Accurately predicting the binding affinity between ribonucleic acid (RNA) and small molecules (RSMA) is crucial for RNA-targeted drug discovery, yet existing computational methods face challenges in fully leveraging multisource feature information and modeling complex interactions. To address these challenges, this paper presents DeepMIF, an innovative deep learning framework based on a novel multiview interactive fusion paradigm. Initially, the framework employs a hybrid RNA representation combining a Localized Enhanced Scalable k-mer (L-ESKmer) strategy with pretrained embeddings to capture multiscale sequence patterns, while simultaneously extracting small molecule features from both sequence and graph views, yielding four distinct feature channels. At its core is an advanced multiview interactive fusion module wherein fine-grained interactions among multiple molecular modalities are modeled. Information is subsequently exchanged through a multihead cross-attention network equipped with a fused value vector. This mechanism transforms the attention process from simple information retrieval into an intelligent information synthesis, dynamically building a shared value vector from the context of all modalities. In a rigorous 5-fold cross-validation (CV) on a benchmark data set of 1439 RNA–small molecule pairs, DeepMIF demonstrates state-of-the-art performance, achieving a Pearson correlation coefficient (PCC) of 0.796 and a root-mean-square error (RMSE) of 0.874. More importantly, the model exhibits a strong generalization ability and robustness in challenging cold-start scenarios. The capability of DeepMIF to capture biologically meaningful, critical binding sites is further confirmed through interpretability analysis and case studies, highlighting its potential to guide structure-based RNA-targeted drug design.
Song et al. (Wed,) studied this question.