Bioinformatics Advance Access originally published online on October 25, 2005
Bioinformatics 2006 22(1):77-87; doi:10.1093/bioinformatics/bti737
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Integrating time-course microarray gene expression profiles with cytotoxicity for identification of biomarkers in primary rat hepatocytes exposed to cadmium
,*
1Cedars-Sinai Research Institute, Cedars-Sinai Medical Center Los Angeles, CA 90048, USA
2Center for Toxicoinformatics, National Center for Toxicological Research FDA, Jefferson, AR 72079, USA
3Applied Biotechnology Branch, US Air Force Research Laboratory WPAFB, OH 45434, USA
4David Geffen School of Medicine, UCLA Los Angeles, CA 90048, USA
*To whom correspondence should be addressed.
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMETHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
Motivation: DNA microarrays can provide information about the expression levels of thousands of genes simultaneously at the transcriptomic level, while conventional cell viability and cytotoxicity measurement methods provide information about the biological functions at the cellular level. Integrating these data at different levels provides a promising approach for evaluating or predicting how cells respond to chemical exposure. It is important to investigate the multi-scale biological system in a systematic way to better understand the gene regulation networks and signal transduction pathways involved in the cellular responses to environmental factors.
Results: Primary rat hepatocytes were exposed to cadmium acetate at 0, 1.25 and 2 µM. mRNA expression profiles at 0, 3, 6, 12 and 24 h were measured using the Affymetrix RatTox U34 GeneChip® arrays. Simultaneously, cytotoxicity was assessed by lactase dehydrogenase leakage assay. Gene expression profiles at different time points were used to evaluate cytotoxicity at subsequent time points using partial least squares, and it was found that gene expression profiles at 0 h had the best prediction accuracy for the cytotoxicity observed at 12 h. Some biomarkers whose expression profiles showed strong relationship with cytotoxicity were identified and the underlying pathways were reconstructed to illustrate how hepatocytes respond to cadmium exposure. Permutation studies were also applied to assess the reliability of the predictive models.
Availability: Matlab source code is available upon request and DNA microarray data are available at GEO (http://www.ncbi.nlm.nih.gov/geo).
Contact: cwang61{at}ucla.edu
Supplementary information: Supplementary data are available at Bioinformatics online.
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
The heavy metal cadmium is ranked number 7 on the top 20 hazardous substances priority list by the Agency for Toxic Substances and Disease Registry (Fay and Mumtaz, 1996; Liao and Freeman, 1998). It has also been categorized as a group 1 human carcinogen (International Agency for Research on Cancer, 1993) and a potent multi-tissue animal carcinogen (Waalkes and Misra, 1996). Cadmium has an extremely long biological half-life (10–30 years) that makes it a cumulative toxin and carcinogen primarily in liver and kidney (Goering et al., 1994). Occupational or environmental exposure to cadmium is associated with lung cancers in humans (International Agency for Research on Cancer Monographs, 1993; Waalkes, 2000) and with the development of cancers of the prostate, kidney, liver, hematopoietic system and stomach (International Agency for Research on Cancer Monographs, 1993; Waalkes, 1995; Waalkes and Misra, 1996; Waalkes and Rehm, 1994).
Traditional toxicological testing utilizes whole animals to test the potential toxicity of chemicals. These tests are expensive, time-consuming, and provide little information on early, low dose effects in target cells. Primary rat hepatocytes have been used extensively for testing chemical and drug toxicity by pharmaceutical companies. It is believed that any toxic chemical will disturb the gene expression pattern and, as a result, may give rise to a specific ‘fingerprint’ of chemical toxicity. By monitoring thousands of genes in cells simultaneously, DNA microarray technology provides the opportunities to characterize the patterns of gene expression induced by chemicals, to understand the biological effects and mechanisms on a genome-wide scale and to make tailored therapeutics to specific pathologies possible (Brown and Botstein, 1999; Young, 2000). Since changes in the transcriptomic level usually precede the toxic outcomes, i.e. there is a time delay in the signal transduction from the gene expression at mRNA level to toxic endpoint at the cellular level, gene expression profiling offers an opportunity for us to identify the early, sensitive indicators of cellular toxicity resulting from chemical exposure, and it also provides a more comprehensive view of molecular responses induced by chemicals and drugs (Aardema and MacGregor, 2002; Kier et al., 2004).
The advances in high-throughput technologies provide unprecedented opportunities to investigate the complex biology system at multi-scales, from genetic and transcriptomic levels to proteomic and metabonomic levels. There is an increasing need to link the large amount of information from different scales to gain insight into the biological functions, pathways and dynamics at the system level, leading to the emergence of an exciting new field called systems biology (Hood et al., 2004; Hood and Perlmutter, 2004; Li and Chan, 2004; Oltvai and Barabasi, 2002). The use of gene expression to evaluate the phenotypic endpoints such as chemical and drug toxicity and efficacy of drugs is an emerging interdiscipline of systems biology and toxicology called toxicogenomics (Aardema and MacGregor, 2002; Hamadeh et al., 2002; Lehmann, 2003; de Longueville et al., 2004; Irwin et al., 2004; Morgan et al., 2004). It is expected that the gene expression profiles will allow the discrimination among mechanisms of toxic damage, the prediction of specific toxicities and the identification of molecular biomarkers that signal a particular toxic or carcinogenic mechanism (Aardema and MacGregor, 2002; Irwin et al., 2004; Suter et al., 2004). The results from toxicogenomics will not only result in considerable savings with regard to time, cost and animal use relative to conventional methods but also make certain human studies that could not be carried out at overtly toxic exposures possible (Aardema and MacGregor, 2002). Toxicogenomics has found promising applications in terms of chemical or drug toxicity and efficacy (Kier et al., 2004; Staunton et al., 2001; Vekris et al., 2004).
One challenge associated with the application of microarray gene expression data is that the number of variables (genes) is much larger than the number of observations (samples), resulting in a high degree of collinearity in the gene expression profiles. Direct application of conventional statistical techniques often results in ill-posed or computationally infeasible problems (Park et al., 2002). To deal with this kind of severely ill-conditioned problems, one common way is to reduce the dimensionality of the gene expression data. Principal component analysis (PCA) is a popular method to serve this purpose by finding a set of orthogonal principal components (linear combinations of the original variables) to account for the maximum variations in gene expression data. Since the information in the response variables such as cytotoxicity is not taken into account in constructing the principal components in PCA, the performance of PCA in prediction or classification may not be optimal (Tan et al., 2005). To overcome this problem, partial least squares (PLS), developed by Wold et al. (1984, 2001), was used to find orthogonal linear combinations of the original predictor variables, which highly correlate with the response variables, while accounting for as much variance in predictors as possible. PLS is powerful in analyzing data with strong collinear (correlated), noisy and numerous X-variables (Wold et al., 2001), and has been widely used in microarray data analysis in terms of tumor classification and survival time prediction (Ding and Gentleman, 2004; Fort and Lambert-Lacroix, 2005; Huang and Pan, 2003; Nguyen and Rocke, 2002a, b,c; Park et al., 2002; Pérez-Enciso and Tenenhaus, 2003; Tan et al., 2004).
