Workflow for training ML models to predict miRNA logFC. Matched TCGA samples containing mRNA expression, APA profiles, and miRNA expression are split into random training (80%) and test (20%) sets, with the training set further divided into cross-validation folds. Within each fold, samples are stratified by miRNA expression levels and randomly paired in both directions to prevent data leakage during training. For each miRNA, annotated target genes with valid APA measurements are selected, and differential features (mRNA logFC, ΔPDUI) are computed for each sample pair. The observed miRNA logFC serves as the prediction label, and the same feature computation is applied across folds, the test set, and the final training set after hyperparameter tuning. A separate ML model is trained for each miRNA to capture the relationship between transcriptomic changes and miRNA expression dynamics.
