Overview of the microbiome of the built environment and its occupants
Samples were taken from four representative locations (toilet bowl, kitchen floor, desk in one of the bedrooms, desk in the main room) within the confined built environment and from the skin (front torso) of six isolated crew members. Sampling was performed at 27 time points spanning one year. Besides amplicon sequencing, all samples plus laboratory controls (n=186 in total) were subjected to quantitative PCR (qPCR) on the 16S rRNA gene to assess the overall bacterial load, and on four representative resistance genes (blaOXA, int1a, qacE∆1, tetM) to assess the progression of microbial resistances on skin and surfaces over time.
Along with the samples, 16 types of metadata of environment and crew members were recorded (selected numerical metadata is listed in Supplementary Table S4). The crew was composed of three male and three female members (crewID A-F), with an average age of 30+/-4 and an average body size of 176 cm +/-9 cm. During the isolation and confinement, hygiene practices were restricted. On average, crew members showered preferentially on Saturdays, every 5.4 +/- 1.8 days (60.67 +/- 15.7 times) for about 1 min and 42 seconds +/- 47 seconds. However, individual showering practices differed. For instance, some crew members showered shorter, but more frequently and others showered longer, but only a few times during the isolation period. The diet was mostly composed of dehydrated food, but the crew was allowed to bring beneficial microbes into the habitat, e.g., starters for sourdough bread, tempeh, cream cheese, kombucha and yoghurt. Toilets were composting toilets. Cyanobacteria (Anabaena sp. PCC 7120 and Chroococcidiopsis sp. CCMEE 029) were used for research purposes. Part of the kitchen waste was processed in a bokashi composting system.
No direct or real-time contact to other humans except to crew members was allowed, and the extravehicular activities included donning of a mock spacesuit that prevented exposure to open air and direct sunlight (described in [79]. Nine resupply events happened during the isolation period (on days 15, 43, 79, 107, 148, 185, 223, 258, 303, 335). A total of 132 samples were processed before and 43 samples after a resupply event. Temperature was stable over time (mean temperature 18 °C +/- 1 °C). CO2 levels were always in the recommended range for indoor environments (400 - 1,000 ppm) with an average of 662 ppm +/- 62 ppm.
Denoising of demultiplexed amplicon data with DADA2 resulted in 10,016 unique features (ASVs). In an initial step, we analyzed the processed controls (sampling blanks, process controls, no-template controls), which showed significantly lower microbial Shannon diversity than actual samples (pairwise Kruskal-Wallis P = 1.8 x 10-4; Shannon H’ ~5.7 vs. 7.0; rarefaction depth of 7,850 sequences). Moreover, microbial composition differed significantly between samples and controls, according to weighted UniFrac metrics (PERMANOVA, ANOSIM R = 0.42, Adonis R2 = 0.06, for all three tests P = 0.001). To clean the dataset, contaminants were identified from processed controls with decontam [49] and subsequently removed from the dataset. All subsequent analyses were performed with the cleaned dataset, which contained 3,077,780 sequences (median frequency was 17,533 sequences per sample). According to rarefaction curves, sequencing depth was of sufficient quality, as the Shannon diversity metric (H) plateaued at ~2,500 sequences.
Since we were interested in the characteristics of the microbiome profile of the different sample groups, the dataset was divided into different surface types (crew [skin], built environment) and sample locations (e.g., individual crew members and locations within the facility).
The overall microbial diversity and composition of biotic (skin) and abiotic surfaces differs significantly
In the first step, we compared the alpha diversities (based on Shannon index) of all sample types. Samples from the crew’s skin showed significantly lower diversity than samples from surfaces of the built environment (pairwise Kruskal-Wallis P = 7.3 x 10-16; mean Shannon H’ ~6.2 vs. 7.5) (Fig. 1A). Significant differences were also detected in the diversity index of five crew members (pairwise Kruskal-Wallis of crew member A and B: q-value 1.6 x 10-3; A and D: q-value 1.3 x 10-3; A and F: 7.5 x 10-3; C and F: 2.7 x 10-2; Fig. 1C).
Notably, the microbial diversity on the crew’s skin varied more (mean Shannon H’ ~5.0 in samples from crew member F to ~6.7 for crew member A) than that on abiotic surfaces (mean Shannon H’ ~7.2 in samples of the kitchen floor to ~7.6 in bedroom samples; Fig. 1B and Fig. 1C).
With respect to alpha diversity, the built environment surfaces showed only significant differences between bedroom and kitchen floor samples (pairwise Kruskal-Wallis q-value 4.5 x 10-2; Fig. 1B), while richness was variable (Supplementary Fig. S3). Alpha diversity was also significantly different according to type of surface material (plywood vs. polymer; pairwise Kruskal-Wallis q-value 4.5 x 10-2; Supplementary Fig. S4).
