NCRDLLM: Predicting ncRNA-Drug Response Associations via Multimodal Feature Fusion and Large Language Models

Noncoding RNAs (ncRNAs) play critical regulatory roles in cancer drug response. However, most existing methods are limited to predicting a single type of ncRNA, failing to fully capture the complex semantic associations between multimodal biological features, and thus exhibit weak generalizability and robustness. To overcome these limitations, this study proposes NCRDLLM, a unified framework that leverages large language models (LLMs) to predict associations between three types of ncRNA (circular RNA, microRNA, and long noncoding RNA) and drugs. The method integrates 19,020 experimentally validated associations and 120,009 disease association records. Three types of multimodal features are constructed: sequence features extracted using pretrained foundation models RNA-FM and ChemBERTa, structural features generated through Graph2Vec for RNA secondary structures and AttentiveFP combined with ECFP for drug molecules, and association features obtained via disease-associated coding and semantic similarity. These features are subsequently mapped into the hidden space of LLaMA-3.2-3B through adapter modules, with LoRA employed for parameter-efficient fine-tuning. Experimental results demonstrate that NCRDLLM achieves AUC-ROC values of 0.9665, 0.9832, and 0.9676 on miRNA-drug, lncRNA-drug, and circRNA-drug data sets, respectively. Ablation studies confirm the contribution of each module, while literature evidence and tissue-specific expression profiling further support the biological relevance of the predictions. NCRDLLM provides an effective strategy for identifying potential ncRNA-drug response associations.