Pathway level analysis of gene expression using singular value decomposition
© Tomfohr et al; licensee BioMed Central Ltd. 2005
Received: 27 January 2005
Accepted: 12 September 2005
Published: 12 September 2005
A promising direction in the analysis of gene expression focuses on the changes in expression of specific predefined sets of genes that are known in advance to be related (e.g., genes coding for proteins involved in cellular pathways or complexes). Such an analysis can reveal features that are not easily visible from the variations in the individual genes and can lead to a picture of expression that is more biologically transparent and accessible to interpretation. In this article, we present a new method of this kind that operates by quantifying the level of 'activity' of each pathway in different samples. The activity levels, which are derived from singular value decompositions, form the basis for statistical comparisons and other applications.
We demonstrate our approach using expression data from a study of type 2 diabetes and another of the influence of cigarette smoke on gene expression in airway epithelia. A number of interesting pathways are identified in comparisons between smokers and non-smokers including ones related to nicotine metabolism, mucus production, and glutathione metabolism. A comparison with results from the related approach, 'gene-set enrichment analysis', is also provided.
Our method offers a flexible basis for identifying differentially expressed pathways from gene expression data. The results of a pathway-based analysis can be complementary to those obtained from one more focused on individual genes. A web program PLAGE (Pathway Level Analysis of Gene Expression) for performing the kinds of analyses described here is accessible at http://dulci.biostat.duke.edu/pathways.
Gene expression microarrays provide a snapshot of the expression levels of thousands of genes within a cell or tissue sample. A persistent challenge is to interpret this data: to identify key genes or patterns of expression associated with some condition and so to gain valuable clues about the biological processes related to that condition.
While a variety of methods have been developed to identify significant changes in the expression of individual genes [1–4], another useful perspective can be gained by viewing expression data at the level of groups of related genes. One approach along these lines identifies similarities, such as shared pathways or GO annotations , between genes previously identified in an individual gene analysis [6, 7]. A potential problem is that this approach relies on the individual genes within a category of interest to stand out. Modest but consistent changes in the expression of a group of related genes could be missed if relatively few of the individual genes appear significant.
A promising alternative focuses at the outset on identifying significantly differently expressed groups of genes from a collection of predefined sets of genes (e.g., pathways and complexes) [8, 9]. The usefulness of such an approach was strikingly demonstrated by Mootha et al.  in a study of gene expression profiles of muscle in type 2 diabetics (DM2). As reported by them, no single gene showed up as significant in a comparison between DM2s and subjects with normal glucose tolerance (NGT). Their 'gene-set enrichment analysis' (GSEA), however, uncovered a set of genes involved in oxidative phosphorylation as being significantly downregulated in DM2 vs. NGT.
In this article we present a new pathway based approach to the analysis of gene expression that, while similar in spirit to GSEA, has a number of potential advantages. Briefly, GSEA involves ranking all the genes (for example, by significance level in a two-group comparison) and then calculating an 'enrichment score' (ES) for each pathway that depends on the rankings of its member genes. Our method instead begins by translating gene expression levels into pathway 'activity' levels, which are derived from singular value decompositions (SVD). The activity levels are used for making comparisons and in general can be used in the same kinds of applications as gene expression levels.
We demonstrate the approach using the same expression data analyzed by Mootha et al.  in their study of type 2 diabetes, and also with expression data from airway epithelia of smokers and non-smokers . Our analysis leads us to conclusions similar to those obtained using GSEA in the diabetes set, but overall appears to perform better in identifying differentially expressed pathways in comparisons between smokers and non-smokers.
The results presented in this article, including statistics for pathways and colormaps of expression profiles, were obtained using a web program we have developed called PLAGE (Pathway Level Analysis of Gene Expression) .
Results and discussion
Outline of the method
In the next two sections we analyze gene expression data from skeletal muscle of type 2 diabetics and airway epithelia of different types of smokers. Here we give a brief overview of our approach. A more detailed description is given in the Methods section.
