Target-site dynamics explain a large share of apparent microRNA differential expression

Prediction performance of ML models for miRNA logFC. (A) Line plot of mean R² values across different ML algorithms as a function of the number of input features (tick marks represent logarithmically spaced values). Algorithms are abbreviated as follows: (EN) elastic net, (LA) lasso, (RI) ridge regression, (HB) HistGradientBoost, (MLP) multilayer perceptron, (RF) random forest, (OLS) ordinary least squares. (B) Distribution of R² values from EN trained with 1000 features across all evaluated miRNAs. (C) Fraction of miRNAs achieving R² > 0.5 as a function of the number of input features (tick marks represent logarithmically spaced values). (D) Dot plot of Pearson correlation between predicted and observed logFC values versus miRNA-specific R². (E) Comparison of cross-validation R² against test set R² across miRNAs. (F) Mean absolute error (MAE) and mean squared error (MSE) across algorithms (for 1000 features). (G) Relationship between the standard deviation of predicted logFC values and relative error (RMSE/standard deviation). The distribution of standard deviation values is shown as a histogram.