DNA Enhancer-promoter Interaction Identification Based On Multi-task Learning

The distal cis-acting DNA regulatory elements play important roles in controlling gene expression.They are the key factor to activate and maintain gene transcription.Accurately predicting enhancers and identifying enhancer-promoter interactions(EPI)are critical for understanding the gene expression regulation and discovering the causes of some diseases.Therefore,enhancer prediction and EPI identification have attracted more researchers attention,recently.Existing enhancer prediction methods mainly use the genomic features obtained from high-throughput experiments to represent DNA sequences.As the high-throughput data are unavailable for many tissues or cell-types,these methods are inapplicable to many tissues or cell-types.At the same time,the existing studies on the EPI identification always ignore the trigger effect of enhancer on EPI,which affects the performance of EPI identification.Targeted to the above problems,this thesis first investigates effective enhancer prediction methods,and then studies the EPI identification method based on multi-task learning.The main contents of this thesis are as follows.The research on enhancer prediction method based on attention mechanism and adversarial network.In order to solve the problem that the existing methods are highly dependent on high-throughput genomic features,the chromatin features predicted by the Deep Sea model are utilized to represent DNA sequences.In order to solve the problem that existing enhancer prediction methods have the shortcomings to model long DNA sequence,the context information of DNA sequence are learned by bidirectional long short-term memory network,while the feature association relationships within DNA sequence are modeled by attention mechanism.The experimental results on public datasets show that the proposed method achieved 1.70%,1.35% and 2.75% improvements in AUC,AUPR and F1 compared to the existing state-of-the-art(SOTA)methods,respectively.Aiming at the species-specific problem of enhancers,an adversarial network is introduced to capture the common features between different species for improving the performance of cross-species enhancer prediction.The experimental results on cross-species enhancer prediction show that this method achieved 2.85%,2.70%and 2.60% improvements,compared to the existing SOTA method in AUC,AUPR and F1,respectively.It achieves the highest known performance.Research on DNA enhancer-promoter interaction identification based on multitask learning.In order to solve the problem of insufficient representation capability of DNA sequence expressed by dna2 vec,we further introduce the high-throughput genomic information.Aiming at the problem that existing EPI identification methods ignore the trigger effect of enhancer,a multi task learning method is proposed to jointly model the enhancer prediction task and EPI identification task.By setting the task private layer and the shared layer.The common features and task-specific features between two tasks are extracted to improve EPI identification.The experimental results show that compared with the SOTA method,the proposed multi-task learning method improved 2.0% and 2.0% in AUC and AUPR,respectively.It shows that the combining the multi-task learning method which incorporates enhancer prediction effectively improve the performance of EPI identification.