Graph-learning model reports high drug-response accuracy in cancer cell lines

A new computational framework called HGACL-DRP achieved a reported mean area under the receiver operating characteristic curve of 95.48% for drug-response prediction on the Cancer Cell Line Encyclopedia dataset and 98.99% on the Genomics of Drug Sensitivity in Cancer dataset, according to a research article published in BMC Bioinformatics. These results concern computational predictions for cancer cell-line data and do not establish performance in patients or clinical treatment selection.

The model was designed to address the marked heterogeneity of cancer, which can lead to substantial variation in responses to the same treatment. Its developers sought to improve drug-sensitivity prediction while reducing reliance on labeled data, a limitation they associated with existing methods. HGACL-DRP stands for Heterogeneous Graph Attention Dual-Perturbation Contrastive Learning Network for Drug Response Prediction.

At the core of the framework is an optimized heterogeneous graph structure that integrates multi-omics features. It also incorporates a neighbor-filtering mechanism. The architecture combines these elements with a graph-attention mechanism and a dual-perturbation contrastive-learning paradigm. According to the study abstract, the two perturbations operate on features and graph structure and are intended to support self-supervised learning of more robust heterogeneous graph node embeddings.

The researchers tested the approach on two major public datasets: GDSC and CCLE. Performance was assessed across multiple classification metrics. The reported mean AUC was higher on GDSC than on CCLE, with values of 98.99% and 95.48%, respectively. The article also reports that HGACL-DRP produced statistically significant improvements in predictive accuracy compared with existing state-of-the-art drug-response prediction algorithms.

The authors describe the model’s multi-view graph augmentation strategy and attention mechanism as tools intended to support personalized drug selection and novel drug discovery. Those proposed applications remain distinct from demonstrated clinical benefits: the supplied study evidence reports a computational evaluation on public cancer cell-line datasets, not a patient trial or prospective clinical study. The article therefore provides evidence that HGACL-DRP can perform drug-response prediction in the evaluated datasets, while its usefulness in clinical decision-making remains unestablished.

Subject of Research: Computational prediction of cancer drug response using heterogeneous graph learning

Article Title: HGACL-DRP: Heterogeneous graph attention dual-perturbation contrastive learning network for drug response prediction

Article References: Jin, M., Lou, M., Wang, P., Xu, W., Hou, G., Wang, L., & Li, M. (2026). HGACL-DRP: Heterogeneous graph attention dual-perturbation contrastive learning network for drug response prediction. BMC Bioinformatics. https://doi.org/10.1186/s12859-026-06723-6

Image Credits: AI Generated

DOI: 10.1186/s12859-026-06723-6

Keywords: drug-response prediction, cancer cell lines, HGACL-DRP, heterogeneous graphs, graph attention, contrastive learning, multi-omics, computational biology, personalized medicine, Heterogeneous, graph, attention