結合變異數數量性狀基因座篩選與懲罰張量迴歸於基因基因交互作用分析

Abstract

許多基因遺傳變異已經被發現與人類表型或複雜疾病的形成有關。然而，基因與基因間（G×G）的交互作用卻難以被偵測出來，更遑論其效果量的估計。現今，檢測變異數數量性狀基因座（variance quantitative trait loci, vQTL）以及使用懲罰張量迴歸（penalized tensor regression, PTensor）有助於同時達成上述兩項目標。考量傳統估計G×G交互作用的效果量其時間成本較高，本次研究目的在於結合這兩種方法，建立一套兩階段分析策略。首先，在第一階段時，分別使用單變量尺度檢定（univariate scale test, UST）、分位數整合線性模型（quantile integral linear model, QUAIL）、離差迴歸模型（deviation regression model, DRM）、Levene’s檢定（Levene’s test, LT），以及雙重廣義線性模型（double generalized linear model, DGLM）等方法來篩選vQTL。接著，於第二階段時，將篩選出來的vQTL透過PTensor來估計G×G交互作用的效果量。 我們將此策略應用在臺灣人體生物資料庫（Taiwan Biobank, TWB），並將其區分為TWB1（作為複製性群體）與TWB2（作為探索性群體）兩個子群體進行分析，樣本數分別為25,200以及93,708位受試者。本次分析主要探討血紅素（hemoglobin, HB）、血球比容（hematocrit, HCT）、紅血球數量（red blood cells, RBC）、白血球數量（white blood cells, WBC）、血小板數量（platelets）等性狀的G×G交互作用。此外，也會根據運算時間和G×G交互作用從TWB2成功複製到TWB1的比例來評估各分析策略的表現。我們的結果顯示，兩階段分析策略可以減少PTensor需要估計的係數數量，因此能有效減少運算時間。同時，由於兩階段分析策略確保G×G交互作用皆是由兩個vQTL所組成，故在血紅素和紅血球數量等性狀中，能提高G×G交互作用在不同的群體中成功複製的比例。在這些兩階段分析中，特別是在血紅素和紅血球數量等性狀上，DRM-PTensor和DGLM-PTensor這兩個方法表現出最高的複製成功率。然而，如果同時考量到計算效率，則以DRM-PTensor的表現較佳，使其成為兼具高效能與實用性的首選方法。

Numerous genetic variants have been identified as contributing factors to human phenotypes and the development of complex diseases. However, detecting gene-gene (G×G) interactions is much more difficult, not to mention the challenge of estimating their effect size. Nowadays, identifying variance quantitative trait loci (vQTL) and using penalized tensor regression (PTensor) can simultaneously achieve the two goals mentioned above. Given the computational burden associated with traditional G×G effect size estimation, this study proposes a two-stage analytical strategy integrating these two methods. In the first stage, we screened for variance quantitative trait loci (vQTL) using the Univariate Scale Test (UST), Deviation Regression Model (DRM), Levene’s Test (LT), Double Generalized Linear Model (DGLM), and Quantile Integral Linear Model (QUAIL). Subsequently, in the second stage, PTensor was employed to estimate the G×G interaction effect sizes for the identified vQTL. We applied our strategies to the Taiwan Biobank (TWB) data, which were divided into two cohorts, TWB1 (as a replication cohort) and TWB2 (as a discovery cohort), containing 25,200 and 93,708 individuals, respectively. The analysis aimed to explore G×G interactions for hemoglobin (HB), hematocrit (HCT), red blood cells (RBC), white blood cells (WBC), and platelets. In addition, the performance of each analysis strategy was evaluated based on the computation time and the replication rate of G×G interactions from TWB2 to TWB1. Our results indicated that the two-stage strategies reduced the number of coefficients to be estimated by PTensor, thereby effectively decreasing the computation time. Moreover, because the two-stage strategy ensured that G×G interactions were composed of two vQTL, it improved the replication rate of G×G interactions for traits such as HB and RBC across different cohorts. Among the two-stage analyses, the DRM-PTensor and DGLM-PTensor methods achieved the highest replication rate across multiple traits, specifically HB and RBC. However, when computational efficiency was also taken into consideration, the DRM-PTensor method performed more optimally, making it the preferred approach for both high efficacy and practical utility.