Firmicutes

Ideally, our positive data set should consist of a large number of proteins secreted via non-classical pathways. Unfortunately, it was not possible to obtain a sufficiently large data set as only a small number of proteins undergoing non-classical secretion are known. Since we are looking for features shared among extracellular proteins, the mechanism by which a protein is secreted should not be important. We therefore used for training the large number of proteins known to be secreted via the classical Sec-dependent secretion mediated mechanism. All sequence data was extracted from Swiss-Prot release 44.0. Two individual training sets were created for Firmicutes and Proteobacteria, respectively.
A set of 690 extracellular proteins from Firmicutes (Gram-positive) and a set of 2185 extracellular proteins from Proteobacteria (Gram-negative) were extracted from the Swiss-Prot database based on annotations in the feature table (FT) and comments line (CC) [52 (link)]. Partial sequences were excluded from the data set. As we wanted to train a predictor that works in the absence of signal peptides, the signal peptide part of each sequence was removed according to the Swiss-Prot annotation. These lists of secreted proteins formed our positive data sets. Negative training sets were constructed by extracting 1084 proteins for Firmicutes and 2098 proteins for Proteobacteria from Swiss-Prot, which were annotated as localised to the cytoplasm. After redundancy reduction of the data sets based on a structural similarity criteria [53 (link)], 152 and 350 extracellular sequences were left in the positive data sets for Firmicutes and Proteobacteria, respectively. In the negative data sets, 140 and 334 sequences remained for Firmicutes and Proteobacteria, respectively. For Gram-positive bacteria (Firmicutes and Actinobacteria) a set of non-classically secreted proteins was retrieved from Swiss-Prot based on literature searches (see Table 1).
All data sets used are available as supplementary information from our website [37 ].
For identification of putative non-classically secreted proteins in E. coli and B. subtilis, we used the following accession numbers to extract the annotated and translated proteomes: [Genbank:NC_000913] for E. coli and [Genbank:NC_000964] for B. subtilis.

We collected 1562 virus RefSeq genomes infecting prokaryotes and 31,986 prokaryotic host RefSeq genomes from NCBI in May 2015. The NCBI accession numbers of the RefSeq sequences are provided in the Additional file 2: Table S2. To mimic fragmented metagenomic sequences, for a given length L = 500, 1000, 3000, 5000, and 10000 bp, viruses were split into non-overlapping fragments of length L and the same number of non-overlapping fragments of length L were randomly subsampled from the prokaryotic genomes. Fragments were generated for virus genomes discovered before 1 January 2014 and after 1 January 2014 and were separately used as training and testing sets, respectively (Table 1). To generate evaluation datasets containing 10, 50, and 90% viral contigs, the number of viral contigs was set as in Table 1 and was combined with 9 times more, equal numbers, or 9-fold less randomly sampled host contigs, respectively.
Highly represented host phyla (Actinobacteria, Cyanobacteria, Firmicutes, Proteobacteria) and genera (Mycobacterium, Escherichia, Pseudomonas, Staphylococcus, Bacillus, Vibrio, and Streptococcus) were selected for the analyses where viruses infecting these taxa were excluded from the training of VirFinder. For evaluation of the different trained VirFinder models, equal numbers of contigs of the excluded viruses and all other viruses were selected and then combined with randomly selected host contigs such that total virus and host contigs were equal in number.
For the analysis of VirFinder trained with 14,722 prokaryotic genomes with or without proviruses removed, these genomes were downloaded from the database cited in [6 (link)]. Likewise, the positions of proviruses predicted by VirSorter in these 14,722 genomes were obtained from the published data of [6 (link)] and were used to remove theses sequence from their corresponding host genomes.

