利用深度學習預測微核醣核酸可能結合之信使核糖核酸結合位

Abstract

次世代定序、高通量技術使得定序的成本大幅下降、大量的定序資料產生，再加上電腦運算加快與演算法的開發，讓生物學家能和電腦科學家合作探討不同問題。基因調控是眾多生物資訊問題中重要的一環，基因的調控會影響生物體產生蛋白質的能力，不當的調控會使生物體產生疾病，在人類和哺乳類中，microRNA約調控60%的基因，microRNA的失調也與眾多疾病有關。本研究著重於預測microRNA(微核醣核酸)與mRNA(信使核醣核酸)之間的結合關係，旨於協助基因調控網路的建立。 過往研究對於microRNA與mRNA之間的結合關係，主要是將microRNA和mRNA之間的種子區域互補關係、序列保守性、結合穩定性等等資料當成特徵來預測可能的結合位，但這種使用人類定義的生物特徵方法無法完全掌握序列未知的結合關係，有鑒於此，本研究使用深度學習方法進行預測，利用深度學習強大的特徵擷取能力，以microRNA序列和mRNA結合位序列當成輸入進行microRNA與mRNA結合位點的預測。本論文所提出的深度學習模型於兩個測試集上的表現結果，V1資料集達到98%的準確度、V2資料集則達到95%的準確度，相較於其他現有基於規則或深度學習方法均有所提升。相信這個研究的成果與所累績的經驗，將對microRNA調控網路的建立有莫大的幫助。

With the advanced of Next Generation Sequencing (NGS) and high-throughput technologies, the cost of sequencing greatly reduced and a large amount of sequencing data produced. In addition, the acceleration of computer operations and the development of algorithms, allowing biologists and computer scientists to collaborate on different issues. Gene regulation is an important part of many biological information issues, improper regulation may cause diseases. In humans and mammals, microRNA regulates approximately 60% of genes, and the dysregulation of microRNA is also related to many diseases. This research focuses on predicting the binding relationship between microRNA and mRNA, aiming to help the establishment of gene regulatory networks. Previous research mainly considered the complementary relationship between microRNA and mRNA, including seed regions, sequence conservation, binding stability and other data as features to predict possible binding sites. But the human defined biometric methods cannot fully interpret the binding relationship of all kinds of interaction pairs. In view of this, this research used a deep learning approach, utilizomg the powerful feature extraction capabilities of deep learning to overcome the problem. In this research, we constructed an end-to-end deep learning framework, given a microRNA sequence and mRNA binding site sequence as input, the model will predict whether it’s a functional interaction pair or not. The performance on the V1 testing set achieved 98% accuracy, and on the V2 testing set achieved 95% accuracy, which is a huge improvement when compared with other existing rule-based or deep learning method tools. It is believed that the proposed method and the accumulated experiences of this study can largely benefit the future research of constructing microRNA regulatory networks.