Reducing the Bottleneck in Discovery of Novel Antibiotics.

Most antibiotics were discovered by screening soil actinomycetes, but the efficiency of the discovery platform collapsed in the 1960s. By now, more than 3000 antibiotics have been described and most of the current discovery effort is focused on the rediscovery of known compounds, making the approach impractical. The last marketed broad-spectrum antibiotics discovered were daptomycin, linezolid, and fidaxomicin. The current state of the art in the development of new anti-infectives is a non-existent pipeline in the absence of a discovery platform. This is particularly troubling given the emergence of pan-resistant pathogens. The current practice in dealing with the problem of the background of known compounds is to use chemical dereplication of extracts to assess the relative novelty of a compound it contains. Dereplication typically requires scale-up, extraction, and often fractionation before an accurate mass and structure can be produced by MS analysis in combination with 2D NMR. Here, we describe a transcriptome analysis approach using RNA sequencing (RNASeq) to identify promising novel antimicrobial compounds from microbial extracts. Our pipeline permits identification of antimicrobial compounds that produce distinct transcription profiles using unfractionated cell extracts. This efficient pipeline will eliminate the requirement for purification and structure determination of compounds from extracts and will facilitate high-throughput screen of cell extracts for identification of novel compounds.

Genomic and transcriptomic differences in community acquired methicillin resistant Staphylococcus aureus USA300 and USA400 strains.

BACKGROUND: Staphylococcus aureus is a human pathogen responsible for substantial morbidity and mortality through its ability to cause a number of human infections including bacteremia, pneumonia and soft tissue infections. Of great concern is the emergence and dissemination of methicillin-resistant Staphylococcus aureus strains (MRSA) that are resistant to nearly all ß-lactams. The emergence of the USA300 MRSA genetic background among community associated S. aureus infections (CA-MRSA) in the USA was followed by the disappearance of USA400 CA-MRSA isolates. RESULTS: To gain a greater understanding of the potential fitness advantages and virulence capacity of S. aureus USA300 clones, we performed whole genome sequencing of 15 USA300 and 4 USA400 clinical isolates. A comparison of representative genomes of the USA300 and USA400 pulsotypes indicates a number of differences in mobile genome elements. We examined the in vitro gene expression profiles by microarray hybridization and the in vivo transcriptomes during lung infection in mice of a USA300 and a USA400 MRSA strain by performing complete genome qRT-PCR analysis. The unique presence and increased expression of 6 exotoxins in USA300 (12- to 600-fold) compared to USA400 may contribute to the increased virulence of USA300 clones. Importantly, we also observed the up-regulation of prophage genes in USA300 (compared with USA400) during mouse lung infection (including genes encoded by both prophages ΦSa2usa and ΦSa3usa), suggesting that these prophages may play an important role in vivo by contributing to the elevated virulence characteristic of the USA300 clone. CONCLUSIONS: We observed differences in the genetic content of USA300 and USA400 strains, as well as significant differences of in vitro and in vivo gene expression of mobile elements in a lung pneumonia model. This is the first study to document the global transcription differences between USA300 and USA400 strains during both in vitro and in vivo growth.

A Multi-Omic View of Host-Pathogen-Commensal Interplay in Salmonella-Mediated Intestinal Infection.

The potential for commensal microorganisms indigenous to a host (the 'microbiome' or 'microbiota') to alter infection outcome by influencing host-pathogen interplay is largely unknown. We used a multi-omics "systems" approach, incorporating proteomics, metabolomics, glycomics, and metagenomics, to explore the molecular interplay between the murine host, the pathogen Salmonella enterica serovar Typhimurium (S. Typhimurium), and commensal gut microorganisms during intestinal infection with S. Typhimurium. We find proteomic evidence that S. Typhimurium thrives within the infected 129/SvJ mouse gut without antibiotic pre-treatment, inducing inflammation and disrupting the intestinal microbiome (e.g., suppressing Bacteroidetes and Firmicutes while promoting growth of Salmonella and Enterococcus). Alteration of the host microbiome population structure was highly correlated with gut environmental changes, including the accumulation of metabolites normally consumed by commensal microbiota. Finally, the less characterized phase of S. Typhimurium's lifecycle was investigated, and both proteomic and glycomic evidence suggests S. Typhimurium may take advantage of increased fucose moieties to metabolize fucose while growing in the gut. The application of multiple omics measurements to Salmonella-induced intestinal inflammation provides insights into complex molecular strategies employed during pathogenesis between host, pathogen, and the microbiome.

