Settled surface dust was collected from two different swine confinement facilities (housing 400–600 hogs), the storage facilities at two different grain elevators, and two different pet-free domestic homes as a control. The dust samples were obtained from surfaces approximately 3–5 feet above the floor to ensure the sampled dust had been airborne and potentially inhaled by a worker. Permission was granted by the owners of the swine confinement facilities, grain elevator facilities, and households to obtain samples in an anonymous manner. Total genomic DNA was isolated by bead beating following the manufacturer’s instructions (Mo Bio, PowerSoil Kit, Carlsbad, CA), then assayed using a Nanodrop ND-1000 UV spectrophotometer (NanoDrop Technologies, Wilmington, DE). Each DNA sample (3–5 µg) was used to prepare a shotgun pyrosequencing library using a kit for this purpose (Roche Applied Science, Indianapolis, IN) according to the manufacturer’s protocol. The multiple id, barcoded template DNAs were combined in equal amounts and titrated to obtain the optimal copies per bead (3 copies per bead). Emulsion PCR and pyrosequencing were performed with the Roche/454 Life Sciences’ Lib-L (LV) and XLR70 kits, respectively. Multiplexed shotgun metagenomic DNA pyrosequencing was performed by the Core for Applied Genomics and Ecology Laboratory at the University of Nebraska, Lincoln, using a Roche/454 Life Sciences’ GS FLX Titanium instrument (Branford, CT). GS FLX Off-Instrument Software was used to de-multiplex raw pyrosequencing reads into sample-specific bins. These six shotgun pyrosequencing metagenomic read datasets are publicly available at MG-RAST [DNAdustGrainTLV2011s (4465551.3), DNAdustGrainTLV2011e (4465547.3), DNAdustSwineTLV2011f (4465549.3), DNAdustSwineTLV2011n (4465550.3), DNAdustHouseTLV2011n (4465546.3), DNAdustHouseTLV2011p (4465548.3)].

Three bacterial 16S rRNA gene variable region amplicon pyrosequencing read datasets (SRX065852, SRX065863 and SRX065864) and three shotgun pyrosequencing metagenomic read datasets (SRX065862, SRX065867 and SRX065871) from the same study of swine feces microbiota (SRA037229) were downloaded from the NCBI’s Sequence Read Archives [28]. These six relevant publicly available datasets were used to create two MG-RAST dataset collections (16S and shotgun) that were analyzed as described below and which served as controls to estimate the concordance of phylogenetic profiling results (MG-RAST’s “organism abundance profiles”) between 16S rRNA gene amplicon and shotgun pyrosequencing metagenomic approaches. The three individual shotgun pyrosequencing metagenomic read datasets (SRX065862, SRX065867 and SRX065871) from the study of swine feces microbiota were also used as described below to determine if shed lumenal cells from the swine digestive tract might contribute significant amounts of swine DNA in swine feces metagenomes and hence possibly also in swine dust metagenomes.

Unexpectedly, we found that additional bioinformatic processing steps had to be taken with dust-derived read datasets from environments with a resident mammal, i.e., swine facility (swine and human) and household (human) datasets. Each of these datasets had to be computationally partitioned into “swine/human” and “filtered” data subsets in order to maintain the original focus of this study, which was only on the latter subset, i.e., agricultural dust microbiota. Thus, post-QC reads were aligned using the blastn program in BLAST+ v.2.2.25 against the unmasked swine (Sus scrofa) draft genome sequence (ssc_ref_Sscrofa10) and/or the unmasked human genome sequence (hs_ref_GRCh37.p5). The seed size used was six nucleotides and only the best BLASTn hit per read was considered. The NCBI BLASTn program reports expect values <1e⁻¹⁷⁹ as zero; hence zero expect values were converted to 1e⁻¹⁷⁹ before log₁₀ transformation.

