ICOR: improving codon optimization with recurrent neural networks.

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• Synonymous codons, while not altering amino acid sequences, significantly impact protein expression levels.
• Existing codon optimization methods often focus on high-frequency codons, neglecting the influence of rare codons.
• Heterologous gene expression relies on effective codon optimization for synthetic DNA sequences.

Purpose of the Study:

• To develop a novel codon optimization tool, ICOR, utilizing recurrent neural networks (RNNs).
• To learn codon usage bias from a genomic dataset of Escherichia coli for improved protein expression.
• To predict synonymous codons for synthetic genes to optimize expression in Escherichia coli.

Main Methods:

• Compiled a deep learning dataset of over 7,000 non-redundant, high-expression genes from Escherichia coli.
• Employed a bidirectional long short-term memory (LSTM)-based RNN architecture to learn sequential codon context.
• Developed ICOR, a tool predicting synonymous codons for optimal expression.

Main Results:

• ICOR's sequential context-based codon selection demonstrated greater similarity to the host genome.
• Computational metrics predicted superior theoretical protein expression optimization compared to frequency-based approaches.
• Evaluated ICOR on 1,481 Escherichia coli genes and 40 characterized DNA sequences using five performance metrics.

Conclusions:

• In silico results suggest ICOR theoretically enhances recombinant protein expression more effectively than existing techniques.
• ICOR provides a novel approach to codon optimization by incorporating sequential context.
• The ICOR tool is available as an open-source software package.