Evolving fisher kernels for biological sequence classification.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Fisher kernels are effective in bioinformatics but rely on high-quality generative models.
• Creating accurate, domain-specific generative models for novel problems remains a challenge.
• Existing kernel methods often lack mechanisms for incorporating prior knowledge or domain specificity.

Purpose of the Study:

• To present a novel framework for automatically creating domain-specific generative models.
• To enable the generation of Fisher kernels for support vector machines (SVMs) and other kernel methods.
• To improve the discriminative power of kernel-based methods by capturing prior knowledge.

Main Methods:

• Utilizing genetic algorithms (GAs) to evolve the structure of hidden Markov models (HMMs) for generative model creation.
• Generating Fisher kernels from the evolved HMMs.
• Integrating the Fisher kernels with SVMs (termed GA-SVM) for enhanced classification.

Main Results:

• The GA-SVM method demonstrates performance comparable or superior to state-of-the-art methods in classifying secretory protein sequences.
• GA-SVM outperforms sequence-similarity-based approaches in protein enzyme family classification without requiring homologous sequence data.
• The framework effectively identifies high-performing features from biological sequences without complex model tuning.

Conclusions:

• The proposed GA-SVM framework offers a novel and effective approach for generating domain-specific kernels in bioinformatics.
• This method addresses limitations in current kernel-based techniques by incorporating prior knowledge and domain specificity.
• GA-SVM provides a powerful tool for biological sequence analysis, enhancing classification accuracy and reducing reliance on extensive parameter tuning.