Building high level knowledge from high dimensionality biological dataset (NCI60) using Genetic Algorithms and feature selection strategies

Abstract

When it comes to complex biological problems the use of conventional computation techniques has shown not to be the best approach. With the aim of selecting small sets of genes, that have strong predictive correlations with a disease, the Genetic Algorithms (GAs) are being increasingly used. In this paper, we propose a hybrid approach, using methods of feature selection and a classifier based on GA as a tool to identifying a subset of relevant genes and developing high-level classification rules for the cancer dataset NCI60, revealing concise and relevant information about the application domain. As a result it was obtained a set of IF-THEN rules with few genes per class and high predictive power that can be used as a classifier and assist experts to understand the biologic relationship between the genes and the classes of cancer. Moreover, the accuracy of the proposed approach overcame the results obtained by traditional classification methods such as PART, J48, Naive Bayes, Random Forest and IBK, demonstrating that the rules balance interpretability, comprehensibility and prediction precision.