Accurate prediction of human essential genes using only nucleotide composition and association information

Abstract

AbstractThree groups recently identified essential genes in human cancer cell lines using wet experiments, and these genes are of high values. Herein, we improved the widely used Z curve method by creating a λ-interval Z curve, which considered interval association information. With this method and recursive feature elimination technology, a computational model was developed to predict human gene essentiality. The 5-fold cross-validation test based on our benchmark dataset obtained an area under the receiver operating characteristic curve (AUC) of 0.8814. For the rigorous jackknife test, the AUC score was 0.8854. These results demonstrated that the essentiality of human genes could be reliably reflected by only sequence information. However, previous classifiers in three eukaryotes can gave satisfactory prediction only combining sequence with other features. It is also demonstrated that although the information contributed by interval association is less than adjacent nucleotides, this information can still play an independent role. Integrating the interval information into adjacent ones can significantly improve our classifier’s prediction capacity. We re-predicted the benchmark negative dataset by Pheg server (https://cefg.uestc.edu.cn/Pheg), and 118 genes were additionally predicted as essential. Among them, 21 were found to be homologues in mouse essential genes, indicating that at least a part of the 118 genes were indeed essential, however previous experiments overlooked them. As the first available server, Pheg could predict essentiality for anonymous gene sequences of human. It is also hoped the λ-interval Z curve method could be effectively extended to classification issues of other DNA elements.