Abstract:
|
Real-time quantitative polymerase chain reaction (qPCR) remains the gold standard for sensitive and accurate quantification of nucleic acid sequences. A major challenge in the analysis of qPCR data is the presence of non-detects, reactions failing to produce a measurable signal. While most current software replaces these non-detects with the maximum possible value, this introduces large biases in estimation of both absolute and differential expression. Recent approaches to address these biases underestimate the variability, leading to anti-conservative inference. We propose treating non-detects as non-random missing data, modeling the missing data mechanism, and using this model to implement a multiple imputation procedure. This procedure incorporates three sources of variability - uncertainty in the missing data mechanism, model uncertainty, and residual variance. The benefits of this approach are demonstrated using three real data sets. These data sets and the proposed methodology are freely available in the open-source R/Bioconductor package, nondetects.
|