A novel logic-based approach for quantitative toxicology prediction.

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Accurate in silico prediction of molecular toxicity is crucial for environmental safety and pharmaceutical development.
• Structure-activity relationship (SAR) methods are key to predictive toxicology.
• Inductive logic programming (ILP) offers a powerful, rule-based approach to SAR, overcoming limitations like molecular superposition.

Purpose of the Study:

• To develop a novel quantitative modeling approach by extending ILP for predictive toxicology.
• To introduce support vector inductive logic programming (SVILP) for enhanced molecular toxicity prediction.
• To evaluate the performance of SVILP against existing methods using real-world datasets.

Main Methods:

• Inductive logic programming (ILP) was employed to learn chemical substructure rules.
• A novel kernel was developed to integrate ILP-derived rules into a support vector machine (SVM) model.
• The SVILP approach was applied to predict fathead minnow fish toxicity for diverse molecular datasets.

Main Results:

• SVILP achieved a cross-validated correlation coefficient (R2CV) of 0.66, outperforming the chemical descriptor method (CHEM) at 0.52.
• SVILP correctly identified 73% of toxic molecules, compared to 58% for CHEM and 55% for ILP alone.
• On unseen data, SVILP yielded an R2 value of 0.57, significantly better than the commercial software TOPKAT (0.26).

Conclusions:

• SVILP represents a significant advancement in quantitative predictive toxicology, offering improved accuracy and interpretability.
• The method automatically derives novel, chemically relevant rules identifying toxicity alerts.
• SVILP is a versatile machine learning approach with broad applications in chemoinformatics and in silico drug design.