The objectives of this study include (1) evaluating and predicting the cytotoxicity of primary rat hepatocytes exposed to cadmium acetate using gene expression profiles; (2) identifying early biomarkers at mRNA level that are strongly related to Cd-induced toxic outcomes and can be used to predict the cytotoxicity at a later stage and (3) elucidating the roles of these genes involving in the cellular responses to Cd exposure by reconstructing the underlying signal transduction network.
|
METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
Experiment design
The experimental design is shown in Figure 1. Briefly, primary rat hepatocytes were isolated according to a reported procedure with minor modifications (Delraso and Frazier, 1999; Hussain and Frazier, 2003; Karpinets et al., 2004) and were exposed to cadmium acetate (0, 1.25 and 2.0 µM) for 2 h. The doses of cadmium exposure were chosen based on the dose–response studies of cytotoxicity assessed by lactase dehydrogenase (LDH) leakage as well as 3-(4,5-dimethylthiazol-2-yl)-2,5-diphenyltetrazolium bromide (MTT) reduction assays. The lower dose (1.25 µM) was the lowest observable adverse effect level based on cytotoxicity as measured by LDH leakage and MTT reduction. At this dose, a slight depression in mitochondrial function at the end of the 2 h exposure was observed, but the cells recovered rapidly to normal after the removal of the media containing Cd. The higher dose (2.0 µM) resulted in a significant depression of mitochondrial function, showing
20% of mitochondrial function at 24 h post-exposure. As shown in Figure 1, cells were collected at -2 (pretreatment control), 0, 3, 6, 12 and 24 h in all three groups (0, 1.25 and 2.0 µM Cd) for cytotoxicity evaluation by LDH leakage as well as for mRNA expression profiling by DNA microarray. The parallel baseline controls (0 dose) were run at each time point with cells exposed to cadmium-free medium and they served as time-matched controls.
|
DNA microarray
Affymetrix GeneChip® oligonucleotide arrays (RatTox U34) were used for mRNA expression profiling. The microarray experiment was repeated with hepatocytes from 3 animals, each with 2 replicates (independent cultures) for each dosage (3 dosage) at each time point (5 time points), resulting in a total of 90 chips (3 animals x 2 replicates x 3 doses x 5 time points). Total RNA was isolated from cultured primary hepatocytes (2 million cells pooled from two wells of a 6-well plate at each dosage of each time point) using Qiagen RNeasy Mini kit (Qiagen, Valencia, CA). DNA microarray experiment was performed according to a procedure as previously described (Wang et al., 2005) with some modifications.
LDH leakage
Membrane damage that results in LDH leakage is generally considered irreversible; therefore, LDH leakage was used as an indicator of cellular viability. LDH leakage was assessed by measuring the activity of LDH in the cells and in the media as described previously (Moldeus et al., 1978; Hussain and Frazier, 2003; Kikkawa et al., 2005). The percentage of LDH activity was calculated by dividing the amount of activity in the medium by the total activity (medium and cell lysate). No significant difference in LDH leakage was found in controls (cadmium-free medium) at each time point.
Partial Least Squares
The basic idea of various projection or dimension-reduction approaches, e.g. PCA and PLS, is to project the observations (samples) from the high-dimensional variables (genes) space to a low-dimensional subspace spanned by several linear combinations of the original variables to satisfy a certain objective criterion (Tan et al., 2005). PCA attempts to find a set of orthogonal principal components to account for the maximum variance in predictors (X). Since there is no guarantee that the principal components representing the maximum variance in X should necessarily be the components strongly related to response variables (Y), the performance of PCA may be not satisfactory from the predictive point of view. To overcome the problem with PCA, PLS was developed by constructing a set of orthogonal components that maximize the sample covariance between response and the linear combination of predictor variables (Wold et al., 1984, 2001).
The objective criterion of PLS can be written as
 | (1) |
 | (2) |
Although PLS was originated from chemometrics, it has been widely implemented in bioinformatics, especially in microarray data analysis, because of its powerful capability in analyzing data with strong collinear (correlated), noisy and numerous X-variables (Wold et al., 2001). For example, Nguyen and Rock used PLS as a dimension reduction method for binary and multiple cancer classification (Nguyen and Rocke, 2002b,c). PLS discriminant analysis was used for prediction of clinical outcome with microarray data by Pérez-Enciso and Tenenhaus (2003) and Tan et al. (2004). Generalized PLS has been used to analyze microarray gene expression data for cancer classification (Ding and Gentleman, 2004; Fort and Lambert-Lacroix, 2005); and PLS has also been applied for prediction of survival time of cancer patients (Park et al., 2002; Nguyen and Rocke, 2002a).
When PLS is used in toxicogenomics, the response matrix (Y) contains the information about the toxic or pathological outcome such as cytotoxicity, whereas the predictor matrix (X) represents the microarray gene expression profiles with each column corresponding to the expression levels of a gene across different samples. The PLS model can be formulated as follows:
 | (3) |
The PLS estimation of regression coefficients (BPLS) can be calculated as follows:
 | (4) |
The prediction of response variables on a new set of samples is made by the following equation:
 | (5) |
The PLS algorithm used in this paper can be found in the paper of Höskuldsson (1988) and it was implemented using Matlab 7.0 R14 (The Mathworks Inc., Natick, MA). The selection of the optimal number of genes and the optimal number of PLS components was based on leave-one-out cross validation (LOOCV) (according to the maximum 
.
Gene preselection
Variable (gene) selection is important for the successful analysis of gene expression data since most of the genes are supposedly unchanged by chemical treatment, thus they are irrelevant to the prediction of phenotypic measurement. The inclusion of those non-informative genes in the modeling process will degrade the performance of the model. Variables that do not contribute to response should be removed before the analysis. There are different approaches for gene selection such as neighborhood analysis (Golub et al., 1999), significance analysis of microarrays (Tusher et al., 2001), Wilks' lambda (Stephanopoulos et al., 2002), t-score and critical score (Nguyen and Rocke, 2002b,c), classifier feedback approach (Bicciato et al., 2003) and nearest shrunken centroids (Tibshirani et al., 2002).
