Statistical Methods For Genetic Epidemiology

我们以前已经开发了一种方法，用一种称为Trimm(三联体多标记)的方法，对受影响的个体及其父母使用多个SNP基因型来检测质量性状。该测试方法对由于人口分层而产生的偏差具有很强的鲁棒性。我们进一步扩展了这种方法，通过一种我们称为GEI-Trimm的方法，允许测试单倍型与环境的交互作用。描述这种方法并通过模拟来表征其性能的论文于今年发表。 在另一个项目中，我们正在估计家族史数据中存在的不对称性，这种不对称性是母亲中介的遗传效应存在的次要因素。我们将这一策略应用于Sister研究的家族史数据，发现有证据表明，年轻(50岁以下)乳腺癌患者的祖母比父亲的祖母更有可能患乳腺癌。这表明可能存在母体介导的乳腺癌遗传风险因素，可能存在与风险相关的印记基因，或者线粒体变异起到了一定作用。表观遗传学对乳腺癌也可能很重要。 我们现在考虑的一个特别重要的设计涉及一个“四分体”结构，除了父母之外，还有一个受影响的后代和一个未受影响的后代。这一设计已经在两姐妹研究中实施，该研究正在评估遗传和环境风险因素在年轻发病(50岁以下)乳腺癌中的联合作用。不一致的同胞对允许估计暴露的影响，而嵌入的病例-亲本三联体允许检测提供保护或风险的单倍型。共同分析的四分体应该为评估基因与环境的交互作用提供一个强有力的设计。我们一直致力于开发和评估用于四分体设计的方法。两姐妹研究继续纳入核心家庭，其中一个女儿在50岁之前患上乳腺癌，而另一个女儿没有受到影响。我们目前已经招收了1300多个这样的家庭。这在另一个单独的项目中进行了描述。将需要探索遗传基因类型以及肿瘤特征，以调查预测治疗后临床过程的因素，在这方面也需要开发改进的统计方法。 我们已经开发了一种使用四分体结构来研究基因与环境相互作用的方法，并进行了广泛的模拟来记录它在一系列场景下的表现，其中一些场景有暴露相关的种群结构，另一些没有暴露相关的种群结构。令我们惊讶的是，如果种群具有暴露涉及的种群结构，所有现有的基因与环境相互作用的方法都会受到偏差。当研究中的标记等位基因的频率和暴露流行率都不同时，就会发生这种情况。由此产生的偏差可以很好地理解为反映了这样一个事实，即在这种结构下，暴露可以替代被研究标记与致病SNP/单倍型之间的连锁不平衡程度(这在不同的亚群中也是不同的)。这种偏见可能是极端的。我们现在已经在最近发表的研究中找到了一些避免这种情况的补救措施，同时保留了良好的统计能力。 在另一个项目中，我们考虑了进行全基因组关联研究分析的问题，其中一个人希望结合来自使用不同平台进行基因分型的许多研究的信息。现有的软件Mach被用来根据观察到的基因数据对缺失的基因类型进行建模，但输出生成的是一个分数，而不是离散的等位基因计数。对于家庭数据，现有的方法不符合这种格式，分配的分数的质量也没有得到很好的研究。我们利用墨西哥城儿童哮喘研究的数据(Stephanie London，PI)，开发了将推测的基因得分纳入病例-父母三联体分析的方法，并使用墨西哥城儿童哮喘研究的数据(Stephanie London，PI)对大约7000个同时存在这两种基因的SNPs的基于推测基因的结果和基于实际基因的结果之间的一致性进行了评估。这篇论文是最近发表的。

We had previously developed methods for qualitative traits using multiple-SNP genotypes for affected individuals and their parents in a method called TRIMM (triad multi-marker). The testing approach is robust against bias due to population stratification. We further extended the approach to allow testing for haplotype-by-environment interaction, via a method we call GEI-TRIMM. The paper describing this approach and characterizing its performance through simulations was published this year. In another project, we are estimating the asymmetry that would exist in family history data secondary to the existence of a maternally-mediated genetic effect. We applied this strategy to family history data from the Sister Study, and found evidence that maternal grandmothers of young-onset (under 50) cases of breast cancer were more likely to have had breast cancer than were paternal grandmothers. This suggests there may be maternally-mediated genetic risk factors for breast cancer, that there may be imprinted genes related to risk or that mitochondrial variants play a role. Epigenetics could also be important for breast cancer. A particularly important design we are now considering involves a "tetrad" structure, with one affected and one unaffected offspring, in addition to the two parents. This design has been implemented in the Two Sister Study, which is assessing the joint role of genetic and environmental risk factors in young-onset (under age 50) breast cancer. The discordant sib pair allows estimation of effects of exposures, while the embedded case-parent triad allows detection of haplotypes that confer either protection or risk. The tetrad analyzed together should provide a powerful design for assessing gene-by-environment interaction. We have been working on developing and evaluating methods for use with the tetrad design. The Two Sister Study is continuing to enroll nuclear families where one daughter developed breast cancer before age 50 and the other daughter is unaffected. We currently have enrolled over 1,300 such families. This is described under a separate project. Inherited genotypes, together with tumor characteristics, will need to be explored to investigate factors that predict the clinical course following treatment, and improved statistical methods will also need to be developed in that context. We have developed a method for studying gene-by-environment interaction using the tetrad structure and we carried out extensive simulations to document its performance under a range of scenarios, some with and some without exposure-involved population structure. We learned to our surprise that all of the existing gene-by-environment interaction methods are subject to bias if the population has exposured-involved population structure. This happens when there are subpopulations that differ both in their frequency of the marker allele under study and in their exposure prevalence. The resulting bias can best be understood as reflecting the fact that with that kind of structure the exposure can serve as a surrogate for the degree of linkage disequilibrium (which also varies across subpopulations) between the marker under study and a causative SNP/haplotype. This bias can be extreme. We have now developed some remedies for avoiding it, while preserving good statistical power, in work that was recently published. In another project, we considered the problem of carrying out genome-wide association study analyses where one wishes to combine information from a number of studies that have used different platforms for doing the genotyping. Existing software called MACH is used to model missing genotypes based on observed genotype data, but the output generates a score, rather than a discrete allele count. With family data the existing methods are not amenable to that format and the quality of the assigned score has not been well studied. We developed methods to incorporate the imputed genotype scores into a case-parent triad analysis and we assessed the agreement between results based on imputed genotypes and results based on actual genotypes for some 7000 SNPs where we had both, using data (Stephanie London, PI) from the Mexico City Childhood Asthma Study. This paper was recently published.