Fig. 4
From: UTR-DynaPro: a CNN–transformer multimodal language model for decoding 5′UTR regulatory mechanisms
Ablation study of UTR-Dynapro on MRL prediction using the Random_Vary dataset (random 5′ UTRs). Performance was evaluated under different ablation settings, including removing specific modules (w/o Fusion, w/o MoE, w/o ExpInfo), restricting to Transformer-only or CNN-only architectures, and varying kernel sizes. Evaluation metrics include: (a) coefficient of determination (R2), (b) Pearson correlation coefficient, (c) Spearman correlation coefficient, (d) root mean square error (RMSE), and (e) mean absolute error (MAE). (f) Radar chart summarizing overall performance across all metrics. Results indicate that each module contributes to predictive accuracy, with the full UTR-Dynapro model achieving the best overall balance.
