A hybrid multimodal network for cancer drug response prediction using delayed pooling and cross attention at the atom level

Accurate prediction of cancer drug response remains a major problem in precision oncology, particularly when prediction is required for therapeutic compounds or tumor profiles that are poorly represented in training data. In this study, the Hybrid Multi-Modal Drug Response Network (HMM-DRN) is presented for joint learning from transcriptomic profiles and molecular graph structure. Each compound is represented as an explicit molecular graph and encoded by graph convolutional layers, while cellular state is represented using split-specific transcriptomic features selected and standardized without using held-out cell lines. The model is evaluated under four leakage-controlled scenarios: mixed drug-cell pairs, unseen cell lines, unseen drugs, and simultaneously unseen drugs and unseen cell lines. Across five random seeds, HMM-DRN shows its clearest numerical advantage in the unseen-drug setting, where it obtains RMSE = 2.4902 ± 0.2598, PCC = 0.5422 ± 0.0882, and SRCC = 0.4550 ± 0.0269. However, the advantage over the early-pooling Static-GNN is not uniform across all scenarios and is not statistically significant in the available paired tests. The double-unseen setting remains challenging, with Static-GNN numerically outperforming HMM-DRN in that scenario. These results support a cautious conclusion delayed atom-level fusion is a useful design hypothesis for selected out-of-distribution drug-response prediction settings, but it does not establish uniform superiority over early-pooling graph baselines. The split files and repeated-seed result tables are provided to support reproducibility.