Introduction: Cancer is the second leading cause of death worldwide. Although substantial efforts have been devoted to developing effective treatments, conventional chemotherapy and radiotherapy remain limited by systemic toxicity, drug resistance, and high recurrence rates. Anticancer agents offer an alternative therapeutic approach, yet experimental discovery strategies are expensive and time-consuming. Methods: This study developed a computational model, AntiCanNet, to predict Anti-Cancer Small Molecules (ACSMs). Molecular features were generated using PaDEL and ChemGPT, a chemical large language model. A small-molecule network was constructed based on structural similarity. These features and the network were processed through a graph neural network to obtain high-level representations, followed by a fully connected neural network for prediction. Results: AntiCanNet was evaluated on one training dataset and two test datasets. Cross-validation on the training dataset yielded an AUC of 0.971, and both test datasets achieved AUC values above 0.9, demonstrating strong predictive performance Discussion: Ablation analyses supported the effectiveness of the model design. Several latent ACSMs identified by AntiCanNet showed potential associations with cancer-related pathways. Conclusion: AntiCanNet provides an efficient computational approach for identifying ACSMs and may facilitate the discovery of previously unrecognized anti-cancer agents.
Chen et al. (2026) studied this question.