Answers are grounded in this paper and cite their sources.
Enhancing the precision of carcinogen classification is essential for improving chemical safety evaluation and risk assessment. Machine learning techniques have increasingly been used in cheminformatics to identify relationships between molecular descriptors and carcinogenic outcomes. However, many existing studies rely on limited descriptor sets or evaluate a small number of predictive models. In this study, a comprehensive dataset was developed by integrating diverse molecular fingerprints and descriptors, including radial basis function descriptors, MaxsolH, MaxsolO, MaxssO, and several bond‑based features such as B01(C–O), B02(C–C), and B07(C–O). The integration of multiple descriptor categories provides a more detailed representation of molecular structure and enables a more systematic evaluation of their influence on carcinogenicity prediction. To investigate the predictive capability of these descriptors, several machine learning algorithms were applied to classify carcinogenic potential as a binary outcome. These algorithms include k nearest neighbors, decision tree, extreme gradient boosting, light gradient boosting machine, logistic regression, bagging classifier, adaptive boosting, gradient boosting, categorical boosting, multilayer perceptron, voting classifier, and quadratic discriminant analysis. A comparative modeling framework was implemented to evaluate the predictive performance of these algorithms and to identify the most suitable models for carcinogenicity prediction. The results demonstrate that ensemble boosting algorithms achieved the highest predictive performance. In particular, extreme gradient boosting achieved a training accuracy of 0.9656, while light gradient boosting machine obtained a testing accuracy of 0.9394. In addition to model comparison, Pearson correlation analysis was conducted to identify the most influential molecular descriptors. Among the investigated features, B07(C–O) exhibited a moderate negative correlation with carcinogenicity (−0.43), followed by B08(C–O) (−0.41) and B10(C–C) (−0.40). These values suggest a mild-to-moderate inverse statistical trend within this dataset, rather than a direct mechanistic causative link, serving as useful exploratory features for the predictive models. Overall, this study provides a comprehensive evaluation of multiple machine learning algorithms using an integrated set of molecular descriptors and highlights the predictive relevance of bond‑based features for carcinogenicity assessment. The proposed approach contributes to the development of data‑driven computational tools that can support carcinogen classification and assist researchers in prioritizing chemical compounds for further toxicological evaluation in cheminformatics and computational toxicology.
No takes yet. Share an insight, caveat, or question.
Yadav et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: