Transfer Learning for Designing Efficient Signal Peptides to Improve the Secretion Level of Recombinant Protein in Bacillus amyloliquefaciens

Signal peptides (SPs) play an essential role in determining
the
secretion efficiency of proteins of interest (POIs). However, the
manual identification of SPs with a high secretion potential is both
time-consuming and labor-intensive. Recently, many advanced machine
learning (ML) techniques have emerged in biology and food research.
This research aimed to utilize experimental SP-POI secretion data
to create ML models that could predict how SPs influence the POI secretion
efficiency. Given the limitations of the available data, which affected
model accuracy, this study introduced transfer learning and confirmed
its effectiveness through model selection experiments, leading to
the development of more precise ML models. Utilizing the SP generator
and ML models developed, high-quality SPs were successfully designed.
Experimental validation confirmed that 80% of ML-designed SPs secreted
the POI, with 60% achieving high-level secretion.