Abstract
Genes are termed to be essential if their loss of function compromises viability or results in profound loss of fitness. On the genome scale, these genes can be determined experimentally employing RNAi or knockout screens, but this is very resource intensive. Computational methods for essential gene prediction can overcome this drawback, particularly when intrinsic (e.g. from the protein sequence) as well as extrinsic features (e.g. from transcription profiles) are considered. In this work, we employed machine learning to predict essential genes in Drosophila melanogaster. A total of 27,340 features were generated based on a large variety of different aspects comprising nucleotide and protein sequences, gene networks, protein-protein interactions, evolutionary conservation and functional annotations. Employing cross-validation, we obtained an excellent prediction performance. The best model achieved in D. melanogaster a ROC-AUC of 0.90, a PR-AUC of 0.30 and a F1 score of 0.34. Our approach considerably outperformed a benchmark method in which only features derived from the protein sequences were used (P < 0.001). Investigating which features contributed to this success, we found all categories of features, most prominently network topological, functional and sequence-based features. To evaluate our approach we performed the same workflow for essential gene prediction in human and achieved an ROC-AUC = 0.97, PR-AUC = 0.73, and F1 = 0.64. In summary, this study shows that using our well-elaborated assembly of features covering a broad range of intrinsic and extrinsic gene and protein features enabled intelligent systems to predict well the essentiality of genes in an organism.