The main goal is to determine, based on the gene expression data, which (if any) of the pathways are associated with some variable of interest such as disease status. To address this, we start by calculating activity levels for each pathway within the samples (in this article, each sample is the gene expression profile in a tissue sample from one individual).
As a gene expression value represents the level of expression of a gene in some sample, the activity level represents the 'level' of the first metagene in a sample. The first metagene is simply a vector of weights, one weight for each gene, and a positive activity level indicates the relatively high (low) expression of genes with positive (negative) weight; a negative activity level indicates the reverse. The activity level in a given sample can be thought of as specifying the position of an expression profile (one column of the expression matrix) in a range of possible profiles as shown in Fig. 2C.
The main motivation for using the first metagene from SVD to define the activity level is that the weights (first metagene) and associated activity levels together capture the main component of the variation in the full expression matrix Y (Fig. 2B depicts the main component of the variation for the expression matrix in Fig. 2A). Generally, the higher metagenes may also contain meaningful structure but in this article we focus only on the first component of the variation. It may be useful in the future to devise a scheme for extending the analysis to higher metagenes.
Once the activity levels are determined, they can be used in the same kinds of applications as gene expression levels. For example, we might ask which pathways have activity levels that are significantly higher in samples from a case group (e.g., diabetic) than those from a control group. In this article, we mainly perform simple two-group comparisons using t statistics (see Methods) but, in principle, any statistical model for the level of expression of individual genes is adaptable to one for activity levels.
Type 2 diabetes
The data set analyzed by Mootha et al.  contains gene expression profiles in muscle tissue for each of 17 type 2 diabetics (DM2), 17 subjects with normal glucose tolerance (NGT), and 9 with impaired glucose tolerance (IGT). The gene expression data is available at the Whitehead Institute Center for Genome Research website  along with phenotype data including, for example, ages, body size measurements, blood glucose levels after oral glucose tolerance test (OGTT), and other information. More details can be found in Ref. .
The calculations of activity levels were done using all 43 gene expression profiles. Using the t statistic to compare mean activity levels identified no pathways showing apparently different expression between any of the groups DM2, IGT, and NGT. Specifically, the pathway with the highest significance level, 'Activation of cAMP-dependent protein kinase, PKA (protein kinase A)' (Biocarta) was found in comparing DM2 and NGT (upregulated in DM2); the significance level, however, was calculated to be only p = 0.4.
Pathways correlated with a type 2 diabetic phenotype. The table shows p-values for the three pathways most correlated with blood glucose concentration as measured two hours after an oral glucose tolerance test (OGTT). r is the Pearson correlation. Also shown are p-values, generally indicating low significance levels, for these pathways determined from t-statistic comparisons between DM2 (type 2 diabetic) and NGT (normal glucose tolerance).
(genes in data set/total genes in pathway) Pathway
glucose after OGTT
DM2 vs. NGT
(96/123) Oxidative phosphorylation
0.031 (r = -0.517)
0.565 (down in DM2)
(6/6) Activation of cAMP-dependent protein kinase
0.062 (r = +0.474)
0.395 (up in DM2)
(35/40) ATP synthesis
0.166 (r = -0.432)
0.855 (down in DM2)
The connection between blood glucose levels and oxidative phosphorylation (and ATP synthesis) seems biologically reasonable. Oxidative phosphorylation makes up the last few steps in the series of reactions leading to the synthesis of ATP from the oxidation of glucose. Unusually high blood glucose levels imply a relatively low activity of this pathway. Protein kinase A (PKA) facilitates the breakdown of glycogen into glucose in skeletal muscle cells ; an elevated PKA pathway in skeletal muscle of subjects exhibiting a type 2 diabetic phenotype may then possibly be interpreted as a cellular reaction to glucose starvation.
