Prediction of structural features and application to outer membrane protein identification

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in biology

Background:

• Protein three-dimensional (3D) structures offer critical biological insights.
• One-dimensional structural properties (secondary structure, solvent accessibility, residue depth, torsion angles) are vital for predicting protein function, fold recognition, and ab initio folding.

Purpose of the Study:

• To predict various protein structural features using neural network learning.
• To improve outer membrane protein identification by incorporating predicted structural features.

Main Methods:

• Neural network learning was employed for predicting secondary structure, residue solvent accessibility, residue depth, and backbone torsion angles.
• A profile-to-profile alignment method was used to enhance outer membrane protein identification by integrating predicted structural features into a scoring function.

Main Results:

• Protein secondary structure prediction achieved approximately 80% Q3 accuracy on an independent test dataset.
• Prediction of relative solvent accessibility yielded a mean absolute error of 0.164, while residue depth prediction had a lowest mean absolute error of 0.062.
• Outer membrane protein identification accuracy improved by ~3% at a 1% false positive level with the inclusion of predicted structural features.

Conclusions:

• Neural network-based prediction of protein structural features is effective and accurate.
• Incorporating predicted structural features significantly enhances outer membrane protein identification.
• The study provides two user-friendly programs, PSSM-2-Features and PPA-OMP, for predicting structural features and identifying outer membrane proteins, respectively.