The possibility remained that some reads that aligned to the swine/human genomes may align with an even lower BLASTn expect value to a known bacterial genome sequence. Thus, the 1,480 complete and 1,659 draft bacterial genome sequences that were available on the NCBI FTP site on 11/16/2011 were downloaded and formatted as a BLASTn database for a second round of alignments, using the mammalian best-hit reads as queries. Very few of these reads yielded a significant alignment to any of the available bacterial genome sequences, but in cases where such an alignment yielded a lower BLASTn expect value than was obtained with the same read’s best mammalian genome BLASTn hit, the read was re-classified and added to the “filtered” dataset. For the control read alignments against mammalian genome sequences using the swine feces shotgun metagenomic read datasets (post-QC read datasets comprised of 127,088; 427,661; and 563,638 reads for the swine feces 1, 2, and 3 datasets, respectively), the alignment workload was reduced by using an evenly sampled subset of 20,000 reads for each of these three datasets. Except where explicitly indicated, all results reported for the swine facility and household dust-derived datasets are based on analyses of only their respective “filtered” read subsets.

During MG-RAST (v. 3.0) read dataset upload, the default options for quality control (QC) were selected, i.e., base-call quality filtering, read-length filtering, and de-replication of reads, but screening against a model organism genome sequence was not selected. Individual read datasets were then used for MG-RAST organism. These individual read dataset MG-RAST abundance profiles were then used as input for Principal Component Analysis (PCA) performed at multiple levels of the relevant classification hierarchy, as well as two-group statistical tests performed at the lowest or “leaf” level of the relevant classification hierarchy, using the “Statistical Analysis of Metagenomic Profiles” (STAMP v. 2.0) software [29].

Read dataset collections (e.g., swine confinement facility dust [n = 2 samples], grain elevator dust [n = 2 samples] or household dust [n = 2 samples]) were also created in MG-RAST, and these collections were also used for MG-RAST organism. These read dataset collection results were then used as input for summary histograms at all levels in the relevant classification hierarchy. Read dataset collection results were also used for comparisons of MG-RAST’s organism abundance profiles between swine feces control datasets obtained using either 16S rRNA amplicon-based or shotgun metagenomic-based approaches.

Organism abundance profiling using shotgun metagenomic read datasets was carried out using the “best hit classification” alignment procedure against the M5 non-redundant protein database (M5NR), using the following parameter values: Max. e-Value Cutoff: 1e⁻⁵; Min. % Identity Cutoff: 60%; Min. Alignment Length Cutoff: 50. MG-RAST’s Lowest Common Ancestor (LCA) organism abundance profiling procedure for shotgun reads did not produce profiles that could be used with STAMP, and hence the LCA results were not compared.

Organism abundance profiling using the swine feces control “16S” read dataset collection was carried out using the “best hit classification” alignment procedure against the Ribosomal Database Project database (RDP, University of Michigan) [30], using the following parameter values: Max. e-Value Cutoff: 1e⁻⁵; Min. % Identity Cutoff: 97%; Min. Alignment Length Cutoff: 50. The top 10 taxa in the RDP-based organism abundance profiles were ranked based on their relative abundances (taxon-specific abundance/total abundance). Relative abundances values for the same taxa were obtained from the organism abundance profiles carried out using the swine feces control shotgun read dataset collection and the M5NR-based database. Relative abundance ratios were then calculated as the ratio of M5NR-based relative abundance divided by RDP-based relative abundance. Perfect concordance between the RDP and M5NR organism abundance profiles would yield a relative abundance ratio of 1.

STAMP profile files (.spf) were created from MG-RAST abundance profiles (.tsv files) that had been created as described above using individual shotgun metagenomic read datasets. Candidate taxonomic biomarkers were identified at different taxonomic levels using STAMP Extended Error Graphs and pair-wise statistical tests with dust type as the group field. Welch’s t-test and Welch’s inverted method were used for estimation of the difference between group mean proportions and a 95% confidence interval. A minimum difference between group mean proportions of 0.1% was used to minimize spurious results due to low read counts. Significance (q) values were corrected for multiple testing using Storey’s FDR (false discovery rate). A significance threshold (ceiling) was chosen that selected the top-ranked group-distinguishing taxa (typically around 10 taxa), and then these most significantly ranked taxa were ordered by effect size (difference in mean proportions in the two-group statistical test).

