Original Article

The Pharmacogenomics Journal (2009) 9, 137–146; doi:10.1038/tpj.2008.17; published online 23 December 2008

A new statistical screening approach for finding pharmacokinetics-related genes in genome-wide studies

Y Sato1,2, N M Laird1, K Nagashima3, R Kato3, T Hamano4, A Yafune5, N Kaniwa6,7, Y Saito6,8, E Sugiyama6,7, S-R Kim6, J Furuse9, H Ishii9, H Ueno10, T Okusaka10, N Saijo11, J-i Sawada6,8 and T Yoshida2

  1. 1Department of Biostatistics, Harvard School of Public Health, Boston, MA, USA
  2. 2Genetics Division, National Cancer Center Research Institute, Tokyo, Japan
  3. 3Faculty of Engineering, Tokyo University of Science, Tokyo, Japan
  4. 4Hamano Statistical Analysis Ltd., Tokyo, Japan
  5. 5Clinic Sendagaya, Tokyo, Japan
  6. 6Project Team for Pharmacogenetics, National Institute of Health Sciences, Tokyo, Japan
  7. 7Division of Medicinal Safety Science, National Institute of Health Sciences, Tokyo, Japan
  8. 8Division of Functional Biochemistry and Genomics, National Institute of Health Sciences, Tokyo, Japan
  9. 9Hepatobiliary and Pancreatic Oncology Division, National Cancer Center Hospital East, Chiba, Japan
  10. 10Hepatobiliary and Pancreatic Oncology Division, National Cancer Center Hospital, Tokyo, Japan
  11. 11National Cancer Center Hospital East, Chiba, Japan

Correspondence: Dr T Yoshida, Genetics Division, National Cancer Center Research Institute, 5-1-1 Tsukiji, Chuo-ku, Tokyo 104-0045, Japan. E-mail: tyoshida@ncc.go.jp

Received 7 August 2008; Revised 11 November 2008; Accepted 19 November 2008; Published online 23 December 2008.

Top

Abstract

Biomedical researchers usually test the null hypothesis that there is no difference of the population mean of pharmacokinetics (PK) parameters between genotypes by the Kruskal–Wallis test. Although a monotone increasing pattern with a number of alleles is expected for PK-related genes, the Kruskal–Wallis test does not consider a monotonic response pattern. For detecting such patterns in clinical and toxicological trials, a maximum contrast method has been proposed. We show how that method can be used with pharmacogenomics data to a develop test of association. Further, using simulation studies, we compare the power of the modified maximum contrast method to those of the maximum contrast method and the Kruskal–Wallis test. On the basis of the results of those studies, we suggest rules of thumb for which statistics to use in a given situation. An application of all three methods to an actual genome-wide pharmacogenomics study illustrates the practical relevance of our discussion.

Keywords:

gemcitabine, genome-wide study, maximum contrast method, pharmacokinetics-related gene, statistical screening method

Top

Introduction

Interindividual variation in drug response among patients is well known and poses a serious problem in medicine. The variation could be because of multiple factors such as disease phenotypes, genetic and environmental factors and the variability in drug target response (pharmacodynamic response) or allergic response, all factors that affect drug absorption, distribution, metabolism and excretion, side effects or efficacy.1, 2, 3 However, at present, few biomarkers can predict, which group of patients will respond positively, which patients are nonresponders and who might experience adverse reactions for the same medication and dose. To realize personalized medicine, it is critically important to observe individual differences in drug response and the role of genetic polymorphisms that are relevant to the pathways of drug metabolism and the biology of drug responses in the pharmacogenomics of common diseases.4

With this background, the Food and Drug Administration (FDA) recognizes the importance of pharmacogenomics and has issued a guidance that encourages pharmacogenomics during drug development.5 Many pharmacogenomics studies have been launched worldwide, such as a combination of a pharmacokinetic (PK) study and analyses of single nucleotide polymorphisms (SNPs) in a candidate gene or in a genome-wide approach. Following the completion of the HapMap project,6 the advent of the powerful array-based SNP typing platforms has heralded an era in which a genome-wide approach is a popular or standard strategy for identifying disease susceptibility or drug response genes for common diseases.7, 8, 9

