Combined use of network inference tools identifies ecologically meaningful bacterial associations in a paddy soil

Abstract

High-throughput sequencing technologies have recently made it possible to interrogate the phylogenetic diversity of soils at considerable depth. This ability has led to the development of many computational tools to infer interaction networks from environmental samples. Although such tools have widely been used, they have more often served as a visual means to compare microbial communities across environmental gradients than as a means to appreciate microbial interactions associated with certain ecological processes. Previous studies have often regarded a subnetwork (module) as a functional unit but its functionality in ecological context has never been evidenced. To make better use of these tools in appreciating microbial interactions, we propose the combinational use of different inference tools. This ensemble approach permits the use of more independent predictors and the removal of tool-specific predictions in order to increase prediction accuracy. The purpose of the present study is to identify ecologically meaningful bacterial associations using multi-tool approach. Soil samples were collected in time series from experimental paddy rice plots. Bacterial communities were characterized by high-throughput tag sequencing of 16S rRNA gene fragments. We used three tools, Co-occurrence Network inference (CoNet), Molecular Ecological Network Analysis (MENA) and extended Local Similarity Analysis (eLSA), to infer networks from abundance profiles, partitioned the networks into modules, screened for the modules with ≥50% of genus-/species-level nodes, captured the modules that were derived from different tools and shared ≥ 50% of order-level nodes (tool-agreed modules) and tested their robustness against the changes in the tool parameters. By these procedures, two three-tool-agreed modules were found. One represented a guild that is phenotypically associated with aerobic respiration and fermentation and the other represented a guild phenotypically associated with metal/sulphur cycles, all of which are essential processes of water-submerged paddy soils that are mediated by bacteria. These data suggested that the linked members in a module were functionally associated taxa that work together to achieve a distinct function or an ecological process, and thus were ecologically meaningful to the environment. We selected three linked species from a three-tool-agreed module and validated their interactions using co-culture methods. Results showed that the interaction type between Janthinobacterium lividum and Leuconostoc lactis in the two-species mixture was validated to be ambivalent, positive for one partner and negative for the other. However, this type of interaction was not retained when a third party Lactococcus piscium was introduced, signifying the complexity of multi-species interactions. Validation results suggested that the selected species were interacting partners in laboratory but the validated interaction types were different from those inferred. By multi-tool approach, we also found that highly linked nodes, which are often referred to as “keystone species” and are frequently interpreted as the species playing important roles in soils, are tool dependent. Among top ten highly linked nodes, only four are conserved across three tools. These results suggest more research is required on the ecological significance of degree-based identification of keystone species. Overall, the present study highlights the potential utility of combined use of inference tools to identify ecologically meaningful bacterial associations in soils and other environmental samples. It is interesting to see what type of ecologically meaningful bacterial associations can be found in other soils.