Statistical Methods For Gene/environment Interaction And Genetic Susceptibility

我们正在开发方法，利用从几个人的合并标本中测量的暴露量，以及对每个人单独测量的基因型，来研究基因-环境相互作用。假设有一个病例对照研究，并在一组SNP（单核苷酸多态性）上对每个个体进行基因分型。假设一个人也有生物标本（例如，血清或尿液），但是缺乏预算来分析每个个体样本的感兴趣的暴露。合并样本并分析所得合并样本不仅可以节省分析成本，还可以保留样本量以备将来使用。在过去，我们已经开发了用于分析病例对照研究的方法，在合并样本中测量暴露。这些方法合理地假设，合并样本的测量值是单个样本值的平均值。使用这些方法，在单个SNP上测试基因-环境相互作用需要在具有相同SNP基因型的个体层内创建样本池。为了研究一组SNP的基因-环境相互作用，我们以前的方法需要为研究的每个SNP创建新的合并样本，并且潜在的检测成本节省将消失。我们正在开发的方法将个体测量值视为缺失数据，并以原则性的方式使用合并样本来插补这些缺失数据。有了一组给定的插补数据，我们可以使用病例对照数据的标准统计方法来估计基因-环境相互作用。在实践中，我们使用多重插补方法：创建多组插补数据，对每组数据进行病例对照分析，并结合多重分析的结果。这一办法已显示出一些希望，但仍有一些问题有待解决。关于这一问题的工作正在进行中。 在全基因组研究中，当单个SNP的边际效应很小时，识别致病SNP可能具有挑战性，因为测试阈值必须反映大量的SNP。对于复杂的疾病，SNP的特定组合可能会显着增加一种上位性或基因-基因相互作用的风险。我们目前正在研究使用机器学习技术来发现病例父母数据中共同导致疾病的SNP集（致病SNP）。首先，我们设计了一种方法来使用实际的情况下，父母的三联体基因型，以创建模拟的全基因组数据集，反映现实的连锁不平衡结构，并与已知的致病SNP集播种。 我们目前正在努力更好地描述以这种方式模拟的种群的遗传特性。 其次，我们实现了一种现有的随机搜索算法（称为GA-KNN），该算法基于进化算法来找到多组预测疾病的k个SNP（这里k是一个小数字，比如2或4）。通过对预测疾病的SNPs中出现频率最高的SNPs进行分类，我们希望能够发现致病SNPs。在对模拟数据进行的初步试验中，每个模拟数据都包含两组相互作用的四个SNP，我们的方法显示出了希望。在正在进行的工作中，我们正试图加快算法的速度，看看在更复杂的情况下是否保持有希望的性能。 (see也是Z 01 ES 040007; PI克莱尔温伯格; Min Shi也是该项目的实验室内合作者;她的时间分配在温伯格的项目中，但不在本项目中。）

We are in the process of developing methods for using exposures measured in pooled specimens from several individuals, together with genotypes measured separately on each individual, to study gene-environment interactions. Suppose one has case-control study and genotyped each individual at a panel of SNPs (single nucleotide polymorphisms). Suppose that one also has biological specimens (e.g., serum or urine) from the same individuals but lacks the budget to assay each individual specimen for an exposure of interest. Pooling specimens and assaying the resulting pooled specimens will not only save assay costs but preserve specimen volume for future uses. In the past, we have developed methods for analyzing case-control studies with exposures measured in pooled specimens. Those methods assume, reasonably, that the measured value on the pooled specimen is the average of the values for the individual specimens. With those methods, testing gene-environment interactions at a single SNP required creating specimen pools within strata of individuals who all had the same genotype for that SNP. To study gene-environment interactions for a panel of SNPs, our previous methods would require creating new pooled specimens for each SNP studied and the potential savings in assay costs would disappear. The approach that we are developing regards the individual measurements as missing data and uses the pooled specimens in a principled way to impute those missing data. With a give set of imputed data in hand, we can use standard statistical methods for case-control data to estimate gene-environment interactions. In practice, we use a multiple-imputation approach: creating multiple sets of imputed data, doing a case-control analysis for each set, and combining the results from the multiple analyses. This approach has shown some promise but some problems remain to be resolved. Work on this problem is ongoing. Identification of causative SNPs in a genome-wide study can be challenging when individual SNPs have small marginal effects because testing thresholds must reflect the large number of SNPs under study. For complex diseases, particular combinations of SNPs may dramatically increase risk a kind of epistasis or gene-gene interaction. We are currently investigating the use of a machine learning technique for the discovery of sets of SNPs that together cause disease (causative SNPs) in case-parents data. First, we devised a way to use actual case-parent triad genotypes to create simulated genome-wide data sets that reflect realistic linkage disequilibrium structure and are seeded with known sets of causative SNPs. We are currently working to better characterize the genetic properties of populations simulated in this way. Second, we implemented an existing stochastic search algorithm (called GA-KNN) that is based on an evolutionary algorithm to find multiple sets of k SNPs that are predictive of disease (here k is a small number, say 2 or 4). By cataloguing those SNPs which appear most frequently among the sets that are predictive of disease, we hope to uncover the sets of causative SNPS. In preliminary trials on simulated data seeded with two interacting sets of four SNPs each, our approach shows promise. In ongoing work, we are attempting to speed up the algorithm and to see whether the promising performance is maintained in more complex situations. (see also Z01 ES040007; PI Clare Weinberg; Min Shi is also a within-lab collaborator on this project; her time is allocated in Weinberg's project but not in this one.)