To identify the SNPs, which relate to the pathways of drug metabolism, biomedical researchers usually test the null hypothesis (H0) that there is no difference of the population mean of PK parameters (that is, area under the curve (AUC), maximum drug concentration (Cmax), half-life period (t1/2) and so on) between genotypes by using mainly the nonparametric analysis of variance, that is, the Kruskal–Wallis test.10, 11 On the basis of the statistical significance of the results from that test, researchers check the PK-genotype response patterns by sight, and then detect an additive, recessive or dominant model. The PK-related genes indicate a monotone increasing pattern in the number of alleles such as genetic models (additive, recessive and dominant pattern) as shown in Figure 1. Because a genome-wide association study is often designed as a multistage process, a relatively relaxed type I error rate is adopted in the first stage screening to assure the overall power of the study. If the Kruskal–Wallis test is applied to a significance level of P=0.05 in the first stage of a genome-wide association study, it is expected that as many as 5000–50000 SNPs should be visually inspected for the current standard genome scan typing platform. Development of more objective and efficient screening method than the current subjective procedure is needed.

Figure 1.
Figure 1 - Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

Pharmacokinetics parameter–genotype response patterns.

Full figure and legend (27K)

As a statistical method for detecting a monotonic dose–response relationship, a maximum contrast method has been proposed in clinical trials and toxicological trials.12, 13, 14, 15, 16, 17 This method, formed by taking the maximum over multiple contrast statistics for detecting a monotonic increase with the dose levels, is useful. The dose–response curves are modeled as the response patterns, and the set of contrast statistics is determined based on the set of contrast coefficients that correspond to these patterns. The contrast statistics should consist of contrast coefficients that are highly correlated with the population means for realistic dose–response relationships. The maximum contrast method is then applied to this set, and the pattern of the contrast coefficients for the contrast statistic that takes the maximum value is then selected as the true response pattern (see Materials and methods section).

In a typical pharmacogenomics study, the association of between PK parameter and genotype is modeled as the response patterns in Figure 1, and then the maximum contrast method is applied to this study on three contrast statistics with the following coefficients: c1=(−1 0 1)t, c2=(−2 1 1)t, c3=(−1 −1 2)t. The first contrast statistics corresponds to an additive model, the second to a recessive model and the third to a dominant model. As a result, the pattern of the coefficients for the contrast statistic that takes the maximum value is then selected as the true response pattern. However, the sample size of each genotype was enormously unbalanced, because the minor allele frequency (MAF) is less than 0.5, and the population is in Hardy–Weinberg equilibrium (HWE). Here we propose a modified maximum contrast method to detect an association between a PK parameter and genotype that accounts for the imbalance in genotype groups (see Materials and methods section).

In this paper, we proposed a modified maximum contrast method for detecting an association between PK parameter and genotype. Further, using simulation studies, we compare the Kruskal–Wallis test, the maximum contrast method and the modified maximum contrast method for pharmacogenomics data. We also suggest rules of thumb for choosing a powerful maximum contrast statistic for such data. Finally, the discussed methodology is illustrated by its application to a pharmacogenomics study to antitumor drugs on Japanese cancer patients.

Top

Results

Simulation study

To assess the power of the Kruskal–Wallis test, the maximum contrast method and the modified maximum contrast method, we first performed a Monte-Carlo simulation. Summary results of the simulation for each method are shown in Table 1 for various values of Δ and MAF when sample size (n) is set at 300. Note first that the diagonal box shows the positive predictive value (RTP) for detection of true response patterns, whereas the row at the far right is the power (Rp) for detection of PK-related SNPs and includes the misidentification of true response patterns, where the result of RTP by the Kruskal–Wallis test is blank because that test is an overall test and rejects the null hypothesis that there is no difference of the population mean of PK parameters. In a case where Δ=0, the Rp corresponds to a type I error rate of 5%, and is mostly controlled below 5%, but is inflated slightly for some models. The Rp and RTP increase with increasing Δ; on the other hand, both Rp and RTP decrease as the MAF decreases. We evaluated each method for detecting true response patterns. As a result, when the MAF is less than or equal to 0.25, in the additive and dominant pattern, the modified maximum contrast method is about 0.1–0.4 higher than the maximum contrast method in the RTP. However, in the recessive model, the modified maximum contrast method is about 0.4 lower than the maximum contrast method in the RTP. Therefore, under unbalanced sample size, the modified maximum contrast method is more powerful for detecting true response patterns than the maximum contrast method in the additive and dominant model, whereas the maximum contrast method is more powerful than the modified maximum contrast method in the recessive model.


