Learned random-walk kernels and empirical-map kernels for protein sequence classification.

Area of Science:

• Computational biology
• Bioinformatics
• Genomics

Background:

• Biological sequence classification, particularly protein remote homology detection, is crucial in the genomics era.
• Existing Support Vector Machines (SVMs) with mismatch string kernels show limitations on difficult protein families.

Purpose of the Study:

• To propose novel computational approaches for enhanced protein remote homology detection.
• To improve prediction performance beyond current profile kernel methods.

Main Methods:

• Developed a convex combination of random-walk kernels to approximate an optimal random-step kernel.
• Constructed an empirical-map kernel utilizing a profile kernel.
• Leveraged pairwise sequence similarity information and unlabeled data.

Main Results:

• Both proposed kernels demonstrated significantly better prediction performance than the best profile kernel.
• Achieved overall mean ROC(50) scores above 0.90 across 54 protein families on a SCOP benchmark dataset.
• Outperformed previously published results on protein remote homology detection.

Conclusions:

• The novel kernel methods offer a substantial advancement in biological sequence classification.
• These approaches effectively address the challenges of remote homology detection using sequence data.
• The findings have significant implications for computational biology and genomics research.