FDA perspectives on potential microarray-based clinical diagnostics

The US Food and Drug Administration (FDA) encourages the development of new technologies such as microarrays which may improve and streamline assessments of safety and the effectiveness of medical products for the benefit of public health. The FDA anticipates that these new technologies may offer the potential for more effective approaches to medical treatment and disease prevention and management. This paper discusses issues associated with the translation of nucleic acid microarray-based devices from basic research and target discovery to in vitro clinical diagnostic use, which the Office of In Vitro Diagnostic Device Evaluation and Safety in the Center for Devices and Radiological Health foresees will be important for assurance of safety and effectiveness of these types of devices. General technological points, assessment of potential concerns for transitioning microarrays into clinical diagnostic use and approaches for evaluating the performance of these types of devices will be discussed.


Introduction
The question of whether high-throughput genep latforms liken ucleic acid microarrays arer eady to be used as clinical diagnostic devices 1 is the focus of many ongoing discussions and efforts in the scientific and medical communities. Specific concerns include standardisation, reproducibility and accuracy, appropriate statisticala pproachesf or data analysis and validation and the availability of control materials to ensure that theseplatformscan be reliably used in clinical diagnostics.
In this paper,w ew illa ttempt to assess areas of concern in transitioning microarrays from research into the clinical diagnostic arena as in vitro diagnostic devices (IVDs), 2 and providei nformation on the scientific issues that need to be considered to demonstrate whether thesed evices are safe and effectivea sc linical diagnostic tools.
'The Critical Path to New Medical Products' of the US Food and Drug Administration (FDA) has identifiedgenomic technologies as being crucialinadvancing medical product developmentand personalised medicine. 3 Themicroarray community and regulatoryagencies are coordinating efforts to develop standards to supportmicroarraydata for regulatory decisionmaking. The FDAhas participated in aseries of public workshops, 4-6 drafted several guidances 7,8 related to genomics and pharmacogenomics and released aconceptpaper addressing the co-developmentofpharmacogenomic-based therapeutic agents and diagnostic tests necessaryfor therapeutic decision making. 9 Recently,ajoint FDA/Johns HopkinsUniversity/ Pharmaceutical Research and Manufacturers of America (PhRMA) workshop entitled 'Microarrays in Transcriptional Profiling' focused on expression profiling issues as they may relate to nucleic acid microarrays for diagnostic use. 10 Issues that were addressed included microarraycontrols, variation between laboratories and platforms, datanormalisation, microarray scanner performance and clinicalvalidation.
Te chnology overviewa nd potential concerns for translation to clinical diagnostics We will limit the scopeo ft his paper to nucleica cid microarrays and will not address tissuea nd protein microarrays. We begin with ab rief technologyo verview,e mphasising potentiala reas of concerni nt ransitioning microarrays into the clinical IVD arena for three types of nucleica cid arrays: (i) genee xpression arrays, (ii) genotyping arrays and (iii) comparativeg enomic hybridisationa rrays( array-CGH).

Gene expression arrays
Gene expression profiling as ag enomic technique is intended to determine the fraction of genes that are expressedthat is, actively transcribed into mRNAu nder specific circumstances in certain cells. Gene expression varies depending on factorss uch as cell lineage,s tage of differentiation,e xtracellular stimuli or intracellular regulation. Most microarrays used in exploratorye xpression profiling measure the relativee xpression of tens of thousands of genes at at ime, creating am olecularp rofile of the mRNAi naspecific sample.I ti sb elieved that expression profiling will become a useful tool for evaluating disease susceptibility,t om akee arlier or more reliable diagnoses and to classify tumoursb yt heir molecular signature. 11,12 For example,w em ay expect a potentiald iagnostic microarrayd evice to be used for class prediction among existing classesofd isease types -thatis, to determine in which categoryapatientb elongs. In the case of diseases where the expression of multiple genes mayb e needed to distinguish between the disease states or types, such as certain types of cancer, 12 there is an implied need for a diagnostic platformt hat can reliably and quantitatively measure expression differences of af ew hundred genes, mirroring clinically important diagnostic and prognostic differences. 12 Although gene expression microarraya nalysisi satremendously promisingfield, many problems remain to be addressed. Comparing gene expression levels usingm icroarrays is av ery complex process,h ighly prone to variation. 13,14 Specific concerns include the reproducibility of the results between different laboratories and operatorsusingthe same platform, as well as the ability to reproduce results generated for the same samplesu singad ifferent platform. The number and type of probes on amicroarraymay alter its performance and henceits reproducibility.
