Predicting gene sequences with AI to study codon usage patterns.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Selective pressures shape codon usage, optimizing biological signals that are not fully understood.
• Codon usage patterns are influenced by evolutionary processes and gene expression levels.

Purpose of the Study:

• To train AI models to predict codon usage based on amino acid sequences in various organisms.
• To investigate the extent to which naturally occurring codon patterns can be learned and utilized for improved prediction.
• To explore the relationship between prediction accuracy, gene expression, protein length, and evolutionary factors.

Main Methods:

• Trained artificial intelligence (AI) models on protein sequences from *Saccharomyces cerevisiae*, *Schizosaccharomyces pombe*, *Escherichia coli*, and *Bacillus subtilis*.
• Evaluated model predictions on separate datasets of proteins with varying lengths and expression levels.
• Compared AI model performance against naïve frequency-based prediction methods.

Main Results:

• AI models significantly outperformed frequency-based approaches, indicating learnable dependencies in evolutionary-selected codon usage.
• Prediction accuracy was higher for highly expressed genes and greater in bacteria than eukaryotes, supporting a link between selection pressure and effective population size.
• Models showed improved accuracy for longer proteins in *S. cerevisiae* and bacteria, suggesting a connection to cotranslational folding; gene functionality and conservation also impacted performance.

Conclusions:

• Contemporary AI methods can effectively learn complex, evolutionary-selected codon usage patterns.
• The developed deep-learning-based prediction tool offers insights into codon optimization for endogenous and heterologous protein expression.
• Findings support hypotheses linking codon usage complexity to selective pressure and effective population size.