Using recursive partitioning for exploration and follow-up of linkage and association analyses

Recursive partitioning analysis of complex disease pharmacogenetic studies. I. Motivation and overview

Identifying genetic variation predictive of important phenotypes, including disease susceptibility, drug efficacy, and adverse events, is a challenging task, and theory and computer science work is being carried out in an attempt to tackle this issue. For many important diseases, such as diabetes, schizophrenia, and depression, the etiology is complex; either the disease is a result of several multiple mechanisms or is caused by an interaction among multiple genes or gene-environment interactions, or both. There is a need for statistical methods to deal with the large, complex data sets that will be used to disentangle these diseases. Each putative genetic polymorphism can be tested for association sequentially. The most difficult problem, however, is the identification of combinations of polymorphisms or genetic markers with increased predictive characteristics. Data from clinical trials, where patients with a particular disease are treated with certain drugs, can be retrospectively assembled using a case-control design. Such data will typically include treatment assignment, demographics, medical history, and genotypes for a large number of genetic markers. The number of variables in such data is expected to be much larger than the number of subjects. This report focuses on some of the methods being employed to deal with this complex data and covers, in some detail, a data-mining method--recursive partitioning--to analyze such data. The methods are demonstrated using a complex simulated data set, as there are few available public data sets. This explication of recursive partitioning should provide researchers with a better idea of the current available analysis techniques, in order to allow them to plan their experiments more effectively.

Genetic variants in the epithelial sodium channel associate with oedema in type 2 diabetic patients receiving the peroxisome proliferator-activated receptor gamma agonist farglitazar

Peroxisome proliferator-activated receptor gamma (PPARgamma) agonists are highly effective in the treatment of type 2 diabetes. In some patients, PPARgamma ligands are associated with fluid retention/oedema, for which the mechanism is not fully understood. A pharmacogenetic study was undertaken to investigate effects of variations in 21 candidate genes related to epithelial sodium channel (ENaC) pathways on oedema. This study used DNA samples collected from type 2 diabetes phase III clinical trials of the PPARgamma agonist farglitazar (administered alone or in combination with insulin or glyburide) and investigated oedema reported as an adverse event as phenotype. Initial case-control analysis of oedema identified candidate gene single nucleotide polymorphisms with significant associations. These included three polymorphisms in ENaCbeta subunit (SCNN1B) that showed significant associations (P<0.05) with the two combination treatments in discrete regions of the gene, but not farglitazar treatment alone. Sequencing of SCNN1B in 207 Caucasian participants receiving farglitazar plus insulin or glyburide combination therapies, identified additional polymorphisms that were also significantly associated with oedema (P<0.0005) and maintained the treatment-regional associations. Further covariate analysis accounting for clinical factors influencing oedema supported these observations. One of the SCNN1B polymorphisms, at position -405 of the 5' flanking region (rs34241435), was predicted to modify transcriptional interactions and in a transfected COS cell luciferase reporter gene assay exhibited higher promoter activity. These exploratory studies provide clinical pharmacogenetic and functional genomic evidence to support a pivotal role for ENaC regulation in PPARgamma-induced oedema and provide insight into mechanisms and possible management of this side effect.

Pharmacogenetics and obesity: common gene variants influence weight loss response of the norepinephrine/dopamine transporter inhibitor GW320659 in obese subjects

BACKGROUND

GW320659, a highly selective neuronal norepinephrine and dopamine re-uptake inhibitor, has been evaluated for the treatment of obesity. Scrutiny of the weight loss data from a phase II study (GlaxoSmithKline study OBS20001) showed a wide variation in weight loss response following GW320659 treatment and the possibility that the study population might include subgroups with enhanced weight loss response.

METHODS

Pharmacogenetic analysis was performed in 191 subjects prospectively ascertained from a Phase II dose ranging study to evaluate the influence of genotype on weight loss efficacy and safety of GW320659 in obese subjects.

RESULTS

Common genetic polymorphisms in the drug target (norepinephrine transporter protein 1, SLC6A2) and mechanism pathway (NMDA receptor channel NR1 subunit, GRIN1) were associated with increased weight loss following GW320659 treatment in a proportion (36%) of the study population. In the patient subgroup selected for these genotypes, GW320659 (15 mg/day) produced a significant difference in mean weight loss of 7.84 kg (SD 5.23, n = 14), compared to 2.53 kg (SD 5.17, n = 24) in the subgroup that did not possess the genotypes (P = 0.006). This subgroup also showed a highly significant weight loss response for GW320659 compared to placebo (+0.31 kg, SD 3.32, n = 16) with the same genotypes (P < 0.0001). In addition, there was no difference in placebo response between either subgroup.

CONCLUSIONS

Polymorphisms in SLC6A2 and GRIN1 could be used to maximize effective obesity pharmacotherapy by norepinephrine/dopamine transporter inhibitors by identifying patients that may be predisposed to particularly good treatment weight loss response.

