DB-IRES: a deep learning model based on ensemble learning for predicting internal ribosome entry sites

Area of Science:

• Molecular Biology
• Bioinformatics

Background:

• Internal ribosome entry sites (IRES) are crucial for cap-independent translation initiation in viral and cellular mRNAs.
• Accurate IRES identification is vital for understanding viral pathogenesis, cellular translation, and developing therapeutics.
• Current manual and computational methods for IRES identification are limited in efficiency, cost, and accuracy.

Purpose of the Study:

• To develop a novel, accurate, and robust computational tool for predicting internal ribosome entry sites (IRES).
• To overcome the limitations of existing methods in IRES identification.

Main Methods:

• Developed DB-IRES, a deep learning model utilizing densely connected 1D convolutional neural network blocks, bi-directional gated recurrent units, and a self-attention mechanism.
• Employed an ensemble learning framework and a five-fold cross-validation strategy for model training.
• Evaluated model performance on an independent test set using multiple metrics.

Main Results:

• The DB-IRES model demonstrated superior and more robust predictive performance compared to existing computational methods.
• The hybrid architecture and ensemble design of DB-IRES contributed to its enhanced discriminatory capabilities.
• The model achieved high accuracy and reliability in predicting IRES elements.

Conclusions:

• DB-IRES provides a reliable and precise computational tool for IRES element prediction.
• The enhanced performance of DB-IRES facilitates deeper functional studies of IRES biology.
• This tool supports broader applications in RNA research and the development of RNA-targeted therapeutics.