Scholarlink

<p>Drug-induced nephrotoxicity (DIN) is an infrequent adverse reaction to medications and represents a complex clinical outcome influenced by multiple factors. Predicting DIN using preclinical animal models remains challenging, and in silico approaches have emerged as promising alternatives for DIN risk assessment. A high-quality dataset consisting of 1,018 compounds was constructed in this study. Compounds in this dataset were systematically collected from five authoritative sources: the SIDER, FDA, ChEMBL, DrugBank, and literature on “drug-induced nephrotoxicity” published in the past decade (screened via keyword search on PubMed). Clear criteria were followed for compound screening and label annotation: using “kidney,” “nephrotoxicity,” “kidney injury,” and “kidney disease” as core search terms, compounds retrieved that were clearly associated with kidney injury or could induce kidney disease were classified into the positive set (DIN = 1); compounds with no records of renal adverse reactions, or those explicitly having renal protective effects or used for treating renal diseases, were classified into the negative set (DIN = 0). Ultimately, a dataset of 1018 compounds with clear labels and reliable sources was integrated. The 42 classification models, which depended on six different molecular fingerprints, were built via deep neural network (DNN) and six machine learning algorithms. A comparative study demonstrated that models utilizing DNN consistently surpassed traditional machine learning approaches across six molecular fingerprint types. Notably, the ECFP_6 fingerprint exhibited the highest performance, achieving an area under the receiver operating characteristic curve (AUC) of 75.9%, an accuracy (ACC) of 71.4%, and an F1-score of 76.0%. Furthermore, the SHapley Additive exPlanations (SHAP) algorithm was applied to interpret the predictions of the high-performing models, identifying key structural fragments associated with DIN. The ten most influential substructures, identified based on their impact on model predictions, were chosen as early warning markers for future DIN screening research. Overall, these results suggest that DNN models utilizing molecular fingerprints can function as dependable and efficient tools for assessing nephrotoxicity risk in potential drug candidates during the initial phases of drug development.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prediction of drug-induced nephrotoxicity based on deep learning algorithm and molecular fingerprints

  • 期刊论文

  • 发表日期：

    2025年10月12日

查看全文
  • Shuailong Wang,
  • Yan Li

摘要

Drug-induced nephrotoxicity (DIN) is an infrequent adverse reaction to medications and represents a complex clinical outcome influenced by multiple factors. Predicting DIN using preclinical animal models remains challenging, and in silico approaches have emerged as promising alternatives for DIN risk assessment. A high-quality dataset consisting of 1,018 compounds was constructed in this study. Compounds in this dataset were systematically collected from five authoritative sources: the SIDER, FDA, ChEMBL, DrugBank, and literature on “drug-induced nephrotoxicity” published in the past decade (screened via keyword search on PubMed). Clear criteria were followed for compound screening and label annotation: using “kidney,” “nephrotoxicity,” “kidney injury,” and “kidney disease” as core search terms, compounds retrieved that were clearly associated with kidney injury or could induce kidney disease were classified into the positive set (DIN = 1); compounds with no records of renal adverse reactions, or those explicitly having renal protective effects or used for treating renal diseases, were classified into the negative set (DIN = 0). Ultimately, a dataset of 1018 compounds with clear labels and reliable sources was integrated. The 42 classification models, which depended on six different molecular fingerprints, were built via deep neural network (DNN) and six machine learning algorithms. A comparative study demonstrated that models utilizing DNN consistently surpassed traditional machine learning approaches across six molecular fingerprint types. Notably, the ECFP_6 fingerprint exhibited the highest performance, achieving an area under the receiver operating characteristic curve (AUC) of 75.9%, an accuracy (ACC) of 71.4%, and an F1-score of 76.0%. Furthermore, the SHapley Additive exPlanations (SHAP) algorithm was applied to interpret the predictions of the high-performing models, identifying key structural fragments associated with DIN. The ten most influential substructures, identified based on their impact on model predictions, were chosen as early warning markers for future DIN screening research. Overall, these results suggest that DNN models utilizing molecular fingerprints can function as dependable and efficient tools for assessing nephrotoxicity risk in potential drug candidates during the initial phases of drug development.