Predicting Dose-Range Chemical Toxicity using Novel Hybrid Deep Machine-Learning Method.

Area of Science:

• Environmental Science
• Toxicology
• Computational Chemistry

Background:

• Assessing chemical toxicity is resource-intensive and challenging.
• Limited understanding of potential toxicity for numerous environmental chemicals.
• In vivo testing is time-consuming and requires substantial resources.

Purpose of the Study:

• To develop a novel deep learning method for predicting chemical toxicity at various doses.
• To create a Hybrid Neural Network (HNN) model, HNN-Tox, integrating Convolutional Neural Networks (CNN) and Feed-Forward Neural Networks (FFNN).
• To provide an alternative to animal-based toxicity assessments.

Main Methods:

• Developed a Hybrid Neural Network (HNN) model (HNN-Tox) by combining CNN and MLP-type FFNN.
• Trained models on a large dataset of 59,373 chemicals with known LD50 and exposure routes.
• Compared HNN-Tox performance against Random Forest, Bagging, and Adaptive Boosting algorithms.

Main Results:

• HNN-Tox achieved high predictive accuracy (84.9% and 84.1%) even with reduced feature sets.
• The model demonstrated strong performance with an Area Under the ROC Curve (AUC) of 0.89 and 0.88.
• HNN-Tox outperformed other machine learning methods on external validation datasets.

Conclusions:

• HNN-Tox is a novel and effective deep learning approach for predicting dose-range chemical toxicity.
• The model shows broad applicability for diverse chemicals and can reduce reliance on animal testing.
• This study represents a significant advancement in large-scale, feature-diverse chemical toxicity prediction.