| The rapid development of single cell transcriptome sequencing technology and the emergence of a large number of analytical tools have enabled researchers to analyze transcriptome sequencing data at a single cell level,thereby better revealing the heterogeneity between cells.We can use cell clustering algorithms to gather similar cells and divide them into different types to obtain relevant information about cell types.Finally,we can use trajectory inference algorithms to obtain cell differentiation trajectories to better explain the development process of organisms and the occurrence of diseases.This thesis proposes a single cell trajectory inference method based on graph automatic encoders,which combines the powerful node embedding and representation capabilities of graph self coding neural networks with the powerful feature aggregation and feature extraction capabilities of graph convolutional neural networks,enabling effective feature extraction and cell clustering of high-dimensional single cell transcriptome sequencing data,providing a good foundation for downstream analysis trajectory inference.In this thesis,a graph convolutional neural network is embedded in the encoder of a graph automatic encoder.In the graph convolutional neural network,by defining compatibility classes,neighborhood similarity,and the similarity between neighboring nodes and central nodes that share close neighbors,it is determined whether the next layer of aggregation is necessary,thereby constructing a graph automatic encoder with an adaptive aggregation mechanism.Finally,the cross entropy is used as a loss function,By making the reconstructed matrix of the decoder as similar as possible to the original matrix,we can obtain hidden variables with high dimensional information and low dimensional representation.Then,hidden variables are used to cluster and construct cell trajectories.The clustering results and trajectory inference results of fourteen real unit cell transcriptome datasets show that the model proposed in this thesis can accurately extract important features and aggregate them to obtain a relatively compact cell cluster,thereby obtaining a clearer cell development trajectory,which demonstrates the effectiveness of the model. |
