Guild-based analysis for understanding gut microbiome in human health and diseases

Abstract

To demonstrate the causative role of gut microbiome in human health and diseases, we first need to identify, via next-generation sequencing, potentially important functional members associated with specific health outcomes and disease phenotypes. However, due to the strain-level genetic complexity of the gut microbiota, microbiome datasets are highly dimensional and highly sparse in nature, making it challenging to identify putative causative agents of a particular disease phenotype. Members of an ecosystem seldomly live independently from each other. Instead, they develop local interactions and form inter-member organizations to influence the ecosystem's higher-level patterns and functions. In the ecological study of macro-organisms, members are defined as belonging to the same "guild" if they exploit the same class of resources in a similar way or work together as a coherent functional group. Translating the concept of "guild" to the study of gut microbiota, we redefine guild as a group of bacteria that show consistent co-abundant behavior and likely to work together to contribute to the same ecological function. In this opinion article, we discuss how to use guilds as the aggregation unit to reduce dimensionality and sparsity in microbiome-wide association studies for identifying candidate gut bacteria that may causatively contribute to human health and diseases.

Keywords: Guild; Gut microbiota; High dimensionality; High sparsity.

Fig. 1

The reduction of dimensionality and sparsity from raw metagenomic dataset to genes, genomes, and guilds. In our PWS example, ~ 2 million non-redundant microbial genes were predicted from the 109 metagenomes. Seventy-nine percent of values in the corresponding abundance matrix of these genes were zeros. These non-redundant microbial genes were further binned into ~ 28,000 draft genomes based on their abundance correlations across the 109 samples. In the corresponding abundance matrix of these draft genomes, 72% of values were zeros. We then selected 161 prevalent bacterial genomes, each with more than 700 bacterial genes and shared by more than 20% of the samples. In the corresponding abundance matrix of these 161 genomes, 52% of values were zeros. Eighteen guilds were identified by clustering these prevalent bacterial genomes. In the corresponding abundance matrix of these 18 guilds, 16% values were zeros

Fig. 2

The taxonomic heterogeneity of guilds identified in the PWS study. a This stacked bar plot shows the phylum assignment of genomes belonging to each guild. b This table presents the distribution of species across guilds. The numbers in the table represent the number of genomes belonging to each species found in each guild. For example, 5 different genomes of the Eubacterium eligens species were found in guild#1, guild#12, and guild#13. A blank entry means that no genome from this species was found in this guild. “Up” denotes the guilds increased after the intervention, while “Down” indicates the guilds decreased after the intervention

Fig. 3

Guild-based aggregation overcomes the pitfall of taxon-based analysis to reflect the variations in strain-specific responses. a and b together illustrate why guild-based aggregation method produces a more accurate representation of strain-level microbiome response to dietary intervention in a PWS study compared with taxon-based aggregation. a shows the abundance change of the 5 Eubacterium eligens strains over time. If taxon-based aggregation is used, all 5 Eubacterium eligens strains could be collapsed into one species-level unit and represented by the black line in b. In contrast, using the guild-based aggregation method, the same 5 Eubacterium eligens strains are grouped into 3 different guilds (#1, #12, and #13). Each of the colored lines in b represents the abundance change over time of one guild. Abundance change pattern of the three guilds in b accurately captures the three types of abundance change patterns among the 5 Eubacterium eligens illustrated in a. The dots on each line in a and b represent the mean abundance (see S.E.M in Supplementary Table 1)

Fig. 4

Comparing taxon-based and guild-based analysis in the PCOS study. a shows that the correlations between clinical parameters and prevalent genera are significantly different among PCOS patients and non-obese controls. The color of spots represents R value of the Spearman correlation between each genus and clinical parameter (+FDR < 0.05, ++FDR < 0.01, +++FDR < 0.001). b and c show the different abundance distributions of Bacteroides genus and 3 Bacteroides OTUs or Alistipes genus and 2 Alistipes OTUs in different patient groups. Bacteroides OTU4 belonged to a guild that was positively correlated with disease phenotype, while Bacteroides OTU7 and Bacteroides OTU63 belonged to a negatively correlated guild. Alistipes OTU200 belonged to a guild that was positively correlated with disease phenotype, and Alistipes OTU130 belonged to a guild that was negatively correlated with disease phenotype. a For leucocyte, neutrocyte, lymphocyte, and hirsutism, n = 46; for the other parameters, n = 48. BMI, body mass index; WHR, waist hip ratio; FSH, follicular stimulating hormone; LH, luteinizing hormone; FPG, fasting plasma glucose; PPG, 2 h postprandial plasma glucose; FINS, fasting plasma insulin; P2hINS, 2 h postprandial plasma insulin; HbA1c, hemoglobin A1c; ALT, alanine aminotransferase; AST, aspartate transaminase; GGT, γ-glutamyltransferase; TCH, total cholesterol; TG, triglyceride; PYY, peptide YY; SDS, self-rating depression scale; SAS, self-rating anxiety scale. b, c CN, non-obese control group (n = 9); CO, obese control group (n = 6); PN, non-obese PCOS group (n = 12); PO, obese PCOS group (n = 21)