Predicting circRNA-Drug Sensitivity Using Integrated Hybrid Graph and Molecular Features

Circular RNAs (circRNAs) are a class of covalently closed RNA molecules that exhibit high stability and play crucial roles in tumorigenesis and therapeutic response. Increasing evidence highlights their potential as biomarkers for drug sensitivity and resistance. However, uncovering circRNA-drug interactions remain challenging: experimental approaches are costly and time-consuming, while many existing computational methods rely on limited similarity features, leading to incomplete utilization of available data and reduced predictive accuracy. We propose HMCDSP, a hybrid molecular and network-based circRNA-Drug sensitivity prediction framework. HMCDSP integrates BERT-derived circRNA sequence embeddings and SMILES-based drug representations with network-level features learned through a two-layer hybrid graph neural network combining Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT). To address data sparsity, circRNA and drug similarity networks were constructed by fusing sequence-derived metrics with Gaussian Interaction Profile kernels. On the NcRNADrug benchmark, HMCDSP consistently outperformed six state-of-the-art methods, achieving an average Area Under Curve (AUC) of 0.9588 and Area Under the Precision-Recall Curve‌ (AUPRC) of 0.9622. Ablation analyses confirmed the complementary contributions of molecular and network features, while classifier comparisons established the superiority of the Multilayer Perceptron (MLP) predictor. Case studies further validated predicted circRNA-drug associations, including cisplatin, enzalutamide, and sorafenib, against published experimental findings, demonstrating HMCDSP's capacity to reveal clinically relevant biomarkers and inform personalized therapeutic strategies.