Beyond alpha diversity, the microbiome profile of samples from built environment surfaces was significantly different from that of crew skin samples (weighted UniFrac distances, PERMANOVA q-value = 3 x 10-3; ANOSIM R = 0.3, P = 3.3 x 10-3; Adonis R2 = 0.15, P = 1 x 10-3) and samples clustered separately in PCoA plot analysis (Fig. 1G and Fig. 1H). Further significant differences were found between all four locations of the HI-SEAS habitat (ranging from q-values of 1.8 x 10-3 to 3.3 x 10-3, PERMANOVA pairwise testing; ANOSIM R = 0.14 to 0.85, highest for kitchen floor vs. toilet bowl; Adonis R2 = 0.33). However, the microbial composition was not significantly different between individual crew members despite highly explained variability along PCoA Axis 1, indicating dynamic changes of microbial composition on skin samples over time (see below).
Supported by extensive metadata analysis, the factors time (P = 0.04, Spearman rank correlation of Shannon diversity with time) and sampling location (P = 9.4 x 10-19, Kruskal-Wallis test of all groups, see above for more details) were identified to have a significant impact on microbial diversity and the microbial profile, whereas the microclimate of the habitat revealed no significant influence.
Metadata predictions based on Random Forest classifiers and regressors showed high overall accuracy estimates of 95% for the sampling environment (skin samples vs. built environment samples), and the day of sampling (R=0.77, P = 2.3 x 10-8). Thus, our subsequent analyses focused on the impacts of time and sampling location.
Each surface was characterized by a specific set of microbial signatures, which can be predicted with high accuracy
As selected surfaces of the built environment (desk in a bedroom, kitchen floor, desk in the main room, toilet bowl) and the crew’s skin showed a significantly different composition (see below), we were interested in a detailed analysis of the characteristic features.
Overall, the skin samples were characterized by high abundance of Staphylococcus, Propionibacterium, Enterobacteriaceae, Enhydrobacter, and Methanobrevibacter signatures (LEfSe analysis, Fig. 1D), whereas the built surfaces were characterized by the presence of Chryseobacterium, Lactobacillus, Gardnerella, Prevotella, and Acinetobacter.
Indicative microbial signatures were identified for the toilet bowl (Staphylococcus, Anaerococcus), the main room desk (Acinetobacter, Streptococcus), the kitchen floor (Brevundimonas, Achromobacter), and the bedroom desk surface (Enhydrobacter, Micrococcus; Fig. 1E).
As observed for habitat surfaces, individual crew members revealed typical microbiome profiles with Propionibacterium being indicative for crew member D, Peptoniphilus for crew member C, Staphylococcus for crew member B and Kocuria for crew member A (Fig. 1F). Remarkably, accuracy of metadata prediction based on RandomForest classifications was possible for certain individuals (e.g. crew member D with 100%) or distinct surfaces of the built environment (e.g. kitchen floor and the toilet bowl both 100% accuracy) (Supplementary Figure S5).
Microbial diversity on skin increased during the isolation over time
Overall, the longitudinal microbial diversity in samples from skin showed a steady increase over time (mean Shannon H’ 4.9 to 6.4), whereas the increase in microbial diversity on built environment surfaces was lower (mean Shannon H’ 6.4 to 7.3). The microbial diversity from built environment surfaces was subject to greater fluctuations throughout the time period (Fig. 2A). This observation, however, could be due to a higher number of analyzed built environment samples. Increasing microbial diversity on skin was also confirmed by linear mixed effect models which tested whether Shannon diversity changed over time in response to the sampling locations (Supplementary Fig. S6). An increase in microbial diversity on skin was observed for most crew members (C, D, E, and F; mean Shannon H’ 5.1 to 6.5; highest increase for individual C from 5 to 7.8). However, almost no change was visible for individual A, and a slight decrease was observed for individual B (mean Shannon H’ 5.5. to 4.8; Fig. 2B).
The microbial diversity on different locations inside the HI-SEAS habitat changed as well (Fig. 2B). The largest fluctuations were detected for samples of the main room, and a slight increase in microbial diversity was visible for all locations apart from the toilet bowl. In the latter case, microbial diversity decreased by 1 log (mean Shannon H’ 7.9 to 6.8), possibly due to more rigorous cleaning procedures.
Other metrics describing the alpha diversity of all samples, such as richness (92 to 213.67) and phylogenetic diversity estimates (7.6 to 14.7), followed a similar pattern, while Pielou’s evenness remained constant over the entire isolation period (0.8 to 0.84; Supplementary Fig. S7 and Supplementary Fig. S8).
Temporary dynamics of microbial diversity were investigated by pairwise difference comparisons of samples from individual time points. Significant differences were only evident between day 210 and day 252 for skin samples (Kruskal-Wallis test for multiple groups, P = 0.02) and between skin and built environment surface samples (Mann-Whitney U test, q-value = 0.03).
Furthermore, correlating patterns of microbial diversity were analyzed by Spearman rank correlations. After FDR correction significant positive correlations of microbial diversity were only evident between crew member C and E (q - value = 0.05-4, rho = 0.9).
