Exploiting physico-chemical properties in string kernels.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• String kernels are widely used for biological sequence classification.
• Standard string kernels overlook amino acid physico-chemical properties (e.g., size, hydrophobicity, charge).
• Incorporating these properties is crucial, particularly for datasets with limited training data.

Purpose of the Study:

• To develop novel string kernels that integrate physico-chemical properties of amino acids.
• To evaluate the performance of these new kernels in biological sequence classification tasks.

Main Methods:

• Proposed novel string kernels combining physico-chemical descriptors with standard string kernel approaches.
• Assessed kernel performance on MHC-peptide binding classification using position-specific kernels.
• Evaluated kernels on protein classification using substring spectrum analysis.

Main Results:

• The proposed string kernels incorporating amino acid properties demonstrated improved performance over standard string kernels.
• Performance gains were observed in both MHC-peptide binding and protein classification tasks.
• The RBF substring kernel combination showed consistent improvements without increasing computational complexity.

Conclusions:

• The novel string kernels effectively leverage amino acid physico-chemical properties for enhanced classification.
• These modified kernels offer superior performance for protein sequence-based inference.
• The proposed kernels are recommended for protein sequence analysis due to improved accuracy and maintained computational efficiency.