Combining evolutionary information and neural networks to predict protein secondary structure

Area of Science:

• Computational biology
• Bioinformatics
• Protein structure prediction

Background:

• Accurate prediction of protein secondary structure is crucial for understanding protein function and design.
• Previous methods using neural networks and multiple sequence alignments showed promise but required further refinement.

Purpose of the Study:

• To improve the accuracy of protein secondary structure prediction by incorporating additional evolutionary information into a neural network system.
• To evaluate the performance of the enhanced prediction method against existing techniques and experimental methods.

Main Methods:

• Utilized a three-level neural network system augmented with evolutionary information from multiple sequence alignments.
• Incorporated position-specific conservation weights, insertion/deletion counts, and global amino acid content as input features.
• Conducted rigorous cross-validation tests on diverse protein datasets, including recently solved structures.

Main Results:

• Achieved a sustained overall accuracy of 71.6% in cross-validation, with an average accuracy above 72% for 250 unique protein chains.
• Demonstrated a performance advantage of at least 6 percentage points in three-state accuracy compared to other methods.
• Developed a position-specific reliability index, achieving 88% accuracy for 40% of residues, comparable to homology modeling.

Conclusions:

• The enhanced neural network system significantly improves protein secondary structure prediction accuracy and reliability.
• The method offers a powerful computational tool for structural biology, comparable in accuracy to experimental techniques for secondary structure estimation.
• The developed reliability index provides valuable confidence measures for predicted secondary structure elements.