Prediction of Enhancer Regions using Deep Learning Techniques
| dc.contributor.author | Wong Yu Shiong | |
| dc.date.accessioned | 2026-10-16T08:04:04Z | |
| dc.date.issued | 2018 | |
| dc.description | Identification of transcription factor proteins binding locations (motifs) genome-wide and their specificities is a critical task for the understanding of gene regulation and associated diseases. Computational methods for motif prediction have been well studied and many tools have been proposed to target different categories of motifs. With more labelled datasets are available, machine learning techniques have become more useful for model construction. Deep-learning neural network has become a very prominent machine learning technique because of its capability to model complex datasets in different problem domains. This thesis aspires to develop machine learning systems to predict the locations and specificities of transcription factor binding sites (TFBSs) in DNA enhancer regions. In addition, the deep-learning is also employed to learn features associated with ChIP-datasets. For that, this thesis work has developed two models called DeepFinder and GMFR-CNN for motif prediction. First deep learning model called DeepFinder, which formulates the problem as enhancer classification, where an initial subset of the whole sequences can be performed by motif discovery tools and these tools select most potential candidate motifs as seed motif profile to predict TFBSs in the remaining sequences. The second deep learning model is called GMFR-CNN, which considers the importance of gapped motif features and integrates with deep convolutional neural networks (CNNs) that are able to compose higher level features from lower level ones. This thesis contributes to the problem solving in the following aspects: (1) a framework called three-stage method is proposed to tackle large search space task by using seed motif to search for TFBSs in the remaining sequences and this framework is employed in design of enhancer prediction using stacked autoencoders classifier; and (2) to take into account the significance of gapped motif feature, we proposed a novel representation to describe DNA sequence as twodimensional gapped motif feature representation (GMFR) and implements with CNNs as v classifier to obtain highly significant gapped features. A thorough assessment of the proposed frameworks has been done using five-fold cross validation method. Empirical results indicate that DeepFinder and GMFR-CNN have good potential in identifying the locations of TFBSs and specifying the enhancer binding regions in DNA sequences. Genes and enhancer regions annotation protocol for proboscis monkey genome was introduced. Analysis of coverage has been carried among different enhancer prediction tools where GMFR-CNN has shown highest coverage as compared to the other tools. Other than that, the prior knowledge regarding motifs and enhancer sequences is advantageous to improve prediction performance. | |
| dc.identifier.uri | https://scholarhub.unimas.my/handle/123456789/3133 | |
| dc.language.iso | English | |
| dc.publisher | UNIVERSITI MALAYSIA SARAWAK | |
| dc.relation.ispartofseries | Faculty of Cognitive Sciences and Human Development | |
| dc.subject | motif prediction, deep-learning neural network, enhancer, classification | |
| dc.title | Prediction of Enhancer Regions using Deep Learning Techniques | |
| dc.type | PhD |