A multi-omic systems approach to elucidating Yersinia virulence mechanisms.

The underlying mechanisms that lead to dramatic differences between closely related pathogens are not always readily apparent. For example, the genomes of Yersinia pestis (YP) the causative agent of plague with a high mortality rate and Yersinia pseudotuberculosis (YPT) an enteric pathogen with a modest mortality rate are highly similar with some species specific differences; however the molecular causes of their distinct clinical outcomes remain poorly understood. In this study, a temporal multi-omic analysis of YP and YPT at physiologically relevant temperatures was performed to gain insights into how an acute and highly lethal bacterial pathogen, YP, differs from its less virulent progenitor, YPT. This analysis revealed higher gene and protein expression levels of conserved major virulence factors in YP relative to YPT, including the Yop virulon and the pH6 antigen. This suggests that adaptation in the regulatory architecture, in addition to the presence of unique genetic material, may contribute to the increased pathogenecity of YP relative to YPT. Additionally, global transcriptome and proteome responses of YP and YPT revealed conserved post-transcriptional control of metabolism and the translational machinery including the modulation of glutamate levels in Yersiniae. Finally, the omics data was coupled with a computational network analysis, allowing an efficient prediction of novel Yersinia virulence factors based on gene and protein expression patterns.

Model-driven multi-omic data analysis elucidates metabolic immunomodulators of macrophage activation.

Macrophages are central players in immune response, manifesting divergent phenotypes to control inflammation and innate immunity through release of cytokines and other signaling factors. Recently, the focus on metabolism has been reemphasized as critical signaling and regulatory pathways of human pathophysiology, ranging from cancer to aging, often converge on metabolic responses. Here, we used genome-scale modeling and multi-omics (transcriptomics, proteomics, and metabolomics) analysis to assess metabolic features that are critical for macrophage activation. We constructed a genome-scale metabolic network for the RAW 264.7 cell line to determine metabolic modulators of activation. Metabolites well-known to be associated with immunoactivation (glucose and arginine) and immunosuppression (tryptophan and vitamin D3) were among the most critical effectors. Intracellular metabolic mechanisms were assessed, identifying a suppressive role for de-novo nucleotide synthesis. Finally, underlying metabolic mechanisms of macrophage activation are identified by analyzing multi-omic data obtained from LPS-stimulated RAW cells in the context of our flux-based predictions. Our study demonstrates metabolism's role in regulating activation may be greater than previously anticipated and elucidates underlying connections between activation and metabolic effectors.

Comparative omics-driven genome annotation refinement: application across Yersiniae.

Genome sequencing continues to be a rapidly evolving technology, yet most downstream aspects of genome annotation pipelines remain relatively stable or are even being abandoned. The annotation process is now performed almost exclusively in an automated fashion to balance the large number of sequences generated. One possible way of reducing errors inherent to automated computational annotations is to apply data from omics measurements (i.e. transcriptional and proteomic) to the un-annotated genome with a proteogenomic-based approach. Here, the concept of annotation refinement has been extended to include a comparative assessment of genomes across closely related species. Transcriptomic and proteomic data derived from highly similar pathogenic Yersiniae (Y. pestis CO92, Y. pestis Pestoides F, and Y. pseudotuberculosis PB1/+) was used to demonstrate a comprehensive comparative omic-based annotation methodology. Peptide and oligo measurements experimentally validated the expression of nearly 40% of each strain's predicted proteome and revealed the identification of 28 novel and 68 incorrect (i.e., observed frameshifts, extended start sites, and translated pseudogenes) protein-coding sequences within the three current genome annotations. Gene loss is presumed to play a major role in Y. pestis acquiring its niche as a virulent pathogen, thus the discovery of many translated pseudogenes, including the insertion-ablated argD, underscores a need for functional analyses to investigate hypotheses related to divergence. Refinements included the discovery of a seemingly essential ribosomal protein, several virulence-associated factors, a transcriptional regulator, and many hypothetical proteins that were missed during annotation.

Positional identification of variants of Adamts16 linked to inherited hypertension.