In this study, the sum of squared correlation coefficients between gene expression and cytotoxicity was used to select a subset of genes for subsequent analysis with PLS (Tan et al., 2004; Tan et al., 2005). For example, the g* = 100 genes are taken as the first 100 genes with the largest values of sum of squared correlation coefficients. The numbers of genes, g*, preselected for PLS analysis in this study are 2–100 (with an interval of 1), 500 and all the genes (972).
Leave-one-out cross-validation
LOOCV has become a standard procedure to evaluate the performance of various prediction or classification methods in microarray data analysis. Note that when gene selection or dimension reduction is used together with a LOOCV procedure, a pitfall was to perform gene selection or dimension reduction before CV loop. However, such incomplete LOOCV procedure is known to be substantially biased and prone to generating spuriously good results since the information about all samples is used for gene selection or dimension reduction before CV loop (Ambroise and McLachlan, 2002; Simon et al., 2003a,b; Tan et al., 2004). In this study, the complete LOOCV (the gene selection and dimension reduction within the CV loop) was applied.
Evaluation of cytotoxicity using gene expression at earlier time points
Changes in gene expression at transcriptomic level are thought to precede the toxic outcome (Aardema and MacGregor, 2002) and are often more sensitive and characteristic of the toxic response than other cytotoxicity indicators at cellular level such as commonly used LDH release. There is a time delay during the signal cascade from the gene expression at transcriptomic level to cytotoxicity measured by LDH release at cellular level as a response to toxic damage (Kier et al., 2004). Therefore, in this study, the microarray gene expression profiles at a certain time point was used to predict the cytotoxicity at subsequent time points to take into account this kind of time delay in the information flow from transcriptomic level to cellular level. For example, gene expression profiles at time 0 h was used to evaluate the cytotoxicity measured at time 0, 3, 6, 12 and 24 h. And there were 15 combinations of biology-relevant prediction models (Table 1). For each of the 15 prediction models, microarray gene expression data at a certain time point (0, 3, 6, 12, 24 h), including 9 samples (3 animals and 3 dosages, mRNA expression level from two replicate arrays of same animal at the same dosage were averaged) and 972 variables (representing about 800 genes), were used to predict the cytotoxicity measured by LDH release at later stages (0, 3, 6, 12, 24 h); the optimal number of genes [g*: 2–100 (with interval of 1), 500 and all the genes] and the optimal number of PLS components (ranging from 1 to 20) were chosen based on LOOCV procedure (the highest R2cv).
|
|
RESULTS AND DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
Cytotoxicity measured by LDH release
LDH leakage that resulted from membrane damage because of chemical exposure is generally considered irreversible, therefore LDH leakage has been used as an indicator of cellular viability and cytotoxicity (Hussain and Frazier, 2003). Figure 2 shows the effects of cadmium exposure on LDH release in primary rat hepatocytes. There was no significant LDH leakage in cultured hepatocytes exposed to Cd at the two dosages up to 6 h. However, a significant percentage of LDH leakage was observed at both 12 and 24 h after Cd exposure at the higher dosage (2.0 µM) in all three rats, with a dramatic increase of LDH release at 24 h (Fig. 2). There was no significant change in LDH releases in cultured primary rat hepatocytes exposed to the lower dosage of Cd (1.25 µM). Overall, there was a clear time- and dose-dependent effect of Cd-induced cytotoxicity in cultured primary rat hepatocytes.
|
Prediction of cytotoxicity (LDH leakage) using gene expression profiles
Microarray gene expression data included 45 samples (3 animals, 3 dosages and 5 time points). mRNA expression derived from two replicate arrays of independently cultured cells for each animal with the same treatment were averaged in the analyses. There are 972 probe sets representing
800 important toxicology-related genes in the RT U34 array. As mentioned previously, DNA microarray gene expression profiles at certain time points, including 9 samples (3 animals and 3 dosages), were used to evaluate the cytotoxicity (LDH leakage) at subsequent time points, resulting in 15 biological relevant prediction models as shown in Table 1. Since the objective of PLS method is to maximize the sample covariance between response (Y) and linear combination of predictor variables (X), the relationship or correlation between the gene expression (X) and the effect of cadmium treatment (cytotoxicity) measured by LDH release at different dosage (Y) was taken into account intrinsically in the PLS model. And the variations across time course was also taken into account using gene expression at a certain time point to predict the toxic endpoint at subsequent time points. The best prediction result (goodness of prediction, R2cv), the optimal number of genes and the optimal number of PLS components for each prediction model were determined based on the LOOCV procedure (according to the maximum R2cv) and is presented in Table 1. Although the R2cv values ranged from 0.29 to 0.92, only 4 out of the 15 R2cv values were lower than 0.5. In most cases, the gene expression profiles at a certain time point (except for time 12 h) predicted well the cytotoxicity at the subsequent time points with relatively high R2cv. This means that the idea of using gene expression at a certain time point to predict the cytotoxicity at later stages did work. In other words, the changes of mRNA expression profiles due to Cd exposure are highly related to the cytotoxicity (LDH leakage) in a time-delay manner. As shown in Table 1, for gene expression profiles at a fixed time point, the highest R2cv generally came from the model that used gene expression data at this time point to predict the cytotoxicity at a later time point, rather than from using the same gene expression data to predict the cytotoxicity at the same time point. There may be several possible biological explanations behind this. Firstly, time delay during the information flow from the gene expression at mRNA level to the toxic endpoint at cellular level does exist and can be captured by a well-designed time-course microarray gene expression experiment. Secondly, such an important time-delay response can be characterized to some degree using the gene expression profiles at a certain time point to predict the cytotoxicity at subsequent time points. Interestingly, the best prediction performance (R2cv = 0.92) came from the model using the gene expression profiles at time 0 h (right after 2 h of exposure) to evaluate the cytotoxicity at time 12 h, suggesting that the genes selected from the gene expression profiles at time 0 h may serve as a set of early biomarkers for evaluating cytotoxicity at time 12 h. DNA microarray gene expression profiles at 3 and 6 h also showed a good prediction performance for the cytotoxicity at subsequent time points.