For detecting PK-related genes, the Kruskal–Wallis test is lower in power than both maximum contrast methods in the Rp, under equivalent sample size (MAF=0.5), whereas the Kruskal–Wallis test is about 0.05–0.10 higher in power than both maximum contrast methods under the unbalanced sample size (MAFless than or equal to0.25). On the other hand, the Rp of the Kruskal–Wallis test is about 0.10–0.50 higher than both maximum contrast methods under the simulation condition without the PK-related SNPs in Table 2. Therefore the simulation studies suggest that the Kruskal–Wallis test detects many PK-nonrelated SNPs because of not considering the response patterns.


Questionnaire survey on judgment

We next conducted a questionnaire survey to compare the judgment by each statistical method (modified maximum contrast method and maximum contrast method) with the judgment by experts. The results are shown in Table 3. For the recessive and no-response pattern, the judgment of both statistical methods well accords with the judgment by expert. For the additive and dominant pattern, the maximum contrast method detects as a recessive pattern by mistake, and the modified maximum contrast method well accords with the judgment by expert. The Kendall’s rank correlation coefficient between modified maximum contrast method and expert’s judgment is 0.731; on the other hand, the correlation coefficient between the maximum contrast method and expert’s judgment is 0.423. The conclusion is that the judgment of modified maximum contrast method is closer to the judgment by expert than is the maximum contrast method.


Application to actual genome-wide pharmacogenomics study

In this paper, we focus on the elimination rate constant (Kel), which is the first order rate constant describing drug elimination from the body, and report the results of PK-related SNPs associated with Kel. The histogram for the Kel is shown in Figure 2. The PK data seem to be skewed to the left tail of the distribution; therefore we transformed data before applying both maximum contrast methods by taking the natural logarithm of the observed values.

Figure 2.
Figure 2 - Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

Histogram of elimination rate constant (Kel) for the gemcitabine pharmacogenomic (PGx) study.

Full figure and legend (8K)

Table 4 summarizes the number of significant SNPs for various P-value cutoffs by each method. For the P-value cutoff of 0.001, 84 SNPs are significant by the Kruskal–Wallis test, 77 SNPs by the maximum contrast method and 84 SNPs by the modified maximum contrast method. Two SNPs (GLT25D1: rs3848643 and BCNP1: rs6512201) are detected as additive model by all three methods in Table 5. In considering a multiple testing problem, we assume the existence of about 10000 linkage disequilibrium blocks within 100000 gene-centric SNPs, which are concentrated in about 2% of the human genome (that is, average interval of two SNPs is 600bp). It follows that the P-value cutoff is set at 5.0 × 10−6 as Bonferroni correlation. As a result, two SNPs (rs2148582, rs699) in angiotensinogen (AGT) gene are detected as a dominant model by both the maximum contrast method and modified maximum contrast method, but no SNPs are detected by the Kruskal–Wallis test. We show all of the results for association between PK parameters and SNPs at Genome Medicine Database of Japan (http://gemdbj.nibio.go.jp).



Top

Discussion

We reviewed the current methodology for analyzing PK-SNP data in pharmacogenomics studies. We proposed a modified maximum contrast method for detecting a PK-genotype response pattern under unbalanced sample size. The simulation study suggested that both the maximum contrast method and the modified maximum contrast method were more useful for screening PK-related SNPs in consideration of genetic models than is the Kruskal–Wallis test, and that under unbalanced sample size, the modified maximum contrast method was more powerful for detecting true response patterns than the maximum contrast method in the additive and dominant pattern, whereas the maximum contrast method is more powerful than the modified maximum contrast method in the recessive model.

