Scholarlink

<p>This study investigates the application of advanced machine learning and deep learning techniques to predict the acute oral toxicity (LC50) of 200 pesticides on Oncorhynchus mykiss (Rainbow Trout) using Quantitative Structure-Activity Relationship (QSAR) approach. A Support Vector Machine technique optimized with the Dragonfly Algorithm (DA-SVM) was used to model LC50. Two independent molecular feature types were used: 208 molecular descriptors and 1024 Morgan molecular fingerprints, generated using RDKit and the selected features were optimized via Genetic Algorithm and Convolutional Neural Networks (CNNs) methodologies. three models were developed namely Model 1: GA-DA SVMr used with 19 descriptors, model 2: CNN-MLP used with 64 features from descriptors, and model 3: CNN-DA SVMr used with 72 optimized features derived from Morgan fingerprints). Model 3 presents the best performance by achieving high predictive global accuracy in terms of RMSE of 0.1755, R² of 0.9860, AARD of 2.9765% and MAE of 0.0590. Results show the superiority and the potential of Morgan molecular fingerprints over traditional descriptors. This study highlights the potential of integrating advanced machine learning techniques and molecular feature selection for ecotoxicological assessments.</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

Predicting Pesticide LC50 Toxicity Using Convolutional Neural Networks and Dragonfly-Support Vector Machine Methods with Quantitative Structure-Activity Relationship and 2D Molecular Fingerprints: A Case Study on Oncorhynchus Mykiss

  • 期刊论文

  • 发表日期：

    2025年3月6日

查看全文
  • Badiet El Djamel Nedjma Mahtal,
  • Maamar Laidi,
  • Soufiane Rahal,
  • Mohamed Hentabli,
  • Mabrouk Hamadache

摘要

This study investigates the application of advanced machine learning and deep learning techniques to predict the acute oral toxicity (LC50) of 200 pesticides on Oncorhynchus mykiss (Rainbow Trout) using Quantitative Structure-Activity Relationship (QSAR) approach. A Support Vector Machine technique optimized with the Dragonfly Algorithm (DA-SVM) was used to model LC50. Two independent molecular feature types were used: 208 molecular descriptors and 1024 Morgan molecular fingerprints, generated using RDKit and the selected features were optimized via Genetic Algorithm and Convolutional Neural Networks (CNNs) methodologies. three models were developed namely Model 1: GA-DA SVMr used with 19 descriptors, model 2: CNN-MLP used with 64 features from descriptors, and model 3: CNN-DA SVMr used with 72 optimized features derived from Morgan fingerprints). Model 3 presents the best performance by achieving high predictive global accuracy in terms of RMSE of 0.1755, R² of 0.9860, AARD of 2.9765% and MAE of 0.0590. Results show the superiority and the potential of Morgan molecular fingerprints over traditional descriptors. This study highlights the potential of integrating advanced machine learning techniques and molecular feature selection for ecotoxicological assessments.