16S rRNA gene sequencing has been widely used for phylogenetic profiling of microbial communities. Alternatively, shotgun metagenomic sequence data can be used for this purpose. Both approaches have their limitations; 16S rRNA sequencing may be biased because of unequal amplification of 16S rRNA genes, whereas shotgun sequencing may not be deep enough or the phylogenetic diversity of databases like the M5NR may be inadequate to detect rare taxa in a microbial community. To determine if these two approaches give largely similar phylogenetic profiles, we conducted a comparison of these two approaches using publicly available high-quality metagenomic pyrosequencing read datasets obtained from a study of swine feces microbiota [28]. Using data from this study, we created one MG-RAST dataset collection from a triplicate set of shotgun metagenomic pyrosequencing reads, and a second dataset from of a triplicate set of 16S rRNA gene amplicon pyrosequencing reads. We compared the organism abundance profiles obtained using the M5NR or the Ribosomal Database Project (RDP) databases, respectively, and the MG-RAST “Best-Hit” protocol. We defined a “relative abundance ratio” as the ratio of the relative taxon abundance (RTA) values determined by the two protocols (RTA_M5NR/RTA_RDP), and then plotted the relative abundance ratios obtained, together with the values for RTA_RDP, for the ten most abundant taxa at five different taxonomic levels (Phylum to Genus; Figure S1A–E). For these two types of swine feces-derived read dataset collections and their respective analysis methodologies, of the top ten taxa, six Phyla, four Classes, three Orders, four Families, and four Genera have relative abundance ratios between 0.5 to 2.0, indicating that the two analysis methodologies for characterizing microbial communities yield reasonably concordant results. We observed that when a species is rare as assessed by 16S rRNA gene amplicon sequencing, more discordance is observed.

Swine facility and household datasets each had to be carefully partitioned into “swine/human” and “filtered” data subsets using a BLAST-based protocol to remove mammalian reads of swine and human origin (Figure 1), as such reads were not relevant to the focus of this study. The three different dust types (swine facility, grain elevator and household), which differ substantially in their yields of metagenomic DNA (Figure 2) and post QC metagenomic shotgun pyrosequencing reads, also differ substantially in their content of swine/human reads (Table S1). Individual read datasets derived from grain elevator dust or swine feces yield negligible percentages of their reads aligning to the human or swine genome (2% ±0 SEM and 2% ±1 SEM, respectively) compared to the other two dust types, which are from environments with a resident mammal, i.e., swine confinement facilities (19% ±5 SEM) and households (62% ±14 SEM). Box plots of BLAST “Expect” values show that the alignments of post-QC pyrosequencing reads to the swine or human genome are of very high quality when the query sequences are reads derived from dusts with a high resident mammalian contribution, i.e. swine and household (Figure S2). Despite the source of the swine feces control data, each of these three metagenomic shotgun read datasets can be described as having a low resident mammalian contribution and poor quality alignment to the human and swine genomes, indicating that shed lumenal cells in swine fecal debris is not the source of the mammalian DNA present in dust isolated from swine confinement facilities. Finally, except where noted, in all of our subsequent analyses using shotgun metagenomic pyrosequencing reads generated from dust samples, the “swine/human” reads from the swine facility and household dust samples were excluded as our study focus was agricultural dust microorganisms.

