PiPred - a deep-learning method for prediction of π-helices in protein sequences

Area of Science:

• Protein structure and bioinformatics
• Computational biology and structural bioinformatics

Background:

• Canonical π-helices are short, unstable protein secondary structures found in 15% of known protein structures.
• These elements are often located in functionally critical regions like ligand- and ion-binding sites.
• Predicting π-helices is challenging due to their similarity to α-helices, and current methods do not address this.

Purpose of the Study:

• To develop and present PiPred, a novel neural network-based tool for predicting π-helices in protein sequences.
• To rigorously benchmark PiPred's performance in π-helix prediction.

Main Methods:

• Development of a neural network-based prediction tool, PiPred.
• Rigorous benchmarking of PiPred's performance using established datasets (CB6133, CB513, CASP10, CASP11).

Main Results:

• PiPred achieves a per-residue precision of 48% and sensitivity of 46% for π-helix prediction.
• The tool correctly identifies α/π-bulges, even though trained only on canonical π-helices.
• Mispredictions of α-helices as π-helices suggest shared geometric characteristics.

Conclusions:

• PiPred demonstrates effective prediction of π-helices and related helical structures.
• The findings suggest that π-helices, α/π-bulges, and other helical deformations share sequence constraints.
• PiPred offers a valuable tool for analyzing protein secondary structures, with accessible online and standalone versions.