New public QSAR model for carcinogenicity.

Area of Science:

• Computational toxicology
• Cheminformatics
• Regulatory science

Background:

• The Registration, Evaluation, and Authorization of Chemicals (REACH) regulation necessitates comprehensive data on human health-affecting chemical properties.
• Quantitative structure-activity relationship (QSAR) models are recognized as valuable tools for filling data gaps in chemical safety assessments.
• The EU-funded CAESAR project focused on developing predictive models for five key endpoints relevant to regulatory purposes, including carcinogenicity.

Purpose of the Study:

• To develop and validate QSAR models for predicting the carcinogenic potency of chemicals, meeting specific regulatory requirements.
• To create two distinct models utilizing different sets of molecular descriptors for carcinogenicity prediction.
• To assess the performance and reliability of these models according to established validation principles.

Main Methods:

• Utilized a dataset of 805 non-congeneric chemicals from the Carcinogenic Potency Database (CPDBAS).
• Implemented the Counter Propagation Artificial Neural Network (CP ANN) algorithm.
• Developed two models: Model A using eight MDL descriptors and Model B using twelve Dragon descriptors.

Main Results:

• Model A achieved 91% training set accuracy and 73% test set accuracy, with 69% specificity.
• Model B achieved 89% training set accuracy and 69% test set accuracy, with 61% specificity.
• External validation on 738 compounds yielded accuracies of 61.4% (Model A) and 60.0% (Model B).

Conclusions:

• QSAR models for carcinogenicity prediction offer valuable support for expert evaluation and conventional testing methods.
• The developed MDL and Dragon descriptor-based models can assist in prioritizing chemicals for further toxicological testing.
• The CAESAR QSAR models are publicly accessible via a Java implementation on the CAESAR website.