Abstract

Low-density quantitative real-time PCR (qPCR) arrays are often used to profile expression patterns of microRNAs in various biological milieus. To achieve accurate analysis of expression of miRNAs, non-biological sources of variation in data should be removed through precise normalization of data. We have systematically compared the performance of 19 normalization methods on different subsets of a real miRNA qPCR dataset that covers 40 human tissues. After robustly modeling the mean squared error (MSE) in normalized data, we demonstrate lower variability between replicates is achieved using various methods not applied to high-throughput miRNA qPCR data yet. Normalization methods that use splines or wavelets smoothing to estimate and remove Cq dependent non-linearity between pairs of samples best reduced the MSE of differences in Cq values of replicate samples. These methods also retained between-group variability in different subsets of the dataset.