LOCUSTRA: accurate prediction of local protein structure using a two-layer support vector machine approach

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning

Background:

• Accurate prediction of local protein structure is crucial for advancing 3D structure prediction and structure-based database searches.
• Existing methods for secondary structure prediction and dihedral angle prediction have limitations in fine-grained local structure analysis.

Purpose of the Study:

• To introduce LOCUSTRA, a novel scheme for multiclass prediction of local protein structure.
• To evaluate the performance of LOCUSTRA against established secondary structure prediction methods.

Main Methods:

• Utilized a two-layer support vector machine (SVM) architecture.
• Employed a 16-letter structural alphabet for detailed local structure classification.
• Assessed prediction accuracy on an independent test set of 222 proteins.

Main Results:

• Achieved a prediction accuracy of Q16=61.0% for the 16 local structure classes.
• Obtained Q3=79.2% accuracy when mapping to three secondary structure classes (helix, sheet, coil).
• Reported a mean phi(psi) error of 24.74 degrees and a median RMSDA of 52.1 degrees per protein chain.

Conclusions:

• LOCUSTRA demonstrates competitive performance compared to related local structure prediction approaches.
• The developed method provides valuable insights into local protein structural elements.
• The LOCUSTRA web server is publicly accessible for research purposes.