewatertreatment plant in East Lansing, MI) were sampled and filteredthrough 0.45-µm nitrocellulose filters (Millipore, Billerica,MA). gDNA was extracted from the filters using a MegaPrep UltraCleansoil DNA kit (Mo Bio Laboratories, Carlsbad, CA) according tothe manufacturer's instructions. The amount and quality of theextracted DNA were determined with a NanoDrop ND-1000 spectrophotometer(NanoDrop Technologies, Wilmington, DE).

Primer design and PCR conditions.
A total of 47 gene-specific primer sets (see Table S1 in thesupplemental material) were designed for 35 VMGs flanking generegions with high probe density. Primers were selected so thatthey had similar annealing temperatures (one set had an annealingtemperature of 53°C, and another set had an annealing temperatureof 58°C) and covered most alleles of a given VMG availablein the GenBank database. The uniqueness of the primers was confirmedby a BLAST search against the GenBank database. To ensure primerspecificity, mismatches with related nontarget sequences werelocated near the 3' ends of the primers. For multiplex PCR,primers were segregated into five primer combinations, and eachcombination contained 9 or 10 primer pairs (see Table S2 inthe supplemental material). Mixtures of primer sets were selectedso that each pathogen (except P. aeruginosa) was targeted inat least two different multiplex PCRs. Primers were synthesizedby Integrated DNA Technologies (Coralville, IA).

PCR mixtures (25 µl) consisted of 1x PCR buffer, 2 mMMgCl₂, 1.5 U (for monoplex PCR) or 3 U (for multiplex PCR) ofAmpliTaq Gold (Roche Molecular Systems, Pleasanton, CA), eachdeoxynucleoside triphosphate at a concentration of 200 µM(Invitrogen), each primer at a concentration of 500 nM, 200ng of bovine serum albumin (New England BioLabs, Beverly, MA),and 1 µl of DNA. After the initial enzyme activation at94°C for 10 min, 35 cycles of the following temperatureregimen was used for amplification: denaturation at 94°Cfor 60 s, annealing at 53 or 58°C for 60 s, and elongationat 72°C for 60 s. This was followed by a final elongationstep at 72°C for 7 min. PCR mixtures were purified usinga QIAquick PCR purification kit (Qiagen), and the amplificationproducts were quantified using the NanoDrop ND-1000 spectrophotometer(NanoDrop Technologies).

Fluorescent labeling of PCR products.
PCR products pooled from different PCRs were labeled using anaminoallyl dUTP (aa-dUTP) incorporation and cyanine dye couplingprotocol as described previously (56, 64), with minor modifications.Briefly, PCR products were amplified, and aa-dUTP was incorporatedinto the amplification products with a Bioprime DNA labelingkit (Invitrogen, San Diego, CA) using a ratio of aa-dUTP todTTP of 5:1 and an incubation time of 120 min. Purified productswere then coupled with cyanine dye by incubation for 80 minin a 1:1 mixture of 0.1 M sodium carbonate buffer (pH 9.3) andN-hydroxysuccinimide ester cyanine dye (Amersham Biosciences).

Biochip hybridization and melting curve development.
Hybridization and melting profile generation were performedusing previously established protocols (56, 64). The methodused for melting profile analysis was based on methods usedin previous studies utilizing this approach for microarraysusing 16S rRNA targets (15, 29, 34, 60). Briefly, after primingof the biochips, target DNA (200 pmol of Cy dye) was hybridizedovernight at 20°C using an M-2 hybridization station (Invitrogen[formerly Xeotron Corporation, Houston, TX]) with hybridizationbuffer containing 35% deionized formamide (Ambion), 6x SSPE(pH 6.6; Invitrogen), and 0.4% Triton X-100 (Sigma) (1x SSPEis 0.18 M NaCl, 10 mM NaH₂PO₄, and 1 mM EDTA [pH 7.7]). Afterthe chips were washed, the initial point of the melting curveprofile was obtained by washing each chip with high-stringencywash buffer (20 mM NaCl, 10 mM Na₂PO₄, 5 mM Na₂EDTA; pH adjustedto 6.6 with HCl) for 1.4 min at 25°C and imaging the chip.High-stringency wash buffer was degassed under a vacuum to preventformation of air bubbles during the wash steps. Subsequent developmentof the melting profile was performed by using manual cyclesof washing and scanning of the chip that were repeated at 1°Cintervals until 60°C was reached. At the end of this procedure,the chip was additionally stripped at 60, 45 and 30°C (2.5min each) with nuclease-free water (Sigma). The tubing of thehybridization station was washed for 20 min before and aftereach experiment to prevent carryover between experiments.

