Prediction of essential genes in prokaryote based on artificial neural network

Background Rapid identification of new essential genes is necessary to understand biological mechanisms and identify potential targets for antimicrobial drugs. Many computational methods have been proposed. Objectives To construct an essential genes classifier which satisfies more different organisms, and to study the redundancy of features used in the prediction of essential genes. Methods We designed a 57-12-1 artificial neural network model to predict the essential genes of 31 prokaryotic genomes. Four methods including self-predictions of each organism, the leave-one-genome-out method, predicting all by one organism, and self-predictions of all organisms were applied to assess the predictive performance. Additionally, the 57 features used in the artificial neural network model were analyzed by weighted principal component analysis to screen the key features strongly related to the essentiality of genes. Results Our results compared with previous researches indicate that our models had better generalizability. Furthermore, this method reduced the features to 29 while maintaining stable prediction performance overall, suggesting that some features are redundant for gene essentiality, and the screened features contained more important biological information for gene essentiality. Conclusion This study showed the effectiveness and generalizability of our artificial neural network model. In addition, the screened features could be used as key features in computational analysis and biological experiments.