Scholarlink

<p>Acute intraperitoneal toxicity is a critical endpoint in drug safety evaluation, yet traditional animal-based methods face ethical concerns and interspecies translational limitations. The study developed machine learning (ML) models to predict acute intraperitoneal toxicity (LD<sub>50</sub>) using 12,480 standard compounds from the ChemIDplus database. Using RDKit (Research Data Kit) descriptors and MACCS keys (Molecular Access System Class keys) to encode molecular structures as input, two integrated models are constructed: RF (Random Forests) and GBDT (Gradient Boosted Decision Trees). Results demonstrated that the MACCS-GBDT (M-G) model achieved superior performance with a test R² of 0.914 and RMSE of 0.218, outperforming other combinations. Feature importance analysis revealed molecular weight (MolWt), topological indices (Chi1n), and specific functional groups (e.g., MACCS bit42) as dominant predictors. These features correlate with bioavailability, metabolic stability, and receptor interactions, explaining species-specific toxicity variations. The study demonstrates the potential of machine learning in accelerating toxicity assessments, offering an ethical alternative to traditional animal testing. It aids in predicting acute toxicity, guiding safer chemical and drug development, and promoting sustainable toxicology research following the 3Rs principles.</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

Acute intraperitoneal toxicity prediction using molecular descriptors and feature importance analysis

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Leilei Xin,
  • Yifan Zhang,
  • Xuefeng Liu,
  • Ronghua Liang,
  • Ruru Ma,
  • Yinglong Wang,
  • Peizhe Cui

摘要

Acute intraperitoneal toxicity is a critical endpoint in drug safety evaluation, yet traditional animal-based methods face ethical concerns and interspecies translational limitations. The study developed machine learning (ML) models to predict acute intraperitoneal toxicity (LD50) using 12,480 standard compounds from the ChemIDplus database. Using RDKit (Research Data Kit) descriptors and MACCS keys (Molecular Access System Class keys) to encode molecular structures as input, two integrated models are constructed: RF (Random Forests) and GBDT (Gradient Boosted Decision Trees). Results demonstrated that the MACCS-GBDT (M-G) model achieved superior performance with a test R² of 0.914 and RMSE of 0.218, outperforming other combinations. Feature importance analysis revealed molecular weight (MolWt), topological indices (Chi1n), and specific functional groups (e.g., MACCS bit42) as dominant predictors. These features correlate with bioavailability, metabolic stability, and receptor interactions, explaining species-specific toxicity variations. The study demonstrates the potential of machine learning in accelerating toxicity assessments, offering an ethical alternative to traditional animal testing. It aids in predicting acute toxicity, guiding safer chemical and drug development, and promoting sustainable toxicology research following the 3Rs principles.