Experimental design.
Specificity and sensitivity were evaluated with samples consistingof pathogen gDNA spiked into DNA obtained from different watersources (tap water, river water, and tertiary effluent froma wastewater treatment plant), as has been done in numerousother validation studies (43, 66). A first set of samples wasprepared by spiking 10 pg of gDNA of each pathogen (equivalentto approximately 1,400 to 5,500 genome copies depending on thepathogen) into 10 ng of DNA from the different water samples.This yielded a relative abundance of 0.1% for each of the pathogens,calculated as a mass-based percentage of pathogen gDNA in thetotal community gDNA. Both spiked and unspiked samples weresubjected to the split multiplex PCR amplification step. Theamplicons obtained from spiked water samples were labeled withCy3, and the amplicons obtained from unspiked water sampleswere labeled with Cy5. Equal amounts of the two labeled products(100 pmol of each dye or $~$ 2.5 µg of DNA) were mixed andhybridized in duplicate. For the unspiked water samples, 2,034nontargeted probes were evaluated to examine probe specificity.For the spiked samples, 673 probes were evaluated to examineprobe sensitivity. A second set of samples contained pathogengDNA spiked at a relative abundance of 0.01% into gDNA fromriver water and at a relative abundance of 0.001% in gDNA fromtertiary effluent.

Data acquisition and processing.
Microarrays were scanned with a GenePix 4000B 16-bit laser scanner(Axon instruments, Union City, CA). Fluorescence signal intensitieswere extracted from the scanned images using GenePix5.0 (AxonInstruments, Union City, CA), which yielded values between 0and 65,535 arbitrary units. The median of all pixel intensitieswithin a spot was used as the raw spot intensity. Subsequentdata analysis was done with Microsoft Excel (Microsoft, Redmond,WA), and plotting was done with SigmaPlot 9.0 (Systat Software,Point Richmond, CA).

Raw spot intensities were divided by the mean signal intensityof 22 empty control spots to obtain signal-to-noise ratios (SNRs)(56). The SNR was computed for each wash step between 30 and45°C, and the values were subsequently averaged. Averagingof SNRs is equivalent to calculating an SNR based on the areaunder the melting curve within this temperature interval. TheSNR was subsequently divided by the median SNR of the 2,034nontargeted probes. A probe signal was considered positive ifthe SNR was greater than 3. The PF was computed by dividingthe number of positive signal probes within a set by the numberof probes within the set (66).

Gibbs free energy estimation.
The ${Delta}$ G°_duplex was calculated with the DINAMelt web server(14, 40, 53). This parameter reflects the energy released duringprobe-target duplex formation and is sequence dependent (53).It provides a thermodynamic measure of the affinity betweena probe and a target, with more negative ${Delta}$ G°_duplex valuesbeing indicative of higher binding affinities. For all calculations,perfectly matching duplexes were assumed, and effects of targetdangling ends were neglected. The temperature used for the calculationswas 43°C instead of the actual hybridization temperature(20°C) to account for the presence of 35% formamide in thehybridization solution (3, 60), and the Na⁺ concentration usedwas 1 M. A linear relationship between the G+C content of aprobe and its ${Delta}$ G°_duplex was observed. This relationship wasused to convert an estimated ${Delta}$ G°_duplex into a correspondingG+C content.