We note that our results differ slightly from those of Mootha et al.  who, using GSEA, found a significant (p = 0.029) downregulation of oxidative phosphorylation genes in diabetics compared to non. In contrast, we find reasonable significance only in the level of correlation with blood glucose levels after OGTT. This and other differences with GSEA are discussed in a comparison below.
Effects of smoking on airway epithelia
A recent study examined the effects of smoking on gene expression in airway epithelia . Expression data were obtained from a large number of subjects including former and current smokers, and those who have never smoked. The study identified genes differentially expressed between the different groups and some of the general functional categories represented by these genes. A pathway based analysis can be complementary to the kind already given by drawing attention to groups of genes involved in more specific cellular processes.
We obtained the gene expression data for the smoking study from the Airway Gene Expression Database (AGED) . The data consist of gene expression profiles from 75 subjects including 34 current smokers, 18 former smokers, and 23 subjects who have never been smokers. The data set has undergone some preprocessing steps including normalization and filtering for genes detected on the microarrays. More details can be found at the AGED website  and in Ref. .
Top pathways identified in comparisons between current (C), former (F), and never (N) smokers. Pathways identified as differentially expressed with p < 0.05 using pathway activity levels in comparisons between smokers and non-smokers. The p-values were determined using 10,000 random permutations as described in the Methods section. p < 0.0001 means no pathway in any of the 10,000 permutations showed higher significance. We note that the change (up or down) is determined, somewhat arbitrarily, by the average expression level captured by the first metagene. Specifically, the pathway is called 'up' if the average of ∑ i c j w i is greater in the first group (e.g., C in 'C vs. N') than in the second. A given pathway, however, will typically have some genes with higher and some with lower mean expression in one group as compared to another.
(genes in data set/total genes in pathway) Pathway
C vs. N
(14/45) gamma-Hexachlorocyclohexane degradation
(15/39) Prostaglandin and leukotriene metabolism
(11/24) O-Glycans biosynthesis
(6/21) Pentose and glucuronate interconversions
(24/34) Glutathione metabolism
(3/12) Lectin Induced Complement Pathway
(11/19) Chaperones modulate interferon Signaling Pathway
(6/15) TACI and BCMA stimulation of B cell immune responses.
(3/6) Tetrachloroethene degradation
(3/6) FXR and LXR Regulation of Cholesterol Metabolism
(4/7) TSP-1 Induced Apoptosis in Microvascular Endothelial Cell
(16/28) Galactose metabolism
(13/20) Biosynthesis of steroids
(7/11) Map Kinase Inactivation of SMRT Corepressor
(25/68) Nicotinate and nicotinamide metabolism
(4/14) Classical Complement Pathway
(6/19) Complement Pathway
(9/13) Nucleotide sugars metabolism
(3/3) Degradation of the RAR and RXR by the proteasome
F vs. C
(3/12) Lectin Induced Complement Pathway
(13/20) Biosynthesis of steroids
(24/34) Glutathione metabolism
F vs. N
(14/45) gamma-Hexachlorocyclohexane degradation
We also performed one-way ANOVA using pathway activity levels and found the top pathways from this analysis (those with the smallest p-values) to be virtually the same as those identified using t statistics.
Weights for genes in two pathways that show an association with smoking status. SVD weights for genes in pathways identified as significantly differentially expressed between current and never smokers. The gene with highest absolute weight in the gamma-Hexachlorocyclohexane degradation pathway, CYP2A6, plays a key role in nicotine metabolism and has been linked to nicotine dependence . GALNT3, a gene with relatively high weight in the O-Glycans biosynthesis pathway, initiates mucin-type O-glycosylation , suggesting a connection with the increased sputum production observed in smokers. The overall sign of the weights has here been chosen so that a positive weight implies relatively higher expression in current as compared to never smokers.
Other pathways upregulated in current vs. never smokers include metabolism of prostaglandins and leukotrienes (associated with pain response and inflammation), O-Glycans biosynthesis, and glutathione metabolism.
