Machine Learning Models for Identification and Prediction of Toxic Organic Compounds Using Daphnia magna Transcriptomic Profiles

Abstract

A wide range of environmental factors heavily impact aquatic ecosystems, in turn, affecting human health. Toxic organic compounds resulting from anthropogenic activity are a source of pollution in aquatic ecosystems. To evaluate these contaminants, current approaches mainly rely on acute and chronic toxicity tests, but cannot provide explicit insights into the causes of toxicity. As an alternative, genome-wide gene expression systems allow the identification of contaminants causing toxicity by monitoring the organisms’ response to toxic substances. In this study, we selected 22 toxic organic compounds, classified as pesticides, herbicides, or industrial chemicals, that induce environmental problems in aquatic ecosystems and affect human-health. To identify toxic organic compounds using gene expression data from Daphnia magna, we evaluated the performance of three machine learning based feature-ranking algorithms (Learning Vector Quantization, Random Forest, and Support Vector Machines with a Linear kernel), and nine classifiers (Linear Discriminant Analysis, Classification And Regression Trees, K-nearest neighbors, Support Vector Machines with a Linear kernel, Random Forest, Boosted C5.0, Gradient Boosting Machine, eXtreme Gradient Boosting with tree, and eXtreme Gradient Boosting with DART booster). Our analysis revealed that a combination of feature selection based on feature-ranking and a random forest classification algorithm had the best model performance, with an accuracy of 95.7%. This is a preliminary study to establish a model for the monitoring of aquatic toxic substances by machine learning. This model could be an effective tool to manage contaminants and toxic organic compounds in aquatic systems.

The first report on the assessment of maximum acceptable daily intake (MADI) of pesticides for humans using intelligent consensus predictions

“…Diéguez et al 23 reported linear and nonlinear QSAR modeling approaches for the prediction of the acute toxicity of pesticides towards Americamysis bahia , and they have reported that the Random Forest (RF) regression model was a superior model having R 2 = 0.812, lower values of RMSE = 0.595, and MAE = 0.462 in the cross-validation training set, and also similar values for the external validation set. Choi et al 24 reported different machine-learning models for the identification and prediction of toxic organic compounds using Daphnia magna transcriptomic profiles. From their work, they have concluded that a combination of feature selection based on feature ranking and a random forest classification algorithm had the best model performance, with an accuracy of 95.7%.…”

“…[16][17][18][19][20] There has been no QSAR work reported previously on toxicity assessment of pesticides (prediction of maximum acceptable daily intake) specically against humans but some similar studies focusing on pesticide risk assessment embodying similar objectivesexploring less toxic and safer pesticides, though without specically targeting their MADIwere previously reported. 15,[21][22][23][24][25][26][27][28][29][30][31][32] These studies include, for example, several diverse indicator speciescrucial for better understanding the toxicity of the pesticide in different ecosystems, pesticidal mechanisms, and so forth, utilizing either linear or machine learning (ML) non-linear models. However, some of the previous studies reported neither different internal and external validation metrics nor mechanistic interpretations.…”

Systematic approaches to machine learning models for predicting pesticide toxicity

“…To predict the toxicity DT [ 42 ] 2 2022 LVQ; RF, and SVML 95.7% 22 toxic organic compounds which includes herbicides, pesticides, and industrial chemicals This approach allows them to thoroughly evaluate the performance of each family and reduce the chances of bias or overfitting. Additionally, by repeating the cross-validation process three times, they can ensure the robustness and reliability of their results RF [ 44 ] 3 2023 RF, ANN, and SVR – 7858 compounds were chosen from a pool of 9403 chemicals. ToxCast Database of the US EPA The study selected the chemical bonds (CBs) of 216 compounds using RF, outperforming ANN and SVF models.…”

“…ML can be used to predict the toxicity of pesticides and assess their potential risk to human health [ 28 , [42] , [43] , [44] ]. These methods can analyze a wide range of data, such as chemical properties and molecular structures, to identify the patterns and relationships between pesticides and their toxicity levels.…”