Transfer learning with pre-trained language models for protein expression level prediction in Escherichia coli

Area of Science:

• Molecular Biology
• Computational Biology
• Bioinformatics

Background:

• Predicting recombinant protein expression in Escherichia coli is complex due to gene regulation and translation factors.
• Current computational methods often focus on either codon usage or protein sequence, limiting prediction accuracy and scope.

Purpose of the Study:

• To develop an advanced computational framework for accurate prediction of recombinant protein expression levels.
• To integrate codon and protein sequence information using transfer learning for enhanced predictive power.

Main Methods:

• Introduced TLCP-EPE, a transfer learning framework fusing codon-level (CaLM) and protein-level (ProtT5) pre-trained language models.
• Utilized low-rank adaptation (LoRA) for fine-tuning models and a BiGRU-MLP predictor to integrate embeddings.
• Evaluated performance on two independent test datasets.

Main Results:

• TLCP-EPE demonstrated superior performance compared to state-of-the-art methods.
• Achieved robust predictive accuracy with AUC 0.835 on codon data and AUC 0.713 on protein data.
• Outperformed conventional codon-based metrics and existing deep learning baselines.

Conclusions:

• Dual-modal modeling of codon and protein sequences significantly improves expression level prediction accuracy and generalizability.
• The TLCP-EPE framework offers a powerful tool for rational protein design and biomanufacturing.
• This approach advances the field of computational protein engineering.