RESULTS AND DISCUSSION

Top

RESULTS AND DISCUSSION

Probe behavior with amplicon mixtures generated by monoplex PCR.
The performance of the VMG biochip was evaluated by hybridizationof a target mixture containing all 47 VMG amplicons obtainedin individual PCR amplifications. Of the 47 amplicons, 24 werelabeled with Cy3 and 23 were labeled with Cy5. For the Cy3-labeledamplicons, 337 of 393 (85.5%) targeted probes displayed positivesignals in all three replicates. Furthermore, only 10 of the393 (2.5%) probes displayed positive signals in at least oneof the replicates with the Cy5-labeled amplicons (i.e., nontargetedamplicons), with a median SNR of 3.4. For the Cy5-labeled amplicons,383 of 398 (96.2%) targeted probes yielded positive signalsin all three replicates. Furthermore, only 18 of 398 (4.5%)probes displayed positive signals in at least one of the replicateswith the Cy3-labeled amplicons (i.e., nontargeted amplicons),with a median SNR of 3.5. Of the 2,034 nontargeted probes, 27(1.3%) and 34 (1.7%) displayed positive signals with the Cy3-and Cy5-labeled amplicons, respectively, in at least one ofthe replicates.

Overall, of the 791 targeted probes, 720 (91.0%) produced positivesignals with SNRs up to 1,000. Of the 2,034 nontargeted probes,61 (3.0%) yielded positive signals with either the Cy3- or Cy5-labeledamplicons. The signal intensities for targeted and nontargetedprobes were well separated, except for a small fraction of theprobes (Fig. 1b), indicating the good discriminatory power ofthe probes. Altogether, 95% of the 5,650 individual probe-targetinteractions (twice the total number of targeted and nontargetedprobes) were the expected interactions in terms of positive/negativeprobe signals.

The PF was calculated for each probe set (and correspondingamplicon) by dividing the number of positive probes by the numberof probes within the set (66). A PF based on the number of initiallydesigned probes is equivalent to the success level of probeselection for individual probe sets. Of the 47 targeted probesets, 22 displayed a PF of 1, implying that 100% of the designedprobes yielded positive signals. Of the remaining 25 sets, 18displayed a PF between 0.8 to 1 and 7 displayed a PF between0.6 to 0.8 (Fig. 1c). Of the 67 nontargeted probe sets, 34 displayeda PF of 0 and 31 displayed a PF between 0 and 0.1. The two remainingsets had PFs of 0.125 and 0.3. The success level of probe selectioncould also be illustrated by plotting the number of designedprobes versus the number of positive probes (Fig. 1d). As Fig.1d shows, the PFs for both targeted and nontargeted probes werenot related to the number of designed probes.

Substantial variation in signal intensity was observed amongprobes for a given VMG amplicon (Fig. 2b). Extensive unevennessin signal intensity among probes for a given target is welldocumented (5, 44, 46, 56). This variability may be attributedto various factors, including probe and target secondary structure(9, 27, 47), target length (33), ${Delta}$ G°_duplex (23), and eventhe position of fluorescent labels (68). Also, sequence dissimilaritiesbetween the probes and hybridized targets may have contributedto this variability. When probe sets were sorted according totheir median SNRs, a trend toward increasing PF with a highermedian SNR was apparent (Fig. 2a). Interestingly, the G+C contentof the VMG amplicons displayed an analogous tendency (Fig. 2c).This may also partially explain the overall lower success ratefor the probe sets hybridized with the Cy3-labeled amplicons(85.5%) than for the probe sets targeted by the Cy5-labeledamplicons (96.2%). The G+C content for Cy3-labeled amplicons(36.5% ± 5.4%) was, on average, lower than the G+C contentfor the Cy5-labeled amplicons (48.6% ± 7.9%). These trendsimply that probe design for genes or genomes (and hence detectionof the corresponding microorganisms) with a considerably lowerG+C content may be more challenging when continuous regionswith a higher G+C content are not present in the selected genetargets.

View larger version (24K):
[in this window]
[in a new window]

FIG. 2. PFs (a), SNRs (b), and G+C contents (c) for all 47 targeted VMG amplicons. The probe sets are sorted from bottom to top according to increasing median SNR. The boundaries of the boxes in panel b indicate the 25th and the 75th percentiles, and the whiskers indicate the 10th and the 90th percentiles. The median is given as a solid line, and outlying data points are shown as open symbols.