Several recent publications have provided valuable insights and suggestedw ayso fo vercoming some of the sources of variationi nm icroarraye xperiments. 15 -21 These studies demonstrated that better reproducibilityc an be achieved between arrayp latformsa nd among laboratories if experiments arecarefully designed, controlled and executed. 16 Implementation of standardised protocolsf or both experimental and computational aspectsofthe comparison study led to ad ramatic increase in reproducibility and also underlined the need to makeraw data available,sothat data normalisation can be understood and possibly reassessed usingam ore appropriate algorithm. 15 Many microarrayn ormalisation methods currently used in research and discovery involvet he use of data from all arrays within ag iven experiment. 21,22 This approach might not be practicalfor acommercial device, where data normalisationm ay need to be performed for a single patienta rray,p ossibly requiring ar eference database. 23 Concordance between twoarray platformsl argely depends on the exact gene sequences measured by each platform and howt he measured genes maptoe ach other. 18,24 Discordant measurements between platformsm ay,inm anyc ases, be a consequence of asignificant level of cross-hybridisation 24 or differential use of splice variants inherent in the probe design. If quantitativereverse transcriptase polymerase chain reaction (QRT-PCR)i sused to verify discordant measurements between the microarrayplatforms, this verificationmay depend on the region of the gene from which the QRT-PCR assayisdesigned. Thus, QRT-PCR maynot be able to validate all microarrayp latformr esults, whether discordant or concordant among different microarrayp latforms. 17,19 Microarrayt echnology allowst he analysis of large quantities of data without as tarting hypothesis and the resulting patterns mayh aven oc lear biological meaning. Therefore, assessing the clinical and analyticalp erformance of the potential diagnostic test mayb ee venm ore important for RNA expression analysis than in other fields.M anyo ft he published studies, however, have not appropriately demonstrated analytical reproducibility or clinicalv alidity. 25,26 Often,RNA expression data maynot be generalisable,such as when the method is developed on one group of patients but subsequently evaluatedo na nother patient group with different characteristics. 27 Biasi so ften present in patientselectiona nd result interpretation. Bias maya lso be due to sample collection,p rocessinga nd storage,o rt of actorsr elated to operator performance or instrument use. 25 Another statistical issue related to analysiso ft he complex datasets from expression microarrayexperiments is over-fitting of the data when usingmultivariate models for alarge number of potentialp redictorsw hile trying to discriminateu singa relatively smalln umber of patients. 26,28 While models based on large numberso fp redictorsa ppeart od ow elli natraining set, this performance is often not duplicated when using a separate validation set. Traininga nd validation (testing) sets of samples shouldb ed istinct and independentt oa void these pitfalls. 29 It should be noted that subdividing ag roup of patients in two-for example using half of the data from each of four hospitals rather than using dataf romd istinct hospitals for training and validation parts -may not yield ar eliably independent test of ad evice performance.I fr eproducibility of the study results has not been demonstrated usinga n independent validation set of adequate size,t he resulting predictorsc annot be accepted as giving definitivec onclusions for use in clinical practices since they mayl ead to inappropriate treatment. 26 For example,d ata used to develop ac omposite biomarker to predict drug response should be distinct from the data generated and used to evaluate the actual drug response within patients ubsets having the specific pharmacogenomic profile. 28,30 Genotyping arrays Microarrays can be used to identify heritable variationo r somatic changesofDNA in individuals and across populations (ie genotype). 31,32 Arrays can detect many forms of heritable genetic variation, including single nucleotide polymorphisms (SNPs), single or multiple base insertions and deletions, gene conversions and repeats. Polymorphism analysis using microarrays can potentially be used to detect codingc hanges associated with ab iologically relevant phenotype (eg drug metabolism enzyme polymorphisms), for linkage analysisi f performed on ag enome-wide scale or in ad ense chromosomal region,o rf or assessment of loss of heterozygosity( LOH) in cancer. 23 Most genotyping arrays fall into one of twog eneral categories -S NP arrays or resequencing arrays. 