Identification of candidate biomarkers at mRNA level
As shown in Table 1, cytotoxicity (LDH release) can be evaluated and predicted using only a subset of genes, suggesting that this subset of genes could serve as biomarkers at transcriptomic level in terms of evaluating cytotoxicity or cell viability and may also shed lights on the mechanism of cellular response to toxic chemicals. Note that during the LOOCV procedure, different subsets of genes (with the number of genes fixed) were selected when each sample was taken out from the total samples and reserved for prediction. A frequency number for each gene was calculated based on the number of times the gene was selected during LOOCV and was used to represent the importance of the gene in the prediction of cytotoxicity. Genes with frequency number larger than 6 were arbitrarily selected as potential biomarkers (note that the maximum frequency for each gene being selected during the LOOCV is 9, the total number of samples for each time point). The procedure of selecting candidate biomarkers based on the frequency of occurrence of variables has been widely used in bioinformatics, especially in high-throughput data analysis for biomarker discovery, and the underlying assumption is that the most frequently selected features are the most relevant candidate biomarkers (Zhou and Mao, 2005; Somorjai et al., 2003; Li and Chan, 2004; Liu et al., 2005; Kim et al., 2005). If a gene was only selected for 1 or 2 times during the LOOCV, the chance of its relevance to the prediction of cytotoxicity would be low. However, if a gene was selected for most of the times during LOOCV, it would have a better chance to be related to the cytotoxicity. In other words, the response of this gene is more possibly due to cadmium exposure. Since not all of the prediction models listed in Table 1 showed good prediction performance, only eight models with R2cv larger than 0.65 were used for selection of biomarkers. The cut-off value of R2cv 0.65 was chosen based on a trade-off between the prediction accuracy and the number of selected candidate biomarkers, as well as some biological knowledge about the possible cadmium-related biomarker. The prediction models with R2cv lower than or equal to 0.50 could not be trusted. If the cut-off value is chosen as 0.62 (3 h 
6 h) or 0.64 (6 h >12 h), too many candidate biomarkers would be produced. If the cut-off value is chosen as 0.66 (3 h 24 h), a few important candidate biomarkers, such as heme oxygenase, may be missed. Table 2 showed a portion of the selected toxicological-relevant biomarkers which may be actively involved in stress response of primary rat hepatocyte to cadmium exposure. The functional classification was partially based on the Rat Toxicology (RT) gene array database (Karpinets et al., 2004) and the complete list of genes selected by PLS can be found in the supplementary table.
|
Assessment of the reliability of prediction models by permutation analysis
It is important to evaluate the stability and reliability of a prediction model, given the relatively small sample size of microarray data compared with the number of variable. There are various statistical methods available to assess the reliability when there are no enough samples available to perform external validation (Bicciato et al., 2003; Nguyen and Rocke, 2002c; Tan et al., 2004, 2005). In this study, permutation or shuffle studies were applied to compare the observed prediction performance, R2cv, with those expected at random. At first, the cytotoxicity values (LDH leakage) from all samples (the rows of dependent variable Y) were permuted while keeping the microarray gene expression profiles (X matrix) unchanged. Then, the newly generated random dataset with shuffled Y and unchanged X was analyzed by PLS using exactly the same LOOCV procedure as applied to the original dataset (number of genes and number of PLS components were chosen exactly in the same way as those chosen to obtain the best prediction performance for original datasets, as shown in boldface in Table 1). This procedure was carried out for 100 times. The distributions of R2cv over 100 permutations for situations from which the 8 prediction models with R2cv higher than 0.65 were developed are plotted in Figure 3 and compared with the prediction performance obtained from original datasets. It is obvious that, in all cases, R2cv obtained by PLS for the original dataset is significantly higher than those from the permuted random datasets.
|
Relevance and significance of candidate biomarkers identified by PLS modeling
To assess the biological relevance of the identified candidate biomarkers, we evaluated the biological roles or functions of the candidate biomarkers to see if they are truly related to the biological problem being addressed (Li and Chan, 2004).
The detoxification process plays a critical role in getting rid of or neutralizing a wide range of toxic chemicals either produced internally or absorbed from the environment (Rushmore and Kong, 2002; Schiaffonati and Tiberio, 1997; Sheweita, 2000). This enzymatic process usually occurs in two steps involving both phase I and phase II enzyme systems. Generally speaking, phase I enzymes such as cytochrome P450s (CYP450s) are involved in the activation of chemicals, converting chemicals to more active intermediates (Sheweita, 2000), whereas phase II enzymes such as glutathione S-transferases (GSTs) are involved in the detoxification by conjugating an endogenous substances (e.g. glutathione, GSH) to the toxic chemical, leading to derivatives that are more water-soluble and can be easily excreted from urine or bile (Hinson and Forkert, 1995; Townsend and Tew, 2003). As shown in Table 2, many genes identified by our modeling belong to the phase I and phase II enzymes. CYP450s are a group of enzymes that play a critical role in the phase I detoxification, although in most time, they convert chemicals to more active forms. Ten probe sets of CYP450s were found to be related to the LDH release in the primary rat hepatocytes exposed to Cd (Table 2). PLS model also identified many phase II enzymes such as GSTs, sulfotransferase (ST) and UDP-glucuronosyltransferase (UGT). GSTM1 and GSTA5 are the members of GST superfamily of enzymes, catalyzing conjugation of GSH with chemicals. The GSH conjugation produces water-soluble mercaptates that are excreted via kidneys (Hinson and Forkert, 1995; Tawe et al., 1998; Townsend and Tew, 2003). Sulfotransferases (SULT1A1, SULT1A2 and STE) encode another family of phase II detoxification enzymes that are involved in the conjugation of toxins with sulfur-containing compounds to detoxify the toxic chemicals (Goldstein and Faletto, 1993). UGT2B12 and UGT1A6 are members of the UGT superfamily of phase II enzymes that catalyze the glucuronidation of endogenous compounds as well as xenobiotics including toxin and carcinogens (Gueraud and Paris, 1998; Zheng et al., 2002). Some other detoxification enzymes such as EPHX1, EPHX2, NAD(P)H, and quinone reductase were also identified (Dicker and Cederbaum, 1993; Fretland and Omiecinski, 2000).
In response to stresses, including heat, oxidizing conditions and exposure to toxic compounds, cells can produce heat shock proteins (HSPs) to protect the cells against stress-induced damage. Upon exposure to cadmium, cellular proteins are abnormally denatured, and meanwhile certain cell stress response genes such as HSPs are stimulated (Papaconstantinou et al., 2003; Boone and Vijayan, 2002). Most of the time, HSPs act as molecular chaperones to help denatured proteins refold. Seven probe sets representing HSPs such as Hsp8, Hspb1, Hspa1a and Hspca were identified as a response to the cadmium toxicity in primary rat hepatocytes by our modeling.
In addition, many mitogen-activated protein kinases (MAPKs) such as MAPK3k1, MAPK6 and MAPK14 were identified by our modeling. These genes are important components of the MAPK pathway, which is apparently regulated by Cd exposure, mostly likely due to Cd-induced oxidative stress. MAPKs include a large number of serine/threonine kinases involved in regulating a wide range of cellular processes such as proliferation, stress adaptation and apoptosis (Martindale and Holbrook, 2002). The c-jun N-terminal kinases (JNK) and the p38 kinases are often related to the stress adaptation and referred to as stress-activated protein kinases (Kolch, 2000; Martindale and Holbrook, 2002). They can be activated by a wide range of stresses such as radiation, osmotic shock, mechanical injury, heat stress and oxidative damage (Kolch, 2000; Martindale and Holbrook, 2002).
