MicroRNA Array Normalization: An Evaluation Using a Randomized Dataset as the Benchmark

MicroRNA arrays possess a number of unique data features that challenge the assumption key to many normalization methods. We assessed the performance of existing normalization methods using two microRNA array datasets derived from the same set of tumor samples: one dataset was generated using a blocked randomization design when assigning arrays to samples and hence was free of confounding array effects; the second dataset was generated without blocking or randomization and exhibited array effects. The randomized dataset was assessed for differential expression between two tumor groups and treated as the benchmark. The non-randomized dataset was assessed for differential expression after normalization and compared against the benchmark. Normalization improved the true positive rate significantly in the non-randomized data but still possessed a false discovery rate as high as 50%. Adding a batch adjustment step before normalization further reduced the number of false positive markers while maintaining a similar number of true positive markers, which resulted in a false discovery rate of 32% to 48%, depending on the specific normalization method. We concluded the paper with some insights on possible causes of false discoveries to shed light on how to improve normalization for microRNA arrays.

微RNA芯片（microRNA array）具备诸多独特的数据特征，这些特征对诸多归一化（normalization）方法的核心假设构成了挑战。我们采用源自同一批肿瘤样本的两套微RNA芯片数据集，对现有归一化方法的性能开展评估：其中一套数据集在芯片与样本的分配环节采用了区组随机化设计（blocked randomization design），因此不存在芯片混杂效应；另一套数据集未实施区组分组或随机分配，呈现出显著的芯片效应。该随机化数据集被用于评估两组肿瘤间的差异表达情况，并作为基准参照；而非随机化数据集则在完成归一化处理后进行差异表达分析，并与该基准参照进行比对。归一化处理可显著提升非随机化数据中的真阳性率，但错误发现率仍高达50%。在归一化前增设批次校正步骤，可在保留相近数量真阳性标记物的同时，进一步减少假阳性标记物的数量，此时根据具体归一化方法的不同，错误发现率可降至32%至48%区间。本研究文末就假阳性发现的潜在成因展开探讨，以期为优化微RNA芯片的归一化方法提供参考思路。