Genome Structural Variation Detection With Deep Neural Network

As a major genetic variation,structural variation is widely found in human genes.Compared to other types of genetic variation,structural variation affects a larger region of the gene and many variations are thought to be associated with disease.The complete and accurate detection of structural variation is of great importance for both theoretical research and clinical applications.Currently,neural networks have achieved excellent performance in many fields,especially in handling complex unstructured data.Considering the complexity of structural variation itself and the unstructured nature of genetic data,this paper investigates the detection of structural variation using a combination of neural networks and long-read sequencing technology.Based on the difference of the extracted information,two different methods are proposed in this paper.(1)In this paper,we propose a method for the detection of deletion in gene fragments,which analyzes the data,extracts the information of deletion and generates the corresponding variation feature matrix.Meanwhile,convolutional neural network is used to map the variant feature matrix into variant feature vectors,and a bidirectional longshort memory network is used to analyze and predict multiple consecutive variant feature vectors.The results show that this method has better performance in detecting deletion than other structural variation detection methods,and it has better sensitivity for detecting deletion in more complex regions.(2)In this paper,we propose a method for the detection and genotype of deletion and insertion.The estimation of variation information is obtained by random sampling of the data.Subsequently,the genetic loci in the data are traversed and calculate the variation information.Finally,the neural network algorithm is used to detect the variation and determine the genotype of the extracted variation information.The results show that this method has better variant detection and genotype performance than other structural variant detection methods.In particular,the method has a greater performance advantage over other methods on more challenging low-coverage data.The method also has good support for multi-threaded accelerated computation,and the method has faster running speed and more significant linear acceleration performance with similar memory consumption as other methods.