| Transcription factors are a group of proteins identified in the genome and able to bind to specific DNA regions to regulate gene expression.With the development of genomic technology,many related studies have shown that transcription factors are a kind of protein that plays a critical role in cellular signaling and is a crucial component of transcriptional regulation,and is necessary for regulating gene expression.Transcription factor binding sites are a specific type of DNA sequence that can bind to transcription factors.However,due to the richness of biological data,identifying the hidden transcription factor binding sites have been an important yet complex task.Biochemical experiments such as the electrophoretic mobility shift assay and the Ch IP-seq determine fewer transcription factor binding sites,and these methods are inefficient and expensive.Therefore,using computational biology techniques to achieve transcription factor binding site prediction has become a necessary option.The popular transcription factor binding site predicting algorithms now include sequence computation-based and machine learning-based transcription factor binding site predicting algorithms.These prediction algorithms have some advantages,but no one can predict transcription factor binding sites with high efficiency and precision.Therefore,the main effort of this thesis is to improve and optimize the related prediction algorithms.The main difficulty of transcription factor binding site prediction is feature selection fusion and prediction model construction.For the feature selection and fusion,the combined feature encoding effectively improves feature extraction,and the fusion of DNA shape data captures more discrepancy information in the original feature set.As for the prediction model construction,this thesis uses the related machine learning methods to overcome the drawbacks of traditional prediction algorithms.In addition,using the weighted multi-grained scanning strategy and the attention mechanism enhances the accuracy of prediction results and the efficiency of prediction of binding sites of transcription factors.In summary,for transcription factor binding site prediction algorithms,this thesis analyzed and summarized the current research status and problems of this type of algorithms,introduced the traditional biological test methods,sequence-based prediction methods,feature-based machine learning prediction methods,and other related DNA site prediction algorithms,and then clarified the content of this thesis and the innovation points of the algorithms.The main efforts of this thesis are as follows:1.The thesis proposes the WMS＿TF to predict the transcription factor binding site.To better extract DNA sequence features,WMS＿TF discards the idea of using only a single-nucleobase feature and combines multi-nucleobase features coding to extract inter-nucleobase signal features,improving the accuracy of classification prediction results.Moreover,to break the limitation of the traditional deep forest to view all features identically in the multi-granularity scanning stage,this thesis proposes the weighted multi-grained scanning strategy.The weight vectors were also scanned while the feature vectors,and the scanned vectors multiplied,ensuring the rigor of the model training to reduce the classification prediction error.Finally,this thesis analyzes the features for higher weights to justify the necessity of multi-nucleobase feature coding and establish the research foundation for other transcription factors binding site prediction.The experimental results demonstrate that the WMS＿TF can achieve highly accurate transcription factor binding site prediction.2.This thesis also proposes the LAM＿TF to predict the transcription factor binding site.To better represent the features of binding sites,besides using the DNA sequence data,this thesis also incorporates DNA shape data as the initial data for predicting transcription factor binding sites.In addition,also using the LSTM captures the long-term dependence between DNA sequences.Meanwhile,the attention mechanism used in LAM＿TF makes it capable of autonomously learning the importance of individual nucleobases or fragments in DNA sequences,overcoming the difficulty of existing algorithms efficiently to capture the effects of high-value nucleobases on gene regulation.Finally,using the prediction output module exports the binding affinity scores of the corresponding samples.The experimental results demonstrated that the LAM＿TF algorithm enhanced the prediction ability of transcription factor binding sites.This thesis provides an in-depth study of feature-selecting fusion and prediction model construction in transcription factor binding site prediction.The study optimized the traditional representation methods of DNA sequence feature to extract single-nucleobase features while capturing inter-nucleobase signal features.In addition,this thesis also designed two efficient prediction algorithms to improve the accuracy of the prediction results.This thesis experimentally verifies that the WMS＿TF algorithm and LAM＿TF algorithm can enhance the accuracy and efficiency of transcription factor binding site prediction,providing a basis for further research on transcription factor binding site prediction. |
