A Novel Decomposing Model with Evolutionary Algorithms for Feature Selection in Long Non-Coding RNAs

Machine learning algorithms have been applied to numerous transcript datasets to identify Long non-coding RNAs (lncRNAs). Nevertheless, before these algorithms are applied to RNA data, features must be extracted from the original sequences. As many of these features can be redundant or irrelevant, the predictive performance of the algorithms can be improved by performing feature selection. However, the most current approaches usually select features independently, ignoring possible relations. In this paper, we propose a new model, which identifies the best subsets, removing unnecessary, irrelevant, and redundant predictive features, taking the importance of their co-occurrence into account. The proposed model is based on decomposing solutions and is called $k$ -rounds of decomposition features. In this model, the least relevant features are suppressed according to their contribution to a classification task. To evaluate our proposal, we extract from 5 plant species datasets, a set of features based on sequence structures, using GC content, k-mer (1-6), sequence length, and Open Reading Frame. Next, we apply 5 metaheuristics approaches (Genetic Algorithm, ( $\mu + \lambda $ ) Evolutionary Algorithm, Artificial Bee Colony, Ant Colony Optimization, and Particle Swarm Optimization) to select the best feature subsets. The main contribution of this work was to include in each metaheuristic a decomposition model that uses round and voting scheme. To investigate its relevance, we select the REPTree classifier to assess the predictive capacity of each subset of features selected in 8 plant species. We identified that the inclusion of the proposed decomposition model significantly reduces the dimensions of the datasets and improves predictive performance, regardless of the metaheuristic. Furthermore, the resulting pipeline has been compared with five approaches in the literature, for lncRNA, when it also showed superior predictive performance. Finally, this study generated a new pipeline to find a minimum number of features in lncRNAs and biological sequences.