Maximum margin classifier working in a set of strings.

Area of Science:

• Bioinformatics
• Machine Learning
• Computational Biology

Background:

• String data classification is crucial across many fields, with current methods converting strings to numerical vectors via string kernels.
• This conversion process leads to information loss and hinders theoretical generalization error evaluation using probability theory.
• Existing approaches struggle with the increasing volume and complexity of string data in modern applications.

Purpose of the Study:

• To develop a novel string data classifier that operates directly within the string space.
• To establish a theoretical framework for evaluating the generalization error of string classifiers using probability theory.
• To demonstrate the practical utility of the proposed classifier in biological sequence analysis.

Main Methods:

• Extension of a limit theorem for consensus sequences of strings to enable theoretical analysis.
• Construction of a learning machine that directly classifies string data without intermediate numerical conversion.
• Application of the classifier to predict protein-protein interactions and classify RNA secondary structures.

Main Results:

• The developed learning machine demonstrates asymptotically optimal string classification performance.
• The theoretical framework allows for probability-based evaluation of generalization error for string classifiers.
• Successful application in predicting protein-protein interactions and classifying RNA structures validates the method's practical utility.

Conclusions:

• The proposed string classifier overcomes limitations of traditional numerical conversion methods.
• This approach enables robust theoretical analysis and practical application in bioinformatics.
• The study advances the field of string data classification with a theoretically sound and practically effective method.