张强锋/汪阳明合作开发AI工具SMRTnet，无需RNA三级结构，精准预测小分子-RNA相互作用

Core Viewpoint - RNA-targeted small molecule drugs are emerging as a new frontier in the biopharmaceutical field, addressing the challenge of lacking precise tertiary structure information for most disease-related RNAs, which traditional computational methods rely on for predicting interactions with small molecules [2][5]. Group 1: SMRTnet Development - The research teams from Tsinghua University and Peking University developed a deep learning tool named SMRTnet, which predicts small molecule-RNA interactions without relying on RNA tertiary structures, thus expanding the range of targetable RNA [2][5]. - SMRTnet integrates two large language models, convolutional neural networks, and graph attention networks, utilizing only RNA sequence and secondary structure information to predict binding capabilities [5]. Group 2: Performance and Validation - SMRTnet demonstrated robust performance, achieving an average area under the receiver operating characteristic curve (auROC) of 0.830-0.844 in five-fold cross-validation, significantly outperforming existing tools [9]. - In bait evaluation tasks, SMRTnet's average ranking reached 92.6%, far exceeding four molecular docking tools (27.3%-46.6%) and two deep learning tools (16.0%-23.8%), indicating its superior ability to identify true binding molecules [9]. - The model's performance was notably affected when using predicted secondary structures instead of experimentally determined ones, highlighting the importance of accurate experimental data [10]. Group 3: Binding Site Prediction - SMRTnet can also identify small molecule binding sites on RNA, quantifying the contribution of each nucleotide to the predicted binding score using the Grad-CAM algorithm [12]. - The accuracy of binding site predictions was validated against experimentally determined sites, achieving an average auROC of 0.695-0.793 across multiple datasets [13]. - Out of 190 predicted small molecule-RNA interactions, 40 were experimentally validated as binding molecules, yielding an average validation rate of 21.1% [14]. Group 4: Case Study on MYC IRES - The research focused on MYC IRES, a target considered "undruggable," showing a positive correlation between predicted binding scores and experimental validation rates, with a validation rate of 28.6% for scores between 0.9-1.0 [16]. - Among 15 identified MYC IRES binders, the team highlighted IHT (Irinotecan Hydrochloride Trihydrate) for its favorable drug development characteristics [17]. - IHT was predicted to bind at a specific site on MYC IRES, and experimental validation confirmed the reliability of SMRTnet's predictions, demonstrating significant reductions in MYC mRNA and protein levels in HeLa cells [19]. Group 5: Future Prospects - SMRTnet represents a significant advancement in predicting small molecule-RNA interactions, overcoming traditional method limitations and broadening the scope of disease-related RNA targets [21]. - The accumulation of multi-omics data, including chemical-RNA interaction genomics and functional screening, is expected to enhance AI methods for predicting binding interactions and downstream biological effects, accelerating RNA-targeted drug development [21]. - The study illustrates the potential of AI-driven approaches in RNA-targeted small molecule therapy, with expectations for breakthrough advancements in the coming years as data quality and algorithm optimization improve [21].