基于信息融合与图神经网络的单细胞测序数据插补模型OA

To address the problem of low gene expression counts and missing events faced by single-cell sequencing technologies,the paper proposes the model scFGImpute,a graph-convolution autoencoder that utilizes integrated distance and depth information fusion.Firstly,an integrated distance is used to construct a similarity network between cells,avoiding the bias and uncertainty of a single distance metric.Secondly,a graph attention convolutional structure aggregates similarity information from multi-layer neighboring cells,fully leveraging high-dimensional gene features and high-order cellular topology.Through a deep in-formation fusion network,a more complete consensus representation is generated.Experimental results demonstrate that compared to competing methods,scFGImpute can effectively utilize the original gene fea-ture information and the topological relationships between cells to uncover gene expression patterns hidden by noise from both global and local perspectives.It improves missing events,imputes excessive zero values,and reduces noise effects while maintaining robust and stable performance.