Genes from the O-Glycans pathway are listed in Table 3. Several of these are also within the list of genes identified by Spira et al.  The present analysis, however, reveals more directly the functional relationship shared by the genes and draws special attention to the role of this pathway in the physiological response to smoking. O-linked glycosylation plays an important role in the production of proteoglycans, some of which are constituents of mucus . In particular, one of the O-Glycans genes GALNT3 initiates mucin-type O-glycosylation . It therefore seems plausible that the elevated activity of the O-Glycans pathway is linked to the increased sputum production observed in smokers .
Glutathione is an important antioxidant known to be increased in the lungs of smokers . A colormap of the expression levels of the genes in the glutathione metabolism pathway is shown in Fig. 4; the higher expression of many of these genes in the current smokers is clear. The increased expression of glutathione metabolism genes in smokers was also noted by Spira et al. . Interestingly, the two genes with the highest weights are GCLM and GCLC; these genes encode the subunits of glutamate cysteine ligase (GCL), which is the rate-limiting enzyme in the synthesis of glutathione .
Comparison with gene-set enrichment analysis
We used GSEA to identify significant pathways in comparisons between the different groups of smokers. The ranking of genes required to evaluate enrichment scores was done, following Ref. , using the signal to noise ratio (the absolute value of the difference of the means of the two groups divided by the sum of the within-group standard deviations). For this analysis we used our collection of pathways and complexes. In comparing current and never smokers, the gene-set with the highest enrichment score (ES) was determined to be the KEGG ribosome genes with an ES of 162. However, using 1000 random permutations as described in Ref.  to evaluate the significance of this ES yielded a p-value of only 0.17. Comparing former and never smokers, GSEA also finds the ribosome gene-set to have the highest ES (304) and with a reasonably high significance level of p = 0.047 (1000 permutations). Finally, the comparison between current and former smokers identifies 'Lectin Induced Complement Pathway' (Biocarta) as the pathway with the highest ES (128) but the significance level is very low (p = 0.52).
We also used GSEA with our collection of gene-sets to compare DM2 and NGT in the diabetes data discussed above. Here the ribosome gene-set is found to have the highest ES (278) and p = 0.073 (1000 permutations). Oxidative phosphorylation has the next highest ES (261) with p = 0.092.
To summarize, our approach suggests a number of potentially interesting pathways and hypotheses from comparisons between smokers and non-smokers. The results of our analysis seem biologically reasonable and are more consistent with the findings of Spira et al.  than those obtained by GSEA. As a specific example, we identified glutathione metabolism genes as having significantly (p < 0.0001) higher expression levels in current vs. never smokers. This feature was also noted by Spira et al.  in their original analysis and seems evident in the colormap shown in Fig. 4. GSEA, however, does not find this or any other pathway to be differentially expressed below p = 0.05 between current and never smokers.
We have introduced a method for analyzing gene expression data in terms of a collection of predefined pathways and complexes. In summary, the approach quantifies the level of activity of each pathway within each sample and uses the activity levels as the basis for making comparisons. Mootha et al.  previously demonstrated that looking at expression data in terms of predefined pathways can provide valuable insights not easily attainable by methods more focused on individual genes and we feel that the results in this article reinforce their view.
The main applications we have presented are simple two-group comparisons. We suggest that pathway activity levels may also be useful in other areas, for example, as variables in models for gene regulation (e.g., Bayesian networks ), for clustering and classification [15, 26], and in computational efforts to discover novel gene-pathway associations. One possibility is to use SVD on the activity levels themselves to determine 'metapathway' signatures for the status of some disease following the approach described in Ref. . There are a number of potential advantages to be gained in these and other cases. For one, features defined by groups of genes will tend to be more robust in the face of variation at the level of individual genes. In addition, the features here (pathways) are based on a large amount of prior biological knowledge and so are potentially more directly biologically meaningful.