33 Both types primarily use highly redundant hybridisation-based discrimination. Genotypinga rraysm ay querym ultiple polymorphisms or mutations in as ingle gene (for example, cystic fibrosis transmembrane conductance regulator, cytochrome P450 2D6L OH)o rm ultipleg enetic loci throughout the genome.M any humanS NPs that are diseasecausing or arei nvolved in response to drug therapies arew ell known and can be assessedi nahighly parallel fashion using high-density oligonucleotidem icroarrays. If used in infectious disease,t his type of arrayw ould allowf or the detection of genotypic variants of microbial pathogens. 34 -36 Ar apidly developing clinicala pplication of microarray genotyping is pharmacogenetic testing. 37,38 Ap harmacogenetic test can be defined as an assayi ntended to study inter-individual variations in DNA sequence related to drug absorption and disposition (pharmacokinetics) or to drug action (pharmacodynamics). 39 This includes polymorphisms within the genes encoding metabolising enzymes, transporters, receptorsa nd other proteins. The ultimate promise of pharmacogenetics is the possibility that knowledge of a patient'sD NA sequencem ightb eu sed in drug therapyt o maximise treatment efficacy,a void adverse drug reactions, optimise doses and target drug treatment only to patients who are likelyt or espond.
Depending on the intended use of the potentialIVD,many of the areas of concerni dentifiedf or gene expression array technologies, such as bias, over-fitting and generalisability,may also apply to genotyping arrays.

Comparativeg enomic hybridisation using arrays (array-CGH)
Changes in DNA copyn umberi nagenome are associated with gains and losses of chromosomes and chromosomal segments due to deletions, insertions and duplications. These can be germline mutations or somatic events. DNA alterations can lead to copyn umber polymorphisms in normali ndividuals, 40 -42 but can also be the cause of various disease states. Detection of these alterations facilitates the identification of crucialg enes and pathwaysi nvolved in biological processes and diseases, and mayaid in diagnosis and therapyf or genetic and somatic diseases. 43,44 Array-CGH is a method for identifying variations in genomic DNA copy number betweennormal and pathological samples. In atypical array-CGH experiment, total genomic DNA isolated from reference and test samples is differentially labelled witht wo fluorescent molecules withd ifferent excitation and emission properties and hybridised to am icroarrayc ontaining probes representingd ifferent regionso fg enomic DNA. Information on relative DNA copyn umber differences is generated based on the intensity of the fluorescence signal generated by each labellingm olecule.
Depending on the amount of genome coverage,array-CGH can be either marker based or contig based. Marker-based arrays target specific regions selected based on previous knowledge of widely known chromosomal alterations. Contig-based arrays target the whole genome derived from probe sequences representings horti ntervals (as lowasakilobase scale) of the entire genome for the purpose of an ew biomarker discovery for agiven condition. 45 These twoarray types mayi nclude platformsu sing clones of bacterial artificial chromosomes, fosmid clones, cDNA and shorto ligonucleotides. These different platformsp rovide different levelsofp erformance, such that someare more suitable for particular applications than for others. Factorst hat determine performance requirements include the magnitudes of the copyn umber changes, the state and composition of the specimen, howm uch material is available for analysisand howthe results of the analysiswill be used. There area dvantages and disadvantages, as well as inherent technicald ifficulties, associated witheach of these platforms, which impact on the accuracy of data generated.
Alternativep roceduresa re being developed to tackle technicali ssues such as highly complex genomic material, the presence of repetitive sequences, method of sample processing (eg fresh tissues versus formalin-fixed tissues) and data managementa nd extrapolationo fd ata generatedf rom array-CGH. 46 -48 Once these technical hurdles areo vercome, array-CGH mayp otentially be used for the analysis of clinical samples to identifyg enomic alterations, as well as for measuring the loss of allelic heterozygosity and improving quantitativeaccuracy,r esolution and the dynamic range of the detection of genomic copyn umber variations compared with existing cytogenetic methods. 44,45,47 High analytical performance is ap rerequisite for array-CGH to have clinical utility.For example,noiselevel of the arraymeasurements and difference in behaviour of different arraye lements mayn eed to be appropriately addressed. Reliable detection of single copyc hangesm ay be difficult in heterogeneous cell populations (egi ntermixed tumour and normal cells); however, hybridisations of defined specimens with known aberrations mayh elp in establishing the performance characteristics of array-CGH. 23 Tež ak et al.

