Modelling and Recognition of Protein Contact Networks by Multiple Kernel Learning and Dissimilarity Representations.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Multiple kernel learning (MKL) integrates diverse data representations for enhanced analysis.
• Predicting protein functional roles from structural data is crucial in bioinformatics.

Purpose of the Study:

• To develop a hybrid classification system using MKL for protein functional role prediction.
• To enable joint optimization of kernel weights and representative selection for knowledge discovery.

Main Methods:

• A hybrid classification system employing a linear combination of multiple kernels over dissimilarity spaces.
• Joint optimization of kernel weights and selection of pivotal patterns as representatives.
• Application to real proteomic data using eight representations from graph-based protein structure descriptions.

Main Results:

• The system demonstrated remarkable classification capabilities on proteomic data.
• Analysis of kernel weights identified suitable representations for classification.
• Selected representative patterns provided insights into the modeled biological system.
• Performance was benchmarked against a clustering-based classification system.

Conclusions:

• The proposed MKL system reliably predicts protein functional roles from folded structures.
• The knowledge discovery phase offers valuable insights for biologists and domain experts.
• The approach shows significant potential for advancing protein function prediction and biological data analysis.