DFT_ANPD: A dual-feature two-sided attention network for anticancer natural products detection.

The exploration of anticancer drugs has aimed at more effective, adaptable, and less harmful treatments, with natural products pivotal in cancer research. In addition to experimental techniques for identifying anticancer drug candidates, computational methods have been developed to virtually screen for potential anticancer compounds. This study introduces DFT_ANPD, a deep-learning framework for predicting anticancer properties in natural compounds by integrating molecular structural information with embeddings generated by large language models (LLMs). Its architecture combines structural and semantic insights to form enriched molecular representations. SMILES strings are processed by a 1D-CNN to uncover intricate chemical patterns, while a fine-tuned SMILES-BERT extracts semantic information. A two-sided attention mechanism fuses these features, selectively prioritizing chemically relevant features for a holistic molecular representation. Based on this comprehensive embedding, a sigmoid activation function predicts anticancer potential. Benchmarked against iANP-EC with the NPACT and CancerHSP datasets, DFT_ANPD demonstrated superior performance in AUC-ROC, AUC-PR, balanced accuracy, and MCC. Its generalization ability was further validated using the independent NPASS dataset, identifying and ranking the top 10 compounds with the highest anticancer potential, including the FDA-approved dactinomycin. Docking analysis of NPC479359, the sixth-ranked compound, revealed significant binding affinities with top five cancer-related proteins, such as the CB2 receptor and DNA lyase. These results underscore DFT_ANPD's utility in accelerating anticancer drug discovery across experimental-computational pipelines.