seq2ribo: structure-aware integration of machine learning and simulation to predict ribosome location profiles from...

Area of Science:

• Computational Biology
• Bioinformatics
• Synthetic Biology

Background:

• Ribosome dynamics are crucial for protein expression but current methods require genomic context, limiting applications like mRNA vaccine design.
• Existing simulation methods like TASEP oversimplify translation by focusing only on codon elongation times.

Purpose of the Study:

• To develop a computational framework, seq2ribo, that predicts ribosome A-site locations using only mRNA sequence as input.
• To enable de novo design and optimization of mRNA sequences for applications such as mRNA vaccines.

Main Methods:

• Implemented a hybrid simulation and machine learning approach combining a structure-aware TASEP (sTASEP) with a refinement model.
• sTASEP incorporates codon wait times and structural features (local angles, base-pairing, positional buckets) to model translation.
• A polisher model refines simulated ribosome distributions for high-fidelity predictions.

Main Results:

• seq2ribo accurately predicts ribosome locations across diverse cell types, outperforming existing methods.
• Achieved high transcript-level (up to 0.920) and within-transcript (up to 0.186) correlations with experimental ribosome profiles from sequence alone.
• Demonstrated strong correlations with translation efficiency (up to 0.732) and protein expression (up to 0.903) by adding a task-specific head.

Conclusions:

• seq2ribo is the first method to achieve meaningful ribosome profile prediction from mRNA sequence alone.
• This framework offers a powerful new tool for synthetic biology, facilitating rational mRNA sequence design without expression data or genomic context.