Statistical Methods For Genetic Epidemiology

尽管由于已经进行了广泛的GWAS研究，将遗传变异与疾病联系起来，因此取得了一些进展，但已知的变异只解释了大多数可遗传健康状况的遗传性的一小部分。出生缺陷，口腔裂，就是一个很好的例子。已知在随后的兄弟姐妹中，复发风险约为30-40倍，但已知的变异对复发的解释很少。考虑到许多生物通路中设计的保护性冗余，可能在这种缺陷表现出来之前，某些特定的SNPs必须联合存在。考虑到这一点，暴露在环境中可能也很重要。我们用于模拟具有真实连锁结构的基因组数据的算法的可用性为我们提供了一个实验室，用于开发评估GxGxG和GxGxE相互作用的新方法。该方法现在能够识别多条上位性通路，其中某些变异SNP的同时存在会对结果产生风险。我们现在已经回到了分裂数据，并分别在亚洲数据和欧洲数据中挖掘了上位性。 使用我们的(公开可用的)模拟算法创建玩具数据集，我们改进了我们的机器学习方法，以随机搜索多SNP效应。这篇论文代表了现有方法的一个重要进步，因为它可以处理更大的候选SNPs集，并且还可以以一种根本不需要模型的方式专门测试上位性。这种方法首次发表在《生物信息学》杂志上，并赢得了拉里·库珀奖，因为今年北卡罗来纳大学生物统计学的一篇论文获得了最佳论文奖。我们正在修订我们的论文，扩展了该方法，允许将母亲的SNPs作为病因的一部分。我们还扩展了这项工作，允许遗传途径与环境暴露相互作用，并将新方法应用于与怀孕期间母亲吸烟有关的口腔唇裂。那篇论文很快就会提交。该方法还可用于区分与更具体的疾病亚类有关的可能不同的上位性效应，例如乳腺癌的分子亚型，或与结局的严重性有关的可能不同的上位效应，例如极端早产与不太极端早产。未来的计划包括评估上位性，以获得血压等量化结果。

Despite some advances due to the extensive GWAS studies that have been done to associate genetic variants with diseases, the known variants have only explained a small proportion of the heritability of most heritable health conditions. The birth defect, oral cleft, is a good example. The recurrence risk is known to be about 30-40 fold in a subsequent sibling, but known variants explain little of that recurrence. Given the protective redundancy designed into many biologic pathways, it may be that certain particular SNPs must be jointly present before this defect is expressed. Given that combination, an environmental exposure might also be important. The availability of our algorithm for simulating genomic data with realistic linkage structure provides us with a laboratory for development of new methods for assessing both GxGxG interactions and GxGxE interactions. The method is now able to identify multiple epistatic pathways where the simultaneous presence of certain variant SNPs confers risk for the outcome. We have now gone back to the clefting data and mined for epistasis there, separately in the Asian data and the European data. Using our (publicly available) simulation algorithm to create toy data sets, we have refined our machine-learning approach to stochastically search for multi-SNP effects. This paper represents an important advance over existing methods because it can handle a much larger set of candidate SNPs and can also test specifically for epistasis in a fundamentally model-free way. This method was first published in Bioinformatics and earned the Larry Kupper Award for the best paper arising from a dissertation this year at UNC Biostatistics. We are revising our paper that extended the method to enable inclusion of maternal SNPs as part of the etiology. We have also extended the work to allow the genetic pathways to interact with environmental exposures, and have applied the new methods to oral clefting in relation to maternal smoking during pregnancy. That paper will be submitted soon. The approach can also be used to disentangle possibly distinct epistatic effects in relation to more specific disease subcategories, e.g. the molecular subtypes of breast cancer, or to severity of the outcome, e.g. extreme prematurity versus less extreme prematurity. Future plans include evaluating epistasis for quantitative outcomes like blood pressure.