利用深度學習來預測阿拉伯芥DNA序列中編碼基因的基因結構

Abstract

基因的結構可以使我們了解其功能，它可以透過如Augustus等模型的預測來獲得。這些模型為了註解DNA序列，需事先對其特徵組成進行分析並設計多個子模型來偵測。深度學習不需要事先分析其特徵組成並可以學習它所需要的特徵，使之容易應用在多個領域。本研究的目的為建立一個深度學習模型來對阿拉伯芥DNA序列上編碼基因的基因結構進行預測。本研究藉由global run-on sequencing和Poly (A)-Test RNA-sequencing的資料來清洗與重新註解現有的轉錄資料，並得到含有977編碼基因的註解。本研究提出一個全新的深度學習模型和新的損失函數。結果顯示深度學習在macro F-score的中位數為0.969，而在Augustus的結果為0.957，且統計結果顯示深度學習在macro F-score顯著優於Augustus。本研究提出兩種後處理方法，一種名為邊界後處理方法（boundary post-processing method）來處理內含子的邊界，另一種名為長度過濾方法（length filtering method）來處理短片段。深度學習的預測結果經處理後在16個評分中有9個評分有顯著進步。深度學習的預測結果經後處理方法處理後顯示在16個評分中有6個顯著好於Augustus和5個顯著落後於Augustus。這些結果顯示深度學習模型結合後處理方法可以和Augustus匹敵。另外，經後處理方法處理的深度學習預測結果可以在部分基因體上預測出平均為18642個含有已知蛋白質結構域的基因結構。整體來講，深度學習模型結合後處理方法可以成為在阿拉伯芥DNA序列上預測編碼基因的基因結構的替代方法。

The structure of the gene can help us to have a better understanding of its function, and it can be predicted by models such as Augustus. In order to annotate the DNA sequence by these models, the feature composition of annotation needed to be analyzed, and many submodels would be designed to detect these features. The deep learning does not need to analyze the feature composition and can learn the features it needs, and this makes it easily be applied in many fields. The purpose of the thesis is to build a deep-learning-based model to directly predict gene structures of coding genes in DNA sequences of Arabidopsis thaliana. Annotation with 977 coding gene structures was created by using data from global run-on sequencing and Poly (A)-Test RNA-sequencing to reannotate and filter the existed transcripts. A new deep learning model and loss were proposed. The median macro F-score of the deep learning model was 0.969, and the value of Augustus was 0.957. The statistical result showed that the result of the deep learning model in the macro F-score was significantly better than Augustus. Two post-processing methods were proposed, one named boundary post-processing method handled the boundary of the intron, and the other named length filtering method filtered out the region with short length. The revised result of the deep learning model showed that there were 9 out of 16 metrics performances were significantly improved. The revised result of the deep learning model showed that 6 out of 16 metrics were significantly better than Augustus, and 5 out of 16 metrics were significantly worse than Augustus. These results show that the deep learning model with the post-processing procedure is competitive to Augustus. Furthermore, the revised result of the deep learning model on the part of the genome showed that it could predict an average of 18642 gene structures that contained existed protein domains. Overall, the proposed deep learning model with the post-processing procedure can be an alternative method to predict gene structures of coding genes on DNA sequences of Arabidopsis thaliana.