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

Designing signal peptides (SPs) for efficient recombinant protein secretion remains challenging, as current approaches depend largely on labor-intensive screening. We developed a transformer-based model trained on 158,768 SP-protein pairs from Gram-positive bacteria to generate type-specific Sec- or Tat-type SPs for given mature proteins. The model uses tailored tokenization strategies, including region delimiter tokens, to enable region-aware sequence design. In silico evaluation showed accurate SP classification and mean pairwise sequence identities above 60% compared with native SPs. Training exclusively on Gram-positive data outperformed training on a universal dataset, highlighting mechanistic differences in SP architectures. Beam-search decoding and additional sampling methods further improved sequence diversity and ensured robust SP generation. Experimental validation in Corynebacterium glutamicum demonstrated successful secretion for 15 of 16 designed SPs across two target proteins (M18 and XynA). This study establishes a practical, data-driven framework for rational SP design, supporting more efficient protein biomanufacturing in Gram-positive hosts.