融合遗传算法与XGBoost的玉米百粒重相关基因挖掘

The method of 100-kernel weight related genes mining in maize mixed with genetic algorithm and XGboost

基于RNA-Seq的转录组测序数据特征维度较高,使用传统生信方法寻找表型相关基因需要大量计算资源,且差异分析所得候选基因范围较大,进一步筛选依赖已有的先验知识。针对这一问题,本文提出了融合遗传算法和XGBoost的转录组分析方法-GA-XGBoost,通过融入机器学习算法缩小了后续分析的候选基因范围。在一组高质量玉米数据集上对基因-百粒重性状的关联进行了对比实验和后续分析,结果显示,相比于分别使用全体基因和差异表达基因直接训练XGBoost模型,所提方法得到的候选基因训练的XGBoost模型在玉米百粒重的预测结果上具有最小的MSE;相比于差异表达分析结果的1542个差异表达基因,GA-XGBoost方法最终将候选基因范围减小至48个,范围缩小了31倍,表明所提方法能够有效提升对转录组数据的分析能力和效率。

The RNA-Seq-based transcriptome sequencing data has a high feature dimension that requires a lot of computing resources when using traditional methods to find phenotype related genes.Moreover,the range of candidate genes obtained by difference analysis is large,and further screening depends on existing a prior knowledge.A transcriptome analysis method combining genetic algorithm and XGBoost,GA-XGBoost,was proposed to narrow the range of candidate genes for subsequent analysis by incorporating machine learning algorithm.A comparative experiment and subsequent analysis of the gene-100-kernel weight trait association on a set of high-quality maize datasets showed that,compared with training the XGBoost model directly with whole genes and differentially expressed genes,the candidate gene training XGBoost model obtained by the proposed method had the minimum MSE in predicting the 100-kernel weight of maize.Compared with 1542 differentially expressed genes in the results of differential expression analysis,the range of candidate genes was reduced to 48 by the GA-XGBoost method,which was reduced by 31 times,indicating that the proposed method could effectively improve the ability and efficiency of transcriptome data analysis.