Active Stacking-Deep Learning with Strategic Sampling for Small and Imbalanced Chemical Toxicity Prediction

Major challenges in toxicity prediction include dealing with imbalanced and limited data sets, especially when evaluating the harmful potential of chemicals. These issues often lead to poor predictive model performance. Stacking ensemble learning enhances performance by combining predictions from multiple base models, enabling the stack model to improve overall generalization. Active learning (AL), on the other hand, reduces the need for large-scale data sets by effectively training models using carefully selected samples. One effective approach to address data imbalance is the use of strategic sampling techniques. Hereby, we introduce an active stacking-deep learning framework that integrates deep neural networks (DNNs), including a convolutional neural network (CNN), a bidirectional long short-term memory (BiLSTM), and an attention mechanism, with strategic data sampling to tackle challenges posed by imbalanced and limited data, ultimately improving the performance of a chemical risk assessment predictive model. In this study, we focused on thyroid-disrupting chemicals (TDCs) that target thyroid peroxidase, as they are linked to thyroid dysfunction, making it essential to evaluate their risks to human health. Using stacking ensemble learning with strategic sampling within an AL framework, our approach achieved an MCC of 0.51, AUROC of 0.824, and AUPRC of 0.851. Although performance decreased across varying test ratios, our uncertainty-based method demonstrated superior stability under severe class imbalance. While a full-data stacking ensemble trained with strategic sampling performs slightly better in MCC, our method achieves marginally higher AUROC and AUPRC, requiring up to 73.3% less labeled data. Molecular docking further 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 strategic sampling can transform toxicity prediction, offering a more accurate and data-efficient alternative to traditional chemical risk assessment methods.