DMAGCL: A dual-masked adaptive graph contrastive learning framework for predicting circRNA-drug sensitivity

Circular RNAs (circRNAs) are a unique class of non-coding RNAs with stable covalently closed structures that play key regulatory roles in gene expression and drug response. However, experimental identification of circRNA-drug sensitivity remains labor-intensive. To overcome these limitations, we introduce DMAGCL, a Dual-Masked Graph Contrastive Learning framework, whose core innovations include: (1) a synergistic dual-masking strategy (path- and edge-level) that forces the model to learn robust representations against both macro-level path disruptions and micro-level edge noise; (2) an adaptive contrastive loss with a scheduled temperature parameter (t) to dynamically balance exploration and exploitation during training; and (3) an attention-based fusion classifier (AFC) that explicitly models complex cross-modal interactions between circRNA sequences and drug molecular graphs for adaptive multi-source information fusion. Comprehensive evaluations demonstrate that DMAGCL achieves state-of-the-art performance, attaining an average AUC of 0.8940 and AUPR of 0.9006 under five-fold cross-validation, and a slightly higher average AUC of 0.8982 under the more stringent ten-fold cross-validation, consistently surpassing strong baselines including GATECDA and MNGACDA. This performance advantage stems from our core design choices, as evidenced by systematic ablation studies confirming the indispensable and complementary roles of the dual-masking strategy and the effectiveness of the adaptive loss and fusion classifier. Case studies on four representative anticancer drugs (doxorubicin, gefitinib, sorafenib, and paclitaxel) achieved an average experimental validation rate of 80%, highlighting the framework's predictive reliability and biological relevance. In conclusion, this study makes three primary contributions: (1) it introduces the novel DMAGCL framework, establishing a new paradigm for circRNA-drug association prediction via its synergistic dual-masking, adaptive learning, and attentive fusion components; (2) it delivers a highly robust and interpretable model with validated predictive reliability through extensive experiments and case studies (80% average validation rate); and (3) it provides a scalable computational tool that offers valuable insights for discovering novel circRNA-drug associations, understanding drug resistance mechanisms, and informing precision therapy design, with clear pathways for extension to other biological interaction tasks.