Quantitative Toxicity Prediction via Meta Ensembling of Multitask Deep Learning Models

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Quantitative structure-activity relationship (QSAR) models are crucial for toxicity prediction.
• Existing machine learning methods often use single feature types and neural networks, limiting performance.
• Current multi-feature methods struggle with effective information aggregation.

Purpose of the Study:

• To develop an advanced deep learning framework for quantitative toxicity prediction.
• To overcome limitations of single-representation models and predetermined aggregation formulas.
• To improve the accuracy and robustness of toxicity prediction models.

Main Methods:

• Proposed a deep learning framework (QuantitativeTox) using five base deep learning models with distinct feature representations.
• Implemented a meta-ensemble approach with a separate deep learning model for aggregating base model outputs.
• Trained models in a weighted multitask fashion using four toxicity datasets (LD50, IGC50, LC50, LC50-DM), minimizing root-mean-square errors.

Main Results:

• QuantitativeTox outperformed the state-of-the-art TopTox method on three out of four datasets (LD50, IGC50, LC50-DM).
• Achieved superior performance with 5.46%, 16.67%, and 6.34% better root-mean-square errors on these datasets.
• Demonstrated significant improvements in mean absolute errors (6.41-12.16%) and coefficients of determination (2.54-7.36%).

Conclusions:

• The proposed meta-ensemble deep learning framework offers a significant advancement in quantitative toxicity prediction.
• QuantitativeTox effectively integrates diverse feature representations for enhanced predictive accuracy.
• The framework provides a robust and accurate alternative to existing toxicity prediction methodologies.