AttenDRP: A Cross-Attention Network for Cancer Drug Response Prediction by Integrating Multi-Omics Data

Background: The rapid advancement of precision medicine in oncology has intensified the need for the accurate prediction of cancer drug responses. Deep learning technologies present a promising pathway toward addressing this challenge. However, because cancer is a complex disease influenced by multiple molecular layers, the integration of multi-omics data and the elucidation of their intricate relationships are essential for improving predictive accuracy. Traditional methods currently have limitations in fully exploiting multi-omics information and capturing complex interactions, highlighting the need for more sophisticated approaches. Objective: This study aims to design a new method for integrating drug and cell line data through cross-attention, thereby improving the accuracy of drug response prediction. Methods: In this study, we leverage a Transformer to encode drug SMILES sequences to capture fine-grained molecular semantics and introduce a cross-attention mechanism to model bidirectional interactions between drugs and multi-omics features, thereby significantly improving the accuracy and robustness of drug response prediction. Results: Experimental evaluations demonstrate that, compared with traditional approaches, the proposed method achieves considerable improvements in accuracy and stability across diverse prediction scenarios. The model exhibits robust performance in managing complex multi-omics inputs and reliably predicting drug responses Discussion: The proposed AttenDRP model demonstrates strong potential in drug response prediction. Future research can expand its applicability by leveraging large and diverse datasets and incorporating patient-derived or clinical data to further strengthen robustness and translational utility Conclusion: AttenDRP integrates heterogeneous drug features with multi-omics cell line data through a cross-attention network, enabling the modeling of fine-grained interactions. This approach improves the accuracy and robustness of drug response prediction and provides a methodological foundation for advancing precision oncology