Efficient discriminative models for proteomics with simple and optimized features

Abstract

The broad diversity in biology offers a panoply of interesting categorization problems for the machine learning community. New challenges arise in modern subjects such as protein classification, where huge and complex datasets are common, and demand the most accurate and fast classifiers to retrieve meaningful biological traits in acceptable time. Although the Support Vector Machine algorithm has been playing a significant role by offering the most precise solutions in diverse domains, the problem of protein classification is far from being solved. Other successful Kernel Methods such as the Relevance Vector Machine and extensions that combine Recursive Feature Elimination in formulations capable of performing feature selection like SVM-RFE and RVM-RFE, were tested in a benchmark environment and compared to other popular statistical models such as Nearest Neighbor, Random Forest, Artificial Neural Networks and Logistics Regression. The results show that SVM-RFE can create classifiers with the highest recognition ability even using a simple compact feature set easily computable from protein primary structure. Plus, these models allow getting predictions in a time scale reduced by orders of magnitude when compared with the standardly used PSI-BLAST.