ACLPred: an explainable machine learning and tree-based ensemble model for anticancer ligand prediction

Several small molecules have been approved for cancer treatment, but the continuously growing cancer cases have further encouraged the identification of new anticancer drug compounds. Experimental methods are costly and time-consuming, thus rapid and cost-effective alternative method is much required. The effective identification of anticancer compounds using machine learning (ML) offers a promising solution, reducing both time and cost. In this study, small molecules with known inhibitory activities, both anticancer and non-anticancer were considered to train classification models. Molecular descriptors were calculated, and multistep feature selection was applied to identify significant features. Multiple ML algorithms were employed to build classification models and evaluated their performance using independent test and external datasets. The tree-based ensemble model, particularly Light Gradient Boosting Machine (LGBM), achieved the highest prediction accuracy of 90.33%, with an area under the receiver operating characteristic curve (AUROC) of 97.31%. Consequently, LGBM model was implemented in our proposed method, ACLPred. The ACLPred demonstrated superior prediction accuracy with good generalizability compared to existing methods. SHapley Additive exPlanations (SHAP) analysis provided model interpretability and revealed that topological features made major contributions to decision-making. ACLPred is available as an open-source, user-friendly graphical interface at https://github.com/ArvindYadav7/ACLPred for the screening of potential anticancer compounds.