Compositional data analysis applied to human microbiome network reconstruction

The comprehension of the human gut microbiome has been made possible by technological advances for performing culture-independent analyzes. Next Generation Sequencing techniques produce discrete counts as a result, describing only the relative abundances of each identified bacterial species: such data are of compositional type. Unfortunately the classic methods of analysis on this type of data can lead to completely wrong conclusions: the development of analysis methods for compositional data is still an open issue. The purpose of this work is the description of several analyzes based on compositional data of human gut microbiome. The first result, obtained with t-SNE dimensionality reduction algorithm, is that a different sample clustering was obtained based on the metrics used to define neighborhood. Next ,I applied a biomarker identification method based on the log-ratio variance, a statistical observable used with compositional data, that allowed to identify bacterial species associated to our case/control study design. In the last part I analyzed the problem of the reconstruction of networks of bacterial species. The main objective of the network analysis was to characterize microbiota ecosystem of healthy and infected subjects in our database. Several methods have been proposed to characterize the complex relationships between bacterial populations: in this work I applied the SPIEC-EASI method to reconstruct the correlation structure of the data, and compared the different results obtained. Finally, I propose a new method inspired by Kendall’s Tau correlation, adapted to the peculiarities of compositional data, that provided promising results. This research was made possible thanks to the collaboration with Prof. George Weinstock at the Jackson Laboratory research center (USA). One of the researche aims of JAX is the study and understanding of the gut microbiota for diagnostic purposes (in our case related to infection and diabetes).