Relation-Aware Heterogeneous Graph Learning for miRNA-Drug Response Association Prediction Under Cold-Start Generalization

(1) Background: Cancer drug resistance can be influenced by microRNAs (miRNAs), which can alter gene expression and cellular pathways that determine how cancer cells respond to treatment. Potential miRNA-drug response associations can be identified using computational models; however, evaluating these models using edge-based evaluation can result in an optimistic view of model generalization since test miRNAs or drugs can appear in other response associations during training. (2) Methods: This paper presents a relation-aware heterogeneous learning framework that integrates miRNA expression profiles, miRNA-gene regulatory interactions, gene-gene functional associations, drug molecular features, and distinct drug-gene relations in a unified miRNA-gene-drug network. A Heterogeneous Graph Transformer (HGT) was used to preserve relation-specific information, and biological paths connecting each miRNA to each drug were used as additional information for the corresponding candidate miRNA-drug pair. A leakage-safe five-fold cross-validation approach was used to evaluate the proposed framework under different settings: edge, miRNA-cold, drug-cold, and double-cold. (3) Results: The proposed framework obtained an AUC and AUPR of 0.9537 ± 0.0044 and 0.9494 ± 0.0069, respectively, under the edge-based setting. The performance remained high under the miRNA-cold-based setting, where the model achieved an AUC and AUPR of 0.9466 ± 0.0061 and 0.9436 ± 0.0049, respectively. A substantial drop in performance was observed under the drug-cold setting, where the proposed framework obtained an AUC and AUPR of 0.6685 ± 0.0354 and 0.6534 ± 0.0569, respectively. The performance was also substantially reduced under the double-cold-based setting. The results also show that the structural resemblance alone is not enough to explain generalization to new drugs, since no significant relationship between cold-start performance and either ChemBERTa molecular similarity or DrugBank target similarity was shown by drug-level analysis. Literature evidence supported 6 out of 20 top-ranked previously unlabeled predictions compared to 1 of 120 randomly selected controls (Fisher's exact test, p = 2.63 × 10^-5; empirical p = 2.80 × 10^-5). (4) Conclusions: The obtained results show that the proposed framework can rank candidate miRNA-drug response associations effectively. They also show that predicting response associations under the drug-cold-based setting remains much harder than predicting for similarly held-out miRNAs, which is an important limitation of conventional edge-based evaluation.