工程纳米颗粒生物负荷量的机器学习预测模型  

Machine learning prediction model on body burden of engineered nanoparticles in organisms

作　　者：徐淑君 李瑞香 伍天翔 朱朗辰 蒲宇杰 杨静媛 陈景文[1] 李雪花[1] XU Shujun;LI Ruixiang;WU Tianxiang;ZHU Langchen;PU Yujie;YANG Jingyuan;CHEN Jingwen;LI Xuehua(Key Laboratory of Industrial Ecology and Environmental Engineering(MOE),School of Environmental Science and Technology,Dalian University of Technology,Dalian,116024,China)

机构地区：[1]工业生态与环境工程教育部重点实验室,大连理工大学环境学院,大连116024

出　　处：《环境化学》2024年第10期3406-3415,共10页Environmental Chemistry

基　　金：国家自然科学基金(22176023);国家重点研发项目(2022YFC3902104);大连理工大学基本科研业务费项目(DUT22QN216)资助.

摘　　要：工程纳米颗粒(ENPs)在生物体内的动态浓度即负荷量(BB),是评估ENPs生物积累的基础参数.大多数研究都通过实验手段来测定BB,其过程繁琐复杂.构建BB预测模型成为代替动物实验的有效途径.现有预测模型涵盖的ENPs以及生物种类较少,而且是基于生物整体的BB数据构建的,很难用于预测关键组织器官的负荷情况.了解ENPs在生物体内的分布,以及在关键靶器官中的积累尤为重要.本研究通过文献挖掘,构建了涵盖17种ENPs以及23种水生生物的BB数据集(n=1303),采用5种机器学习算法构建模型,并结合SHAP方法解析了不同特征对lgBB的影响.结果表明,非线性模型要优于线性模型,极端梯度提升(XGBoost)模型的效果最佳(R_(adj-train)^(2)=0.971,Q_(test)^(2)=0.909,Q210-CV=0.887).通过模型特征重要性分析,发现ENPs的本征特性(密度、粒径、电负性、分子量)、暴露参数(暴露浓度、暴露时间)、生物种类以及组织器官类型是影响ENPs负荷量的关键因素.其中,密度是影响ENPs负荷量的首要本征特征,负荷量会随着材料密度的增加而减少.本研究通过增加数据集中生物种类,扩大了模型应用域,实现特定暴露时间和暴露浓度下ENPs负荷量的预测,并将预测结果精准到组织器官水平(脑、脾脏、肌肉等),为ENPs生物积累的研究提供了有效方法.Dynamic concentration of engineered nanoparticles(ENPs)in an organism,commonly known as body burden(BB),is a fundamental parameter for assessing the bioaccumulation of ENPs.Prediction models on BB are effective alternatives to complex and time-consuming experimental methods.However,existing prediction models involved relatively few types of ENPs and biological species,and were based on BB data of whole organisms,making it difficult to predict BB in specific tissues or organs.Furthermore,the tissue-organ dependence of ENPs bioaccumulation highlights the importance of understanding their distribution in organisms and their bioaccumulation in key target organs.In this study,a BB dataset(n=1303)covering 17 types of ENPs,23 aquatic organisms,and 13 different tissue organs was compiled through literature mining.Five machine learning algorithms were used to build models,and SHapley Additive exPlanations(SHAP)analysis was conducted to examine how different features of ENPs affect BB.The modeling results demonstrated that nonlinear models outperformed linear models,and the extreme gradient boosting(XGBoost)model performed best(R_(adj-train)^(2)=0.971,Q_(test)^(2)=0.909,Q210-CV=0.887).SHAP analysis showed that the intrinsic properties of ENPs(density,particle size,electronegativity and molecular weight),exposure parameters(exposure concentrations and time),biological species,and organ or tissues types were key factors affecting BB.Among them,density was identified as the primary basic feature affecting BB,i.e.,the BB decreased with increasing material density.By increasing the number of biological species and types of ENPs in the dataset,this study expands the application domain of the model,and achieves the prediction of the BB of ENPs under specific exposure time and exposure concentrations.The model developed in this study could predict BB accurately at the tissue/organ level(brain,spleen,muscle,etc.),which provided an effective method for the assessment of ENPs bioaccumulation.

关 键 词：工程纳米颗粒 生物积累 负荷量 组织器官 机器学习. 

分 类 号：X-1[环境科学与工程] O6[理学—化学]