Success rate of probe design as a function of ${Delta}$ G°_duplex.
In order to optimize the probe selection criteria, whether probeswith positive or negative signals could be identified in silicobased on evaluation of their theoretical (thermodynamic) propertieswas evaluated. Based on previous studies, it was anticipatedthat probe-target binding affinity (quantified using ${Delta}$ G°_duplex)would be the most valuable parameter for this screening (23,39, 42, 49). Only a weak linear relationship (r² = 0.34) wasobserved between ${Delta}$ G°_duplex and the natural logarithm of signalintensity (see Fig. S2 in the supplemental material). This variabilityin signal intensity for probes with comparable ${Delta}$ G°_duplexvalues precluded evaluation of probe design criteria in termsof signal intensity based on ${Delta}$ G°_duplex. To determine therelationship between ${Delta}$ G°_duplex and probe behavior, hybridizationpatterns were interpreted in terms of positive/negative probesignals, irrespective of their strength. For this analysis,all 791 targeted probes were sorted according to their ${Delta}$ G°_duplexvalues and binned, and the percentage of positive probes ineach bin was calculated. The latter value is equivalent to thesuccess level for probes with a ${Delta}$ G°_duplex within the binrange. A drastic decrease in the probe design success rate wasobserved for probes with a ${Delta}$ G°_duplex value less negativethan –17 kcal/mol (Fig. 3) and a G+C content less than34.4% (as estimated using the linear relationship between G+Ccontent and ${Delta}$ G°_duplex). This trend was independent of thewindow size used for binning of the probes (see Fig. S3 in thesupplemental material). A similar analysis with increasing SNRthresholds revealed analogous trends, but the results were shiftedtoward more negative ${Delta}$ G°_duplex values (data not shown). Thefactors contributing to the unevenness in probe signals fora given target may also explain the lack of hybridization signalsfor probes with a highly negative ${Delta}$ G°_duplex. In accordancewith our observations, Reyes-Lopez et al. (49) previously reportedthat the proportion of predicted signals observed experimentallyincreased for 9-mer probes with increasingly negative ${Delta}$ G°_duplexvalues.

View larger version (17K):
[in this window]
[in a new window]

FIG. 3. Success of probe design as a function of ${Delta}$ G°_duplex. Each symbol indicates the percentage of positive probes for bins of 40 probes, and the error bars indicate the standard deviations for three replicates. The percentage of positive probes is plotted at the median ${Delta}$ G°_duplex of each bin. The G+C content was derived from the linear relationship between ${Delta}$ G°_duplex and G+C content.

Based on our analysis of the hybridization behavior of all targetedprobes, a ${Delta}$ G°_duplex threshold of –17 kcal/mol couldbe used as a valuable design rule for future oligonucleotideprobe design. Although this criterion was demonstrated onlyfor 18-mers in this study, a similar approach to assess theeffect of ${Delta}$ G°_duplex on the success level of probe designcould be adopted for longer probes. The proposed ${Delta}$ G°_duplexdesign threshold is in accordance with the findings of Loy etal. (38), who proposed a value of –16 kcal/mol for 18-merstargeting the 16S rRNA gene. This value was suggested basedon the observation that the majority of probes displaying positivesignals (including cross-hybridization signals) were attributedto probe-target duplexes with a ${Delta}$ G°_duplex more negative than–16 kcal/mol. It should be noted that small deviationsfrom these ${Delta}$ G°_duplex design thresholds may be observed forhybridizations performed under different experimental conditionsand/or for theoretical ${Delta}$ G°_duplex estimates obtained usingother nearest-neighbor model parameters (38). Finally, the ${Delta}$ G°_duplexcriterion was derived from hybridization patterns with high-abundanceand low-complexity target mixtures. For target mixtures havingvery different compositions in terms of target sequence abundanceand diversity, these rules should be applied with caution.