In addition, some DNA repair enzymes such as Adprt and Polb as well as genes related to protein repair and chaperoning, e.g. Dnaja1 and Cryab, were also identified. These proteins are involved in repairing Cd-induced membrane and DNA damages or function as chaperones to protect cells from injuring (Liao and Freeman, 1998).
In this paper, two methods widely used in bioinformatics were applied to validate the relevance of selected candidate biomarkers: (1) since there is not enough data available for external validation, permutation or randomization studies were applied to compare the observed prediction performance of the selected biomarkers with those expected at random, in order to demonstrate that the high prediction capability of the selected biomarkers is not obtained by chance. And this validation method is a typical one used in bioinformatics (Nguyen and Rocke, 2002b; Simon et al., 2003a; Tan et al., 2005); (2) the biological roles or functions of the candidate biomarkers were also evaluated by reviewing literature to see if they are truly related to the biological problem being addressed, in order to further assess the relevance of the identified candidate biomarkers (Golub et al., 1999; Li and Chan, 2004). However, it should be pointed out that the identification of candidate biomarkers from the viewpoint of bioinformatics and high-throughput data analysis is only the first step in the screening of biomarkers and in the understanding of the complicated molecular mechanism of cadmium toxicity. Further biological studies, such as testing the selected candidate biomarkers in different cell systems or different species, would nicely complement the ‘in-silico’ validation and provide stronger evidence about the relevance of candidate biomarkers to cadmium toxicity.
Reconstruction of pathway or gene network of Cd-induced stress responses
Intracellular damages induced by cadmium exposure include protein denaturation, lipid peroxidation and DNA strand break and damage (Liao and Freeman, 1998). The proposed mechanisms include binding of metal to reduced cysteine residues and production of reactive oxygen species (ROS), possibly by lowering reduced GSH levels (Abe et al., 1994; Chin and Templeton, 1993; Liao and Freeman, 1998; Manca et al., 1991). To attenuate the toxic effects of cadmium, cells respond by activating a variety of genes involved in many processes such as scavenging ROS, chelating the metal to prevent further damage, repairing membrane and DNA damage or functioning as chaperones and degrading unfolded proteins (Liao and Freeman, 1998; Liao et al., 2002). However, little is known about how those stress-activated genes are orchestrated in responding to the cadmium-induced toxicity. In this paper, based on the biological function of the genes selected by our prediction models and the available literature knowledge, a hypothetic pathway or gene network of signal transduction was reconstructed to illustrate the underlying mechanism.
A schematic illustration of the reconstructed pathways or network with respect to how cells respond to the Cd-induced oxidative stress is shown in Figure 4. The whole process of cell response to Cd-induced toxicity is the result of an orchestrated set of events and involved a variety of stress-activated pathways. Mammalian cells can respond to a variety of stresses such as heat, oxidative stress, metabolic disturbance and environmental toxins through necrotic or apoptotic cell death. Stress signals are delivered to p38 MAPK cascade by members of small GTPases of the Rho family (Cdc42). P38 MAPK (Mapk14) is involved in regulation of Hsp27 phosphorylation and several transcription factors such as Myc that regulate genes involved in apoptosis (Ichijo, 1999; Lewis et al., 1998; Tibbles and Woodgett, 1999). The increased expression and phosphorylation of heat shock proteins such as Hsp27 can protect cells from cytotoxic responses and apoptosis through mechanisms such as chaperone activity, decreasing the release of cytochrome C and its binding activity in cytosol (Bruey et al., 2000). Hsp27 appears to play a role in preventing cells from Cd-induced damaging by altering the oxidative environment of cells through induction of GSH expression, as well as by blocking apoptosis induced by ROS (Wyttenbach et al., 2002). Growth arrest and DNA-damage-inducible 45 alpha (Gadd45a) may also be involved in the induction of apoptosis and cell cycle arrest by maintaining p38/JNK MAPK activation (Hildesheim et al., 2002).
|
Apoptosis is a cell suicide mechanism to eliminate individual cells that threaten animal's survival (Ashkenazi and Dixit, 1998). Under the severe damage induced by stress, the death receptors such as tumor necrosis factor receptor superfamily, member 1a (Tnfrsf1a) detect the presence of extracellular death signals and, in response, they rapidly ignite the cell's intrinsic apoptosis machinery by transmitting apoptosis signals and activating death cascades (Ashkenazi and Dixit, 1998, 1999). Binding of tumor necrosis factor (TNF) to TNF receptor activates Map kinase, Map3k1 (Mekk1), which stands at the top of MAPKs pathways leading to transcriptional regulation, i.e. phosphorylation of c-Jun (Chadee et al., 2005; Natoli et al., 1997; Xia et al., 2000). Map3k1 is a potent activator for JNK group of MAPKs such as Mapk10 (Minden et al., 1994; Xia et al., 2000; Morel and Barouki, 1999). The activation of p38 MAPK (Mapk14) by Map3k1 leads to the transcriptional activation of many stress and growth-related genes (Yuasa et al., 1998). The activation of immediately early gene Jun may in turn activate the transcription of detoxification enzymes (Morel and Barouki, 1999; Schiaffonati and Tiberio, 1997) and induce a variety of antioxidant genes such as catalase (Cat), heme oxygenase 2 (Hmox2), Ugtla6 and glutathione reductase (Gsr) which are critical in preventing the accumulation and deleterious effects of ROS induced by stress (Morel and Barouki, 1999). In addition, caspase can be indirectly induced by TNF 1 receptor (Muzio et al., 1997) and is involved in the proteolytic cleavage of ADP-ribosyltransferase 1 (Adprt) during apoptosis (Faleiro et al., 1997; Gu et al., 1995). The bcl211 gene is regulated by signal transducer and activator of transcription 5 (Stat5) (Bui et al., 2001; Socolovsky et al., 1999) and is involved in the activation of JNK (Srivastava et al., 1999) and inhibition of cleavage of Adprt (Datta et al., 1997; Emoto et al., 1995).
|
CONCLUSION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
In this study, microarray gene expression profiles at a certain time point were used to evaluate the cytotoxicity at subsequent time points using PLS. The relatively high R2cv obtained from prediction models as well as the low R2cv obtained from the permutation studies indicate that it is possible to use transcriptomic profiles at early time points to evaluate the cadmium-induced toxicity at later time points. Some genes were identified as potential biomarkers to predict cytotoxicity, i.e. LDH leakage, in cultured primary rat hepatocytes exposed to Cd. Based on the functions of genes selected using PLS, a hypothetic pathway/network was also reconstructed to illustrate how cells respond to the stress induced by Cd.
|
Acknowledgments |
|---|
The authors would like to thank Dr. Walter J. Kozumbo of the US Air Force Office of Scientific Research for his support of this study. The authors also would like to thank Dr. Darol E. Dodd, Program Director of Alion Science and Technology, Dayton, OH, for his support during the process of this study. The authors gratefully acknowledge the assistance of Dr. Victor Chan (Alion Science and Technology, Inc., Dayton, OH) with several phases of this project. The authors also want to thank the reviewers for helpful suggestions. This study was supported by the US Air Force Office of Scientific Research (AFOSR) Project (JON 2312A205) and all laboratory work related to data collection were completed at the Genomics and Proteomics Lab, the US Air Force Research Laboratory, WPAFB, OH.
