CodonBERT: a BERT-based architecture tailored for codon optimization using the cross-attention mechanism.

Area of Science:

• Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Codon optimization is crucial for mRNA vaccine efficacy, influencing protein stability and expression.
• Vast sequence space for mRNA presents challenges for in silico optimization methods.
• Existing deep learning methods struggle with long-term codon dependencies.

Purpose of the Study:

• To develop an advanced deep learning model for mRNA codon optimization.
• To address limitations of current machine translation-inspired approaches.
• To improve in silico prediction of stable and highly expressed mRNA sequences.

Main Methods:

• Developed CodonBERT, a BERT-based architecture utilizing cross-attention for codon optimization.
• Employed a masked codon sequence (key/value) and amino acid sequence (query) approach.
• Trained CodonBERT on high-expression transcripts from the Human Protein Atlas.

Main Results:

• CodonBERT effectively captures long-term dependencies between codons and amino acids.
• Demonstrated the model's capability as a customized training framework for specific optimization targets.
• The model shows promise in enhancing mRNA vaccine design.

Conclusions:

• CodonBERT offers a novel and effective approach to codon optimization for mRNA vaccines.
• The model's ability to handle long-term dependencies surpasses existing methods.
• CodonBERT provides a valuable tool for designing more stable and highly expressed mRNA therapeutics.