As in GSEA, the basis for our analysis is a classification of genes into known pathways and complexes. Here, we use about 400 pathways and complexes characterized on the Biocarta and KEGG (Kyoto Encyclopedia of Genes and Genomes) websites [12, 13]. The KEGG pathways mainly include processes related to metabolism and biosynthesis (e.g., Fatty acid metabolism and Ubiquinone biosynthesis). Those on Biocarta cover a wider variety of cellular processes including a large number of immune signaling pathways (Toll-like receptor pathway, B-cell receptor signaling pathway, complement pathway), as well as metabolic and biosynthetic pathways. Both collections also include complexes such as the T-cell receptor and ribosome and broad gene categories such as cytokines. The KEGG and Biocarta pathways continue to expand and evolve. The versions used for this article are available at the PLAGE website .
Pathway activity levels
Our analysis starts by quantifying, in each sample, the level of 'activity' of each pathway. As described in the Results and discussion section, we define the activity level of a pathway in a given sample as the 'level' of expression of a certain metagene in that sample, i.e., the level of the first metagene from the SVD (e.g., Ref ) of the matrix of expression levels. In this section we provide the details and motivation behind this choice.
We begin by standardizing the gene expression levels to have zero mean and unit variance over samples. For each pathway, we form a matrix Y (rows = genes, columns = samples) containing the standardized expression levels from all samples but for the genes in that pathway only. We write the singular value decomposition of Y as
Y = WDC. (1)
Here the columns of the matrix W are the orthonormal (W⊤W = I, the identity matrix) eigenvectors or metagenes of Y, D is a diagonal matrix containing the associated eigenvalues, and each column of C is a vector of coefficients for one of the samples indicating the level of each metagene in the sample. The rows of C are also orthonormal (CC⊤ = I). Assume the eigenvalues are ordered from highest to lowest going down the diagonal of D. The first metagene w – that associated with the largest eigenvalue – is then the first column of W. We write its eigenvalue as λ and the associated coefficients (first row of C) as c j .
The activity level of a pathway in a given sample j is taken as the coefficient c j for the first metagene. It follows also from the orthonormality of the columns of W and rows of C that
That is, the activity level c j can also be regarded (up to a non-essential scale factor) as a weighted sum of the standardized expression levels of the individual genes, the weights being given by the first metagene w.
One motivation for using the first metagene in SVD is that the resulting combination of activity levels and weights specifies an optimal approximation to the matrix Y (i.e., accounts for the main component of the variation in the data). Specifically, assume the following statistical model for the expression levels
y ij = α i χ j + ε ij (3)
where the vector χ is constrained to have unit norm and the ε ij are independent Gaussian random variables. The estimates for α and χ that minimize the sum of the squared errors are just the first metagene scaled by its eigenvalue, λ w, and the associated vector of activity levels c, respectively. The approximation of a set of expression profiles using the first metagene is illustrated in Fig. 2.
A useful fact about the first eigenvalue is that its square is a measure of the amount of variation accounted for by the first metagene. Specifically, with n g = number of genes and n s = number of samples, the total amount of variation in the data is (recall, the expression levels are standardized so that ) and the variation remaining after subtracting off the profile described by the first metagene is ∑ ij (y ij - λ w i c j )2 = n g (n s - 1) - λ2.
In this article, we mainly perform pairwise comparisons to identify pathways for which the mean activity level in one group (e.g., diabetic) is significantly different from that in the other (non-diabetic). To accomplish this we calculate a t statistic for each pathway:
where A and B are labels for the groups, n A is the number of samples in group A, and μ A and V A are the mean and variance of the activity level in A (similarly for group B). To determine a fair measure of significance in an analysis of this kind, it is important to account for the fact that we are testing a large number of hypotheses . For this purpose, we perform a large number (10,000 for this article) of comparisons by randomly permuting the sample labels and for each permutation recording the t statistic for the most significant pathway (the maximum t-statistic) identified using the permuted labels. Our p-values are computed with reference to the maximum t-statistic, i.e., we calculate a p-value by taking the fraction of the maximum t-statistics that exceed the t-statistic.
