文章摘要
钱一凡,裴晨浩,吕陈,吴同亮,刘存,王玉军.水培到土培体系植物根系PFAS吸收风险的迁移机器学习研究[J].农业环境科学学报,2024,43(11):2516-2524.
水培到土培体系植物根系PFAS吸收风险的迁移机器学习研究
Predicting uptake risk of PFAS in plant roots using transfer learning from hydroponic to soil systems
投稿时间:2024-10-08  
DOI:10.11654/jaes.2024-0854
中文关键词: 全氟与多氟化合物  植物根系吸收积累  机器学习  迁移学习
英文关键词: per- and polyfluoroalkyl substances(PFAS)  plant root absorption and accumulation  machine learning  transfer learning
基金项目:国家重点研发计划项目(2021YFC1809100);国家自然科学基金项目(41977027)
作者单位E-mail
钱一凡 土壤与农业可持续发展国家重点实验室, 中国科学院南京土壤研究所, 南京 211135
中国科学院大学, 北京 100049 
 
裴晨浩 土壤与农业可持续发展国家重点实验室, 中国科学院南京土壤研究所, 南京 211135
中国科学院大学, 北京 100049 
 
吕陈 土壤与农业可持续发展国家重点实验室, 中国科学院南京土壤研究所, 南京 211135  
吴同亮 土壤与农业可持续发展国家重点实验室, 中国科学院南京土壤研究所, 南京 211135  
刘存 土壤与农业可持续发展国家重点实验室, 中国科学院南京土壤研究所, 南京 211135 liucun@issas.ac.cn 
王玉军 土壤与农业可持续发展国家重点实验室, 中国科学院南京土壤研究所, 南京 211135
中国科学院大学, 北京 100049 
yjwang@issas.ac.cn 
摘要点击次数: 516
全文下载次数: 440
中文摘要:
      为实现植物根系对全氟与多氟化合物(Per- and polyfluoroalkyl substances,PFAS)的吸收和积累的精准预测,本研究基于涵盖了19种PFAS的水培或土培体系文献数据,共计668条数据点,利用分子描述符、实验条件以及作物属性等作为特征变量,构建4种机器学习模型分别预测了水培体系和土壤体系的根系富集因子(RCF),效果最佳的均是极端梯度提升树(XGB)模型,测试集决定系数(R2)分别为0.69和0.83,均方根误差(RMSE)分别为0.51和0.28。水培体系中PFAS的吸收、积累更容易研究,因此搭建了从水培体系到土壤体系的迁移学习模型,通过知识共享来提升 RCF 预测的准确度。最优的迁移模型的测试集 R2达到了0.86,RMSE 为 0.25,准确性有显著提升。Shapley加性解释(SHAP)特征重要性分析结果显示,暴露时间、土壤 pH 和 PFAS浓度是影响土壤根系吸收积累最主要的3个因素。本研究通过构建机器学习和迁移学习模型来预测土壤中植物根系PFAS的吸收积累,实现了简单水-植物根表体系向土-水-植物多个界面复杂体系的迁移,为评估土壤PFAS生态环境风险提供了新的视角。
英文摘要:
      In order to precisely predict the absorption and accumulation of per- and polyfluoroalkyl substances(PFAS)in plant roots, a total of 668 data points from PFAS plant uptake studies in hydroponic and soil systems covering 19 PFAS species were collected. The features such as molecular descriptors, experimental conditions, and crop properties were used to construct four machine learning models, which were used to predict the root concentration factor(RCF). The Extreme Gradient Boosting Tree(XGB)model performed the best, with R2 values of 0.69 and 0.83 and RMSE values of 0.51 and 0.28 for the test sets, respectively. Due to the easier study of PFAS absorption and accumulation in hydroponic systems, a transfer learning model was established from hydroponic to soil systems to improve the prediction accuracy of RCF in complex soil systems through knowledge transfer. The optimal transfer model achieved R2 of 0.86 and RMSE of 0.25 for test set, showing a significant improvement in accuracy. Analysis of SHAP feature importance revealed that exposure time, soil pH, and PFAS concentration are the top three factors affecting RCF in soil. This study predicts the PFAS absorption and accumulation of plant root in soil through the construction of machine learning and transfer learning models, achieving the transfer from simple to complex systems. It provides new insights for evaluating the environmental risks of PFAS contamination in soils.
HTML    查看全文   查看/发表评论  下载PDF阅读器