Screening large-scale association study data: exploiting interactions using random forests

BACKGROUND

Genome-wide association studies for complex diseases will produce genotypes on hundreds of thousands of single nucleotide polymorphisms (SNPs). A logical first approach to dealing with massive numbers of SNPs is to use some test to screen the SNPs, retaining only those that meet some criterion for further study. For example, SNPs can be ranked by p-value, and those with the lowest p-values retained. When SNPs have large interaction effects but small marginal effects in a population, they are unlikely to be retained when univariate tests are used for screening. However, model-based screens that pre-specify interactions are impractical for data sets with thousands of SNPs. Random forest analysis is an alternative method that produces a single measure of importance for each predictor variable that takes into account interactions among variables without requiring model specification. Interactions increase the importance for the individual interacting variables, making them more likely to be given high importance relative to other variables. We test the performance of random forests as a screening procedure to identify small numbers of risk-associated SNPs from among large numbers of unassociated SNPs using complex disease models with up to 32 loci, incorporating both genetic heterogeneity and multi-locus interaction.

RESULTS

Keeping other factors constant, if risk SNPs interact, the random forest importance measure significantly outperforms the Fisher Exact test as a screening tool. As the number of interacting SNPs increases, the improvement in performance of random forest analysis relative to Fisher Exact test for screening also increases. Random forests perform similarly to the univariate Fisher Exact test as a screening tool when SNPs in the analysis do not interact.

CONCLUSIONS

In the context of large-scale genetic association studies where unknown interactions exist among true risk-associated SNPs or SNPs and environmental covariates, screening SNPs using random forest analyses can significantly reduce the number of SNPs that need to be retained for further study compared to standard univariate screening methods.

Screening large-scale association study data: exploiting interactions using random forests

BACKGROUND

Genome-wide association studies for complex diseases will produce genotypes on hundreds of thousands of single nucleotide polymorphisms (SNPs). A logical first approach to dealing with massive numbers of SNPs is to use some test to screen the SNPs, retaining only those that meet some criterion for further study. For example, SNPs can be ranked by p-value, and those with the lowest p-values retained. When SNPs have large interaction effects but small marginal effects in a population, they are unlikely to be retained when univariate tests are used for screening. However, model-based screens that pre-specify interactions are impractical for data sets with thousands of SNPs. Random forest analysis is an alternative method that produces a single measure of importance for each predictor variable that takes into account interactions among variables without requiring model specification. Interactions increase the importance for the individual interacting variables, making them more likely to be given high importance relative to other variables. We test the performance of random forests as a screening procedure to identify small numbers of risk-associated SNPs from among large numbers of unassociated SNPs using complex disease models with up to 32 loci, incorporating both genetic heterogeneity and multi-locus interaction.

RESULTS

Keeping other factors constant, if risk SNPs interact, the random forest importance measure significantly outperforms the Fisher Exact test as a screening tool. As the number of interacting SNPs increases, the improvement in performance of random forest analysis relative to Fisher Exact test for screening also increases. Random forests perform similarly to the univariate Fisher Exact test as a screening tool when SNPs in the analysis do not interact.

CONCLUSIONS

In the context of large-scale genetic association studies where unknown interactions exist among true risk-associated SNPs or SNPs and environmental covariates, screening SNPs using random forest analyses can significantly reduce the number of SNPs that need to be retained for further study compared to standard univariate screening methods.

Complex interactions in Parkinson's disease: a two-phased approach

The identification of pathogenic mutations in the three genes alpha-synuclein, parkin, and ubiquitin carboxy-terminal hydrolase L1 (UCHL1) has elucidated the ubiquitin proteasome system (UPS) and its potential role as a causal pathway in Parkinson's disease (PD). In addition, polymorphisms of these three genes have been shown to be independently associated with PD. In a sample of 298 unrelated PD cases and 185 controls, we applied a two-phased approach of recursive partitioning and logistic regression analyses to explore complex interactions. For women only, we observed an epistatic interaction of UCHL1 and alpha-synuclein genotypes with significant effects on PD risk (odds ratio = 2.42; P = 0.003). Our findings are consistent with the hypothesis that PD is a multigenic disorder of the UPS.

Classification methods for confronting heterogeneity

Recursive partitioning/tree models are discussed as a method of dissecting the complex nature of traits with different causal mechanisms operating in different subsets of the data (e.g., different genes operating in different subsets of families). In addition to the straightforward application of classification and regression trees to define more homogeneous subsets of the data on which to conduct further analysis, developments incorporating linkage analysis into the definition of the regression trees (Shannon et al., 2000) are discussed. The pros and cons of recursive partitioning vs. the related approach of context-dependent analysis (Turner et al., 1999) are also reviewed as two promising analysis strategies that may be useful for genetic dissection of complex traits.