Detecting sequence signals in targeting peptides using deep learning.

Area of Science:

• Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Machine learning in bioinformatics typically predicts sequence features.
• These methods can also uncover novel biological mechanisms.
• Protein targeting signals direct proteins to specific cellular compartments.

Purpose of the Study:

• Introduce TargetP 2.0, a novel method for identifying N-terminal protein sorting signals.
• Utilize machine learning to gain new insights into the biological basis of protein targeting.
• Investigate the influence of specific amino acid residues on protein localization predictions.

Main Methods:

• Developed TargetP 2.0, a state-of-the-art machine learning model.
• Employed attention mechanisms within the neural network to identify key predictive features.
• Analyzed sequence data from proteins targeted to the secretory pathway, mitochondria, and plastids.

Main Results:

• The second amino acid residue (following methionine) strongly influences protein targeting predictions.
• Two-thirds of chloroplast/thylakoid transit peptides feature alanine at position 2, versus 20% in other plant proteins.
• In fungi and single-celled eukaryotes, only 30% of targeting peptides allow N-terminal methionine removal, compared to 60% for non-targeted proteins.

Conclusions:

• Machine learning models like TargetP 2.0 provide significant biological insights beyond mere prediction.
• The second residue's identity is a crucial, previously underappreciated feature for N-terminal sorting signal classification.
• These findings enhance our understanding of protein trafficking and cellular organization.