TSignal: a transformer model for signal peptide prediction

Area of Science:

• Computational Biology
• Molecular Biology
• Bioinformatics

Background:

• Signal peptides (SPs) are essential for protein translocation into the endoplasmic reticulum.
• Predicting SPs and their cleavage sites is challenging due to sequence variability and lack of conserved motifs.

Purpose of the Study:

• To develop a deep learning model for accurate signal peptide prediction.
• To identify the cleavage site between signal peptides and mature proteins.

Main Methods:

• Utilized a deep transformer-based neural network architecture (TSignal).
• Employed BERT language models and dot-product attention techniques.
• Trained and evaluated on common benchmark datasets.

Main Results:

• TSignal achieved competitive accuracy in predicting signal peptide presence.
• TSignal demonstrated state-of-the-art accuracy in predicting cleavage sites across various SP types and organisms.
• The model identified biologically relevant information from diverse sequences.

Conclusions:

• TSignal offers a powerful, data-driven approach for signal peptide and cleavage site prediction.
• The model's performance advances the field of protein secretion analysis.
• TSignal is publicly available for research use.