Mammalian genomes encode for thousands of microRNAs, which can potentially regulate the majority of protein-coding genes. They have been implicated in development and disease, leading to great interest in understanding their function, with computational methods being widely used to predict their targets. Most computational methods rely on sequence features, thermodynamics, and conservation filters; essentially scanning the whole transcriptome to predict one set of targets for each microRNA. This has the limitation of not considering that the same microRNA could have different sets of targets, and thus different functions, when expressed in different types of cells.
To address this problem, researchers at CINVESTAV, Mexico combined popular target prediction methods with expression profiles, via machine learning, to produce a new predictor: TargetExpress. Using independent data from microarrays and high-throughput RNA sequencing, they show that TargetExpress outperforms existing methods, and that their predictions are enriched in functions that are coherent with the added expression profile and literature reports.
Leave one out cross-validation
The area under the curve (AUC) for each “left out” experiment (indicated in the top legend) given different prediction models: TargetScan, TargetScan-intersect and TargetExpress-TS (green boxplots); microT, microT-intersect and TargetExpress-MT (orange); Sum, Sum-intersect and TargetExpress-Sum (blue)
This method should be particularly useful for anyone studying the functions and targets of miRNAs in specific tissues or cells.
Availability – TargetExpress is available at: http://targetexpress.ceiabreulab.org/ .