In addition, the survey on judgment for PK-related SNPs makes evident that the modified maximum contrast method gave a judgment closer to the judgment of experts than did the maximum contrast method. The expert’s judgment was subjective, and also different by each expert. The modified maximum contrast method and the maximum contrast method were able to objectively detect the response pattern and the PK-related SNPs under unbalanced sample size.

The application of the maximum contrast method, modified maximum contrast method and Kruskal–Wallis test to a pharmacogenomics study supported these considerations. When the P-value cutoff was set at 0.05 in this application, about 5000 SNPs were detected as statistical significance, and both maximum contrast methods were able to show each genetic model to the biomedical researcher. We showed the list of identified SNPs by each method. The two SNPs in the AGT gene showed a statistical significant difference in Kel by both the maximum contrast method and modified maximum contrast method, but the Kruskal–Wallis test was not able to detect the gene after multiple testing adjustment. Polymorphic variations in the human AGT gene have been shown to be associated with increased circulating AGT concentrations,18 an increased risk of essential hypertension,18, 19, 20, 21, 22 and a decline in renal function.23 Although it is possible that the AGT gene polymorphisms influence gemcitabine PK through yet unknown mechanisms, the result and the suggested hypothesis in our study should be evaluated both in replication study using another sample set and also in biological functional analyses.

Many biomedical researchers have used the Kruskal–Wallis test for finding significant SNPs, and visually checked PK-genotype response patterns. Recently, powerful array-based SNP typing platforms have heralded an era in which a genome-wide association study is a popular or standard strategy for identifying disease associated genes or drug response genes for common diseases, and genotype data on 100000–1000000 SNPs are increasingly available to researchers. It is virtually impossible for biomedical researchers to visually check the response patterns on such a genome scan data. Our study has proposed an alternative by showing that either a modified maximum contrast method or a maximum contrast method can readily be applied to genome scans as a statistical screening method.

Top

Materials and methods

Maximum contrast method and modified maximum contrast method

We first describe a general form for a maximum contrast method as discussed in Yoshimura et al.,14 and Wakana et al.,17 in a pharmacogenomics study. We assume Yij (i=1, 2, 3; j=1, 2, …, ni) as an observed response for jth individual in ith genotype group (AA, Aa and aa, where ‘A’ is major allele, and ‘a’ is minor allele), and Yijs are independently and normally distributed with E(Yij)=μi and Var(Yij)=σ2. Under these conditions, bold ital Ymacr=(italic Y macr1, italic Y macr2, italic Y macr3)t follows the trivariate normal distribution N(μ, σ2D) where

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

diag(•) being a diagonal matrix, and the superscript t indicates the transpose of a matrix or vector. In the maximum contrast method, some contrast statistics are set according to the approach. Let contrast statistic Tk with a coefficient ck=(ck1, ck2, ck3)t where ck1+ck2+ck3=0, and k=1, 2 and 3 correspond to an additive, recessive and dominant response pattern, respectively, as shown in Figure 1. The following coefficients were used for each genetic model:

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

The contrast statistics is defined in terms of the contrast vector ck,

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

where ital sigma circ2 is Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author and Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author is degree of freedom of sigma circ2. Then the statistic is defined as:

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

The statistics can be used to test over null hypotheses, H0: μ1=μ2=μ3, on the other hand, in pharmacogenomics studies, alternative hypotheses such as H1: μ1<μ2<μ3, μ1=μ2<μ3, μ1<μ2=μ3. The P-value for the probability distribution of Tmax under the over null hypotheses is calculated by using the complex property of integration in the multivariate t-distribution with singular correlation matrix,24 and the P-value is defined as in the following formula:

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

where tmax* is the observed value of the test statistics.

This method can select the response pattern, which best fits the observed data among a set of patterns. In clinical and toxicological trials, the sample size of each group is almost equivalent; therefore the maximum contrast method assumes an equivalent sample size of each group. However, the sample size of each group is not equivalent in pharmacogenomics studies, because the MAF is less than 0.5 and is generally around 0.2. For this reason, under the unbalanced sample size, the adjoining denominator of contrast statistic Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author is overestimated although the studentized statistics by this variance estimate is robust. Therefore, the maximum contrast method cannot be applied for detecting a true response pattern in pharmacogenomics studies.

