Learning the native-like codons with a 5’UTR and secondary RNA structure aided species-informed transformer model

Efficient protein expression across heterologous hosts remains a major challenge in synthetic biology, largely due to species-specific differences in codon usage and regulatory sequence context. A key difficulty lies in reconstructing the codon landscape of the target expression system within a foreign host with a native-like codon preference. To address this, we present TransCodon, a Transformer-based deep learning model that leverages both 5’ untranslated regions (5’UTRs) and coding sequences (CDS), along with explicit species identifiers and RNA secondary structure information, to learn nuanced codon usage patterns across diverse organisms. By incorporating multisource genomic data and modeling sequence dependencies in a masked language modeling paradigm, TransCodon effectively captures both local and global determinants of codon preference. Our experiments demonstrate that integrating species-level information during training significantly improves the model’s ability to predict optimal synonymous codons when considering different evaluation metrics. More importantly it identifies native-like codons with less divergence from natural sequences compared to other methods. Besides, TransCodon could capture more low-frequency codons which are often omitted by other deep learning-based methods. The results thus indicate that TransCodon as a robust codon language model has the potential for generating native-like CDS with high translational efficiency in target hosts.