Miao Yousi, Chen Xiujuan, Xie Ping, Liang Yemei, Wei Yuanyi, Deng Houliang, Huang Qiongbo, Qiu Haojie, Li Huiyi, Zhou Shi, Liang Huiying, Huang Min, Li Jiali, Gao Xia, Mo Xiaolan
Department of Pharmacy, Guangzhou Women and Children's Medical Center, Guangzhou Medical University, Guangzhou, China.
Department of Medical Big Data Center, Guangdong Provincial People's Hospital (Guangdong Academy of Medical Sciences), Southern Medical University, Guangzhou, China.
Transl Pediatr. 2025 Aug 31;14(8):1873-1887. doi: 10.21037/tp-2025-127. Epub 2025 Aug 25.
The use of tacrolimus (TAC) in clinical settings is hindered by its nephrotoxic effects, which can vary significantly among individuals. Urine retinol-binding protein (RP), as a novel biochemical marker, is a potential indicator for early detection of renal tubular injury caused by TAC. The objective was to develop and validate a machine learning model that combines clinical features with genetic markers for predicting TAC nephrotoxicity in children with nephrotic syndrome (NS).
A retrospective cohort of 203 children diagnosed with NS who were admitted between June 2013 and December 2018 was used for model development, while 12 children were prospectively recruited for external validation. The model incorporated 38 clinical features and 80 genetic variables, with changes in urine RP levels pre- and post-TAC administration indicating renal tubular toxicity. Five machine learning algorithms were employed: Extra Random Trees (ET), Gradient Boosting Decision Tree (GBDT), random forests (RF), and eXtreme Gradient Boosting (XGBoost), and logistic regression (LR).
The LR model, including six genetic markers (*3 rs776746_*3/*3, rs1660144_AA, rs230526_AG, rs696_TC, rs12664637_CT and rs2274223_AG), exhibited the best performance with a sensitivity of 78.6%, specificity of 63.8%, accuracy of 67.2%, and area under the curve (AUC) of 76.1%.
By employing RP as a marker of renal toxicity, we established and validated the renal tubular toxicity prediction model for the use of TAC using machine learning incorporating genetic factors of NS patients. This model allows physicians to evaluate the risk of nephrotoxic effects and adjust treatment plans accordingly to prevent kidney injury.
他克莫司（TAC）在临床应用中受到其肾毒性的限制，而肾毒性在个体间差异显著。尿视黄醇结合蛋白（RP）作为一种新型生化标志物，是早期检测TAC所致肾小管损伤的潜在指标。目的是开发并验证一种将临床特征与基因标志物相结合的机器学习模型，用于预测肾病综合征（NS）患儿的TAC肾毒性。
采用2013年6月至2018年12月期间收治的203例确诊为NS的儿童回顾性队列进行模型开发，前瞻性招募12例儿童进行外部验证。该模型纳入了38项临床特征和80个基因变量，TAC给药前后尿RP水平的变化表明肾小管毒性。采用了五种机器学习算法：Extra Random Trees（ET）、梯度提升决策树（GBDT）、随机森林（RF）、极端梯度提升（XGBoost）和逻辑回归（LR）。
包含六个基因标志物（*3 rs776746_*3/*3、rs1660144_AA、rs230526_AG、rs696_TC、rs12664637_CT和rs2274223_AG）的LR模型表现最佳，灵敏度为78.6%，特异性为63.8%，准确率为67.2%，曲线下面积（AUC）为76.1%。
通过将RP用作肾毒性标志物，我们利用纳入NS患者基因因素的机器学习方法建立并验证了TAC使用的肾小管毒性预测模型。该模型使医生能够评估肾毒性风险，并相应调整治疗方案以预防肾损伤。
