InterActRNA: A Co-Evolutionary Geometric Deep Learning Framework for RNA-Small Molecule Binding Affinity Prediction.

Abstract

RNA is identified as a critical therapeutic target due to its involvement in various cellular functions. Accurate prediction of RNA-small molecule affinity (RSMA) is prioritized for accelerating drug discovery. However, existing deep learning models predominantly treat RNA and small molecules as static entities. The dynamic induced-fit-inspired co-adaptation and multi-scale chemical information are not adequately captured. In this study, InterActRNA is presented as a co-evolutionary geometric deep learning framework. To emulate the induced-fit-inspired dynamic co-adaptation mechanism, an Interleaved Guided Encoding Module is designed. An Iterative Drug Encoder and a Hierarchical RNA Encoder are incorporated to exchange information via cooperative attention. Feature learning in one modality is real-time guided by the contextual information of the partner. Furthermore, a Cross-Modality Substructure Similarity (SSIM) module is introduced. Pairwise interactions between small molecule functional groups and RNA structural motifs are explicitly modeled. The predictive capability is evaluated on the industry-standard R-SIM dataset. InterActRNA consistently outperforms state-of-the-art baselines in both regression and classification tasks. Robustness in rigorous cold-start scenarios is demonstrated. Finally, key pharmacophores are pinpointed through interpretability analysis. The potential to guide rational RNA-targeted drug design is confirmed.