XAI-ACSM: An Ensemble-Based Explainable Artificial Intelligence Framework for the Accurate Prediction of Anticancer Small Molecules.

Cancer continues to be a leading cause of mortality worldwide. While conventional therapies such as chemotherapy, radiation, and immunotherapy remain mainstays in clinical oncology, these approaches often result in systemic toxicity, adverse side effects, and the emergence of drug resistance. Small-molecule drugs have gained prominence as potent anticancer agents due to their favorable drug-like profiles, enabling oral bioavailability and systemic efficacy. The incorporation of computational methodologies has further revolutionized anticancer drug discovery. In particular, machine learning (ML) techniques have shown considerable success in accelerating the identification and optimization of small-molecule candidates. Therefore, we propose a novel ensemble-based explainable artificial intelligence (XAI) framework, termed XAI-ACSM, for the identification and characterization of anticancer small molecules (ACSMs) using only SMILES notation. XAI-ACSM was initially developed through a comprehensive evaluation of five popular ML algorithms in conjunction with 14 molecular descriptors derived from five different feature encoding schemes. Then, these molecular descriptors and ML algorithms were employed to develop 70 baseline models. Finally, the most effective baseline models were selected and integrated to provide high-precision prediction outcomes using a probability averaging strategy. Both cross-validation and independent tests showed that XAI-ACSM outperformed its baseline models and the existing method. Remarkably, XAI-ACSM achieved an accuracy of 0.826, specificity of 0.926, and MCC of 0.666 over the independent test data set, which were 3.65, 9.60, and 8.63% higher than the existing method. Furthermore, XAI-ACSM was applied to identify potential ACSMs among FDA-approved drugs, with predictions validated through molecular docking against the most prevalent cancer targets. XAI-ACSM offers a practical approach for screening large chemical libraries to identify potential ACSMs, particularly among compounds with limited existing characterization, while helping to reduce time and resource requirements.