Koh Hyunwook
Department of Applied Mathematics and Statistics, The State University of New York, Korea, Incheon 21985, South Korea.
NAR Genom Bioinform. 2024 Nov 12;6(4):lqae148. doi: 10.1093/nargab/lqae148. eCollection 2024 Sep.
The effect of a treatment on a health or disease response can be modified by genetic or microbial variants. It is the matter of interaction effects between genetic or microbial variants and a treatment. To powerfully discover genetic or microbial biomarkers, it is crucial to incorporate such interaction effects in addition to the main effects. However, in the context of kernel machine regression analysis of its kind, existing methods cannot be utilized in a situation, where a kernel is available but its underlying real variants are unknown. To address such limitations, I introduce a general kernel machine regression framework using principal component analysis for jointly testing main and interaction effects. It begins with extracting principal components from an input kernel through the singular value decomposition. Then, it employs the principal components as surrogate variants to construct three endogenous kernels for the main effects, interaction effects, and both of them, respectively. Hence, it works with a kernel as an input without knowing its underlying real variants, and also detects either the main effects, interaction effects, or both of them robustly. I also introduce its omnibus testing extension to multiple input kernels, named OmniK. I demonstrate its use for human microbiome studies.
一种治疗方法对健康或疾病反应的影响可能会受到基因或微生物变异的改变。这涉及到基因或微生物变异与治疗之间的相互作用效应问题。为了有力地发现基因或微生物生物标志物，除了主效应外，纳入这种相互作用效应至关重要。然而，在这类核机器回归分析的背景下，现有方法无法应用于一种情况，即存在核但未知其潜在的真实变异。为了解决这些局限性，我引入了一个使用主成分分析的通用核机器回归框架，用于联合检验主效应和相互作用效应。它首先通过奇异值分解从输入核中提取主成分。然后，它将主成分用作替代变异，分别构建用于主效应、相互作用效应以及两者的三个内生核。因此，它在不知道潜在真实变异的情况下以核作为输入进行工作，并且还能稳健地检测主效应、相互作用效应或两者。我还介绍了它对多个输入核的综合检验扩展，称为OmniK。我展示了它在人类微生物组研究中的应用。
