Analysis Of Intestinal Flora Based On Spectral Clusterin

Gut microbiome is related to many major human diseases,such as,cirrhosis,diabetes,obesity,autism,etc.and it is of great significance to study the differences in the structure of gut microbiome under different conditions.When traditional clustering methods were applied to identify different structural patterns in gut microbiome data,not only noise sensitivity and long running time,but also they were unable to the inability to process irregular data.Given that spectral clustering can not only cluster arbitrarily shaped sample data and converge to the global optimal solution,it is more adaptable to the data distribution,and the computational effort of spectral clustering is much smaller,and has higher performance.Therefore,this paper adopted the spectral clustering to analyze the structure of intestinal flora,and Minimal hepatic encephalopathy(MHE)and Type 2 diabetes mellitus(T2DM)datasets were taken as example in our study.Due to the physical lack or insufficient sampling of gut microbiome in the sequencing process,microbiome data contained a large number of zeros.Therefore,geometric mean of pairwise ratios(GMPR)was firstly used to normalize the gut microbiome data,then Spectrum method was used to analyze the structure of the gut microbiome,and lastly the structure of core microflora was compared with the network analysis method.It was found that the Spectrum algorithm performed poorly in terms of clustering performance on a single-class population and also had certain drawbacks,so the Spectrum algorithm was improved with the following improvement process.In order to avoid ignoring the weights occupied by the different eigenvalue sizes corresponding to each sample/feature in that sample when calculating the Euclidean distance,a similarity matrix based on feature weighting is proposed.In order to avoid the sensitivity problem of traditional spectral clustering,the Laplacian matrix is replaced by the Hessian matrix and the sample points are clustered by finding the eigenvectors with negative eigenvalues of the Hessian matrix.In order to be able to adjust the number of cluster centers K,the original Kmeans algorithm is replaced by the ISODATA clustering algorithm.The experimental results showed that the Intraclass correlation coefficient(ICC)of GMPR was 0.919,whose reproducibility was significantly better than other normalization methods.And the running-time,Normalized Mutual Information(NMI),the Davies-Boulding Index(DBI),the Calinski-Harabasz index(CH),Rand Index(RI)and Adjusted Rand Index(ARI)of GMPR+Spectrum were far superior to other clustering algorithms such as M3 C,i Cluster Plus.And GMPR+improved Spectrum was compared with improved Spectrum,Spectrum,GMPR+Spectrum,and it was found that the GMPR+improved Spectrum algorithm was far superior to other algorithms in NMI,DBI,CH,RI and ARI.GMPR+Spectrum and GMPR+improved Spectrum algorithms not only perform well in performance(e.g.,NMI was 0.372,DBI was 4.202,CH was4.429,RI was 0.815,and ARI was-0.000 in MHE for GMPR+improved Spectrum.),but could identify structural differences in the gut flora of different types of sick people and uncover the key bacteria of the gut microbiome(e.g.Lactobacillus in patients minimal hepatic encephalopathy,Blautia,Prevotella in patients hepatic encephalopathy),and the identified key bacteria may provide a new reference for the study of the gut microbiome in disease.