Conflict of Interest: none declared.
|
FOOTNOTES |
|---|
Associate Editor: Alvis Brazma
The authors wish it to be known that, in their opinion, the last two authors should be regarded as joint Senior Authors. 
Received on June 10, 2005; revised on September 28, 2005; accepted on October 20, 2005
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTS AND DISCUSSIONCONCLUSIONREFERENCES |
|---|
Aardema, M.J. and MacGregor, J.T. (2002) Toxicology and genetic toxicology in the new era of ‘toxicogenomics’: impact of ‘-omics’ technologies. Mutat. Res, . 499, 13–25[Web of Science][Medline].
Abe, T., et al. (1994) Induction of heat shock 70 mRNA by cadmium is mediated by glutathione suppressive and non-suppressive triggers. Biochim. Biophys. Acta, 1201, 29–36[Medline].
Ambroise, C. and McLachlan, G.J. (2002) Selection bias in gene extraction on the basis of microarray gene-expression data. Proc. Natl Acad. Sci., USA, 99, 6562–6566
Ashkenazi, A. and Dixit, V.M. (1998) Death receptors: signaling and modulation. Science, 281, 1305–1308
Ashkenazi, A. and Dixit, V.M. (1999) Apoptosis control by death and decoy receptors. Curr. Opin. Cell Biol, . 11, 255–260[CrossRef][Web of Science][Medline].
Bicciato, S., et al. (2003) PCA disjoint models for multiclass cancer a analysis using gene expression data. Bioinformatics, 19, 571–578
Boone, A.N. and Vijayan, M.M. (2002) Constitutive heat shock protein 70 (HSC70) expression in rainbow trout hepatocytes: effect of heat shock and heavy metal exposure. Comp. Biochem. Physiol. C Toxic. Pharmacol, . 132, 223–233[CrossRef].
Brown, P.O. and Botstein, D. (1999) Exploring the new world of the genome with DNA microarrays. Nat. Genet, . 21, 33–37[CrossRef][Web of Science][Medline].
Bruey, J.M., et al. (2000) Hsp27 negatively regulates cell death by interacting with cytochrome c. Nat. Cell Biol, . 2, 645–652[CrossRef][Web of Science][Medline].
Bui, N.T., et al. (2001) Activation of nuclear factor B and bcl-x survival gene expression by nerve growth factor requires tyrosine phosphorylation of IkB
. J. Cell Biol, . 152, 753–764
Chadee, D.N., et al. (2005) Direct activation of mitogen-mctivated protein kinase kinase kinase MEKK1 by the Ste20p homologue GCK and the adapter protein TRAF2. Mol. Cell. Biol, . 22, 737–749.
Chin, T.A. and Templeton, D.M. (1993) Protective elevations of glutathione and metallothionein in cadmium-exposed mesangial cells. Toxicology, 77, 145–156[CrossRef][Web of Science][Medline].
Datta, R., et al. (1997) Activation of a CrmA-insensitive, p35-sensitive pathway in ionizing radiation-induced apoptosis. J. Biol. Chem, . 272, 1965–1969
de Longueville, F., et al. (2004) DNA microarrays as a tool in toxicogenomics. Comb. Chem. High Through. Screen, 7, 207–211.
Delraso, N.J. and Frazier, J.M. (1999) Effect of culture conditions prior to exposure on cadmium cytotoxicity in primary rat hepatocytes. Toxic. Meth, . 9, 97–114[CrossRef].
Dicker, E. and Cederbaum, A.I. (1993) Requirement for iron for the production of hydroxyl radicals by rat liver quinone reductase. J. Pharmacol. Exp. Ther, . 266, 1282–1290
Ding, B.Y. and Gentleman, R. (2005) Classification using generalized partial least squares. J. Comput. Graph. Stat, . 14, , pp. 280–298[CrossRef].
Emoto, Y., et al. (1995) Proteolytic activation of protein kinase C delta by an ICE-like protease in apoptotic cells. EMBO J, . 14, 6148–6156[Web of Science][Medline].
Faleiro, L., et al. (1997) Multiple species of CPP32 and Mch2 are the major active caspases present in apoptotic cells. EMBO J, . 16, 2271–2281[CrossRef][Web of Science][Medline].
Fay, R.M. and Mumtaz, M.M. (1996) Development of a priority list of chemical mixtures occurring at 1188 hazardous waste sites, using the HazDat database. Fd. Chem. Toxic, . 34, 1163–1165[CrossRef][Web of Science][Medline].
Fort, G. and Lambert-Lacroix, S. (2005) Classification using partial least squares with penalized logistic regression. Bioinformatics, 21, 1104–1111
Fretland, A.J. and Omiecinski, C.J. (2000) Epoxide hydrolases: biochemistry and molecular biology. Chem. Biol. Interact, . 129, 41–59[CrossRef][Web of Science][Medline].
Goering, P.L., Waalkes, M.P., Klaassen, C.D. (1994) Toxicology of metals, biochemical effects. In Goyer, R.A. and Cherian, M.G. (Eds.). Handbook of Experimental Pharmacology, , NY Springer Vol. 115, , pp. 189–214.
Goldstein, J.A. and Faletto, M.B. (1993) Advances in mechanisms of activation and deactivation of environmental chemicals. Environ. Health. Perspect, . 100, 169–176[Web of Science][Medline].
Golub, T.R., et al. (1999) Molecular classification of cancer class discovery and class prediction by gene expression monitoring. Science, 286, 531–537
Gu, Y., et al. (1995) Cleavage of poly (ADP-ribose) polymerase by interleukin-1 beta converting enzyme and its homologs TX and Nedd-2. J. Biol. Chem, . 270, 18715–18718
Gueraud, F. and Paris, A. (1998) Glucuronidation: a dual control. Gen. Pharmacol, . 31, 683–688[Web of Science][Medline].