A previously reported blood pressure (BP) quantitative trait locus on rat Chromosome 1 was isolated in a short congenic segment spanning 804.6 kb. The 804.6 kb region contained only two genes, LOC306664 and LOC306665. LOC306664 is predicted to translate into A Disintegrin-like and Metalloproteinase with Thrombospondin Motifs-16 (Adamts16). LOC306665 is a novel gene. All predicted exons of both LOC306664 and LOC306665 were sequenced. Non-synonymous variants were identified in only one of these genes, LOC306664. These variants were naturally existing polymorphisms among inbred, outbred and wild rats. The full-length rat transcript of Adamts16 was detected in multiple tissues. Similar to ADAMTS16 in humans, expression of Adamts16 was prominent in the kidney. Renal transcriptome analysis suggested that a network of genes related to BP was differential between congenic and S rats. These genes were also differentially expressed between kidney cell lines with or without knock-down of Adamts16. Adamts16 is conserved between rats and humans. It is a candidate gene within the homologous region on human Chromosome 5, which is linked to systolic and diastolic BP in the Quebec Family Study. Multiple variants, including an Ala to Pro variant in codon 90 (rs2086310) of human ADAMTS16, were associated with human resting systolic BP (SBP). Replication study in GenNet confirmed the association of two variants of ADAMTS16 with SBP, including rs2086310. Overall, our report represents a high resolution positional cloning and translational study for Adamts16 as a candidate gene controlling BP.

Refinement of light-responsive transcript lists using rice oligonucleotide arrays: evaluation of gene-redundancy.

Studies of gene function are often hampered by gene-redundancy, especially in organisms with large genomes such as rice (Oryza sativa). We present an approach for using transcriptomics data to focus functional studies and address redundancy. To this end, we have constructed and validated an inexpensive and publicly available rice oligonucleotide near-whole genome array, called the rice NSF45K array. We generated expression profiles for light- vs. dark-grown rice leaf tissue and validated the biological significance of the data by analyzing sources of variation and confirming expression trends with reverse transcription polymerase chain reaction. We examined trends in the data by evaluating enrichment of gene ontology terms at multiple false discovery rate thresholds. To compare data generated with the NSF45K array with published results, we developed publicly available, web-based tools (www.ricearray.org). The Oligo and EST Anatomy Viewer enables visualization of EST-based expression profiling data for all genes on the array. The Rice Multi-platform Microarray Search Tool facilitates comparison of gene expression profiles across multiple rice microarray platforms. Finally, we incorporated gene expression and biochemical pathway data to reduce the number of candidate gene products putatively participating in the eight steps of the photorespiration pathway from 52 to 10, based on expression levels of putatively functionally redundant genes. We confirmed the efficacy of this method to cope with redundancy by correctly predicting participation in photorespiration of a gene with five paralogs. Applying these methods will accelerate rice functional genomics.

Molecular networks in Dahl salt-sensitive hypertension based on transcriptome analysis of a panel of consomic rats.

The Dahl salt-sensitive (SS) rat is a widely used model of human salt-sensitive hypertension and renal injury. We studied the molecular networks that underlie the complex disease phenotypes in the SS model, using a design that involved two consomic rat strains that were protected from salt-induced hypertension and one that was not protected. Substitution of Brown Norway (BN) chromosome 13 or 18, but not 20, into the SS genome was found to significantly attenuate salt-induced hypertension and albuminuria. Gene expression profiles were examined in the kidneys of SS and consomic SS-13(BN), SS-18(BN), and SS-20(BN) rats with a total of 240 cDNA microarrays. The substituted chromosome was overrepresented in genes differentially expressed between a consomic strain and SS rats on a 0.4% salt diet. F5, Serpinc1, Slc19a2, and genes represented by three other expressed sequence tags (ESTs), which are located on chromosome 13, were found to be differentially expressed between SS-13(BN) and all other strains examined. Likewise, Acaa2, B4galt6, Colec12, Hsd17b4, and five other ESTs located on chromosome 18 exhibited expression patterns unique to SS-18(BN). On exposure to a 4% salt diet, there were 184 ESTs in the renal cortex and 346 in the renal medulla for which SS-13(BN) and SS-18(BN) shared one expression pattern, while SS and SS-20(BN) shared another, mirroring the phenotypic segregation among the four strains. Molecular networks that might contribute to the development of Dahl salt-sensitive hypertension and albuminuria were constructed with an approach that merged biological knowledge-driven analysis and data-driven Bayesian probabilistic analysis.