Regulatorya pproaches for evaluation of microarray-based IVDs
Since the US Congress enactmento ft he Medical Device Amendments in 1976, 1 the FDAh as evolved risk-based regulations and policiesd esigned to promote and protect the public health by regulating medical devices. TheF DA regulations also set as an equally important goal the encouragement of the discovery and development of new medical products for the benefit and promotiono fp ublic health.I VDs are considered to be medical devices for the purposes of regulatoryo versight, and are defined as reagents, instruments and systems intended for use in the diagnosis of disease or in the determination of the state of health in order to cure,m itigate,t reat or prevent disease. 2 Therefore, by statute, IVD tests that aret ob ec ommercialised for the diagnosis and managemento fp atients are subject to FDAr egulation.
Within the FDA, the Office of In Vitro Diagnostics (OIVD) within the Center for Devices and Radiological Health (CDRH) regulates devices pre-and post-marketing to ensure that they demonstrate ar easonable assuranceo fs afety and effectiveness for the intended use according to the directions for use. 49 Each regulated device is assigned one of three risk-based classes related to the level of FDAo versightp rior to marketing. Class Idevices are generally considered lowrisk and many are exemptf romp re-market notification to the FDA. Class II devices (and non-exemptC lass Id evices) are considered to carry more risk and ares ubjected to pre-market notification review by the FDAtodetermine whether they are similar (in terms of safety and effectiveness) to another legally marketed device intended for the samet ype of use.C lassI II devices are considered the highest risk devices and these devices requirep re-market application approval, involving a more in-depth review and documentationo ft he safety and effectiveness of the device. 50 The FDA/OIVD has put in place as o-called pre-Investi-gationalA dvice Exemption (IDE)m echanism 51 to allowearly communication between the FDAa nd sponsorso fn ew technologya nd to providep rotocolr eviewa nd regulatory guidance with no cost to the sponsor.This process can prevent manufacturers from wasting resources on studies that would not support FDAa pproval for their intended use and allows the FDAt he opportunity to become familiarw itht he test before reviewingt he formal pre-market submission. This may be especially important for devices in emerging fieldss uch as microarray-based diagnostics, where both review policy and regulatorys cience are continually evolving.
As witho ther areas of genetic testing, there areu nique scientific and ethical issues that will need to be addressed for the field of genomics/microarrays to advance and allow utilisationo fm icroarrays in clinicald iagnostics. We will attempttoaddress some of the issues that OIVD has identified in relation to the development of microarrays for use in clinicaldecisionmaking, including intended use and analytical and clinical performance.M ore detailed information referring to some specific types of tests can also be found in Special Controls guidance documents issued by OIVD. 52 -54 It should be noted that since these arer elatively new types of IVD assays, requirements and performance characteristics are evolving and aren ot completely defined for everyt ype of platform and assay.

Intended use
The intendedu se of am icroarray-based IVD,a nd the risk the device poses to the patienta ccording to its intended use,a re the main determinants for the regulatoryc lassification. The intended use of the device for which pre-market approval or clearance is sought shouldspecify the analyte which the device is intended to measure,t he clinical purpose of measuring the analyte and the populations for which the device is indicated, where appropriate.T he exact types of analyticala nd clinical data that have to be submitted depend greatly on the claims made in the intended use.
The risk to ap atient from anyI VD device is primarily related to the use of the information derived from the test results, and is not linked to the complexity of the technology used in thed evice.T he risk to the patient of false-positiveo r false-negativeresults is often tied to the clinical decisions made based on the result. Fore xample,i fadevice is being used to determine whether ap atientw ill benefit from ad rugt hat has significantt oxicity,i tc ould be ah igh-risk device: the patient maye xperience unnecessarys ide-effects with no clinical benefit following an incorrect test result. The opposite case is that ap atient whoc ould potentially benefit from the therapyc ould be sub-optimally treated based on incorrect diagnostic test results predicting adverse events. This emphasises the need for appropriate analytical and clinical validation of the assay.

