$S^2FS$: Single Score Feature Selection Applied to the Problem of Distinguishing Long Non-coding RNAs from Protein Coding Transcripts

Abstract

The task of distinguishing long non-coding RNAs (lncRNAs) from protein coding transcripts (PCTs) has been previously addressed with machine learning (ML) algorithms using hundreds of features. However, the use of a large number of features can negatively affect the predictive performance of these algorithms since it can lead to problems like overfitting due to a phenomenon known as the curse of dimensionality. In order to deal with these problems, dimensionality reduction techniques have been proposed, among them, feature selection. This work proposes and experimentally evaluates a simple and fast feature selection technique, called Single Score Feature Selection - $S^2FS$.

For such, initially, frequencies of 2-mers, 3-mers and 4-mers were extracted from public databases of PCTs and lncRNAs of Homo sapiens, resulting in a dataset composed of two groups of RNA sequences, one for PCTs and the other for lncRNAs, and a large number of features. To reduce the number of features, $S^2FS$ was applied to the dataset. Experimental results showed that relevant features were selected, keeping the predictive accuracy, with a lower processing cost than some existing feature selection techniques.