Abstract Motivation Accurate prediction of drug response remains a major challenge in precision oncology, particularly at the single-cell level and in clinical settings, due to significant distribution shifts between preclinical models and real-world patient data. Existing approaches often rely on transfer learning from cell lines to target domains, but typically require access to target-domain data during training, which is frequently unavailable in practice. Results We propose FourierDrug, a novel domain generalization framework for robust drug response prediction. Given gene expression profiles, the model performs Fourier transformation to project features into the frequency domain and introduces an asymmetric attention mechanism that encourages drug-sensitive samples to form compact clusters while driving resistant samples to be more dispersed. This design facilitates the learning of domain-invariant yet task-relevant representations. Extensive experiments demonstrate that FourierDrug effectively leverages diverse source domains and generalizes well to unseen cancer types. Notably, when evaluated on single-cell and patient-level prediction tasks, our method—trained solely on in vitro cell line data without access to target-domain data—consistently outperforms or matches state-of-the-art approaches. Availability and Implementation The source code and processed datasets are available at: https://github.com/hliulab/FourierDrug. Supplementary Information Supplementary data are available at Bioinformatics online.
Song et al. (Sat,) studied this question.