Analytical performance
Adequate analyticalp erformance of ad iagnostic test is crucial to allowc onfidence in the test results. Microarray-based devices need to be validatedanalytically to determine that they can measure the intended analyte reproducibly and accurately when the assayisperformed by the intended user (eg aclinical laboratory with trained and experienced operators). The analyticalp erformance of the test should be clearly known in order to evaluate the device risk. More technologically or conceptually complex tests mayr equireh igher levelso f analyticale valuation. The userss hould be aware of any analyticall imitationso ft he test, so that test results mayb e interpreted in light of these limitations. High analytical variationm ay decrease the reliability of test results, compromisingt he clinical utility of the test.
Evaluation of microarraya nalyticalp erformance is complicated by the large number of separate analytes evaluateds imultaneously,a sw ell as by the massivea mount of post-analytical data processingr equired to obtain the result; however, pre-analytical steps shouldnot be overlookedasthey are an extensives ource of variationi nm ost microarray experiments. Methods for appropriate sample quality assessment at different steps during pre-analytical processing (eg RNAq uality assessment, labelling, amplification) can be very usefuli nd emonstrating the analyticalp erformance of these tests. Analytical validation requires the use of appropriate controls such as hybridisation controls and samplec ontrols. There ares everal ongoing joint efforts between the FDA, the National Institute of Standards and Te chnology (NIST) and the microarrayc ommunity to maket his process easier: 10,55 -57 this is described in more detail below.
For quality control, both positivea nd negativec ontrols should be included and controls should reflect sample composition and DNA concentration. Controls should show that all steps and critical reactions have proceeded properly and without contamination or cross-hybridisation. For each microarrayf eature or element, the manufacturer of as pecific microarrayd evice should assure the identity of the probe or feature and its reproducible placement on the array. Analytical verificationshould demonstrate that the test adequately detects what it is supposed to detect. This can be relatively straightforwardfor genotyping arrays, but potentially very complex in the case of arrays measuring gene expression levels. The same is true for demonstrating the diagnostic utilityo ft he test.
For example,amicroarray-based test that measures as et of SNPs to aid in the determination of drug therapy will need to demonstrate that the test can detect the polymorphisms it is designed to detect and, ideally,n oo thers. This mayb e straightforward for genes that are easily amplified, without excessiven umberso fr epeats or interfering sequences; however, the inclusion of other genetic modifications (eg duplications, deletions, inversions) or additional SNPs in the assayw ill increase the potential for analyticalc omplications.
The reproducibility of microarray-based devices should be thoroughly evaluated. This becomes especially important as devices become increasingly complex, and reproducibility should be extensively evaluated for each analyte. 8,52 Studies should include assessment of reproducibility between several laboratories (including at least twoe xternal sites), operators, instruments, arrayl ots, reagent batches, scans, days etc,a s appropriate.V alues maybeassessed and reported in anumberof ways -that is, variancecomponent analyses 58 mayberelevant in identifying major sources of variationinthe values observed.
In addition, thea ccuracy of the device should be established. Aw ay to accomplish this mayb et oc ompare the results of the device with the results for thesame samples using ar eference method (ie 'gold standard' test). Forg enotyping tests, double-stranded bi-directional DNA sequencing is the one well-recognised method for establishing performance; however, depending on the clinicalc laimsb eing made -f or example, diagnosing as pecific formo fc ancer,e valuating risk for future developmento fd isease and/or determining drug selectiono rd osing-additionalc linicali nformation, such as biopsy results, imaging data and/or results of blood tests,m ay be required to confirmt he true diagnosis. Alla nalytical parameters( eg fluorescence intensity,s ubset of alleles) should be challenged during testing to substantiate thed evice performance at keyd ecisionl imits.
