Microbiome Data Analysis by Symmetric Non-negative Matrix Factorization With Local and Global Regularization.

Junmin Zhao,Yuanyuan Ma,Lifang Liu

doi:10.3389/fmolb.2021.643014

Abstract

A network is an efficient tool to organize complicated data. The Laplacian graph has attracted more and more attention for its good properties and has been applied to many tasks including clustering, feature selection, and so on. Recently, studies have indicated that though the Laplacian graph can capture the global information of data, it lacks the power to capture fine-grained structure inherent in network. In contrast, a Vicus matrix can make full use of local topological information from the data. Given this consideration, in this paper we simultaneously introduce Laplacian and Vicus graphs into a symmetric non-negative matrix factorization framework (LVSNMF) to seek and exploit the global and local structure patterns that inherent in the original data. Extensive experiments are conducted on three real datasets (cancer, cell populations, and microbiome data). The experimental results show the proposed LVSNMF algorithm significantly outperforms other competing algorithms, suggesting its potential in biological data analysis.

Highlights

With the development of high-throughput metagenomic sequencing and 16S sequencing technologies, more and more biological data have been accumulated
In view of the above considerations, in this paper we introduce Laplacian and Vicus matrices (Wang et al, 2017) to simultaneously model the global and local structure connections residing within the data and compare their performance with the methods only based on Laplacian or Vicus graphs on several real datasets
We can see that LVSNMF outperforms the second best algorithm at 0.49/1.93% points in terms of AC/normalized mutual information (NMI) on the Lung dataset, 3.22/2.17% points on the Pollen dataset, and the 2.09/0.61% points on Human Microbiome Plan (HMP) dataset

Summary

Introduction

With the development of high-throughput metagenomic sequencing and 16S sequencing technologies, more and more biological data have been accumulated. In order to reach a good understanding of the roles that the microbiome plays in the health and disease states of humans, many plans, including the Human Microbiome Plan (HMP) (Turnbaugh et al, 2007), integrative Human Microbiome Plan (iHMP) (The Integrative Hmp (iHMP) Research Network Consortium, 2019), and the Metagenomics of Human Intestinal Tract (MetaHIT) (Qin et al, 2010), have been launched. These actions pave the way for researchers to further explore the complex relationships residing in microbiome data. Clustering approaches and similarity measurements were applied to elucidate the influence that various factors impose on the identification of enterotypes (Koren et al, 2013). Jiang et al (2012) proposed a new approach based on nonnegative matrix factorization (NMF) to identify the structure and functions of complex microbial

Methods

Results

Conclusion