Multiplex amplification of VMG targets.
A split multiplex PCR assay was developed to amplify all VMGamplicons (see Table S2 in the supplemental material). A gelimage of the amplification products obtained with a mixtureof gDNA from all 12 pathogens is shown in Fig. S1 in the supplementalmaterial. Of the 720 probes displaying positive signals afterhybridization of monoplex PCR amplicons, 673 (93.5%) yieldedpositive signals after hybridization of multiplex PCR amplicons(Table 1). For 28 of the 47 probe sets, all probes yielded positivesignals. In general, the probes that displayed negative signalswith the multiplex PCR amplicons also yielded low signals withthe monoplex PCR amplicons (data not shown). For the ctxB geneof V. cholerae, only 2 of 20 probes yielded positive hybridizationsignals after hybridization of amplicons generated by multiplexPCR. This was attributed to poor amplification in the multiplexPCR rather than to low probe hybridization efficiency. Withmonoplex PCR, an amplification product of the expected lengthwas observed, verifying that the designed primers successfullyamplified the ctxB gene of V. cholerae ATCC 39315. In addition,all probes targeting the ctxB gene yielded positive signalsafter hybridization of the ctxB gene amplicon generated by monoplexPCR. The 27 targeted probes (3.9%, excluding probes targetingthe ctxB gene) displaying negative signals were distributedamong multiple VMG amplicons, and hence redesign or furtheroptimization of the multiplex PCR assays was expected to beineffective. In addition, the low levels of the signals of theseprobes with the monoplex PCR amplicons suggested that theirsensitivity may be limited; therefore, these probes were maskedduring further analysis.

Performance of the VMG biochip with spiked water samples.
The performance of the VMG biochip probe sets was further testedwith samples containing pathogen gDNA spiked into gDNA extractedfrom three different water samples: tap water, tertiary effluentfrom a wastewater treatment plant, and river water. For pathogensspiked at a relative abundance level of 0.1%, the median PFsfor 46 targeted VMG amplicons were 0.79, 0.87, and 0.75 fortap water, tertiary effluent, and river water, respectively(Fig. 4). The median PFs for the 67 other probe sets were 0.05for tap water, 0.09 for tertiary effluent, and 0.17 for riverwater (Fig. 4). By applying a PF threshold of 0.5, the numbersof spiked VMG amplicons assigned as present were 39 for tapwater (85%), 40 for tertiary effluent (87%), and 39 for riverwater (85%). For the nontargeted probe sets, the PF was alwaysless than 0.5, and consequently none of the corresponding geneswere identified as present in the water samples. Probe setstargeting S. aureus VMG amplicons indicated the presence ofthis pathogen in all three spiked water samples, although theVMG amplicons assigned as present were in disagreement amongsamples. The VMG amplicons of C. perfringens could not be detectedin any of the spiked water samples.

View larger version (14K):
[in this window]
[in a new window]

FIG. 4. Performance of the VMG biochip with gDNA extracted from various water samples. DNA (10 pg) of each pathogen was spiked into 10 ng of background gDNA, yielding a relative abundance for each pathogen of 0.1%. For spiked samples, 673 probes (46 VMG amplicons) were analyzed. For unspiked samples, 2,034 probes (67 VMG amplicons) were analyzed. The box plots indicate the distribution of the PF among the VMG amplicons. The dashed line at a PF of 0.5 indicates the selected threshold for presence/absence calls. The boundaries of the boxes indicate the 25th and the 75th percentiles, and the whiskers indicate the 10th and the 90th percentiles. The median is given as a solid line, the mean is shown as a dashed line, and outlying data points are shown as open symbols.

