Deng Jingyuan
Division of Epidemiology and Biostatistics, Department of Environmental Health, Cincinnati Children's Hospital, University of Cincinnati Medical Center, 3223 Eden Avenue, ML 56, Cincinnati, OH, 45267-0056, USA,
Methods Mol Biol. 2015;1279:137-51. doi: 10.1007/978-1-4939-2398-4_9.
Essential genes are indispensable for the target organism's survival. Large-scale identification and characterization of essential genes has shown to be beneficial in both fundamental biology and medicine fields. Current existing genome-scale experimental screenings of essential genes are time consuming and costly, also sometimes confer erroneous essential gene annotations. To circumvent these difficulties, many research groups turn to computational approaches as the alternative to identify essential genes. Here, we developed an integrative machine-learning based statistical framework to accurately predict essential genes in microorganisms. First we extracted a variety of relevant features derived from different aspects of an organism's genomic sequences. Then we selected a subset of features have high predictive power of gene essentiality through a carefully designed feature selection system. Using the selected features as input, we constructed an ensemble classifier and trained the model on a well-studied microorganism. After fine-tuning the model parameters in cross-validation, we tested the model on the other microorganism. We found that the tenfold cross-validation results within the same organism achieves a high predictive accuracy (AUC ~0.9), and cross-organism predictions between distant related organisms yield the AUC scores from 0.69 to 0.89, which significantly outperformed homology mapping.
必需基因对于目标生物体的生存不可或缺。对必需基因进行大规模鉴定和表征已证明在基础生物学和医学领域都有益处。当前现有的必需基因全基因组规模实验筛选既耗时又昂贵，有时还会给出错误的必需基因注释。为了规避这些困难，许多研究团队转向计算方法作为鉴定必需基因的替代方法。在此，我们开发了一种基于机器学习的综合统计框架，以准确预测微生物中的必需基因。首先，我们从生物体基因组序列的不同方面提取了各种相关特征。然后，我们通过精心设计的特征选择系统选择了一组对基因必需性具有高预测能力的特征子集。使用所选特征作为输入，我们构建了一个集成分类器，并在一种经过充分研究的微生物上训练模型。在交叉验证中对模型参数进行微调后，我们在其他微生物上测试了该模型。我们发现，在同一生物体内进行的十折交叉验证结果具有很高的预测准确性（AUC约为0.9），并且在远缘相关生物体之间进行的跨生物体预测产生的AUC分数在0.69至0.89之间，这显著优于同源性映射。
