A Comparative Evaluation of Machine Learning and Deep Graph Learning for Chemical Ecotoxicological Prediction

Area of Science:

• Environmental Science
• Computational Chemistry
• Toxicology

Background:

• Chemical ecotoxicity assessment is crucial for environmental protection but faces challenges due to the time and cost of experimental testing.
• Accurate predictive methods are essential to overcome the limitations of traditional ecotoxicity testing.

Purpose of the Study:

• To comprehensively analyze the application of machine learning and graph-based learning for ecotoxicological prediction of chemicals.
• To evaluate the performance of various models in predicting the ecotoxicity across different aquatic species.

Main Methods:

• Constructed 161 models using combinations of molecular representations (Morgan, MACCS, Mol2vec), machine learning algorithms (KNN, NB, RF, SVM, XGB, DNN), and graph neural networks (GAT, GCN, MPNN, Attentive FP, FPGNN).
• Evaluated model performance for predicting ecotoxicity in fish, crustaceans, and algae, including same-species and cross-species predictions.

Main Results:

• Graph convolutional network (GCN) demonstrated the best overall performance in predicting ecotoxicity for fish, crustaceans, and algae.
• GCN models achieved high area under the ROC curve (AUC) values (0.982-0.992) for same-species predictions.
• Cross-species predictions showed reduced performance, with GAT and GCN leading, but DNN with MACCS fingerprinting performed best for unseen chemicals (AUC 0.821).

Conclusions:

• Computational prediction methods, especially GCN, are effective for same-species ecotoxicity assessment.
• Significant challenges remain in accurately predicting chemical ecotoxicity across different species, highlighting the need for further research.
• A web server for ecotoxicology prediction is available, aiding environmental risk assessment.