Abstract: Accurately identifying the anti-cancer activities of small-molecule drugs remains a major challenge in computational drug discovery, owing to the complex and non-linear relationships between molecular structure, biological targets, and heterogeneous cellular contexts. In this study, we introduce DPD-Cancer, a web- and graph-based deep learning framework designed for both binary classification of compound anti-cancer activity and quantitative prediction of cell-line-specific anti-cancer responses. By relying on Graph Attention Transformer (GAT) method, DPD-Cancer demonstrates superior architectural performance against alternative methods, including pdCSM-cancer, ACLPred and MLASM. To address the widespread issue of performance overestimation, we further evaluated the model using a strictly partitioned NCI-60 dataset, where it maintained robust generalisability and provided a more realistic estimate of its utility for de novo drug screening. Across regression tasks spanning 73 cancer cell lines, the framework showed consistent effectiveness, with notable improvements in response prediction for specific tissue types. Collectively, these results highlight the advantages of (graph- and) attention-based mechanisms for learning expressive molecular representations and establish DPD-Cancer as a competitive and reliable tool for prioritising small molecule anti-cancer drug candidates.

Important Information:

• This website is free for all users.
• For cookie policy please click here.