Machine learning models for predicting endocrine disruption potential of environmental chemicals

Area of Science:

• Computational toxicology
• cheminformatics
• Pharmacology

Background:

• Predicting chemical compound activity is crucial for drug discovery and safety assessment.
• Estrogen receptor activity is a key area of toxicological research.
• Existing methods for predicting chemical activity have limitations in sensitivity.

Purpose of the Study:

• To introduce ML4Tox, a novel framework for predicting chemical compound activities.
• To develop and evaluate Deep Learning and Support Vector Machine models for predicting estrogen receptor ligand-binding domain activities (agonist, antagonist, binding).

Main Methods:

• Development of ML4Tox using Deep Learning and Support Vector Machine algorithms.
• Model training and validation using a 10x5-fold cross-validation schema on the CERAPP ToxCast dataset (1677 chemicals, 777 molecular features).
• Evaluation on the CERAPP "All Literature" dataset for agonist, antagonist, and binding activities.

Main Results:

• ML4Tox models demonstrated significantly improved sensitivity compared to published results.
• Achieved sensitivity for agonist prediction: 0.78 (vs. 0.56).
• Achieved sensitivity for antagonist prediction: 0.69 (vs. 0.11).
• Achieved sensitivity for binding prediction: 0.66 (vs. 0.26).

Conclusions:

• ML4Tox provides a robust and sensitive computational approach for predicting chemical activities.
• The framework offers a valuable tool for toxicological assessments and drug discovery.
• The enhanced prediction accuracy highlights the potential of machine learning in chemical safety evaluation.