For all dust types, the domain “Bacteria” predominates with relative abundance (RA) levels of 91.6%, 68.6% and 82.2% in swine, grain and household dust, respectively (Figure 3A). Dusts from grain elevator or pet-free households have a greater RA of “Eukaryota” DNA, 13.1% and 9.6% respectively, compared to swine dust (0.4%). “Eukaryota” DNA identified was primarily from plant and fungal phyla. The predominant fungal genera identified in grain elevator dust are Gibberlla (0.7% RA), Neosartorya (0.5% RA), Saccharomyces (0.4% RA) and Aspergillus (0.3% RA), respectively, which are not found in household dust. In contrast, archaeal DNA is present in low amounts in all of the dust samples with only 1.6% RA in the swine confinement facility dust, 0.03% RA grain elevator dust and 0.3% RA in household dust. The genera Methanosphaera (0.3% RA), Methanobrevibacter (0.3%) and Methanothermobacter (0.2% RA) account for 50% of the archaeal DNA in swine dust. Interestingly, Methanobrevibacter was detected in household dust at a RA of 0.2%. Though not visible on the pie charts, the RA of viral DNA, primarily bacteriophage, in all dust types was <0.2% (Table S2).

Within the domain “Bacteria”, the top four Phyla are Actinobacteria, Bacteroidetes, Firmicutes and Proteobacteria, and the relative distribution of phyla are different for each dust type (Figure 3B). Firmicutes predominate in swine (76% RA) and household (38% RA) dust compared to Proteobacteria in grain elevator dust (73% RA). We also generated relative abundance profiles of the top 15 bacterial genera present in the three dust-type-specific read dataset collections (Figures. S3 A–C).

Relative abundance of genera for the swine facility dust-derived read dataset collection was compared to those obtained for the same genera for the other two dust-type-specific read dataset collections at the genus level (Figure 4). Four of the top five swine facility dust genera are assignable (Clostridium, Lactobacillus, Ruminococcus, and Eubacterium, ranging from 19% to 4% RA), and the RA levels of these four genera in grain elevator dust are all lower than values obtained for the other two dust types. In contrast, among the top 15 swine facility dust genera there are two–Prevotella (9% RA) and Bacteroides (7% RA)–that are much more significantly represented in household dust than agriculture-derived dusts. We also carried out similar analyses at the Phylum, Class, Order, and Family level (Figure S4 A–D). These analyses also highlight the predominance in swine facility dust of Clostridia (44% RA) at the Class level, and Clostridiales at the Order level (41% RA) and Clostridiaceae (20%). Although none of the assignable genera among the top 15 swine facility dust genera are more frequent in grain elevator dust, among the top 15 assignable swine facility dust phyla, Proteobacteria (48% RA) and Actinobacteria (7% RA), are more frequent in grain elevator dust than swine facility dust. At the Class level, Gammaproteobacteria (37% RA) are much more frequent in grain elevator dust than either of the other two dust types.

The Principal Component Analysis (PCA) plots generated using STAMP show good dust-type-specific clustering and resolution of the individual read datasets at all taxonomic levels, especially at the genus level (Figure 5A and Figure S5A–D). Also noteworthy is the distinguishable, separate clustering and resolution of the “swine/human” subset of reads from the “filtered” subset of reads for both the two swine facility dust individual read datasets and two household dust individual read datasets (Figure 5B). These data indicate that the individual read datasets for each sample type are more related in characteristics compared to the other sample types.

We also used STAMP to perform more rigorous two-group statistical comparisons (extended error graphs and bar plots based on Welch’s two-sided t-test) using individual read dataset MG-RAST organism abundance profiles to identify candidate microbial taxonomic biomarkers for swine facility dust samples compared to other dust type samples (grain elevator and household). Given the results of the PCA analyses and using these same individual read dataset organism abundance profiles, we focused initially at the genus level. When we compare the swine facility dust samples to the other dust samples (Figure 6), the genera Ruminococcus and Eubacterium had significant difference between group mean proportions (DP) values and low q values (the minimum False Discovery Rate, or FDR, at which the test is significant). This suggests that these two genera may be candidate biomarkers for swine dust compared to grain elevator or housedust without pets. The genus Bacteroides showed the largest DP between the two groups (swine facility dust vs. all other dust samples) though not statistically significant due to a wide confidence interval, whilst unclassified annotations that were presumed to be derived from Erysipelotrichaceae had the lowest q value, followed by the genera Dorea, Ruminococcus, Alkalaphilus, and Eubacterium.