MIND: a database and web platform for microbial interaction networks

MIND, the Microbial Interaction Network Database, stores microbial interaction and correlation networks from many studies under one schema, keyed on NCBI taxonomy IDs, and provides a browser for querying, comparison, visualisation and export. It came out of the Segrè lab and the Bioinformatics Program at Boston University. Zhenjun Hu designed its architecture and implemented the platform with Yan Wang.

Two networks from different studies cannot be compared directly

Most published microbe-microbe networks are correlation networks, in which an edge means that the presence of one taxon in a sample indicates a raised chance of the presence or absence of another. They are inferred from 16S rRNA amplicon or metagenomic data using a growing and disagreeing set of tools. Placing two of them side by side fails before the comparison starts. The nodes are OTUs defined against whatever reference and clustering threshold that study used, so a node in one network has no reliable counterpart in another, and the two may not even sit at the same taxonomic rank. The datasets are also scattered across different servers and formats, and some exist only as a supplementary table attached to a paper.

There is a second and more heterogeneous body of data on actual interdependencies: Lotka-Volterra coefficients fitted to longitudinal data, pairwise growth measurements, horizontal gene transfer estimated from genomes, and genome-scale model predictions of competition and cross-feeding. None of these share a representation with the correlation networks or with each other. At the time of writing we found no tool that could represent, compare and integrate diverse microbial interaction types in one place.

Every network is stored with the conditions that produced it

A network in MIND is a context, meaning the edges together with the metadata describing the conditions under which they were produced. Each context receives a unique ID such as C0310 or C2000. The organisation is a tree.

Branch	Holds
Metadata	Publication, host, experimental metadata, condition, computational metadata, interaction type, location
Network data	Nodes and edges, each carrying separate visual properties and biological properties

Taxonomy IDs and the level-up function let networks cross ranks

Microbes are identified by NCBI taxonomy ID. The taxonomy hierarchy is integrated into the database and synchronised periodically, which is what makes rank-crossing operations possible. Curation converts OTUs to taxonomy IDs, and the level-up function projects nodes at a given rank up to a higher one. Nodes below the target rank merge, nodes already at or above it are left alone, edges within a merged group are dropped, and edges leaving it become edges between the coarser nodes.

The stack is a three-tier service-oriented architecture inherited from VisANT-Predictome.

Tier	Implementation
Database	PostgreSQL
Services	J2EE web services behind an enterprise service bus, hosted on Apache Tomcat
Client	Single-page application: AngularJS, Bootstrap, d3

The default data has two origins. Some networks were curated by hand from published studies, including correlation networks, lateral gene transfer networks, metabolic influence networks, and networks derived from genome-scale models via COMETS. The rest were generated by running MiCoNE with its recommended settings over published 16S rRNA sequence data. The browser searches microbes and contexts, overlays networks, filters by taxonomy and study metadata, and exports the resulting views. Users can also upload their own networks.

Four IBD networks collapse into one module once rank is controlled

Four inflammatory bowel disease networks (C0310-C0313), from a Swiss cohort and a Bern gastroenterology clinics cohort, load together and lay out as two separate modules. Filtering out phylum-rank nodes and keeping only links present in all four contexts collapses the remainder into a single module. All four contexts used Pearson correlation, so one global weight cutoff of 0.5 is meaningful across them and leaves the strong, reproducible correlations.

In bacterial vaginosis, most of the disagreement was rank mismatch

Study 1 is a 16S co-occurrence network over 220 women using Pearson correlation, with 43 nodes and 900 edges. Study 2 is a qPCR and PCR network over 177 women using Spearman correlation, with 16 nodes and 120 edges. Merged as published they look inconsistent: 23% of study 1’s nodes and 6% of its edges overlap, against 69% of nodes and 34% of edges for study 2. The overlapping edges agree well, and the sign of the correlation is highly conserved between the studies. Levelling up to genus leaves only two nodes specific to study 2 (Ureaplasma and Mobiluncus). At order there is one node specific to study 2 and three to study 1, and at phylum study 2 is a subset of study 1.

In colorectal cancer, module M1 disappears and M2 splits

Reanalysing gut mucosal microbiome data across colorectal carcinogenesis, the normal mucosa network has more nodes (43) and edges (132) than the adenoma (38 nodes, 62 edges) or the carcinoma (37 nodes, 78 edges). The normal network has three modules, M1, M2 and M3, of 6, 10 and 23 nodes. M1 is down to two members in adenoma and absent in carcinoma. M2 keeps its core in adenoma and splits in two in carcinoma, gaining Fusobacterium and Prevotella intermedia. M3 is the most stable.

Overlaying two curated oral biofilm networks (C0007 and C0008) yields only two shared nodes, Prevotella intermedia and Haemophilus parainfluenzae, both of which sit in M2. The small overlap is itself a rank artefact: every node in the oral networks is at species rank, while more than half the colorectal cancer nodes are genus or higher.

Networks are stored as published, and integrated on the client

Curated networks are stored as published, apart from converting OTUs to NCBI taxonomy IDs. Integrating, merging or otherwise modifying networks happens on the client, at the user’s choice. Baking a particular integration into the stored representation would make the data less useful to anyone asking a different question than the curator anticipated.

We examined PSI-MI, BioPAX and SBML for the interaction data and MIGS for the metagenomic metadata, and wrote a simplified JSON format instead that carries network and metadata together. The existing standards were built for molecular interactions rather than for a network plus the study conditions that produced it. The provenance fields are deliberate. Each context records a PubMed ID, a publication date, and the email of the corresponding author or submitter, so a user with a question about a network can contact the person who made it.

Citation

Hu, Z., Kishore, D., Wang, Y., Birzu, G., DeLisi, C., Korolev, K. S., & Segrè, D. (2022). A resource for the comparison and integration of heterogeneous microbiome networks. bioRxiv 2022.08.07.503059.