This basic approach was also used by Mootha et al.  to determine the significance of the ES (enrichment score) for a pathway except that the comparison there is with maximum enrichment scores. Generally, other statistics can be used in similar fashion. For example, to evaluate the significance of the level of correlation of activity levels with blood glucose concentration in the diabetes data, we determined the minimum and maximum (Pearson) correlations between activity levels and randomly permuted glucose concentration values in each of 10,000 permutations. p-values were obtained as the fraction of these extremal correlation values that were stronger (higher if r > 0 and lower if r < 0) than the value obtained using the correct ordering of glucose levels.
Availability and requirements
The SVD-based pathways analysis method has been implemented in a web program  called PLAGE (pathway level analysis of gene expression). PLAGE will run through standard web browsers.
We gratefully acknowledge the financial support of the NIH through the Duke University Center for Translational Research (5 P30 AI051445-03) and of the NSF through a grant to David Bird (NCSU; DBI 0077503) as well as the Duke Center for Bioinformatics and Computational Biology. We also thank two anonymous reviewers for many helpful suggestions.
- Tusher VG, Tibshirani R, Chu G: Significance analysis of microarrays applied to the ionizing radiation response. Proc Natl Acad Sci USA 2001, 98: 5116–21. 10.1073/pnas.091062498PubMed CentralView ArticlePubMed
- Efron B, Tibshirani R, Storey JD, Tusher V: Empirical Bayes analysis of a microarray experiment. J Amer Statist Assoc 2001, 96: 1151–1160. 10.1198/016214501753382129View Article
- Baldi P, Long AD: A Bayesian framework for the analysis of microarray expression data: regularized t-test and statistical inferences of gene changes. Bioinformatics 2001, 17: 509–19. 10.1093/bioinformatics/17.6.509View ArticlePubMed
- Kepler TB, Crosby L, Morgan KT: Normalization and analysis of DNA microarray data by self-consistency and local regression. Genome Biol 2002, 3: research0037. 10.1186/gb-2002-3-7-research0037PubMed CentralView ArticlePubMed
- Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, Harris MA, Hill DP, Issel-Tarver L, Kasarskis A, Lewis S, Matese JC, Richardson JE, Ringwald M, Rubin GM, Sherlock G: Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet 2000, 25: 25–9. 10.1038/75556PubMed CentralView ArticlePubMed
- Doniger SW, Salomonis N, Dahlquist KD, Vranizan K, Lawlor SC, Conklin BR: MAPPFinder: using Gene Ontology and GenMAPP to create a global gene-expression profile from microarray data. Genome Biol 2003, 4: R7. 10.1186/gb-2003-4-1-r7PubMed CentralView ArticlePubMed
- Zeeberg BR, Feng W, Wang G, Wang MD, Fojo AT, Sunshine M, Narasimhan S, Kane DW, Reinhold WC, Lababidi S, Bussey KJ, Riss J, Barrett JC, Weinstein JN: GoMiner: a resource for biological interpretation of genomic and proteomic data. Genome Biol 2003, 4: R28. 10.1186/gb-2003-4-4-r28PubMed CentralView ArticlePubMed
- Kurhekar MP, Adak S, Jhunjhunwala S, Raghupathy K: Genome-wide pathway analysis and visualization using gene expression data. Pac Symp Biocomput 2002, 462–73.
