Protein fold recognition using geometric kernel data fusion.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Protein fold prediction is crucial for understanding protein function.
• Existing methods often rely on integrating diverse protein features using machine learning.
• Current multiple kernel learning techniques, typically using linear combinations, may lose valuable information.

Purpose of the Study:

• To develop novel methods for integrating heterogeneous protein features beyond simple linear combinations.
• To improve the accuracy of protein fold recognition and remote homology detection.

Main Methods:

• Developed geometry-inspired methods to combine kernel matrices, moving beyond convex linear combinations.
• Utilized sequence-based protein features, including position-specific scoring matrices and local sequence alignment.
• Incorporated functional domain composition through a hybridization model.

Main Results:

• Achieved a protein fold recognition accuracy of approximately 86.7% on the SCOP PDB-40D benchmark dataset.
• Further improved accuracy to 89.30% by incorporating a hybridization model for functional domain composition.
• Demonstrated effectiveness in protein remote homology detection by fusing multiple string kernels.

Conclusions:

• Geometry-inspired kernel fusion offers a more effective way to integrate diverse protein features compared to linear methods.
• The proposed hybridization model significantly enhances protein fold recognition accuracy.
• The developed methods provide a valuable advancement for protein structure and function prediction.