Fungi stabilize connectivity in the lung and skin microbial ecosystems

Abstract

Background: No microbe exists in isolation, and few live in environments with only members of their own kingdom or domain. As microbiome studies become increasingly more interested in the interactions between microbes than in cataloging which microbes are present, the variety of microbes in the community should be considered. However, the majority of ecological interaction networks for microbiomes built to date have included only bacteria. Joint association inference across multiple domains of life, e.g., fungal communities (the mycobiome) and bacterial communities, has remained largely elusive.

Results: Here, we present a novel extension of the SParse InversE Covariance estimation for Ecological ASsociation Inference (SPIEC-EASI) framework that allows statistical inference of cross-domain associations from targeted amplicon sequencing data. For human lung and skin micro- and mycobiomes, we show that cross-domain networks exhibit higher connectivity, increased network stability, and similar topological re-organization patterns compared to single-domain networks. We also validate in vitro a small number of cross-domain interactions predicted by the skin association network.

Conclusions: For the human lung and skin micro- and mycobiomes, our findings suggest that fungi play a stabilizing role in ecological network organization. Our study suggests that computational efforts to infer association networks that include all forms of microbial life, paired with large-scale culture-based association validation experiments, will help formulate concrete hypotheses about the underlying biological mechanisms of species interactions and, ultimately, help understand microbial communities as a whole.

Fig. 1

Lung microbiome networks. Networks inferred for the lung microbiome based on a bacteria only, b fungi only, and c bacteria and fungi combined. In all three networks, bacterial nodes are circles and fungal nodes are squares. Each node is colored by phyla. Edges between nodes represented a predicted interaction, either positive or negative

Fig. 2

Robustness curves for all networks. Attack robustness of a network was measured by sequentially removing nodes based on the node’s a betweenness, b degree, or c randomly selected and measuring the percentage of nodes that remain in the central connected component. Measurement of robustness was performed for each of our six networks and the results are plotted here with the percentage of nodes removed on the X axis and the percentage of remaining nodes in the central connected component on the Y axis. Each network is represented by a line on this graph. A larger area under the curve indicates a more robust network

Fig. 3

Subnetwork of exclusive OTUs with HIV+, HIV−, COPD+, and COPD− status and their nearest neighbors. Seventeen fungal OTUs were uniquely present in HIV+ individuals while one bacterial and five fungal OTUs were uniquely present in HIV− individuals. Seven fungal OTUs were uniquely present in COPD+ individuals while eight bacterial and ten fungal OTUs uniquely occurred in COPD− individuals. a The 17 HIV+ OTUs and their 51 nearest neighbors OTUs formed a subnetwork with five components. The HIV− subnetwork was comprised of 6 single-status nodes and 32 adjacent neighbors organized in three components. b The 17 COPD− OTUs with its 51 adjacent OTUs formed a large connected component with 64 members and a small four-node component. The seven COPD+ OTUs with 33 adjacent nodes organized into a disconnected six component network

Fig. 4

Lung microbiome modules and HIV infection/COPD status. a Assignment of OTUs into modules of the lung cross-domain bacterial-fungal (CDBF) network. The CDBF network is comprised of six modules. OTUs uniquely appearing in HIV-infected (dark blue), HIV-uninfected (light blue), COPD negative (yellow), or COPD positive (green) samples are found across all modules with strong enrichment in module 5. The unique species names that appeared in module 5 are listed on the right. b Associations in module 5 of the lung CDBF network. The size of the nodes was scaled by the number of neighbors, the thickness of the edges marks association strength

Fig. 5

Skin microbiome networks. Networks inferred for the skin microbiome based on a bacteria only, b fungi only, and c bacteria and fungi combined. In all three networks, bacterial nodes are circles and fungal nodes are squares. Each node is colored by phyla. Edges between nodes represented a predicted interaction, either positive or negative

Fig. 6

Co-variation pattern and growth curves for co-culture validation experiment. a Emericella nidulans (E, green), Propionibacterium acnes (P, pink), and Rothia dentocariosa (R, blue) form a clique in the skin CDBF network (left). The edge weights are average covariations from the estimated covariance matrix (right) between all species assigned to R. dentocariosa (one member), P. acnes (18 members), and E. nidulans (four members). The microbes were grown in pairs and a trio, and the growth curves for the bacteria were compared to when they were grown in monoculture. Cellular concentration growth curves are based on the average of three biological replicates and the vertical lines indicate their standard deviations. While we were able to grow E. nidulans as a monoculture, no growth curves are available because there is no established method for measuring the growth of filamentous fungi in liquid culture. b R. dentocariosa grown with E. nidulans (cyan line) or alone (blue line). c R. dentocariosa grown with P. acnes (purple line) or alone (blue line). d P. acnes grown with E. nidulans (brown line) or alone (pink line). e Trio of all three organisms grown together (grey line) compared to R. dentocariosa alone (blue line) or P. acnes alone (pink line)

Fig. 7

Keystone species analysis. Betweenness centrality vs. node degree of all species in the cross-domain bacterial-fungal networks of lung (a) and skin (b). Nodes with high betweenness centrality represented potential key connector (or bottleneck) species. Nodes with high degree represented hubs in the network. Both measures were indicators for potential keystone species. Bacterial species (dots) and fungal species (squares) were colored by phylum membership. Bacterial and fungal species that were maximal in either property are highlighted in both plots. One fungus in the Davidiellaceae family (top right) may act as potential keystone species in the skin microbiome. In addition, we highlighted the role of Candida parapsilosis across both networks

Fig. 8

Distribution of interaction strengths (partial correlation coefficients) for all six association networks. All distributions consistently showed a peaked distribution with positive mean and skew. The positive edge percentage (PEP) was > 0.8 for all single-domain networks. Both cross-domain networks showed lower PEP (0.76 for the lung CDBF and 0.71 for the skin CDBF)