Closely linked non-additive blood pressure quantitative trait loci.

There is enough evidence through linkage and substitution mapping to indicate that rat chromosome 1 harbors multiple blood pressure (BP) quantitative trait loci (QTLs). Of these, BP QTL1b was previously reported from our laboratory using congenic strains derived by introgressing normotensive alleles from the LEW rat onto the genetic background of the hypertensive Dahl salt-sensitive (S) rat. The region spanned by QTL1b is quite large (20.92 Mb), thus requiring further mapping with improved resolution so as to facilitate systematic identification of the underlying genetic determinant(s). Using congenic strains containing the LEW rat chromosomal segments on the Dahl salt-sensitive (S) rat background, further iterations of congenic substrains were constructed and characterized. Collective data obtained from this new iteration of congenic substrains provided evidence for further fragmentation of QTL1b with improved resolution. At least two separate genetic determinants of blood pressure underlie QTL1b. These are within 7.40 Mb and 7.31 Mb and are known as the QTL1b1 region and the QTL1b2 region, respectively. A genetic interaction was detected between the two BP QTLs. Interestingly, five of the previously reported differentially expressed genes located within the newly mapped QTL1b1 region remained differentially expressed. The congenic strain S.LEW(D1Mco36-D1Mco101), which harbors the QTL1b1 region alone but not the QTL1b2 region, serves as a genetic tool for further dissection of the QTL1b1 region and validation of Nr2f2 as a positional candidate gene. Overall, this study represents an intermediary yet obligatory progression towards the identification of genetic elements controlling BP.

Cross-talk of expression quantitative trait loci within 2 interacting blood pressure quantitative trait loci.

Genetic dissection of the S rat genome has provided strong evidence for the presence of 2 interacting blood pressure quantitative trait loci (QTLs), termed QTL1 and QTL2, on rat chromosome 5. However, the identities of the underlying interacting genetic factors remain unknown. Further experiments targeted to identify the interacting genetic factors by the substitution mapping approach alone are difficult because of the interdependency of natural recombinations to occur at the 2 QTLs. We hypothesized that the interacting genetic factors underlying these 2 QTLs may interact at the level of gene transcription and thereby represent expression QTLs or eQTLs. To detect these interacting expression QTLs, a custom QTL chip containing the annotated genes within QTL1 and QTL2 was developed and used to conduct a transcriptional profiling study of S and 2 congenic strains that retain either 1 or both of the QTLs. The results uncovered an interaction between 2 transcription factor genes, Dmrta2 and Nfia. Furthermore, the "biological signature" elicited by these 2 transcription factors was differential between the congenic strain that retained Lewis alleles at both QTL1 and QTL2 compared with the congenic strain that retained Lewis alleles at QTL1 alone. A network of transcription factors potentially affecting blood pressure could be traced, lending support to our hypothesis.

Combined application of behavior genetics and microarray analysis to identify regional expression themes and gene-behavior associations.

In this report we link candidate genes to complex behavioral phenotypes by using a behavior genetics approach. Gene expression signatures were generated for the prefrontal cortex, ventral striatum, temporal lobe, periaqueductal gray, and cerebellum in eight inbred strains from priority group A of the Mouse Phenome Project. Bioinformatic analysis of regionally enriched genes that were conserved across all strains revealed both functional and structural specialization of particular brain regions. For example, genes encoding proteins with demonstrated anti-apoptotic function were over-represented in the cerebellum, whereas genes coding for proteins associated with learning and memory were enriched in the ventral striatum, as defined by the Expression Analysis Systematic Explorer (EASE) application. Association of regional gene expression with behavioral phenotypes was exploited to identify candidate behavioral genes. Phenotypes that were investigated included anxiety, drug-naive and ethanol-induced distance traveled across a grid floor, and seizure susceptibility. Several genes within the glutamatergic signaling pathway (i.e., NMDA/glutamate receptor subunit 2C, calmodulin, solute carrier family 1 member 2, and glutamine synthetase) were identified in a phenotype-dependent and region-specific manner. In addition to supporting evidence in the literature, many of the genes that were identified could be mapped in silico to surrogate behavior-related quantitative trait loci. The approaches and data set described herein serve as a valuable resource to investigate the genetic underpinning of complex behaviors.

