Chromatin Contacts Map Prediction Based On Histone Marks And Deep Learning

Chromatin interaction is a phenomenon that interacts between different chromatins base sequences,affecting genetic functions such as biogenetic expression and protein translation.The sequencing experiment data,which named contacts map,costs too much and difficult to migrate between different cell types.The main problems are following:(1)There are few analyses on using histone marks and DNA sequence data to predict chromatin contacts map.The features extraction of histone modification samples lacks the research on the influence of spatial location and the prediction method of chromatin interaction migration for different cell types is difficult to promote.(2)No DNA sequence is introduced to calculate its characteristics affect the prediction results of chromatin contacts map,and there is no effective method for fusion of multiple data features.It is difficult for existing models to migrate to predict chromatin contacts map of other cell types;(3)There is a lack of a complete tool to promote chromatin contacts map on different cell types by multiple data processing and models. In view of the above problems,thesis explores a variety of data processing methods and deep learning models,the main contributions include:(1)Based on the deep learning model on multiple type samples,the paper extracts different model samples from histone modification data,explores the effects of gene distance and data normalization on the results,constructs a feature relationship mapping model modified by histone to chromatin interaction.On the test data under IMR90 cell type,the improved model and normalization method can reach 0.93 in Pearson correlation at best.(2)Proposing a sequence model based on the binding of histone marks with DNA sequence data,and a two-dimensional deconvolution network based on the binding of histone marks correlation matrix with DNA sequence data.Explore DNA sequence affection in prediction process and solve the problem that trained models cannot accept different modes of histone mark input.The final result is that Pearson correlation coefficient index can reach 0.955 and R~2 index can reach 0.892 at best on the test data set under IMR90 cell type.On this basis,the effect of human cell type common histone marks on chromosome interaction prediction is studied,and the migration prediction of different cell types is completely predicted,and the experiment shows that the migration prediction using the proposed model also has good performance,at most PCC can reach 0.95 on average.(3)Proposing the prototype system named Hi CPrediction 1.0 based on histone marks prediction chromatin interaction can be completely analyzed on the entire data prediction process and prediction results of cell types such as IMR90,K562,NHEK,GM12878,etc.,and the forecast scheme can be adjusted according to specific requeirements.