In this article, Nick Fountain-Jones from the University of Tasmania introduces how advances in Bayesian networks can be used to untangle community dynamics and, in particular, the moose microbiome by telling us the #StoryBehindThePaper.
Microbial communities are inherently complex systems with potentially hundreds of millions of interacting species. Every surface of the body is occupied by a diverse set of microbes; interactions between them, mediated by host nutrition and immune function, can determine if outcomes for the host are beneficial or damaging. For example, competition between microbes in the gut determines whether cancer immunotherapies are successful and whether they produce side-effects for the host. Understanding these interactions is important for understanding microbial communities, but identifying how interactions will play out amongst hundreds of millions of species is a unique analytical challenge. Identifying potential interactions, such as competition, out of the vast numbers of interactions plausible offers a unique analytical challenge particularly from real-world observational data. Adding further complexity, some interactions that observationally appear to be competition may actually reflect simple spatial separation or niche partitioning. These interactions not only occur within-hosts but between-hosts and across landscapes; microbes may not co-occur because they actively compete or because they or their hosts inhabit different niches. Understanding these drivers is critical for predicting microbial distributions.
Recently I led a multidisciplinary team that developed an approach that utilizes Markov Random Fields (MRF) to quantify associations between species whilst controlling for spatial artefacts and host factors (e.g., sex or host health) that could influence these relationships. Developed by statistical physicists, MRF considers interacting species as a type of Bayesian network. Note that for these types of analyses ‘associations’ are a more appropriate term than ‘interactions’ as they represent only potential hypotheses for how species interact that need to be further tested (see Blanchet et al 2020). One advantage of the MRF approach over other multivariate regression methods is that co-occurrence probabilities of species can be dynamic and associated with environmental factors. For example, MRF models can capture if species are associated with each other only under certain environmental conditions. Whilst increasingly employed by community ecologists (Clark et al, 2018), MRFs have very rarely been employed to understand microbial communities.
We applied this pipeline to examine how gut microbial communities are structured in a North American moose (Alces alces) population in Minnesota. This moose population has experienced a 55% decline and approximately 68% of deaths were health related, but the role of the gut microbiome was unknown. In general, our understanding of the connections between gut microbial communities and health for wildlife species is poor. Our study not only provided new insights into the types of microbes occupy the gut of moose in this region but also how these microbial communities assemble and correlate with moose health. We sampled the gut microbiome from 52 live-captured moose in which the body condition, location, sex, age and pregnancy status was known. Each of the live-captured moose was anesthetized via a dart from a helicopter and faecal samples were taken to a laboratory for 16S metagenomic sequencing. Serum samples taken from the same individuals were screened for common moose pathogens. Together, these data provided a unique opportunity to apply our machine-learning pipeline to untangle if moose microbial communities could be predicted using host characteristics (e.g. body condition and pathogen exposure), spatial patterns or microbial interactions.
Surprisingly, microbial interactions were the dominant process driving gut microbiome composition; we could predict the distribution of a microbial species with a remarkable 83% accuracy based on the presence or absence of other microbes alone. Incorporating host spatial proximity further improved the predictive performance of our models to 90%, indicating that dispersal opportunities and host spatial organization also plays a role in shaping these microbial communities. In contrast, when we removed microbial interactions from the model, our host and spatial data explained only 67% of microbial occurrences. Our results point to moose gut microbial communities being remarkably resilient to changes in host health and similar results have been shown for other herbivores. In essence, our study demonstrates that microbes were predicting microbes; the impact of the host on the microbes was less important in determining the organization of moose microbial communities. Moreover, the nature of these interactions was more commonly positive than negative, highlighting that processes such as facilitation, rather than competition, are important in shaping moose microbial communities. Interestingly, not all groups of bacteria were involved in these interactions; for example, we found that members of the Bacteroidetes taxa did not have significant interactions with other microbes. In contrast, the Mollicutes (soft skin bacteria), whilst relatively rare in our moose samples, were network hubs involved in many positive interactions. Our Markov random field pipeline enabled us to identify these types of interactions from observational data that can form the basis for future laboratory investigations. Whilst we applied this pipeline to microbial data, this type of approach could provide powerful insights into any community of organisms.
The links below by Nick Clark can help get you started on applying MRF:
One thought on “Untangling community dynamics using spatially explicit Bayesian networks”
Pingback: Graphical network models and moose microbes – Evolutionary disease ecology