Riboexp: an interpretable reinforcement learning framework for ribosome density modeling.

Area of Science:

• Computational Biology
• Molecular Biology
• Genetics

Background:

• Translation elongation is a critical step in protein biosynthesis, characterized by non-uniform ribosome distribution on messenger RNA (mRNA) transcripts.
• Understanding ribosome density determinants is key to deciphering gene expression regulation and optimizing protein synthesis.

Purpose of the Study:

• To develop a novel deep reinforcement learning (DRL) framework, Riboexp, for modeling translation elongation dynamics.
• To predict ribosome density on mRNA transcripts with high accuracy and identify key sequence features influencing this distribution.

Main Methods:

• A deep reinforcement learning-based framework, Riboexp, was developed.
• Riboexp utilizes a policy network for context-dependent feature selection in ribosome density prediction.
• The model was trained and validated on datasets from three different species.

Main Results:

• Riboexp significantly outperforms existing state-of-the-art methods in predicting ribosome density, achieving up to a 5.9% increase in per-gene Pearson correlation coefficient.
• The framework identifies more informative sequence features compared to standard deep learning attribution methods.
• Application of Riboexp for codon optimization led to a ~31% increase in protein production compared to previous methods.

Conclusions:

• Riboexp is a powerful computational tool for studying translation dynamics and protein synthesis.
• The framework provides valuable biological insights into the mechanisms governing ribosome distribution during translation elongation.
• Riboexp facilitates improved protein production through optimized codon usage.