基于放射组学的盆腔肿瘤不同调强放疗技术γ通过率的预测研究

目的采用基于放射组学的机器学习方法,探索盆腔肿瘤不同调强放疗技术下γ通过率(GPR)分类预测模型的可行性,并比较了4种集成树模型的分类性能。方法回顾性收集了409例使用不同调强放疗技术的计划,采用基于模体测量方式的三维剂量验证结果,γ通过率标准为3%/2 mm、10%剂量阈值。提取基于剂量文件的放射组学特征构建预测模型。分别采用随机森林、自适应增强、极端梯度提升树和轻量级梯度提升机4种机器学习算法,并且通过计算灵敏度、特异度、F1分数及曲线下面积(AUC)值来评估它们的分类性能。结果随机森林、自适应增强、极端梯度提升树、轻量级梯度提升机模型的灵敏度和特异度分别为0.96、0.82、0.93、0.89和0.38、0.54、0.62、0.62,F1分数和AUC值分别为0.86、0.81、0.88、0.86和0.81、0.77、0.85、0.83。其中极端梯度提升树模型的灵敏度达到0.93,特异度、F1分数和AUC值均为最高,要优于其他3种模型。结论针对采用不同调强放疗技术的盆腔肿瘤调强计划,使用基于放射组学的机器学习方法来构建伽马通过率分类预测模型具有一定的可行性,能够为将来GPR预测的多机构合作研究提供基础。

Identification of genome-wide nucleotide sites associated with mammalian virulence in influenza A viruses

The virulence of influenza viruses is a complex multigenic trait.Previous studies about the virulence determinants of influenza viruses mainly focused on amino acid sites,ignoring the influence of nucleotide mutations.In this study,we collected>200 viral strains from 21 subtypes of influenza A viruses with virulence in mammals and obtained over 100 mammalian virulence-related nucleotide sites across the genome by computational analysis.Fifty of these nucleotide sites only experienced synonymous mutations.Experiments showed that synonymous mutations in three high-scoring nucleotide sites,i.e.,PB1–2031,PB1–633,and PB1–720,enhanced the pathogenicity of the influenza A(H1N1)viruses in mice.Besides,machine-learning models with accepted accuracy for predicting mammalian virulence of influenza A viruses were built.Overall,this study highlighted the importance of nucleotide mutations,especially synonymous mutations in viral virulence,and provided rapid methods for evaluating the virulence of influenza A viruses.It could be helpful for early warning of newly emerging influenza A viruses.