P-TarPmiR accurately predicts plant-specific miRNA targets

Abstract

microRNAs (miRNAs) are small non-coding ribonucleic acids that post-transcriptionally regulate gene expression through the targeting of messenger RNA (mRNAs). Most miRNA target predictors have focused on animal species and prediction performance drops substantially when applied to plant species. Several rule-based miRNA target predictors have been developed in plant species, but they often fail to discover new miRNA targets with non-canonical miRNA-mRNA binding. Here, the recently published TarDB database of plant miRNA-mRNA data is leveraged to retrain the TarPmiR miRNA target predictor for application on plant species. Rigorous experiment design across four plant test species demonstrates that animal-trained predictors fail to sustain performance on plant species, and that the use of plant-specific training data improves accuracy depending on the quantity of plant training data used. Surprisingly, our results indicate that the complete exclusion of animal training data leads to the most accurate plant-specific miRNA target predictor indicating that animal-based data may detract from miRNA target prediction in plants. Our final plant-specific miRNA prediction method, dubbed P-TarPmiR, is freely available for use at http://ptarpmir.cu-bic.ca . The final P-TarPmiR method is used to predict targets for all miRNA within the soybean genome. Those ranked predictions, together with GO term enrichment, are shared with the research community.

Figure 1

The average value and standard deviation of the performance metrics of each classifier (Human, Human+ath, Human+Plant, and Plant) over the four test species (Glycine max, Oryza sativa, Populus trichocarpa, and Brachypodium distachyon). The performance metrics included are area under the Precision–Recall curve (AUC), recall (Re), Precision (Pr), and accuracy (ACC).

Figure 2

The Precision–Recall curves of each of the four classifiers (Human, Human+ath, Human+Plant, and Plant) for the four plant test species (Glycine max, Oryza sativa, Populus trichocarpa, and Brachypodium distachyon).

Figure 3

Density plots of the prediction scores of the four classifiers (Human, Human+ath, Human+Plant and Plant) on the gma test set. Here, prediction scores for negative test samples are shown in red, while positive test sample scores are shown in blue. A stronger classifier will lead to greater discriminability between the scores generated from positive and negative test miRNA:mRNA pairs.

Figure 4

Screenshots of the P-TarPmiR web server job submission page, (a) for file upload and (b) for direct text input of sequences. Both types of submissions include the ability to add an email so the web server can notify the user when the job is complete. This functionality is particularly useful for large jobs.

Figure 5

A screenshot of example results of a job submission on P-TarPmiR web server. The results page includes the predicted binding of the seed location, the index of the predicted seed location on the miRNA and mRNA, the miRNA and mRNA seed sequences, and the prediction confidence of the miRNA–mRNA pair.