| 摘 要: 牛津纳米孔测序能够直接从单分子电流信号中检测DNA甲基化，相比传统方法具有独特优势。然而，现有检测算法普遍存在适配性不足、准确率不稳定等问题。针对上述问题，我们提出了基于多特征融合与改进Transformer架构的甲基化检测算法SigTrans。该算法融合统计特征与碱基序列构建多模态特征矩阵，并采用 CNN–Transformer–MLP 级联结构以提升空间特征提取与序列建模能力，同时兼容 R9.4 与 R10.4.1 流动池数据。结果表明，SigTrans在人类与幽门螺杆菌样本的单读长检测任务中，F1 分数较 DeepMod2 等主流工具平均提升约 1.2%，为单读长DNA 甲基化的检测提供了一种高效可靠的技术路径。 |
| 关键词: DNA甲基化 注意力机制 深度学习 多特征融合 |
|
中图分类号:
文献标识码:
|
|
| Research on DNA Methylation Modification Prediction Algorithm Based on Oxford Nanopore Sequencing and Transformer Model |
|
wushumin, daiqi
|
zhejiang sci-tech university
|
| Abstract: Oxford Nanopore sequencing can directly detect DNA methylation from single-molecule current signals, which confers unique advantages over traditional methods. However, existing detection algorithms generally suffer from poor adaptability and unstable accuracy. To address these issues, we propose SigTrans, a methylation detection algorithm based on multi-feature fusion and an improved Transformer architecture. This algorithm integrates statistical features and base sequences to construct a multimodal feature matrix, and adopts a cascaded CNN–Transformer–MLP structure to enhance the capabilities of spatial feature extraction and sequence modeling, while being compatible with data generated by both R9.4 and R10.4.1 flow cells. The results demonstrate that SigTrans achieves an average F1-score improvement of approximately 1.2% compared with mainstream tools such as DeepMod2 in the single-read detection tasks of human and Helicobacter pylori samples, providing an efficient and reliable technical approach for single-read DNA methylation detection. |
| Keywords: DNA methylation Attention mechanism Deep learning Multi-feature fusion |
