In Silico Estimation of Chemical Carcinogenicity with Binary and Ternary Classification Methods

Area of Science:

• Computational toxicology
• Cheminformatics
• Machine learning in drug discovery

Background:

• Chemical carcinogenicity poses significant risks to human health.
• Early identification of carcinogens is essential for risk assessment and mitigation.
• Existing methods for carcinogenicity prediction require improvement in accuracy and efficiency.

Purpose of the Study:

• To develop and validate robust machine learning models for predicting chemical carcinogenicity.
• To assess the carcinogenicity of a large set of tobacco smoke components.
• To provide tools for early identification of potential human health hazards from chemical exposure.

Main Methods:

• Collected 829 diverse compounds with known rat carcinogenicity from the Carcinogenic Potency Database (CPDB).
• Utilized six types of molecular fingerprints (e.g., MACCS keys) to represent chemical structures.
• Generated 30 binary and ternary classification models using five machine learning algorithms (e.g., kNN), validated on an external dataset of 87 chemicals.

Main Results:

• The best binary classification model, employing MACCS keys and kNN, achieved a predictive accuracy of 83.91%.
• The optimal ternary model, also using MACCS keys and kNN, demonstrated an overall accuracy of 80.46%.
• Applied to 2251 tobacco smoke components, the models identified 981 potential carcinogens (binary) and classified 110 as strong and 807 as weak carcinogens (ternary).

Conclusions:

• Developed machine learning models effectively predict chemical carcinogenicity with high accuracy.
• The models show utility in identifying hazardous compounds within complex mixtures like tobacco smoke.
• These predictive tools can aid in prioritizing chemicals for further toxicological evaluation and regulatory scrutiny.