DALI-Syn: integrating chemical LLMs with multi-level attention architectures for synergistic drug combination...

Predicting drug synergy remains a critical challenge in personalized cancer treatment. Existing computational methods have made significant progress, offering the potential to accelerate the discovery of novel drug combinations. However, topology-based methods often fail to fully capture the deep chemical semantics within SMILES sequences. Moreover, models employing shallow feature fusion fail to effectively capture the complex interactions between drugs and cell lines.

We propose DALI-Syn, which integrates a chemical large language model and incorporates bidirectional drug-cell line interaction capabilities. The framework leverages ChemBERTa-2 to extract deep semantic representations from SMILES, transcending local topology. Simultaneously, Dynamic Multi-scale Convolution (DMC) and multi-head self-attention mechanisms refine drug and cell line features locally and globally. A symmetric cross-attention module then executes bidirectional feature fusion.

Experiments on the O'Neil dataset demonstrate that DALI-Syn outperforms state-of-the-art methods, achieving an AUC of 0.970 and AUPR of 0.908. Validation on an independent AstraZeneca test set confirms its robust generalization capability.

DALI-Syn advances drug synergy prediction through enhanced chemical semantic understanding and deep feature interaction, positioning it as a reliable tool to accelerate combination therapy development.