Inferring Intra-Community Microbial Interaction Patterns from Metagenomic Datasets Using Associative Rule Mining Techniques

Abstract

The nature of inter-microbial metabolic interactions defines the stability of microbial communities residing in any ecological niche. Deciphering these interaction patterns is crucial for understanding the mode/mechanism(s) through which an individual microbial community transitions from one state to another (e.g. from a healthy to a diseased state). Statistical correlation techniques have been traditionally employed for mining microbial interaction patterns from taxonomic abundance data corresponding to a given microbial community. In spite of their efficiency, these correlation techniques can capture only 'pair-wise interactions'. Moreover, their emphasis on statistical significance can potentially result in missing out on several interactions that are relevant from a biological standpoint. This study explores the applicability of one of the earliest association rule mining algorithm i.e. the 'Apriori algorithm' for deriving 'microbial association rules' from the taxonomic profile of given microbial community. The classical Apriori approach derives association rules by analysing patterns of co-occurrence/co-exclusion between various '(subsets of) features/items' across various samples. Using real-world microbiome data, the efficiency/utility of this rule mining approach in deciphering multiple (biologically meaningful) association patterns between 'subsets/subgroups' of microbes (constituting microbiome samples) is demonstrated. As an example, association rules derived from publicly available gut microbiome datasets indicate an association between a group of microbes (Faecalibacterium, Dorea, and Blautia) that are known to have mutualistic metabolic associations among themselves. Application of the rule mining approach on gut microbiomes (sourced from the Human Microbiome Project) further indicated similar microbial association patterns in gut microbiomes irrespective of the gender of the subjects. A Linux implementation of the Association Rule Mining (ARM) software (customised for deriving 'microbial association rules' from microbiome data) is freely available for download from the following link: http://metagenomics.atc.tcs.com/arm.

Open challenges for microbial network construction and analysis

“…In this case, the interaction between A and B or A and C is nonexistent until the arrival of the third species, which can be seen as an extreme case of interaction modification. Although a few association rules involving more than two microbial taxa have been reported previously [31], it is not clear whether these are due to overfitting (a challenge for all HOI inference algorithms), environmental factors, combinations of pair-wise associations, or true HOIs. Finally, visualizing HOIs is not trivial and requires hypergraphs, i.e., networks where an edge connects more than two nodes.…”

Soil actinobacteria tend to have neutral interactions with other co‐occurring microorganisms, especially under oligotrophic conditions

“…While the interaction patterns in microbial communities can be predicted using metagenomic datasets and mathematical simulations (Freilich et al ., 2011; Tandon et al ., 2016), studying cultivable microbial interactions can provide direct insights into complex communities. By comparing the productivity of species mixtures to that of each species alone, Foster and Bell found that competition, rather than cooperation, is prevalent among culturable bacterial species isolated from aquatic environments (Foster and Bell, 2012).…”

MITRE: inferring features from microbiota time-series data linked to host status

“…Bayesian approaches are powerful, because they provide principled estimates of uncertainty throughout the model, which is an especially important feature in biomedical applications where noisy inputs are the norm. We note that another rule-based method, association rule mining (ARM), has recently been applied to analyzing microbiome data in a different context (finding interaction patterns among OTUs) [14]. Although ARM has some commonalities with Bayesian rule learning approaches, ARM methods tend to employ user-based cutoffs and heuristics, rather than principled probabilistic methods, as their primary function is to mine large databases for putative interactions, rather than build predictive models.…”