It is important that studies be carried out with realc linical samples (either prospectivelycollected or appropriately stored) whenever possible,t oe valuate any biological and matrix effects that mayo ccur and to achieveg reater assurance that there are no unforeseen endogenous interferences that may affect the analyticalp erformance of the device.A nalytical performance of ad iagnostic test should be based on ad ataset that is independent from and prior to the prospectiveo r retrospectives amples on which it is to be clinically validated.

Clinical performance
While analytical validation presents its ownc hallenges, the questionofthe clinical validation 59 of microarray-based tests is likelyt op rove to be an even larger issue for many assays and manufacturers to confront.
Ad iagnostic device has clinical utility when it provides information about ad isease or clinical condition that is useful and/or meaningful to the healthcare provider and the patient. Different types of microarrayt ests mayr equired ifferent types of clinicalv alidation. For some polymorphismsa nd genetic alleles, there mayb easufficient literature base to establish the clinicalu tilityo ft he new test without extensivec linical studies. The FDArecently approved the Roche AmpliChip for Cytochrome P450 60 and the TM Bioscience Cystic Fibrosis assay 61 usinga nalyticals tudies and clinicall iterature.M any microarray-based devices that are in development, however, mayn eed clinical studies to establish the safety and effectiveness of their use in aclinical setting. If certain biomarkershave been extensively studied biologically or clinically,a nd have an underlying biological rationale, their validation mayb ee asier because of support from existing scientific data. Biomarkersor classifiers without an established biological/mechanistic relationship to the disease,however, can also have autilityasa potentiald iagnostic tool. Clinicalv alidation of biomarkers with no known biological mechanism related to the specific disease or drug action can be established usings olid statistical methods and robust clinical trial data. 30 Am ajor hurdle for establishing clinical performance of the test is obtaining and securing adequate numberso fq uality specimens from clinical trial subjects that can be used in support of the clinical utilityofthe device for its intended use. Regardless of whether ac linical trial is performed using prospectives amples or banked samples collected in a prospectivem anner,t he trial should be hypothesis driven and test ap re-specified and well-defined diagnostic biomarker/ classifierc laim (ie all cut-off points, platforms etc should be established in advance of the trial). If usings amplest hatw ere Similarly to analyticalv alidation, the complexityoft he test can influence the types of studies that areneeded.Any patterns or combinations of analytes should be defined prior to the starto ft he clinicals tudies to allowf or verificationo ft he patterna sau nit. In addition, the test population should be clearly defined and considered to support the intended use in the intended population.
Quality control standardisation efforts for microarrays As described above,several types of factorsa ffect theoutcome of am icroarrays tudy; am ajor concerni st oa ssure reproducibility and accuracy.I ntra-and inter-laboratory data consistency is the foundation of reliable knowledge extraction and meaningful comparisons. The microarrayc ommunity and regulatorya gencies arew orking together to establish as et of consensus quality assurance and quality control criteria for assessing and ensuring data quality to identify critical factorsa ffecting data quality and to optimise and standardise microarrayp roceduress ot hat biological interpretation and decisionm akinga re not based on unreliable data. 13 Developing standards and quality measures shoulde nable the successful and reliable use of microarrayt echnology in regulatoryd ecision making and in clinicalp ractice.
Microarrayc ontrols generally can be divided into internal and external. Internal (intrinsic) sample controls, which control for the sample integrity,c an be ag ene or ag roup of genes intrinsic to the sample, chosen because they are known not to vary in the particular condition. Am ajor obstacle to using this type of microarrayc ontrol is findingg enes that would not vary in studied conditions. External controls are generally developed as spike-ins that are not partofthe sample and can be used to characterise the performance of certain steps in the microarraye xperiment. Here,w el ist several current large quality control and standardisation efforts in the microarrayc ommunity.