Gene expression profiling of lymphoblastoid cell lines from monozygotic twins discordant in severity of autism reveals differential regulation of neurologically relevant genes.

BACKGROUND: The autism spectrum encompasses a set of complex multigenic developmental disorders that severely impact the development of language, non-verbal communication, and social skills, and are associated with odd, stereotyped, repetitive behavior and restricted interests. To date, diagnosis of these neurologically based disorders relies predominantly upon behavioral observations often prompted by delayed speech or aberrant behavior, and there are no known genes that can serve as definitive biomarkers for the disorders. RESULTS: Here we demonstrate, for the first time, that lymphoblastoid cell lines from monozygotic twins discordant with respect to severity of autism and/or language impairment exhibit differential gene expression patterns on DNA microarrays. Furthermore, we show that genes important to the development, structure, and/or function of the nervous system are among the most differentially expressed genes, and that many of these genes map closely in silico to chromosomal regions containing previously reported autism candidate genes or quantitative trait loci. CONCLUSION: Our results provide evidence that novel candidate genes for autism may be differentially expressed in lymphoid cell lines from individuals with autism spectrum disorders. This finding further suggests the possibility of developing a molecular screen for autism based on expressed biomarkers in peripheral blood lymphocytes, an easily accessible tissue. In addition, gene networks are identified that may play a role in the pathophysiology of autism.

Independence and reproducibility across microarray platforms.

Microarrays have been widely used for the analysis of gene expression, but the issue of reproducibility across platforms has yet to be fully resolved. To address this apparent problem, we compared gene expression between two microarray platforms: the short oligonucleotide Affymetrix Mouse Genome 430 2.0 GeneChip and a spotted cDNA array using a mouse model of angiotensin II-induced hypertension. RNA extracted from treated mice was analyzed using Affymetrix and cDNA platforms and then by quantitative RT-PCR (qRT-PCR) for validation of specific genes. For the 11,710 genes present on both arrays, we assessed the relative impact of experimental treatment and platform on measured expression and found that biological treatment had a far greater impact on measured expression than did platform for more than 90% of genes, a result validated by qRT-PCR. In the small number of cases in which platforms yielded discrepant results, qRT-PCR generally did not confirm either set of data, suggesting that sequence-specific effects may make expression predictions difficult to make using any technique.

Multiple-laboratory comparison of microarray platforms.

Microarray technology is a powerful tool for measuring RNA expression for thousands of genes at once. Various studies have been published comparing competing platforms with mixed results: some find agreement, others do not. As the number of researchers starting to use microarrays and the number of cross-platform meta-analysis studies rapidly increases, appropriate platform assessments become more important. Here we present results from a comparison study that offers important improvements over those previously described in the literature. In particular, we noticed that none of the previously published papers consider differences between labs. For this study, a consortium of ten laboratories from the Washington, DC-Baltimore, USA, area was formed to compare data obtained from three widely used platforms using identical RNA samples. We used appropriate statistical analysis to demonstrate that there are relatively large differences in data obtained in labs using the same platform, but that the results from the best-performing labs agree rather well.

Iterative microarray and RNA interference-based interrogation of the SRC-induced invasive phenotype.

Src kinase has long been recognized as a factor in the progression of colorectal cancer and seems to play a specific role in the development of the metastatic phenotype. In spite of numerous studies conducted to elucidate the exact role of Src in cancer progression, downstream targets of Src remain poorly understood. Gene expression profiling has permitted the identification of large sets of genes that may be functionally interrelated but it is often unclear as to which molecular pathways they belong. Here we have developed an iterative approach to experimentally reconstruct a network of gene activity regulated by Src and contributing to the invasive phenotype. Our strategy uses a combination of phenotypic anchoring of gene expression profiles and loss-of-function screening by way of RNA-mediated interference. Using a panel of human colon cancer cell lines exhibiting differential Src-specific activity and invasivity, we identify the first two levels of gene transcription responsible for the invasive phenotype, where first-tier genes are controlled by Src activity and the second-tier genes are under the influence of the first tier. Specifically, perturbation of first-tier gene activity by either pharmacologic inhibition of Src or RNA-mediated interference-directed knockdown leads to a loss of invasivity and decline of second-tier gene activity. The targeting of first-tier genes may be bypassed altogether because knockdown of second-tier genes led to a similar loss of invasive potential. In this manner, numerous members of a "transcriptional cascade" pathway for metastatic activity have been identified and functionally validated.

