Active Learning-Optimized Dynamic Sampling Strategy for Enhanced Toxicity Prediction of Chemical Compounds

Toxicity prediction of chemical compounds is critical for environmental risk mitigation and safe material synthesis, particularly in catalytic reactor safety, solvent screening for sustainable processes, industrial material hazard assessment, and chemical product design. Recent advancements in machine learning (ML) have enabled efficient and reliable prediction of chemical compounds through data-driven modeling approaches. However, the generalization ability of models constructed using ML methods is largely constrained by the quality of the original experimental data sets and the complexity of feature engineering. This work proposes a novel dynamic data sampling approach based on the idea of active learning, which is derived from ML methods. The proposed approach entails quantifying chemical space distances and employing visualization methods to detect outlier samples. Subsequently, multiple rounds of weighted training and dynamic sample weight adjustment mechanisms were employed to achieve dynamic sampling. Threshold-based selection was implemented to filter samples for upgrading training models, therefore substantially improving the predictive accuracy of new models. When applied to existing toxicity data sets, the proposed sampling approach resulted in a 25.68% enhancement in the accuracy of random forest models based on molecular fingerprints and a 23.92% improvement in the accuracy of convolutional neural network models. These results illustrate that the optimized toxicity prediction models created with the proposed approach show greater accuracy than conventional approaches, demonstrating the capability of the proposed approach as a dependable tool for toxicity assessment in chemical safety evaluation. The interpretative study of molecular features associated with toxicity, performed with this model, provides significant insights for risk mitigation in chemical process and material development.