Advances in Microbiome Analysis: From the Variance Component Model to Deep Learning

Abstract

Evidence linking microbiome to human health is rapidly growing, suggesting that the human microbiome may serve as novel biomarkers for disease. In microbiome profiling, the direct sequencing outputs are counts or compositions of bacterial taxa at different taxonomic levels. The key research questions in microbiome analysis are to identify the bacterial taxa associated with a clinical outcome and to leverage microbiome aiming to make an accurate prediction on host phenotypes. Although the next generation sequencing has produced extensive microbiome data, data analyses are hindered by several statistical challenges due to the unique characteristics of microbiome profile which includes that 1) the number of taxa greatly exceeds the sample size, 2) most taxa are in extremely low abundance and absent in many samples, and 3) the taxa are related to one another by an evolutionary tree. In this dissertation, three papers are presented to address these challenges. In the first paper, a regularized variance component model is developed for selecting important microbiome taxa. We consider regression analysis by treating bacterial taxa at different levels as multiple random effects. For each taxon, a kernel matrix is calculated based on distance measures in the phylogenetic tree and it acts as one variance component in the joint model. Then, taxonomic selection is achieved by the lasso (least absolute shrinkage and selection operator) penalty on variance components. Our method integrates biological information into the variable selection problem and greatly improves selection accuracies. Simulation studies demonstrate the superiority of our methods versus existing ones, for example, the group-lasso. This method is then applied to a longitudinal microbiome study of Human Immunodeficiency Virus (HIV) infected patients. We implement our method using the high performance computing language Julia. Software and detailed documentation are freely available at https://github.com/JingZhai63/VCselection. In the second paper, we thoroughly investigate the link between lung microbiome composition, pulmonary inflammation, and early lung dysfunction. First, the genus-level taxa are labeled as two pneumotypes: supraglottic and background predominant taxa (i.e., SPT and BPT as previously described by others). Next, multiple statistical approaches are used to characterize these two pneumotype taxa, including dissimilarity-overlap curve (DOC) microbiome dynamics analysis, weighted Spearman correlation analysis, network analysis, etc. We find that previously defined microbial taxa have different effect on inflammation markers and lung function in an HIV positive population. The dynamics of post-ART pneumotype SPT was host-dependent, indicating that the microbiome variability among individuals not only originates from the difference in microbiome species assemblage, but also stems from host specific factors, such as lifestyles. The complex lung function-microbiome-inflammatory network and microbiome dynamics pattern suggest that a healthy lung microbiome may play a critical role in preventing lung function decline of HIV infected individuals. The findings in the second paper suggest that, in addition to identify taxa related to disease, it is also important to uncover the microbiome-phenotype network by understanding microbiome as a whole. In the third paper, we present DeepBiome, a deep learning model, to uncover the network of microbiome and visualize its path to disease. The proposed DeepBiome takes microbiome abundance data as input and uses the phylogenetic tree as a prior knowledge to decide the optimal number of layers and neurons on each of it. By doing so, we are able to relieve the computation burden of tuning DNN hyperparameters. In addition, DeepBiome provides phylogeny regularized weight decay to improve prediction on host phenotypes during training. This deep learning framework not only can analyze a microbiome as a whole to provide a comprehensive network view and but also can identify taxa associate with outcome at each taxonomic level. DeepBiome is designed for both regression and classification problems to support a broader application in microbiome analysis. The simulation studies and real data application show that DeepBiome is a cutting-edge tool in the area of complex microbiome data analysis.