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

Area of Science:

• Computational toxicology and cheminformatics
• Machine learning for chemical safety assessment

Background:

• Toxicity prediction models struggle with imbalanced and limited datasets, hindering accurate chemical risk assessment.
• Existing methods often fail to generalize well due to data scarcity and class imbalance issues.
• Thyroid-disrupting chemicals (TDCs) pose significant human health risks, necessitating reliable prediction methods.

Purpose of the Study:

• To develop and evaluate an active stacking-deep learning framework for enhanced toxicity prediction.
• To address data imbalance and scarcity challenges in predicting chemical harmful potential.
• To improve the accuracy and efficiency of identifying thyroid-disrupting chemicals (TDCs).

Main Methods:

• Integrated deep neural networks (DNNs) including CNN, BiLSTM, and attention mechanisms within a stacking ensemble.
• Employed active learning (AL) with strategic data sampling to optimize model training on limited data.
• Focused on thyroid peroxidase-targeting TDCs for chemical risk assessment validation.

Main Results:

• The active stacking-deep learning framework achieved notable performance with MCC of 0.51, AUROC of 0.824, and AUPRC of 0.851.
• The uncertainty-based AL approach demonstrated superior stability under severe class imbalance.
• The proposed method required up to 73.3% less labeled data compared to full-data stacking ensembles while achieving competitive or superior AUROC and AUPRC.

Conclusions:

• Active stacking-deep learning with strategic sampling offers a data-efficient and accurate solution for toxicity prediction.
• The framework effectively tackles imbalanced and limited data challenges in chemical risk assessment.
• Molecular docking validated the model's reliability in identifying toxic compounds, particularly TDCs.