Enhancing protein expression in humans through codon optimization with transformer and contrastive learning

Following the success of COVID-19 vaccines and therapies, mRNA-based therapeutics have attracted significant attention. Codon optimization is crucial for improving translation efficiency and mRNA stability, both of which are necessary for developing effective mRNA vaccines and drugs. Traditional approaches often rely on codon usage tables and other frequency-based heuristics, which neglect sequence context and show limited scalability. We introduce COformer, a deep learning framework for codon optimization that captures codon-level features related to protein expression in Homo sapiens cells. By integrating convolutional neural networks (CNNs) and transformers, the model learns local and global sequence contexts influencing translation efficiency and stability, while contrastive learning aligns codon representations with amino acid identities. Trained on sequences optimized by commercial tools, COformer generates codon choices that, when validated in in vitro experiments, enhance protein expression through optimized codon usage and improved mRNA stability.