Dynamic Bayesian Networks for Integrating Multi-omics Time Series Microbiome Data

Abstract

A key challenge in the analysis of longitudinal microbiome data is the inference of temporal interactions between microbial taxa, their genes, the metabolites that they consume and produce, and host genes. To address these challenges, we developed a computational pipeline, a pipeline for the analysis of longitudinal multi-omics data (PALM), that first aligns multi-omics data and then uses dynamic Bayesian networks (DBNs) to reconstruct a unified model. Our approach overcomes differences in sampling and progression rates, utilizes a biologically inspired multi-omic framework, reduces the large number of entities and parameters in the DBNs, and validates the learned network. Applying PALM to data collected from inflammatory bowel disease patients, we show that it accurately identifies known and novel interactions. Targeted experimental validations further support a number of the predicted novel metabolite-taxon interactions. IMPORTANCE While a number of large consortia collect and profile several different types of microbiome and genomic time series data, very few methods exist for joint modeling of multi-omics data sets. We developed a new computational pipeline, PALM, which uses dynamic Bayesian networks (DBNs) and is designed to integrate multi-omics data from longitudinal microbiome studies. When used to integrate sequence, expression, and metabolomics data from microbiome samples along with host expression data, the resulting models identify interactions between taxa, their genes, and the metabolites that they produce and consume, as well as their impact on host expression. We tested the models both by using them to predict future changes in microbiome levels and by comparing the learned interactions to known interactions in the literature. Finally, we performed experimental validations for a few of the predicted interactions to demonstrate the ability of the method to identify novel relationships and their impact.

Challenges, Strategies, and Perspectives for Reference-Independent Longitudinal Multi-Omic Microbiome Studies

“…Unfortunately, correlations may be insufficient to assess complex community interactions, whereby the application of modeling approaches would be necessary to resolve those relationships ( Fisher and Mehta, 2014 ; Trosvik et al, 2015 ; Ridenhour et al, 2017 ). Modeling could serve as a means of integrating several layers of omic data ( Lloyd-Price et al, 2019 ; Ruiz-Perez et al, 2021 ) further elucidating microbial interplay beyond species abundances and functional potential.…”

Discovering gene regulatory networks of multiple phenotypic groups using dynamic Bayesian networks

“…For this reason, we suggest learning a range of models based on different assumptions regarding network structure and parameters and likewise select the model for downstream analysis based on cross-validated predictive accuracy. We use mean absolute error (MAE) as a measure of predictive accuracy, which was already used in previous DBN application studies [24,18]. To assess how well different DBN models describe the underlying dynamic process, we predicted the values of all genes in all times slices using the true values of expression of genes at time point t = 0 using leave-one-out CV.…”

“…The Dynamic Bayesian Network (DBN) model overcomes this problem by including dependencies between nodes at different time points and accommodating the possibility of cycles [17]. DBNs were previously used for inference of biological networks [18], including GRNs [19,20,21,22,23] and multi-omics networks [24]. However, none of these studies considered non-homogeneous datasets where DBN structures may differ between groups of samples.…”

“…Structure learning of DBNs is a computationally challenging problem. Existing approaches are either not scalable [25,23,26], use greedy search [20] or restrict the DBN topology to decrease computational costs and accommodate larger networks [18,24,26,27]. However, any restriction may potentially result in the discovery of suboptimal models [28].…”

“…However, these simulations did not compare consensus models obtained via sampling from the posterior and bootstrapping in the case of greedy hillclimbing. In addition, limiting the parent set size was used as a way to prevent overfitting in DBNs [18,24] and to decrease computational costs [26,27]. However, the impact of uniform restrictions of the parent set sizes on the structure fit and predictive accuracy in large networks in the high dimensional setting was never explored.…”