Here we propose a modified maximum contrast method for unbalanced sample size. First, the modified contrast statistic is given by

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

thus the adjoining denominator of the modified contrast statistic is Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author and it is not influenced by unbalanced sample size. The modified maximum contrast statistic is given by

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

In addition, the multiplicity adjusted P-value for the probability distribution of Tmax under H0 was calculated by the permutation procedure.25 Simple random sampling without replacement method generates the sampling distribution of the statistic by drawing repeated samples from the observed sample itself. Consequently, the permutation procedure is used to simulate P-values; the lowest P-value is calculated for each of 100000 permutations for the entire dataset, and from this distribution.

Simulation study

We assess the power of the Kruskal–Wallis test, the maximum contrast method (Equations (1) and (2)) and the modified maximum contrast method (Equations (4) and (5)) by simulation studies. The scenarios discussed here are motivated by the anticancer drug’s pharmacogenomics data that will be introduced below. In particular, we are interested in the performance of all three tests when the MAF decreases. To examine these effects, we simulate PK parameter data from transformed log-normal distribution by each genotype. The distributions of PK parameters are often unknown, but empirically are modeled under the assumption of a log-normal distribution because (1) PK parameter values must be nonnegative and the normal assumption does not enforce this assumption; (2) distribution of estimated PK parameters is often left skewed, which is compatible with a log-normal distribution.26 Here, for applying an intravenous anticancer drug in actual data, we focus on the elimination rate constant (Kel), which is the first order rate constant describing drug elimination from the body. We assume that a sample of patients with any genotype (AA, Aa or aa) and the proportion of MAF is given, and that this population is in HWE. For each patient, the PK parameter, Kel is given. Under these assumptions, we generated Kel by each genotype from the log-normal distribution with mean μi, that is, italic Y macri~LN(μi, 1), i=1, 2, 3 in ith genotype group (AA, Aa and aa), where μi=Δ·cki and Δ is given coefficient and cki corresponded to element of contrast statistics vector ck=(ck1 ck2 ck3)t. Here, Δ is set at 0.0, 0.5 and 1.0, respectively, and the contrast statistics vectors are set as the following coefficients corresponding to the genetic models (k=1, 2 and 3) in Figure 1 (i)–(iii):

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

In addition, the MAF was set at 0.12, 0.25 or 0.50, respectively, where total sample size (n) was 300 or 600 subjects. The criteria to evaluate the performance of each method were two indicators, RP and RTP, defined by Equations (6) and (7).

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

Rp is a proportion of detected PK-related SNPs (power), whereas RTP is the proportion of detected true response patterns among PK-related SNPs (positive predictive value). The Monte-Carlo simulation to calculate Rp and RTP was repeated 10000 times and the mean values of indicators were calculated. Note that N was a constant fixed as the repetition number of the simulation, whereas NR was a random variable realized as the number of rejections by the hypothesis test, and NTP was the number of detected true response patterns.

Questionnaire survey on judgment

To identify PK-related genes, biological experts who are pharmacokinetists, molecular biologists and geneticists used to draw box-and-whisker plots on PK parameters by each genotype, and check these response patterns by sight. This judgment involves much experience, but it is subjective and may differ by expert. We conducted a questionnaire survey to compare statistical judgment with judgment by experts. First, we showed summary statistics and box–whisker plot (Figure 3) to the expert group that consisted of two pharmacokinetists, two molecular biologists and two geneticists. Second, these experts independently gave a decision from six response patterns (Figure 1). Finally, we applied the modified maximum contrast method and the maximum contrast method to this survey dataset, and, by using Kendall’s rank correlation coefficient, evaluated, which method is closer to expert judgment.

Figure 3.
Figure 3 - Unfortunately we are unable to provide accessible alternative text for this. If you require assistance to access this image, please contact help@nature.com or the author

Summary statistics and box–whisker plot in questionnaire survey.

