Background: High throughput molecular-interaction studies using immunoprecipitations (IP) or affinity purifications are powerful and widely used in biology research. One of many important applications of this method is to identify the set of RNAs that interact with a particular RNA-binding protein (RBP). Here, the unique statistical challenge presented is to delineate a specific set of RNAs that are enriched in one sample relative to another, typically a specific IP compared to a non-specific control to model background. The choice of normalization procedure critically impacts the number of RNAs that will be identified as interacting with an RBP at a given significance threshold - yet existing normalization methods make assumptions that are often fundamentally inaccurate when applied to IP enrichment data.
Methods: In this paper, we present a new normalization methodology that is specifically designed for identifying enriched RNA or DNA sequences in an IP. The normalization (called adaptive or AD normalization) uses a basic model of the IP experiment and is not a variant of mean, quantile, or other methodology previously proposed. The approach is evaluated statistically and tested with simulated and empirical data.
Results and conclusions: The adaptive (AD) normalization method results in a greatly increased range in the number of enriched RNAs identified, fewer false positives, and overall better concordance with independent biological evidence, for the RBPs we analyzed, compared to median normalization. The approach is also applicable to the study of pairwise RNA, DNA and protein interactions such as the analysis of transcription factors via chromatin immunoprecipitation (ChIP) or any other experiments where samples from two conditions, one of which contains an enriched subset of the other, are studied.