Fig. 8
From: UTR-DynaPro: a CNN–transformer multimodal language model for decoding 5′UTR regulatory mechanisms
Prediction of mRNA translational efficiency (TE) and expression level (EL) on endogenous datasets. (a–c) Ablation studies of UTR-Dynapro on TE prediction for three cell types: muscle, PC3, and HEK, evaluated by Spearman correlation under different ablation settings (w/o Dynamic Fusion, w/o MoE, w/o ExpInfo, Transformer-only, CNN-only, and kernel sizes. (d) Comparative performance of UTR-Dynapro and baseline models (UTR-LM, Optimus, Cao-RF, Kipoi, MTtrans, RNAFM_MLP, and RNABERT_MLP) on TE prediction across the three cell types. (e) Comparative performance on EL prediction across the same cell types. UTR-Dynapro consistently achieves higher correlation values compared to baseline methods, highlighting its robustness and generalizability in modeling endogenous regulatory mechanisms.