When pathogen gDNA was spiked at a relative abundance of 0.01%into gDNA extracted from river water, the presence/absence callsfor the VMG amplicons varied for a given pathogen (Table 1).A similar observation was made previously by Wilson et al. (66)and was explained by the variable sensitivity among probes (17).These researchers observed that for pathogens spiked at a fractionalabundance of 0.025% (500 fg pathogen gDNA in 2 ng gDNA fromair) only five of eight probe sets for Francisella tularensisand three of eight probe sets for Y. pestis were assigned aspositive. Conceptually, presence/absence calls for a given pathogencould be based upon indices reflecting the PFs of all probesets. Requiring that all probe sets for a given pathogen needto be scored as positive to assign the pathogen as present woulddrastically reduce false-positive calls. However, such an approachmight lead to increased rates of false-negative calls due theunequal distribution of multiple VMGs in clinical and environmentalisolates of most pathogenic species. For pathogen gDNA spikedinto tertiary effluent gDNA at a relative abundance level of0.001% none of the targeted VMG amplicons were assigned as present(data not shown). Hence, the detection limit of the VMG biochipwas between a relative abundance of 0.1% and a relative abundanceof 0.01%, depending on the pathogen and VMG. This detectionlimit is within the range of sensitivity previously reportedfor microbial diagnostic arrays targeting either the 16S rRNAgene (43, 44, 52) or multiple VMGs (66). However, advances thatenhance the sensitivity are required to achieve reliable detectionat abundance levels approaching the minimum infectious dosein environmental samples. These advances include improved experimentaltechniques for target gene enrichment and/or microarray signalenhancement and optimized data analysis algorithms.

The PF threshold applied in this study was less stringent thanthe cutoff values (at least 0.8) used by previous researchers(13, 66). Both target gene sequence diversity and abundanceaffect the selection of an optimal threshold for the PF. Uncharacterizedsequence diversity within the target gene may lead to reducedPFs due to the increased potential for mismatching probe-targetduplexes. Low target gene abundance may also yield decreasedPFs due to variability in probe sensitivity. In both cases,a lower threshold PF needs to be used to reduce false-negativecalls. Although the concept of redundant probe sets for enhancedreliability in presence/absence calls has been exploited widely(1, 12, 31, 32, 37, 38, 54, 61, 66), the effect of target geneabundance on PFs is less well documented. As demonstrated inthis study and by Wilson et al. (66), lower PFs are expectedwith decreasing target gene amounts, and the extent of the decreasemay vary among probe sets.

The use of mismatch probes is a common strategy to identifycross-hybridization signals on microarrays and to enhance thereliability of presence/absence calls (65, 66). In this study,excellent specificity was observed without inclusion of mismatchprobes, even for environmental samples. A combination of variousfactors, including the use of redundant probe sets (with anaverage of 15 probes per VMG amplicon after experimental screening),the use of short probes with increased discriminatory powercompared to long probes, reduction of sample complexity by targetgene enrichment using multiplex PCR, and use of the meltingcurve approach, contributed to the elimination of false-positivecalls.

Probe selectivity as a function of ${Delta}$ G°_duplex.
For selective detection, probes should yield high hybridizationsignals with intended target sequences and low signals withnontarget sequences. However, due to the biochemical natureof the probe-target hybridization process, these two criteriacannot be optimized independently. This translates into a trade-offbetween probe specificity and sensitivity that must be consideredduring probe selection. Based on our analysis shown in Fig.3, ${Delta}$ G°_duplex was selected as the probe design parameter quantifyingthe trade-off between probe specificity and sensitivity. Probesensitivity was derived from the hybridization patterns of thetargeted probes, while probe specificity was quantified basedon the hybridization patterns of the nontargeted probes. Thehybridization results with the unspiked tap water sample wereomitted from this analysis due the minimal amount of nontargetedprobes with positive signals (Fig. 4). The hybridization resultsfor all three water samples spiked with pathogen gDNA at a relativeabundance of 0.1% were included in the analysis.

