A new approach methodology (NAM) for carcinogenicity prediction of organic chemicals using the multiclass ARKA...

The accumulation of organic pollutants in the environment has significantly impacted the lives of flora and fauna, resulting in disruptions in the biological ecosystem. Carcinogenicity has been one of the most alarming adverse effects exhibited by these chemicals, affecting millions worldwide. In this study, we developed simple Quantitative Structure-Activity Relationship (QSAR) models to predict the Oral Slope Factor (OSF) and Inhalation Slope Factor (ISF) of organic pollutants, identifying and prioritizing their carcinogenicity risks. To further enhance the quality of the models, we also considered read-across-derived similarity measures using the quantitative Read-Across Structure-Activity Relationship (q-RASAR) approach, and the response range-specific contribution of different features using the multiclass Arithmetic Residual in K-groups Analysis (ARKA) framework. Various Machine Learning-based stacking regressors were trained using the results from the different modeling workflows, and the best algorithm was selected based on a multi-criteria decision-making approach. The results show that the Linear Support Vector Regression model performed best on the OSF data (MAE_Test=0.907), while the Ridge Regression model performed best on the ISF data (MAE_Test=0.827). They also showed that the improved modeling workflows generate better prediction performance. Predictions for a true external set showed that the results align well with the reported carcinogenic status.