External RNA Controls Consortium (ERCC)
The ERCCi sw orking to develop tools for experiment control and performance evaluation for gene expression analysis. These tools will includes pike-in controls, protocols and informatics tools intended to be useful for various microarrayp latforms and QRT-PCR. 55 These platformindependent control materials are needed for performance evaluation of reproducibility,sensitivity and robustness in gene expression analysis. They will be readily accessible and used as atool for verification of technical performance of amicroarray assay, but not for sample integrity (they aren ot intrinsic controls). There arep lans for the controls to be made commercially available as individual plasmid clones for the synthesis of polyadenylated transcripts. 56

MicroarrayQ uality Control (MAQC)
The purpose of the MAQC project is to provideq uality control tools to the microarrayc ommunity and to develop guidelines for microarrayd ata analysisb yp roviding the public with large reference datasetsa long with readily accessible reference RNA samples. 57 The MAQC project involves six FDAc entres, major providerso fm icroarrayp latformsa nd RNA samples,t he Environmental Protection Agency (EPA), NIST,a cademic laboratories and other stakeholders. The MAQC project aims to establish quality control metrics and thresholds for objectively assessing the performance achievable by various microarrayp latformsa nd evaluating the advantages and disadvantages of variousd ata analysism ethods. 13,57 Tw oh umanR NA samples have been selected and the differential genee xpression levels between the twos amples have been calibrated with both microarrays and QRT-PCR. The resulting microarrayd atasets will be used to assess to the precision and cross-platform/laboratoryc omparability of microarrays. The large QRT-PCR datasets will also enable evaluation of the nature and magnitudeo fa ny systematic biases that maye xist between microarrays and QRT-PCR. The availabilityofthe calibrated RNA samples combinedwith the resulting microarraya nd QRT-PCR datasets, which will be made readily accessible to the microarrayc ommunity, will allowi ndividual laboratories moree asily to identify and correct procedural failures. The MAQC project aims to help improvem icroarrayt echnologya nd foster its proper applications in discovery,d evelopmenta nd review of FDA-regulated products.

Discussion
The developmento fD NA microarrayt echnologyf or use in both exploratorys tudies and as ap otentialm edical diagnostic tool has sparked excitement in the scientific and medical communities. Microarrays have the potential to simultaneously detectm ultiple DNA sequence variations, monitor relative expression levels of thousands of genes in selected tissues and identify infectiousd isease organismso rt he host response to infection in biological specimens. When developed as medical IVDs, the results obtained by thesea ssays mayc omplement or be used in combination with other diagnostic methods and potentially be utilised for patient managementa nd treatment decisions.
Microarrays promise to be ap owerful technology when used correctly,b ut can be misleading without acceptable reproducibility and accuracy.T he inherent strength of microarrays -t heir ability to querym ultiple analytes in a single assay, therebyp roviding results of increased complexity compared with more traditionala ssays -a lso necessitates cautionininterpreting the results for medical decision making. Performance data for use with specific applications, however, is often not sufficiently established scientifically to be suitable for regulatoryd ecision making.
As noted in this paper,animportant issue in the microarray field is platform-to-platform variability,s ince there mayb e inconsistencies between different platforms on agene-by-gene basis. If am anufacturer develops as pecific IVD involving a single microarrayp latform, however, and can demonstrate analyticala nd clinical performance of this diagnostic microarray, the issue of the inter-platformv ariability mayn ot be crucial. To be able to use such platforms clinically,t he variability of that specific microarrayp latforms hould be thoroughly assesseda nd proven to be robust and operating consistently and reproducibly across laboratories and over time. The appropriate evaluation should be performed to establish analyticala nd clinical performance characteristicso ft he IVD test, including whether thet est accurately and reproducibly detects the genemarkersitclaims to detect for its intended use in clinically relevant populations.
The FDAr ecognises the potentialo fm icroarray-based tests for use in clinical practice.W eh avec ollaborated in as eries of public meetings 4-6,10 to obtain inputo nr elevant issues from the scientific community and interested stakeholders. Several guidance documents have been, or areb eing, developed that providei nformation on the FDA'sc urrent thinking on usingn ew technologies in clinicald ecision making. 7-9,52 -54 We arew orking with the whole microarray community towards streamlining this exciting technologya nd developingorassisting in the developmentofstandards to help to validate the performance and expedite transfer of these devices from the research to the clinical use setting.

Disclaimer
This paper represents the current opiniono ft he Office of In Vitro Diagnostic Device Evaluation and Safety (OIVD), Office of Science and Engineering Laboratories (OSEL) and Office of Surveillance and Biometrics (OSB), in the Center for Devices and RegulatoryH ealth (CDRH) within the US Food and Drug Administration (FDA).