基于多尺度图对比学习的空间转录组聚类方法  

A Spatial Transcriptome Clustering Method Based on Multi-Scale Map Contrast Learning

作　　者：阳龙 彭利红 周立前 YANG Long;PENG Lihong;ZHOU Liqian(School of Computer,Hunan University of Technology,Zhuzhou Hunan 412007,China)

机构地区：[1]湖南工业大学计算机学院,湖南株洲412007

出　　处：《湖南工业大学学报》2025年第5期52-57,共6页Journal of Hunan University of Technology

基　　金：国家自然科学基金资助项目(62072172)。

摘　　要：针对图神经网络在空间转录组数据聚类过程中识别的空间域不连续或存在交叉这一问题,提出一种基于多尺度图对比学习的空间转录组聚类方法mcmlST。首先,使用SCANPY和主成分分析法,对空间转录组数据进行预处理。然后对ST数据进行增强处理,形成新的视图。接着,基于图自编码器和辅助自编码器,设计双重编码结构,学习空间转录组数据的嵌入特征。最后,基于嵌入特征,使用k-means算法识别空间转录组数据中的空间域。在3个经典的空间转录组数据集(人类大脑皮层右侧背外侧前额叶皮层、人类乳腺癌Block A Section 1和STARmap)上,该方法与3个基线方法conST、CCST和DeepST相比,计算得到更高的ARI和NMI,表明该方法具有优越的空间转录组聚类性能。In view of the flaw of discontinuous or intersecting spatial domains identified by graph neural networks in the clustering process of spatial transcriptome data,a spatial transcriptome clustering method mcmlST,which is based on multi-scale graph contrastive learning,has thus been proposed.Firstly,the spatial transcriptome data is preprocessed by using SCANPY and principal component analysis,followed by an enhancement of the ST data to form a new view.Next,based on graph autoencoders and auxiliary autoencoders,a dual encoding structure is designed to learn the embedded features of spatial transcriptome data.Finally,the k-means algorithm is used for an identification of spatial domains in spatial transcriptome data on the basis of embedded features.On three classic spatial transcriptome datasets(right dorso lateral prefrontal cortex,human breast cancer Block A Section 1 and STARmap),the proposed method calculates higher ARI and NMI compared with the three baseline methods conST,CCST,and DeepST,indicating a superior spatial transcriptome clustering performance.

关 键 词：多尺度学习 多头注意力 对比图聚类 深度学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]