Both targeted and nontargeted probes were sorted according totheir ${Delta}$ G°_duplex values and binned, and the percentage ofpositive probes was computed for each bin. After the percentageof positive probes was plotted as a function of ${Delta}$ G°_duplex,two distinct regions were apparent (Fig. 5b), and this effectwas independent of the window size used for probe binning (seeFig. S4 in the supplemental material). In region 1 ( ${Delta}$ G°_duplexmore negative than –19.3 kcal/mol), 93.8% ± 2.8%of the targeted probes displayed positive signals, independentof the ${Delta}$ G°_duplex. The percentage of nontargeted probes withpositive signals was influenced more by ${Delta}$ G°_duplex and increasedfrom 14.7% for probes with a ${Delta}$ G°_duplex of –19.9 kcal/molto 42.7% for probes with a ${Delta}$ G°_duplex of –23.2 kcal/mol(average increase, 6.8% per ${Delta}$ G°_duplex). In region 2 ( ${Delta}$ G°_duplexless negative than –19.3 kcal/mol), a rapid decrease inthe percentage of positive targeted probes was observed, andonly 31.3% of the probes yielded positive signals for probeswith a ${Delta}$ G°_duplex of –15.6 kcal/mol (average decrease,–16.1% per ${Delta}$ G°_duplex). The percentage of nontargetedprobes yielding a positive signal was significantly lower inthis region (average, 9.2% ± 3.9%).

View larger version (18K):
[in this window]
[in a new window]

FIG. 5. Derivation of optimal probe design criteria. (a) Selectivity, expressed as the difference in the percentage of positive probes for targeted and nontargeted probes, as a function of ${Delta}$ G°_duplex. (b) Percentage of positive probes as a function of ${Delta}$ G°_duplex for targeted and nontargeted probes (• and ${blacklozenge}$ , respectively). Each symbol indicates the percentage of positive probes for bins of 40 targeted and 80 nontargeted probes, and the error bars indicate the standard deviations for samples and replicates (n = 6 for targeted probes and n = 4 for nontargeted probes). The G+C content was derived from the linear relationship between ${Delta}$ G°_duplex and G+C content.

As evident from Fig. 5b, an increase in the percentage of positivetargeted probes was always associated with an increase in thepercentage of positive nontargeted probes. Probe selectivity,reflecting the ability of a probe to detect intended targetsequences and exclude nontarget sequences, quantifies this trade-offand is the parameter that needs to be maximized during probeselection. Probe selectivity, calculated by determining thedifference in the percentages of positive probes for targetedand nontargeted probes as a function of ${Delta}$ G°_duplex, was estimatedbased on the linear regression lines in Fig. 5b. The highestselectivity ( $~$ 80%) was observed for probes with a ${Delta}$ G°_duplexof –19.3 kcal/mol and a G+C content of 47.2% (Fig. 5a).Thus, probes with a ${Delta}$ G°_duplex of –19.3 kcal/mol providethe best trade-off between sensitivity and specificity. Probeswith a ${Delta}$ G°_duplex deviating from this optimum displayed lowerselectivity. The selectivity was more than 70% for probes witha ${Delta}$ G°_duplex between –18.6 and –21.1 kcal/mol,which corresponds to a G+C content between 42.3 and 56.1% (Fig.5a). Interestingly, the decrease in selectivity in region 2was approximately twofold greater than that in region 1 (–6.6%per ${Delta}$ G°_duplex for region 1 and –14.0% per ${Delta}$ G°_duplexfor region 2). This difference was attributed to the largerdecrease in the percentage of positive targeted probes in region2 than increase in the percentage of positive nontargeted probesin region 1 (Fig. 5b). It should be noted that for DNA samplescontaining more complex nontargeted sequences, the increasein the percentage of positive nontargeted probes (i.e., cross-hybridization)for increasingly negative ${Delta}$ G°_duplex values may be more pronounced,while for more abundant target sequences the decrease in thepercentage of positive targeted probes for decreasingly negative ${Delta}$ G°_duplex values may be suppressed.

