DynaCDR: dynamic graph learning for accurate cancer drug response prediction
Background Accurate prediction of cancer drug response is essential for precision oncology, yet existing computational methods, particularly static graph-based models, suffer from two fundamental limitations: information leakage arising from global graph construction that inadvertently mixes training and test instances, and insufficient integration of multi-modal biological data with context-dependent relational reasoning. Results To overcome these challenges, we propose DynaCDR, a multi-modal framework built upon a dynamically constructed batch-local heterogeneous graph strategy. Unlike static global graphs that risk leaking test information through fixed topological connections, DynaCDR constructs local subgraphs exclusively from samples within each training batch, ensuring strict separation between training and test instances while capturing context-dependent drug-cell line interactions. Extensive experiments on the GDSC dataset demonstrate that DynaCDR achieves superior predictive performance ( RMSE = 0.7836, R 2 = 0.8817), outperforming the leakage-free static baseline, indicating that the gains arise from dynamic, context-dependent relational learning rather than from leakage avoidance alone. The model exhibits remarkable robustness in few-shot scenarios (using only 5–20% of training data) and exceptional generalization to unseen cell lines. More importantly, under the stringent cross-dataset transfer (GDSC-to-TCGA, no fine-tuning) that simulates real-world clinical domain shifts, DynaCDR consistently outperformed all other models by a substantial margin, despite an inevitable overall performance decline caused by the intrinsic in vitro versus in vivo gap. Crucially, ablation studies confirm the critical role of the graph attention mechanism, while biological enrichment analyses reveal that sensitive drug targets are significantly enriched in canonical pathways. Conclusions Overall, DynaCDR offers a general paradigm for precision oncology that addresses fundamental limitations of static graph-based approaches.
Authors
- Lei Liu (ORCID: https://orcid.org/0000-0002-9995-9080)
- Fan Zhong (ORCID: https://orcid.org/0000-0003-2632-011X)
- Liyan Zeng
- Yao Chen
Institutions
- Shanghai Medical College of Fudan University (CN)
- Fudan University (CN)
Publication Details
- Journal
- BMC Biology
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1186/s12915-026-02792-4
- Primary Topic
- Computational Drug Discovery Methods
- Type
- article
- Field-Weighted Citation Impact
- 0.00