A statistical method for flagging weak spots improves normalization and ratio estimates in microarrays
A statistical method for flagging weak spots improves normalization and ratio estimates in microarrays
复制标题
DOI：
10.1152/physiolgenomics.00020.2001
复制
发表时间：
2001-10-10
影响因子：
4.6
通讯作者：
She, JX
中科院分区：
文献类型：
--
作者：
Yang, MCK;Ruan, QG;She, JX
在过去的几年中，有一个显着增加使用cDNA微阵列来监测基因表达的变化在生物系统中。来自这些实验的数据通常被转换成实验样品和共同参考样品之间的表达比率，用于随后的数据分析。这种关键转换的准确性取决于两个主要参数：信号强度和实验与参考信号强度的归一化。在这里，我们描述并验证了一个新的模型，微阵列信号强度，有一个乘法变化和一个添加剂的背景变化。使用重复实验和模拟数据，我们发现，信号强度是最关键的参数，影响性能的归一化，比率估计的准确性，重现性，特异性和灵敏度的微阵列实验。因此，我们开发了一种统计程序，以基于斑点与相邻斑点之间的背景差异的标准偏差（delta（ij））来标记具有弱信号强度的斑点，即，如果信号弱于C δ（ij），则认为光点太弱。我们的研究表明，当这个阈值（c）很小时，归一化和比率估计是不可接受的。我们进一步表明，当应用c（c = 6）的合理折衷时，使用对数比率的修剪平均值的归一化表现略好于全局强度和比率平均值。这些研究表明，降低背景噪声是提高微阵列实验质量的关键。
Over the last few years, there has been a dramatic increase in the use of cDNA microarrays to monitor gene expression changes in biological systems. Data from these experiments are usually transformed into expression ratios between experimental samples and a common reference sample for subsequent data analysis. The accuracy of this critical transformation depends on two major parameters: the signal intensities and the normalization of the experiment vs. reference signal intensities. Here we describe and validate a new model for microarray signal intensity that has one multiplicative variation and one additive background variation. Using replicative experiments and simulated data, we found that the signal intensity is the most critical parameter that influences the performance of normalization, accuracy of ratio estimates, reproducibility, specificity, and sensitivity of microarray experiments. Therefore, we developed a statistical procedure to flag spots with weak signal intensity based on the standard deviation (delta (ij)) of background differences between a spot and the neighboring spots, i.e., a spot is considered as too weak if the signal is weaker than c delta (ij). Our studies suggest that normalization and ratio estimates were unacceptable when this threshold (c) is small. We further showed that when a reasonable compromise of c (c = 6) is applied, normalization using trimmed mean of log ratios performed slightly better than global intensity and mean of ratios. These studies suggest that decreasing the background noise is critical to improve the quality of microarray experiments.