IRESGet: A transformer framework-based model to predict internal ribosome entry sites

Internal ribosome entry site (IRES) is a cap-independent translation element widely used in circular RNA-based protein expression. Current IRES identification relies heavily on labor-intensive experiments, and existing computational methods based on machine learning or CPU-based deep learning suffer from limited accuracy and efficiency.

Here, we present IRESGet, a Transformer-based model for IRES sequence prediction. The model was trained and validated on datasets curated from published literature and public databases. Among five feature encoding strategies, RNA-FM achieved optimal performance. The Transformer architecture was employed to model the preprocessed sequences, followed by hyperparameter optimization using the Optuna library.

IRESGet outperformed five mainstream machine learning frameworks by avoiding overfitting and demonstrated superior accuracy and computational efficiency compared to two established tools, DeepCIP and IRESfinder. Our interpretability analysis reveals that the model primarily attends to the 5' proximal region, and a de novo motif SBNCAGVNVNNN enriched in high‑attention fragments was identified. TOMTOM alignment against the CIS‑BP database (115,667 records) at q‑value < 0.05 returned 176 significant matches (best q‑value = 0.021365), suggesting that the motif represents a statistically enriched and highly specific sequence pattern recognized by a small subset of RBPs.

Our research demonstrates that IRESGet outperforms other machine learning approaches and existing software, offering superior performance. This breakthrough presents a novel solution for IRES prediction. Nevertheless, current dataset limitations may result in certain false positive outcomes. We are confident that with enhanced data quality, the IRES prediction model will achieve substantial improvements in accuracy and effectiveness.