Zhang Meng, Wu Jing, Wang Yulan, Cao Yan, Liu Jingjing, Wang Quan, Song Xiaofeng, Zhao Jian, Wang Yixuan
Department of Biomedical Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China; School of Mathematics and Statistics Science, Ludong University, Yantai 264025, China.
School of Biomedical Engineering and Informatics, Nanjing Medical University, Nanjing 211166, China.
Int J Biol Macromol. 2025 Jul;318(Pt 4):145341. doi: 10.1016/j.ijbiomac.2025.145341. Epub 2025 Jun 18.
N7-methylguanosine (m7G) is one of the most prevalent post-transcriptional modifications in RNA molecules, playing a pivotal role in regulating RNA metabolism and function. Given the complexity of canonical m7G cap-dependent protein synthesis, accurately predicting m7G modification sites facilitates further exploration of translation initiation mechanisms. Hence, we collected the most comprehensive single-nucleotide resolution m7G modification sites from the updated m7GHub v2.0 database. We subsequently developed Deep-m7G, a novel contrastive learning-enhanced deep biological language model, designed for both the full transcript and mature RNA datasets. Our methodological framework incorporates three key innovations: (1) implementation of a Most Distant undersampling strategy to mitigate class imbalance in training data; (2) integration of DNABERT-2 with a parallel convolutional neural network architecture for hierarchical feature extraction; and (3) introduction of a contrastive learning module to enhance feature discriminability and model generalizability. Systematic evaluation through 10-fold cross-validation demonstrated the critical contribution of our contrastive learning component. In rigorous benchmarking against existing tools, Deep-m7G achieved superior predictive performance (Full transcript: AUC = 0.960 vs 0.653-0.898 and Mature RNA: AUC = 0.845 vs 0.684-0.832) on independent test sets. Collectively, this computational advance provides a robust framework for the discovery of epitranscriptomics markers, thereby advancing mechanistic investigations of post-transcriptional regulation.
N7-甲基鸟苷（m7G）是RNA分子中最普遍的转录后修饰之一，在调节RNA代谢和功能方面发挥着关键作用。鉴于经典的m7G帽依赖性蛋白质合成的复杂性，准确预测m7G修饰位点有助于进一步探索翻译起始机制。因此，我们从更新后的m7GHub v2.0数据库中收集了最全面的单核苷酸分辨率m7G修饰位点。随后，我们开发了Deep-m7G，这是一种新颖的对比学习增强型深度生物语言模型，适用于完整转录本和成熟RNA数据集。我们的方法框架包含三项关键创新：（1）实施最远欠采样策略以减轻训练数据中的类别不平衡；（2）将DNABERT-2与并行卷积神经网络架构集成以进行分层特征提取；（3）引入对比学习模块以增强特征可辨别性和模型通用性。通过10折交叉验证进行的系统评估证明了我们对比学习组件的关键作用。在与现有工具的严格基准测试中，Deep-m7G在独立测试集上实现了卓越的预测性能（完整转录本：AUC = 0.960，而现有工具为0.653 - 0.898；成熟RNA：AUC = 0.845，而现有工具为0.684 - 0.832）。总体而言，这一计算进展为发现表观转录组学标记提供了一个强大的框架，从而推动转录后调控的机制研究。
