Abstract
Personalized cancer treatment faces challenges due to tumor heterogeneity and limitations in existing computational methods regarding multi-omics integration. To address these limitations, this study presents MOGA, an advanced framework for drug response prediction. MOGA constructs a heterogeneous graph integrating cell line multi-omics, drug chemical structures, and response data, utilizing a relational graph convolutional network (RGCN) to model complex interactions. Crucially, a type-specific graph augmentation strategy is proposed, which categorizes neighbor influence to preserve critical sensitive signals while reducing noise from non-sensitive nodes. Experimental results demonstrate that MOGA significantly outperforms competitive baselines in both AUROC and AUPR. Ablation studies and case analysis confirm that the integration of multi-omics data and the targeted augmentation mechanism are central to these performance gains, validating MOGA's potential for precision oncology.