MiRTif

MiRTif classifies predicted miRNA:target interactions using a support vector machine to filter false positives and improve the accuracy of miRNA target identification.

Key Features:

• Classifier: Uses a Support Vector Machine (SVM) to distinguish true miRNA:target interactions from false positives.
• Training data: Trained on 195 experimentally validated positive miRNA:target interaction pairs and 38 negative pairs.
• Sequence region modeling: Each interaction is divided into seed and non-seed regions for analysis of positional contributions.
• Feature representation: Constructs feature vectors using k-gram frequencies computed across seed and non-seed regions.
• Feature selection: Selects informative features based on their ability to discriminate true from false targets.
• Evaluation metrics: Assessed by 10-fold cross-validation with reported AUC 0.86, sensitivity 83.59%, and specificity 73.68%.
• False-positive filtering performance: Correctly identifies 28 out of 38 false positive interactions from existing prediction tools.
• Robustness testing: Tested using non-validated and randomly selected targets from miRBase to evaluate overfitting risk from limited negative samples.

Scientific Applications:

• False-positive filtering: Post-processes miRNA target predictions to remove predicted targets likely to be false positives.
• Prediction refinement: Improves the accuracy of downstream analyses that rely on miRNA:target interaction sets.
• Benchmarking and validation: Provides a computational filter for evaluating outputs from existing miRNA target prediction tools.

Methodology:

Implements an SVM trained on 195 positive and 38 negative experimentally validated miRNA:target pairs; splits interactions into seed and non-seed regions; encodes interactions with k-gram frequency feature vectors; selects informative features by discriminative ability and evaluates performance via 10-fold cross-validation; robustness assessed using non-validated and randomly selected targets from miRBase.