- Mootha VK, Lindgren CM, Eriksson KF, Subramanian A, Sihag S, Lehar J, Puigserver P, Carlsson E, Ridderstrale M, Laurila E, Houstis N, Daly MJ, Patterson N, Mesirov JP, Golub TR, Tamayo P, Spiegelman B, Lander ES, Hirschhorn JN, Altshuler D, Groop LC: PGC-1alpha-responsive genes involved in oxidative phosphorylation are coordinately downregulated in human diabetes. Nat Genet 2003, 34: 267–73. 10.1038/ng1180View ArticlePubMed
- Spira A, Beane J, Shah V, Liu G, Schembri F, Yang X, Palma J, Brody JS: Effects of cigarette smoke on the human airway epithelial cell transcriptome. Proc Natl Acad Sci USA 2004, 101: 10143–8. 10.1073/pnas.0401422101PubMed CentralView ArticlePubMed
- Pathway Level Analysis of Gene Expression[http://dulci.biostat.duke.edu/pathways]
- Kyoto Encyclopedia of Genes and Genomes[http://www.genome.jp/kegg]
- Alter O, Brown PO, Botstein D: Singular value decomposition for genome-wide expression data processing and modeling. Proc Natl Acad Sci USA 2000, 97: 10101–6. 10.1073/pnas.97.18.10101PubMed CentralView ArticlePubMed
- West M, Blanchette C, Dressman H, Huang E, Ishida S, Spang R, Zuzan H, Olson JAJ, Marks JR, Nevins JR: Predicting the clinical status of human breast cancer by using gene expression profiles. Proc Natl Acad Sci USA 2001, 98: 11462–7. 10.1073/pnas.201162998PubMed CentralView ArticlePubMed
- Whitehead Institute Center for Genome Research website[http://www.broad.mit.edu/cancer]
- Alberts B, Johnson A, Lewis J, Raff M, Roberts K, Walter P: Molecular Biology of the Cell. New York: Garland; 2002.
- Airway Gene Expression Database[http://pulm.bumc.bu.edu/aged]
- Sellers EM, Tyndale RF: Mimicking gene defects to treat drug dependence. Ann NY Acad Sci 2000, 909: 233–46.View ArticlePubMed
- Ariyoshi N, Miyamoto M, Umetsu Y, Kunitoh H, Dosaka-Akita H, Sawamura Y, Yokota J, Nemoto N, Sato K, Kamataki T: Genetic polymorphism of CYP2A6 gene and tobacco-induced lung cancer risk in male smokers. Cancer Epidemiol Biomarkers Prev 2002, 11: 890–4.PubMed
- Tyndale RF, Sellers EM: Variable CYP2A6-mediated nicotine metabolism alters smoking behavior and risk. Drug Metab Dispos 2001, 29: 548–52.PubMed
- Gu C, Oyama T, Osaki T, Li J, Takenoyama M, Izumi H, Sugio K, Kohno K, K Y: Low expression of polypeptide GalNAc N-acetylgalactosaminyl transferase-3 in lung adenocarcinoma: impact on poor prognosis and early recurrence. Br J Cancer 2004, 90: 436–42. 10.1038/sj.bjc.6601531PubMed CentralView ArticlePubMed
- Dwyer TM: Cigarette smoke-induced airway inflammation as sampled by the expired breath condensate. Am J Med Sci 2003, 326: 174–8. 10.1097/00000441-200310000-00004View ArticlePubMed
- Neurohr C, Lenz AG, Ding I, Leuchte H, Kolbe T, Behr J: Glutamate-cysteine ligase modulatory subunit in BAL alveolar macrophages of healthy smokers. Eur Respir J 2003, 22: 82–7. 10.1183/09031936.03.00080403View ArticlePubMed
- Friedman N: Inferring cellular networks using probabilistic graphical models. Science 2004, 303: 799–805. 10.1126/science.1094068View ArticlePubMed
- Eisen MB, Spellman PT, Brown PO, Botstein D: Cluster analysis and display of genome-wide expression patterns. Proc Natl Acad Sci USA 1998, 95: 14863–8. 10.1073/pnas.95.25.14863PubMed CentralView ArticlePubMed
- Strang G: Introduction to Linear Algebra. Cambridge: Wellesley; 2003.
- Dudoit S, Shaffer JP, Boldrick JC: Multiple hypothesis testing in microarray experiments. Stat Sci 2003, 18: 71–103. 10.1214/ss/1056397487View Article
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.