CoNet offers a series of features that distinguish it from other network inference tools, such as its support for object groups. This feature allows a user to assign objects to different groups (
e.g. metabolites and enzymes). Relationships can then be inferred only between different object types (resulting in a bipartite network) or only within the same object type. CoNet's treatment of two input matrices is built upon this feature.
Furthermore, CoNet can handle row metadata, which allows for instance to infer links between objects at different hierarchical levels (
e.g. between order Lactobacillales and genus Ureaplasma) while preventing links between different levels of the same hierarchy (e.g. Lactobacillales and Lactobacillaceae). CoNet can also read in sample metadata such as temperature or oxygen concentration. When sample metadata are provided, associations among metadata items and between taxa and metadata items are inferred in addition to the taxon associations. Metadata are then represented as additional nodes in the resulting network. In addition, CoNet recognizes abundance tables generated from biom files (
McDonald
et al., 2012) and, in its Cytoscape 3.× version, reads biom files in HDF5 format directly, using the BiomIO Java library (
Ladau ). Taxonomic lineages in biom files or biom-derived tables are automatically parsed and displayed as node attributes of the resulting network. For instance, the lineage "k__Bacteria; p__Firmicutes; c__Bacilli; o__Lactobacillales; f__Lactobacillaceae; g__Lactobacillus; s_Lactobacillus acidophilus" of an operating taxonomic unit with identifier 12 would create a kingdom, phylum, class, order, family, genus and species attribute in the node property table for node OTU-12, filled with the corresponding values from the lineage. CoNet also computes a node's total edge number as well as the number of positive and negative edges, the total row sum and the number of samples in which the object was observed (e.g. was different from zero or a missing value).
To ease the selection of suitable preprocessing steps, CoNet can display input matrix properties and recommendations based on them. Importantly, CoNet can also handle missing values, by omitting sample pairs with missing values from the association strength calculation. Finally, CoNet supports a few input and output network formats absent in Cytoscape, including adjacency matrices (import), dot (the format of GraphViz (
http://www.graphviz.org/)) and VisML (VisANT's format (
Hu
et al., 2013)) (both for export).

We combined prokaryotic genome sequences from two resources to create a panel of reference genomes across which we determined the phylogenetic and taxonomic distribution of ARGs. The first data set consisted of 154,723 metagenome-assembled genomes reconstructed from the same human microbiome samples that we obtained metagenome assemblies for^{26 (link)}. Of these reconstructed genomes 70,178 were labeled as ‘high-quality’ in the original study, based on >90% completeness and < 0.5% strain heterogeneity. The second data set consists of 152,497 bacterial and archaeal genomes from NCBI RefSeq accessed on 19 April 2019. These genomes included representatives from the principal phyla found in the human gut microbiome although were dominated by Proteobacteria (Proteobacteria: 83,445; Firmicutes: 44,484; Actinobacteria: 16,529 and Bacteroidetes: 3563, Others: 3634).
Genome sequences from the two sources were clustered into species-level bins (SGBs) based on 5% average nucleotide identity (ANI) radius according to the method described in Pasolli et al. (2019). The list of reconstructed genomes used in this study and their mapping to SGBs and full-rank taxonomy are provided in Supplementary Data 3. The list of RefSeq genome accession numbers used in this study and their mapping to SGBs and full-rank taxonomy are provided in Supplementary Data 4.

Statistical and bioinformatics analysis of 16S rRNA gene sequencing. The within-community diversity (α-diversity), including Chao1 richness, Shannon diversity index, and Simpson index, were detected using the ASV table in QIIME. ASV-level abundance curves were generated to compare the richness and homogeneity of ASVs in samples. Changes in community composition (β-diversity) was performed by the principal coordinates analysis (PCoA), weighted and unweighted UniFrac NMDS analysis, as well as hierarchical clustering analysis based on weighted UniFrac distances. They were performed by QIIME and R software. The Gram-positive/Gram-negative ratio was calculated by dividing the number of Actinobacteria and Firmicutes sequences by the number of Gram-negative bacteria (here defined as those phyla containing Hsp60 inserts, which are considered to most closely reflect the traditional concept of gram-negative bacteria) (Sutcliffe, 2010 (link); Gupta, 2011 (link)). The relative abundance of Gram-positive and Gram-negative bacteria was also calculated by dividing each group by the total number of sequences found in each group. The linear discriminant analysis (LDA) and LDA effect size (LEfSe) methods were used to detect taxa of different abundant in each group. The Spearman correlation was used for correlation analysis, and the ASV table was centered and log-transformed before calculating the correlation. Use the ggplot package for visualization and plotting. Functional prediction of the intestinal microbiota was performed on Phylogenetic Investigation of Communities by Reconstruction of Unobserved States (PICRUSt 2) (Douglas et al., 2020 (link)).
All the bar and box plots were generated using GraphPad Prism version 8.0 (GraphPad Software, USA). All data were expressed as mean ± SD. One-way ANOVA was used for statistical analysis, followed by the Turkey test for multiple comparisons (*p < 0.05, **p < 0.01).

We also tested the sensitivity of the result that accessory gene contents differed by environment to sampling biases of the plasmid database by repeating our main analyses on subsets of the data. To do this, we first calculated the taxonomic distribution of bacterial hosts in the data set using the R package ggsankey (https://github.com/davidsjoberg/ggsankey). Next, we investigated patterns within the most abundant phyla, as explained above. We then examined trends within E. coli, the most abundant species represented. Finally, we removed the three most abundant genera from Proteobacteria (Acinetobacter, Escherichia, and Klebsiella) and the most abundant genus from Firmicutes (Staphylococcus). We then retested that the results held within these two dominant phyla.