Full figure and legend (60K)

Application to actual gemcitabine pharmacogenomics study

We applied the Kruskal–Wallis test, the maximum contrast method and the modified maximum contrast method to an actual genome-wide pharmacogenomics study on antitumor drugs. The study was performed within the Millennium Genome Project in Japan, and four antitumor drugs were chosen as project targets: irinotecan, fluorouracil, paclitaxel and gemcitabine. One or some combination of the drugs was administered to about 1000 cancer patients with written informed consent at the National Cancer Center, Japan. PK data were also obtained for the drugs except for fluorouracil. In addition, about 1000 DNA samples were extracted from peripheral blood mononuclear cells and 109365 SNPs were genotyped by Illumina Human-1 BeadChip.

In this application, we emphasize a focus on the gemcitabine pharmacogenomics study. Gemcitabine (2′,2′-difluorodeoxycytidine) is a nucleoside anticancer drug that has a broad spectrum of antitumor activity against various solid tumors, such as nonsmall cell lung cancer and pancreatic cancer. The participants consisted of 256 Japanese gemcitabine-naive cancer patients (mainly pancreatic carcinoma) at the National Cancer Center Hospital and National Cancer Center Hospital East. Details of this analysis group were reported previously.27 For gemcitabine PK analysis, 5ml of heparinized blood was sampled before the first gemcitabine administration, and at 0, 15, 30, 60, 90, 120 and 240min after the termination of the infusion. The AUC, mean residence time from 0 to infinity, peak concentration (Cmax), clearance (CL m−2), distribution volume based on the terminal phase (Vz m−2) and elimination rate constant (Kel) were calculated using WinNonlin ver. 4.01 (Pharsight Corporation, Mountain View, CA, USA). For an investigation of the association between these PK parameters and SNPs in Illumina Human-1 BeadChip, we applied the maximum contrast method, the modified maximum contrast method and the Kruskal–Wallis test.

Top

Notes

Duality of Interest

None declared.