Genetic regulation of endotoxin-induced airway disease.

To identify novel genes regulating the biologic response to lipopolysaccharide (LPS), we used a combination of quantitative trait locus (QTL) analysis and microarray-based gene expression studies of C57BL/6J x DBA/2J(BXD) F2 and recombinant inbred (RI) mice. A QTL affecting pulmonary TNF-alpha production was identified on chromosome 2, and a region affecting both polymorphonuclear leukocyte recruitment and TNF-alpha levels was identified on chromosome 11. Microarray analyses of unchallenged and LPS-challenged BXD RI strains identified approximately 500 genes whose expression was significantly changed by inhalation of LPS. Of these genes, 28 reside within the chromosomal regions identified by the QTL analyses, implicating these genes as high priority candidates for functional studies. Additional high priority candidate genes were identified based on their differential expression in mice having high and low responses to LPS. Functional studies of these genes are expected to reveal important molecular mechanisms regulating the magnitude of biologic responses to LPS.

Cardiac transcriptional response to acute and chronic angiotensin II treatments.

Exposure of experimental animals to increased angiotensin II (ANG II) induces hypertension associated with cardiac hypertrophy, inflammation, and myocardial necrosis and fibrosis. Some of the most effective antihypertensive treatments are those that antagonize ANG II. We investigated cardiac gene expression in response to acute (24 h) and chronic (14 day) infusion of ANG II in mice; 24-h treatment induces hypertension, and 14-day treatment induces hypertension and extensive cardiac hypertrophy and necrosis. For genes differentially expressed in response to ANG II treatment, we tested for significant regulation of pathways, based on Kyoto Encyclopedia of Genes and Genomes (KEGG) and Gene Microarray Pathway Profiler (GenMAPP) databases, as well as functional classes based on Gene Ontology (GO) terms. Both acute and chronic ANG II treatments resulted in decreased expression of mitochondrial metabolic genes, notably those for the electron transport chain and Krebs-TCA cycle; chronic ANG II treatment also resulted in decreased expression of genes involved in fatty acid metabolism. In contrast, genes involved in protein translation and ribosomal activity increased expression following both acute and chronic ANG II treatments. Some classes of genes showed differential response between acute and chronic ANG II treatments. Acute treatment increased expression of genes involved in oxidative stress and amino acid metabolism, whereas chronic treatments increased cytoskeletal and extracellular matrix genes, second messenger cascades responsive to ANG II, and amyloidosis genes. Although a functional linkage between Alzheimer disease, hypertension, and high cholesterol has been previously documented in studies of brain tissue, this is the first demonstration of induction of Alzheimer disease pathways by hypertension in heart tissue. This study provides the most comprehensive available survey of gene expression changes in response to acute and chronic ANG II treatment, verifying results from disparate studies, and suggests mechanisms that provide novel insight into the etiology of hypertensive heart disease and possible therapeutic interventions that may help to mitigate its effects.

Within the fold: assessing differential expression measures and reproducibility in microarray assays.

BACKGROUND: 'Fold-change' cutoffs have been widely used in microarray assays to identify genes that are differentially expressed between query and reference samples. More accurate measures of differential expression and effective data-normalization strategies are required to identify high-confidence sets of genes with biologically meaningful changes in transcription. Further, the analysis of a large number of expression profiles is facilitated by a common reference sample, the construction of which must be carefully addressed. RESULTS: We carried out a series of 'self-self' hybridizations in which aliquots of the same RNA sample were labeled separately with Cy3 and Cy5 fluorescent dyes and co-hybridized to the same microarray. From this, we can analyze the intensity-dependent behavior of microarray data, define a statistically significant measure of differential expression that exploits the structure of the fluorescent signals, and measure the inherent reproducibility of the technique. We also devised a simple procedure for identifying and eliminating low-quality data for replicates within and between slides. We examine the properties required of a universal reference RNA sample and show how pooling a small number of samples with a diverse representation of expressed genes can outperform more complex mixtures as a reference sample. CONCLUSION: Analysis of cell-line samples can identify systematic structure in measured gene-expression levels. A general procedure for analyzing cDNA microarray data is proposed and validated. We show that pooled reference samples should be based not only on the expression of individual genes in each cell line but also on the expression levels of genes within cell lines.