- Research Article
1
- 10.1021/acsomega.5c04016
Active Stacking-Deep Learning with Strategic Samplingfor Small and Imbalanced Chemical Toxicity Prediction
- Nov 04, 2025
- ACS Omega
- Darlene Nabila Zetta + 2 more +2
Major challengesin toxicity prediction include dealing with imbalancedand limited data sets, especially when evaluating the harmful potentialof chemicals. These issues often lead to poor predictive model performance.Stacking ensemble learning enhances performance by combining predictionsfrom multiple base models, enabling the stack model to improve overallgeneralization. Active learning (AL), on the other hand, reduces theneed for large-scale data sets by effectively training models usingcarefully selected samples. One effective approach to address dataimbalance is the use of strategic sampling techniques. Hereby, weintroduce an active stacking-deep learning framework that integratesdeep neural networks (DNNs), including a convolutional neural network(CNN), a bidirectional long short-term memory (BiLSTM), and an attentionmechanism, with strategic data sampling to tackle challenges posedby imbalanced and limited data, ultimately improving the performanceof a chemical risk assessment predictive model. In this study, wefocused on thyroid-disrupting chemicals (TDCs) that target thyroidperoxidase, as they are linked to thyroid dysfunction, making it essentialto evaluate their risks to human health. Using stacking ensemble learningwith strategic sampling within an AL framework, our approach achievedan MCC of 0.51, AUROC of 0.824, and AUPRC of 0.851. Although performancedecreased across varying test ratios, our uncertainty-based methoddemonstrated superior stability under severe class imbalance. Whilea full-data stacking ensemble trained with strategic sampling performsslightly better in MCC, our method achieves marginally higher AUROCand AUPRC, requiring up to 73.3% less labeled data. Molecular dockingfurther validated our predictions, especially for highly toxic compounds,reinforcing the reliability of our framework in identifying TDCs.These findings highlight how active stacking-deep learning with strategicsampling can transform toxicity prediction, offering a more accurateand data-efficient alternative to traditional chemical risk assessmentmethods.
Read more