A number of studies have demonstrated the effectiveness of ${Delta}$ G°_duplexas a theoretical probe selection parameter for both long andshort probes and RNA or DNA targets (22, 30, 38, 39, 42, 49,50, 57, 58). In contrast, Pozhitkov et al. (46) recently suggestedthat all theoretical (thermodynamics-based) screening of oligonucleotideprobes should be omitted due to poor correlations between ${Delta}$ G°_duplex(and ${Delta}$ G° for intra- and intermolecular self-structures) andexperimentally observed signal strengths for rRNA targets. Similarly,only a weak linear correlation was observed between probe signalintensity and ${Delta}$ G°_duplex (see Fig. S2 in the supplementalmaterial). This indicates that hybridization signals cannotbe accurately predicted with current thermodynamics models derivedfrom hybridizations in solution. However, thermodynamic indicesmay still be indicative of probe behavior by reflecting theprobability that a probe will yield a signal intensity higherthan a given threshold. Based on the effect of ${Delta}$ G°_duplexon probe selectivity (Fig. 5a), an optimal ${Delta}$ G°_duplex of –19.3kcal/mol is suggested for 18-mer oligonucleotide probes underthe described hybridization conditions. These probes shouldyield the highest confidence in presence/absence calls for agiven number of probes per target gene. For probes with eithermore or less negative ${Delta}$ G°_duplex values with lower selectivity,more probes need to be designed for a given target to achievehigh confidence in presence/absence calls. In general, probeswith a more negative ${Delta}$ G°_duplex displayed better sensitivitybut poorer specificity. Finally, it is interesting that theeffect of ${Delta}$ G°_duplex on the trade-off between probe sensitivityand specificity is analogous to the effect of probe length.In general, long probes (50- to 70-mers) are more sensitivethan short probes (15- to 30-mers) but display lower specificity(28, 48).

Cost and flexibility in microarray probe synthesis are oftenrecognized as two of the major limiting factors in the use ofmicroarray technology for diagnostic purposes (6, 21). In thisstudy, a maskless light-directed in situ microarray synthesistechnology developed at the University of Michigan was employed(19). This technology employs a digital micromirror device togenerate preprogrammed light patterns on the chip surface, triggeringdeprotection of the 5'-hydroxyl group in conventional phosphoramiditemonomers. Synthesis of oligonucleotides using this chemistryprovides high fidelity and a stepwise yield and also allowssynthesis of probes that are up 100 nucleotides long. In addition,synthesis of new chips is low cost, rapid, and flexible andinvolves simply uploading a list of probe sequences in the opticalunit. Other advantages of in situ probe synthesis include highspot uniformity and probe molecule density. In addition, continuousrecycling of the hybridization solution in the microfluidicchips used in this study provides increased signal uniformitywithin a spot and increased reproducibility of the hybridizationsignals. The higher cost per chip compared to conventional glassslides is the major limitation of this platform (21). The abilityto rapidly synthesize biochips with updated and reiterated probesets is considered critical for diagnostic purposes as new genesequence information is appearing almost daily and this informationshould be incorporated in the probe selection exercise and dataanalysis as soon as it becomes available.

Conclusions.
In conclusion, we developed and validated a coupled format formultiplex PCR and an in situ-synthesized DNA biochip for detectionof 12 bacterial pathogens in various water samples. By usingredundant probe sets targeting various VMGs and inferring presence/absencecalls based on the PF, false-positive calls were eliminated.Pathogens could be detected at a relative abundance of 0.1 to0.01%, depending on the pathogen. Analysis of the hybridizationpatterns also showed that probes with a ${Delta}$ G°_duplex providedthe best trade-off between sensitivity and specificity. In futurestudies, the VMG biochip will be applied to additional environmentalsamples, and the presence/absence calls will be verified independentlyusing real-time PCR.

ACKNOWLEDGMENTS

This work was funded in part by grants from the National Institutesof Health (grant 1 R01 RR018625-03), the Michigan Economic DevelopmentCorporation (grant GR-476 PO 085P3000517), and the MSU Foundation.

FOOTNOTES

* Corresponding author. Mailing address: Department of Civil and Environmental Engineering, A126 Research Complex Engineering, Michigan State University, East Lansing, MI 48824. Phone: (517) 355-8241. Fax: (517) 355-0250. E-mail: hashsham{at}egr.msu.edu

Published ahead of print on 1 February 2008.

S.M.M. and D.M.T. contributed equally to this study.

Supplemental material for this article may be found at http://aem.asm.org/.

REFERENCES

Top