Prediction of Essential Genes based on Machine Learning and Information Theoretic Features

Abstract

Computational tools have enabled a relatively simple prediction of essential genes (EGs), which would otherwise be done by costly and tedious gene knockout experimental procedures. We present a machine learning based predictor using information-theoretic features derived exclusively from DNA sequences. We used entropy, mutual information, conditional mutual information, and Markov chain models as features. We employed a support vector machine (SVM) classifier and predicted the EGs in 15 prokaryotic genomes. A fivefold cross-validation on the bacteria E. coli, B. subtilis, and M. pulmonis resulted in AUC score of 0.85, 0.81, and 0.89, respectively. In cross-organism prediction, the EGs of a given bacterium are predicted using a model trained on the rest of the bacteria. AUC scores ranging from 0.66 to 0.9 and averaging 0.8 were obtained. The average AUC of the classifier on a one-to-one prediction among E. coli, B. subtilis, and Acinetobacter is 0.85. The performance of our predictor is comparable with recent and state-of-the art predictors. Considering that we used only sequence information on a problem that is much more complicated, the achieved results are very