Toxicity-indicating structural patterns.

Area of Science:

• Computational chemistry
• Toxicology
• Cheminformatics

Background:

• Assessing the toxicity of uncharacterized compounds is crucial for drug development and safety.
• Existing methods may struggle with novel chemical structures.

Purpose of the Study:

• To develop and evaluate a toxicity alerting system for predicting chemical risks.
• To compare machine learning algorithms for toxicity classification.

Main Methods:

• Developed a toxicity alerting system based on comprehensive tables of toxicity-risk substructure fragments derived from RTECS and World Drug Index databases.
• Employed naive Bayesian clustering, k-nearest neighbor classification, and support vector machines (SVM) for toxicity classification.
• Utilized a large self-organizing map (SOM) to visualize the chemical space of toxic and drug-like molecules.

Main Results:

• Support vector machines demonstrated the best performance in classifying compounds into defined toxicity classes.
• Self-organizing maps effectively separated toxic from nontoxic substances.
• The fragment-based approach successfully extended toxicity predictions to structurally dissimilar compounds.

Conclusions:

• A fragment-based toxicity alerting system, particularly when combined with SVMs and SOMs, offers a robust method for chemical toxicity assessment.
• This system enhances the prediction of toxicity for both known and novel chemical entities.