Intelligent QSAR Approaches: Harnessing Machine Learning for Early Detection of Carcinogenic Agents

Abstract Traditional carcinogenicity testing methods are costly and time-consuming, while existing Quantitative Structure-Activity Relationship (QSAR) models suffer from low accuracy and limited applicability domains. This study addresses these limitations by developing enhanced QSAR models using machine learning (ML) algorithms. A dataset of 805 compounds from the Carcinogenic Potency Database (CPD) was used to train classification and regression models, employing Bayesian classifiers, recursive partitioning, Kernel-based Partial Least Squares (KPLS), and deep learning techniques (Neural Networks, Random Forests). An independent validation dataset (105 compounds) was used to assess model performance. The DeepChem-based classification model achieved 81% test accuracy and 72% external validation accuracy, while the AutoQSAR regression model demonstrated an R² of 0.58 and Q² of 0.51, outperforming existing literature benchmarks. These models exhibit broad chemical space coverage, offering a robust, cost-effective alternative for carcinogenicity prediction.