Hamadeh, H.K., et al. (2002) An overview of toxicogenomics. Curr. Issues Mol. Biol, . 4, 45–56[Medline].
Hildesheim, J., et al. (2002) Gadd45a protects against UV irradiation-induced skin tumors, and promotes apoptosis and stress signaling via MAPK and p53. Cancer Res, . 62, 7305–7315
Hinson, J.A. and Forkert, P.G. (1995) Phase II enzymes and bioactivation. Can. J. Physiol. Pharmacol, . 73, 1407–1413[Web of Science][Medline].
Hood, L., et al. (2004) Systems biology and new technologies enable predictive and preventative medicine. Science, 306, 640–643
Hood, L. and Perlmutter, R.M. (2004) The impact of systems approaches on biological problems in drug discovery. Nat. Biotechnol, . 22, 1215–1217[CrossRef][Web of Science][Medline].
Höskuldsson, A. (1988) PLS regression methods. J. Chemom, . 2, 211–228[CrossRef].
Huang, X.H. and Pan, W. (2003) Linear regression and two-class classification with gene expression data. Bioinformatics, 19, 2072–2078
Hussain, S.M. and Frazier, J.M. (2003) Involvement of apoptosis in hydrazine induced toxicity in rat primary hepatocytes. Toxic. in Vitro, 17, 343–355.
Ichijo, H. (1999) From receptors to stress-activated MAP kinases. Oncogene, 18, 6087–6093[CrossRef][Web of Science][Medline].
International Agency for Research on Cancer Monographs. (1993) Cadmium. International Agency for Research on Cancer Monographs, , Lyon IARC Press Vol. 58, , pp. 119–238.
IARC International Agency for Research on Cancer Monographs on the Evaluation of Carcinogenic Risks to Humans. (1993) Beryllium, cadmium, mercury and exposures in the glass manufacturing industry. , Lyon, France 58, 41–117.
Irwin, R.D., et al. (2004) Application of toxicogenomics to toxicology: basic concepts in the analysis of microarray data. Toxicol. Pathol, . 32, Suppl. 1, 72–83
Karpinets, T., et al. (2004) Tailored gene array databases: applications in mechanistic toxicology. Bioinformatics, 20, 507–517
Kier, L.D., et al. (2004) Applications of microarrays with toxicologically relevant genes (tox genes) for the evaluation of chemical toxicants in Sprague Dawley rats in vivo and human hepatocytes in vitro. Mutat. Res, . 549, 101–113[Web of Science][Medline].
Kikkawa, R., et al. (2005) Investigation of a hepatotoxicity screening system in primary cell cultures—‘what biomarkers would need to be addressed to estimate toxicity in conventional and new approaches?’. J. Toxicol. Sci, . 30, 61–72[CrossRef][Medline].
Kim, B.S., et al. (2005) Statistical methods of translating microarray data into clinically relevant diagnostic information in colorectal cancer. Bioinformatics, . 21, 517–528
Kolch, W. (2000) Meaningful relationships: the regulation of the Ras/Raf/MEK/ERK pathway by protein interactions. Biochem. J, . 351, 289–305[CrossRef][Web of Science][Medline].
Lehmann, J. (2003) Toxicogenomics takes center stage. Drug News Perspect, . 16, 183–186[Medline].
Lewis, T.S., et al. (1998) Signal transduction through MAP kinase cascades. Adv. Cancer Res, . 74, 49–139[Web of Science][Medline].
Li, Z. and Chan, C. (2004) Integrating gene expression and metabolic profiles. J. Biol. Chem, . 279, 27124–27137
Liao, V.H.C., et al. (2002) Molecular characterization of a novel, cadmium-inducible gene from the Nematode Caenorhabditis elegans. J. Biol. Chem, . 277, 42049–42059
Liao, V.H.C. and Freeman, J.H. (1998) Cadmium-regulated genes from the nematode Caenorhabditis elegans. Identification and cloning of new cadmium-responsive genes by differential display. J. Biol. Chem, . 273, 31962–31970
Liu, J.J., et al. (2005) Multiclass cancer classification and biomarker discovery using GA-based algorithms. Bioinformatics, 21, 2691–2697
Manca, D., et al. (1991) Studies on lipid peroxidation in rat tissues following administration of low and moderate doses of cadmium chloride. Toxicology, 67, 303–323[CrossRef][Web of Science][Medline].
Manne, R. (1987) Analysis of two partial-least-squares algorihms for multivariate calibration. Chemo. Intell. Lab. Syst, . 2, 187–197[CrossRef].
Martindale, J.L. and Holbrook, N.J. (2002) Cellular response to oxidative stress: signaling for suicide and survival. J. Cell. Physiol, . 192, 1–15[CrossRef][Web of Science][Medline].
Minden, A., et al. (1994) Differential activation of ERK and JNK mitogen-activated protein kinases by Raf-1 and MEKK. Science, 266, 1719–1723
Moldeus, P., et al. (1978) Isolation and use of liver cells. Meth. Enzymol, . 52, 60–71[Medline].
Morel, Y. and Barouki, R. (1999) Repression of gene expression by oxidative stress. Biochem. J, . 342, 481–496[CrossRef][Web of Science][Medline].
Morgan, K.T., et al. (2004) Complementary roles for toxicologic pathology and mathematics in toxicogenomics, with special reference to data interpretation and oscillatory dynamics. Toxicol. Pathol, 32, Suppl. 1, 13–25
Muzio, M., et al. (1997) FLICE induced apoptosis in a cell-free system. J. Biol. Chem, . 272, 2952–2956
Natoli, G., et al. (1997) Activation of SAPK/JNK by TNF receptor 1 through a noncytotoxic TRAF2-dependent pathway. Science, 275, 200–203
Nguyen, D.V. and Rocke, D.M. (2002a) Partial least squares proportional hazard regression for application to DNA microarray survival data. Bioinformatics, 18, 1625–1632
Nguyen, D.V. and Rocke, D.M. (2002b) Multi-class cancer classification via partial least squares with gene expression profiles. Bioinformatics, 18, 1216–1226
Nguyen, D.V. and Rocke, D.M. (2002c) Tumor classification by partial least squares using microarray gene expression data. Bioinformatics, 18, 39–50
Oltvai, Z.N. and Barabasi, A.L. (2002) Life's complexity pyramid. Science, 298, 763–764
Papaconstantinou, A.D., et al. (2003) Mercury, cadmium, and arsenite enhance heat shock protein synthesis in chick embryos prior to embryotoxicity. Birth Def. Res. B Dev. Reprod. Toxic, . 68, 456–464[CrossRef].
Park, P.J., et al. (2002) Linking gene expression data with patient survival times using partial least squares. Bioinformatics, 18, S120–S127[Abstract].
Pérez-Enciso, M. and Tenenhaus, M. (2003) Prediction of clinical outcome with microarray data: a partial least squares discriminant analysis (PLS-DA) approach. Hum. Genet, . 112, 581–592[Web of Science][Medline].
Rushmore, T.H. and Kong, A.N. (2002) Pharmacogenomics, regulation and signaling pathways of phase I and II drug metabolizing enzymes. Curr. Drug Metab, . 3, 481–490[CrossRef][Web of Science][Medline].
Schiaffonati, L. and Tiberio, L. (1997) Gene expression in liver after toxic injury: analysis of heat shock response and oxidative stress-inducible genes. Liver, 17, 183–191[Web of Science][Medline].
Sheweita, S.A. (2000) Drug-metabolizing enzymes: mechanisms and functions. Curr. Drug Metab, . 1, 107–132[Medline].
Simon, R., et al. (2003a) Pitfalls in the use of DNA microarray data for diagnostic and prognostic classification. J. Natl Cancer Inst, . 95, 14–18
Simon, R. (2003b) Diagnostic and prognostic prediction using gene expression profiles in high-dimensional microarray data. Br. J. Cancer, 89, 1599–1604[CrossRef][Web of Science][Medline].
Socolovsky, M., et al. (1999) Fetal anemia and apoptosis of red cell progenitors in Stat5a-/-5b-/- mice: a direct role for Stat5 in Bcl-X(L) induction. Cell, 98, 181–191[CrossRef][Web of Science][Medline].
Somorjai, R.L., et al. (2003) Class prediction and discovery using gene microarray and proteomics mass spectroscopy data: curses, caveats, cautions. Bioinformatics, 19, 1484–1491
Srivastava, R.K., et al. (1999) Bcl-2 and Bcl-XL Block Thapsigargin-Induced Nitric Oxide Generation, c-Jun NH2-Terminal Kinase Activity, and Apoptosis. Mol. Cell. Biol, . 19, 5659–5674
Staunton, J.E., et al. (2001) Chemosensitivity prediction by transcriptional profiling. Proc. Natl Acad. Sci., USA, 98, 10787–10792
Stephanopoulos, G., et al. (2002) Mapping physiological states from microarray expression measurements. Bioinformatics, 18, 1054–1063
Suter, L., et al. (2004) Toxicogenomics in predictive toxicology in drug development. Chem. Biol, . 11, 161–171[CrossRef][Web of Science][Medline].
Tan, Y.X., et al. (2004) Multi-class tumor classification by discriminant partial least squares using microarray gene expression data and assessment of classification models. Comput. Biol. Chem, . 28, 235–244[CrossRef][Web of Science][Medline].
Tan, Y.X., et al. (2005) Multi-class cancer classification by total principal component regression (TPCR) using microarray gene expression data. Nucleic Acids Res, . 33, 56–65
Tawe, W.N., et al. (1998) Identification of stress-responsive genes in Caenorhabditis elegans using RT–PCR differential display. Nucleic Acids Res, . 26, 1621–1627
Tibbles, L.A. and Woodgett, J.R. (1999) The stress-activated protein kinase pathways. Cell. Mol. Life Sci, . 55, 1230–1254[CrossRef][Web of Science][Medline].
Tibshirani, R., et al. (2002) Diagnosis of multiple cancer types by shrunken centroids of gene expression. Proc. Natl Acad. Sci. USA, 99, 6567–6572
Townsend, D.M. and Tew, K.D. (2003) The role of glutathione-S-transferase in anti-cancer drug resistance. Oncogene, 22, 7369–7375[CrossRef][Web of Science][Medline].
Tusher, V.G., et al. (2001) Significance analysis of microarrays applied to the ionizing radiation response. Proc. Natl Acad. Sci. USA, 98, 5116–5121
Vekris, A., et al. (2004) Molecular determinants of the cytotoxicity of platinum compounds: the contribution of in silico research. Cancer Res, . 64, 356–362
Waalkes, M.P. Handbook on Metal–Ligand Interactions of Biological Fluids, (1995) , NY Marcel Dekker.
Waalkes, M.P. (2000) Cadmium carcinogenesis in review. J. Inorg. Biochem, . 79, 241–244[CrossRef][Web of Science][Medline].
Waalkes, M.P. and Misra, R.R. Toxicology of Metals, (1996) , Boca Raton, FL CRC Press.
Waalkes, M.P. and Rehm, S. (1994) Chronic toxic and carcinogenic effects of cadmium chloride in male DBA/2NCr and NFS/NCr mice: strain-dependent association with tumors of the hematopoietic system, injection site, liver, and lung. Fund. Appl. Toxic, . 23, 21–31[CrossRef].
Wang, C., et al. (2005) Transcriptomic fingerprinting of bone marrow-derived hepatic beta2m-/Thy-1+ stem cells. Biochem. Biophys. Res. Commun, . 327, 252–260[CrossRef][Web of Science][Medline].
Wold, S., et al. (1984) The collinearity problem in linear regression, the partial least squares approach to generalized inverse. SIAM J. Sci. Stat. Comput, . 5, 735–743[CrossRef].
Wold, S., et al. (2001) PLS-regression: a basic tool of chemometrics. Chemom. Intell. Lab. Syst, . 58, 109–130[CrossRef].
Wyttenbach, A., et al. (2002) Heat shock protein 27 prevents cellular polyglutamine toxicity and suppresses the increase of reactive oxygen species caused by huntingtin. Hum. Mol. Genet, . 11, 1137–1151
Xia, Y., et al. (2000) MEK kinase 1 is critically required for c-Jun N-terminal kinase activation by proinflammatory stimuli and growth factor-induced cell migration. Proc. Natl Acad. Sci., USA, 97, 5243–5248
Young, R.A. (2000) Biomedical discovery with DNA arrays. Cell, 102, 9–15[CrossRef][Web of Science][Medline].
Yuasa, T., et al. (1998) Tumor necrosis factor signaling to stress-activated protein kinase (SAPK)/Jun NH2-terminal kinase (JNK) and p38. J. Biol. Chem, . 273, 22681–22692
Zheng, Z., et al. (2002) Glucuronidation: an important mechanism for detoxification of Benzo[a]Pyrene metabolites in aerodigestive tract tissues. Drug Metab. Dispos, . 30, 397–403
Zhou, X. and Mao, K.Z. (2005) LS bound based gene selection for DNA microarray data. Bioinformatics, 21, 1559–1564
CiteULike 
Connotea 
Del.icio.us What's this?
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||