Transformer-Based Prediction of Sec- and Tat-Type Signal Peptides for Enhanced Bacterial Protein Secretion

Area of Science:

• Biotechnology
• Molecular Biology
• Bioinformatics

Background:

• Designing signal peptides (SPs) for efficient recombinant protein secretion is crucial but challenging.
• Current methods rely heavily on labor-intensive screening processes.

Purpose of the Study:

• To develop a computational framework for rational signal peptide design.
• To generate type-specific Sec- or Tat-type SPs for recombinant protein secretion in Gram-positive bacteria.

Main Methods:

• Developed a transformer-based AI model trained on 158,768 SP-protein pairs.
• Utilized tailored tokenization strategies, including region delimiter tokens, for region-aware sequence design.
• Employed beam-search decoding and sampling methods for sequence diversity and robust generation.

Main Results:

• The model achieved accurate in silico SP classification with high sequence identity (>60%) to native SPs.
• Experimental validation in Corynebacterium glutamicum showed successful secretion for 15 out of 16 designed SPs.
• Training exclusively on Gram-positive data demonstrated superior performance, indicating host-specific mechanistic differences.

Conclusions:

• Established a practical, data-driven framework for rational signal peptide design.
• The AI model enables efficient generation of functional signal peptides for protein biomanufacturing.
• This approach supports enhanced protein secretion in Gram-positive hosts.