im6Am-DC：融合DenseNet与注意力机制的RNA序列N6,2′-O-二甲基腺苷修饰预测模型

im6Am-DC: a model for predicting RNA sequence N6,2'-O-dimethyladenosine modification by integrating DenseNet and attention mechanism

摘要:

N6,2′-O-二甲基腺苷(N6,2′-O-dimethyladenosine, m6Am)作为一种重要的RNA表观遗传修饰，其独特的化学结构和生物学功能成为当前研究的热点之一。m6Am不仅能够调控mRNA的稳定性和半衰期，还会调节其与翻译起始因子及RNA结合蛋白的相互作用，从而在转录后调控过程中发挥关键作用。m6Am还与肥胖、2型糖尿病等代谢性疾病以及多种病毒感染和肿瘤的发生密切相关。对m6Am位点的预测研究对于疾病的早期诊断和精准治疗具有重要意义。尽管m6Am修饰的生物学功能正逐渐受到关注，但相关研究仍受限于检测技术的局限与高昂的成本。计算生物学方法为大规模m6Am位点鉴定提供了新的思路。本研究构建了一种非平衡m6Am位点预测模型——im6Am-DC，该模型采用DenseNet网络提取高级局部特征，并引入注意力机制(coordinate attention, CA)以突出重要的特征信息。此外，为解决数据不平衡问题，模型使用了损失函数。结果显示，在full transcript数据集上，im6Am-DC模型的灵敏度(sensitivity, Sn)、特异性(specificity, Sp)、准确度(accuracy, Acc)和马修斯相关系数(Matthews correlation coefficient, MCC)分别为0.523 7、0.988 4、0.946 1和0.629 8；在mature RNA数据集上，同指标表现同样优秀。同时，将im6Am-DC模型分别应用于多个不同的数据集进行预测，包括非平衡m6Am位点数据集、平衡m6Am位点数据集，以及其他类型的RNA修饰位点，从而全面、多维度评估该模型的性能与应用潜力。结果表明，im6Am-DC模型在非平衡m6Am位点预测中具有显著优势，模型的建立为m6Am修饰功能研究及相关疾病机制探索提供了有效的技术支持。

Abstract:

N6,2'-O-dimethyladenosine (m6Am), as a critical RNA epigenetic modification, has become a hot topic in current research due to its unique chemical structure and biological functions. m6Am not only regulates mRNA stability and half-life but also modulates its interactions with translation initiation factors and RNA-binding proteins, thereby playing a crucial role in post-transcriptional regulation. Recent medical research indicates that m6Am is closely associated with metabolic disorders such as obesity and type 2 diabetes mellitus, as well as multiple viral infections and tumor development. Therefore, predicting m6Am site holds significant importance for early diagnosis and precision treatment of diseases. Although the biological functions of m6Am modification are gradually attracting attention, related research is still limited by the limitations of detection technology and high costs. Computational biology methods provide new ideas for large-scale identification of m6Am sites. This paper proposes an imbalanced m6Am site prediction model—im6Am-DC. The model employs a DenseNet architecture to extract high-level local features and introduces a CA attention mechanism to highlight crucial feature information. Furthermore, to solve the data imbalance, the model utilizes the focal loss function. Experimental results showed that on the full transcript dataset, the im6Am-DC model achieved the sensitivity (Sn) of 0.523 7, specificity (Sp) of 0.988 4, accuracy (Acc) of 0.946 1, and Matthews correlation coefficient (MCC) of 0.629 8. On the mature RNA dataset, the model demonstrated equally strong performance across these metrics. Simultaneously, we employed the im6Am-DC model to predict across multiple distinct datasets, including unbalanced m6Am site datasets, balanced m6Am site datasets, and other types of RNA modification sites. This comprehensive, multidimensional approach enabled us to evaluate the performance and application potential of the model. The results demonstrate that the im6Am-DC model exhibits significant advantages in predicting unbalanced m6Am sites, providing effective technical support for studying the roles of m6Am modifications and exploring related disease mechanisms.