Peptide programs: applying fragment programs to protein classification

Functional prediction/classification of proteins is a central problem in bioinformatics. Alignment methods are a useful approach, but have limitations, which have prompted the development and use of machine learning approaches. However, traditional machine learning approaches are unable to exploit sequence data directly, and instead use derived sequence features or Kernel functions to obtain a feature space. Because theoretically all information necessary to predict a protein's structure and function is contained in its sequence, a methodology that could exploit sequence data directly could be advantageous. A novel machine learning methodology for protein classification, inspired in the concept of fragment programs, is presented. This methodology consists in assigning a minimal computer program to each of the 20 amino acids, and then representing a protein as the program resulting from applying sequentially the programs of the amino acids which compose its sequence. The basic concepts of the methodology presented (peptide programs) are discussed and a framework is proposed for their implementation, including instruction set, virtual machine, evaluation procedures and convergence methods. The methodology is tested in the binary classification of 33,500 enzymes into 182 distinct Enzyme Commission (EC) classes. The average Matthews correlation coefficient of the binary classifiers is 0.75 in training and 0.68 in validation. Overall, the results obtained demonstrate the potential of the proposed methodology, and its ability to extract knowledge from sequence data, using very few computational resources