Top

References

  1. Licinio J, Wong M. Pharmacogenomics: The Search for Individualized Therapies. Wiley: Weinheim (Germany), 2002.
  2. Evans WE, McLeod HL. Pharmacogenomics—drug disposition, drug targets, and side effects. N Engl J Med 2003; 348: 538–549. | Article | PubMed | ISI | ChemPort |
  3. Ingelman-Sundberg M. Pharmacogenomic biomarkers for prediction of severe adverse drug reactions. N Engl J Med 2008; 358: 637–639. | Article | PubMed | ChemPort |
  4. Efferth T, Volm M. Pharmacogenetics for individualized cancer chemotherapy. Pharmacol Ther 2005; 107: 155–176. | Article | PubMed | ISI | ChemPort |
  5. US FDA. Guidance for Industry Pharmacogenomic Data Submissions. 2005.
  6. International Human Genome Sequencing Consortium. Finishing the euchromatic sequence of the human genome. Nature 2004; 431: 931–945. | Article | PubMed | ISI | ChemPort |
  7. Risch N, Merikangas K. The future of genetic studies of complex human diseases. Science 1996; 273: 1516–1517. | Article | PubMed | ISI | ChemPort |
  8. Hirschhorn JN, Daly MJ. Genome-wide association studies for common diseases and complex traits. Nat Rev Genet 2005; 6: 95–108. | Article | PubMed | ISI | ChemPort |
  9. Kingsmore SF, Lindquist IE, Mudge J, Gessler DD, Beavis WD. Genome-wide association studies: progress and potential for drug discovery and development. Nat Rev Drug Discov 2008; 7: 221–230. | Article | PubMed | ChemPort |
  10. Kruskal HW, Wallis WA. Use of ranks in one-criterion variance analysis. J Am Stat Assoc 1952; 47: 583–621. | Article |
  11. Siegel S, Castellan NJ. Nonparametric Statistics for the Behavioral Sciences, 2nd edn. McGraw-Hill: New York, 1988.
  12. Mukerjee H, Robertson T, Wright FT. Comparison of several treatments with a control using multiple contrast. J Am Stat Assoc 1987; 82: 902–910. | Article |
  13. Ruberg SJ. Contrasts identifying the minimum effective dose. J Am Stat Assoc 1989; 82: 816–822. | Article |
  14. Yoshimura I, Wakana A, Hamada C. A performance comparison of maximum contrast methods to detect dose dependency. Drug Inf J 1997; 31: 423–432.
  15. Stewart H, Ruberg SJ. Detecting dose response with contrasts. Stat Med 2000; 19: 913–921. | Article | PubMed |
  16. Nishiyama H, Yanagihara H, Yoshimura I. SAS/IML program for computing probabilities related to maximum contrast methods. Jpn J Biometrics 2003; 24: 57–70, (in Japanese).
  17. Wakana A, Yoshimura I, Hamada C. A method for therapeutic dose selection in a phase II clinical trial using contrast statistics. Stat Med 2007; 26: 498–511. | Article | PubMed |
  18. Jeunemaitre X, Soubrier F, Kotelevtsev YV, Lifton RP, Williams CS, Charru A et al. Molecular basis of human hypertension: role of angiotensinogen. Cell 1992; 71: 169–180. | Article | PubMed | ISI | ChemPort |
  19. Caulfield M, Lavender P, Farrall M, Munroe P, Lawson M, Turner P et al. Linkage of the angiotensinogen gene to essential hypertension. N Engl J Med 1994; 330: 1629–1633. | Article | PubMed | ISI | ChemPort |
  20. Hegele RA, Brunt JH, Connelly PW. A polymorphism of the angiotensinogen gene associated with variation in blood pressure in a genetic isolate. Circulation 1994; 90: 2207–2212. | PubMed | ChemPort |
  21. Vasku A, Soucek M, Znojil V, Rihacek I, Tschoplova S, Strelcova L et al. Angiotensin I-converting enzyme and angiotensinogen gene interaction and prediction of essential hypertension. Kidney Int 1998; 53: 1479–1482. | Article | PubMed | ChemPort |
  22. Pereira TV, Nunes AC, Rudnicki M, Yamada Y, Pereira AC, Krieger JE. Meta-analysis of the association of 4 angiotensinogen polymorphisms with essential hypertension: a role beyond M235T? Hypertension 2008; 51: 778–783. | Article | PubMed | ChemPort |
  23. Pei Y, Scholey J, Thai K, Suzuki M, Cattran D. Association of angiotensinogen gene T235 variant with progression of immunoglobulin A nephropathy in Caucasian patients. J Clin Invest 1997; 100: 814–820. | Article | PubMed | ISI | ChemPort |
  24. Genz A, Bretz F. Numerical computation of multivariate t-probabilities with application to power calculation of multiple contrasts. J Stat Comput Simulat 1999; 63: 361–378.
  25. Westfall PH, Young SS. Resampling-Based Multiple Testing: Examples and Methods for P-value Adjustment. Wiley: New York, 1993.
  26. Gabrielsson J, Weiner D. Pharmacokinetic and Pharmacodynamic Data Analysis: Concepts and Applications. Taylor & Francis: Sweden, 2000.
  27. Sugiyama E, Kaniwa N, Kim SR, Kikura-Hanajiri R, Hasegawa R, Maekawa K et al. Pharmacokinetics of gemcitabine in Japanese cancer patients: the impact of a cytidine deaminase polymorphism. J Clin Oncol 2007; 25: 32–42. | Article | PubMed | ISI | ChemPort |
Top

Acknowledgements

We thank Professor Isao Yoshimura and Professor Chikuma Hamada at Tokyo University of Science for their valuable advice, and Dr Hiromi Sakamoto and Ms Sumiko Ohnami for the SNP genotyping. Yasunori Sato is a recipient of Official Trainee of the Foreign Clinical Pharmacology Training Program, Japanese Society of Clinical Pharmacology and Therapeutics. This work was initiated while Yasunori Sato was a postdoctoral fellow at the Department of Biostatistics, Harvard School of Public Health, and partially supported in Japan by the program for promotion of Fundamental Studies in Health Sciences of the National Institute of Biomedical Innovation (NiBio).