微陣列實驗中評估基因表現量資料系統誤差 統計方法之模擬研究

Abstract

微陣列是二十一世紀中一項突破性進展的技術。但是直到現今，美國食品及藥物管理局 (FDA) 才批准了第一項根據微陣列技術所發展的生物晶片產品。其中一項主要的原因，是由於從不同的實驗室及不同的實驗平台所取得之基因表現量的強度測量值有系統誤差的產生。理想情況下，對相同的基因，在不同實驗室或者在不同的平台之間，所取得的重複強度測量值應該是相同的。因此，在方法比較上，迴歸的方法可以適用於評估系統誤差，特別是在於實驗決策閥值上。然而，從不同的實驗室或者不同的實驗平台之間，所取得重複的強度測量值會有隨機變異。因此，傳統線性迴歸 (Ordinary Linear Regression) 就不適合使用。而簡單戴明迴歸 (Simple Deming Regression)以及遞迴再加權戴明迴歸 (Iteratively Reweighted General Deming Regression) 則適合使用於上述的情況。另一方面，對於不同的基因，表現量資料不是獨立的。根據斜率、截距、系統誤差、以及相對應信賴區間的估算，不同基因之中，強度量測值之間相關的影響，是未知的。在不同截距、斜率、系統誤差、實驗決策閥值、隨機誤差的架構、相關、以及樣本大小的幾種組合之下，我們執行一模擬研究，憑經驗依據參數的誤差以及包含率，來比較傳統線性迴歸、簡單戴明迴歸、以及遞迴再加權戴明迴歸的表現。從已出版的論文得來的數值資料說明應用。

Microarray technology is one of the breakthrough technologies in the twenty-first century. But only recently, the US Food and Drug Administration (FDA) approved the first biochip product based on the microarray technology. One of the primary reasons is the systematic bias of intensity measurements on gene expression data obtained from different laboratories and between different array platforms. Ideally the replicated intensity measurements of the same genes obtained from different laboratories or between different platforms should be same. Therefore, regression approaches for method comparison can be applied to assess the systematic bias, especially at the medical decision thresholds. However, replicated intensity measurements from different laboratories or between different platforms are subject to random errors. Consequently, ordinary linear regression (OLR) is not appropriate and simple Deming regression (SDR), and iteratively reweighted general Deming regression (IRGDR) should be used when both measurements contain random error. On the other hand, expression data of different genes are not independent. Impact of correlation of intensity measurements among different genes upon the estimation of slope, intercept, systematic bias, and their corresponding confidence intervals is not known. Under various combinations of intercept, slope, systematic bias, decision points, structures of random errors, correlations, and sample size, we conducted a simulation study to empirically compare performance of OLR, SDR, and IRGDR in estimating bias of the parameters and coverage probability. Numeric data from published papers illustrate the applications.