Observations of microbial composition followed a similar pattern as described for microbial diversity. Hence, composition of skin samples (weighted UniFrac distances) changed to larger magnitudes than those from built environment surfaces along PCoA Axis1 (Fig. 2D). Largest shifts on crew’s skin were visible between day 0 and day 210 (with a maximum at day 84 of -0.3) (Fig. 2D), especially for crew member B. In contrast, almost no changes along PCoA Axis 1 were visible for crew member D and E (Fig. 2E).
Pairwise distance comparisons of microbial composition (weighted UniFrac distances) at individual time points showed significant differences between day 84 and day 126 (Kruskal- Wallis test for multiple groups, P = 0.03), and between skin and built environment surface samples (Mann-Whitney U test, q-value = 0.04).
Comparison with other built environment studies indicates an atypical increase of skin diversity under isolation
For a suitable evaluation of our observations, we performed a meta-analysis of longitudinal microbial diversity patterns inside different built environments. This analysis (see Material & Methods for more details) covered more than 3,400 samples and 10 different sample types (front torso skin, confined habitat surfaces, office dust, room surfaces, desk surfaces, door stoppers, floors, windows, fecal samples and skin samples from the inner elbow) from four longitudinal studies inside the built environment and were all processed in the same way to allow proper comparisons. Public studies from Qiita were selected based on three criteria: first they had to be longitudinal, second they had to be conducted in a built environment setting and third they had to cover samples from human sources beside built environment surfaces. According to these criteria we included a longitudinal analysis of microbial interactions between humans and the indoor environment [80], a longitudinal assessment of the influence of lifestyle homogenization on the microbiome of United States Air Force Cadets [81], and a study which identified geography and location as the primary drivers of office microbiome composition [82]. Our meta-analysis confirmed that microbial diversity progressions in the HI-SEAS habitat were exceptional. While all other sample types showed a decrease in microbial diversity, only samples from the human gut of US air force cadets (mean Shannon H’ 5.1 to 5.5) and skin samples of the HI-SEAS crew (mean Shannon H’ 4.9 to 5.4) showed a steady increase over time (Fig. 2C).
Microbial dynamics during isolation was driven by specific taxa
For a higher resolution of microbial composition in skin and built environment samples over time, the isolation period was grouped into four phases (phase 1: days 0 - 84, phase 2: 84 - 210, phase 3: 210 - 294 and phase 4: 294 - 336). To get insights into microbiome evolution after isolation, skin samples from the post-mission control (day 400) were also studied.
According to differential abundance analysis of all samples using balances in gneiss (Fig. 3A), higher proportions were visible for Staphylococcus, Propionibacterium and Methanobrevibacter (phase 1). Between day 84 and day 210 (phase 2), only Stenotrophomonas and unclassified Enterobacteriaceae showed higher proportions at the latter time point. Later on, unclassified Dermacoccaceae, Propionibacterium, Kocuria and unclassified Rhizobiaceae showed increasing proportions, while Streptococcus and Fusobacterium showed decreasing proportions (phase 3). During phase 4, Methylobacterium populi, Streptococcus, Brevundimonas, Pseudomonas, Lactococcus, Sphingomonas and Cloacibacterium revealed lower proportions than unclassified Enterobacteriaceae and Staphylococcus.
After the isolation period, increasing proportions of Acinetobacter, Propionibacterium, Rhizobium and Methylobacterium populi were prevalent on the skin of the crew, while signatures of Pseudomonas, Corynebacterium or unclassified Intrasporangiaceae decreased (Supplementary Fig. S9).
Dynamics of representative skin, GIT/UGT and environment-associated microbial taxa
In a next step, we selected 15 microbial genera and families which were indicative of either skin (Acinetobacter, Staphylococcus [aureus], Brevundimonas, Kocuria, Propionibacterium, Streptococcus, Kytococcus, Dermacoccacae), gastrointestinal/urogenital tract (Gardnerella, Lactococcus, Methanobrevibacter, Faecalibacterium, Enterobacteriaceae) or environment and water (Pseudomonas, Enhydrobacter), to assess the dynamics of those microbial signatures. Our feature selections were supported by higher rankings in differential abundance tests based on gneiss, aldex2 and songbird, and feature loadings based on deicode. Grouping these representative features into the categories skin, GIT/UGT (gastrointestinal/urogenital tract), and environmental was based on empirical data from literature [10].
Representing the skin microbial taxa, Acinetobacter, despite being recognized as a typical skin microbial taxon, showed higher relative abundances on built environment surfaces (especially in the main room and bedroom). Crew member E showed over proportional prevalence at the beginning and together with crew member A also at the end of the isolation period (Fig. 3B and Supplementary Fig. S10).
Staphylococcus [aureus] was mainly present on human skin (crew members D, E and F), but has also been detected on surfaces inside the habitat. (Fig. 3B and Supplementary Fig. S11). Brevundimonas was clearly associated with the kitchen surfaces (Supplementary Fig. S12) and showed higher proportions on the toilet bowl (between day 70 and day 84), the main room (between day 238 and day 252) and on the skin of crew member F between day 238 and day 252 and again between day 294 and day 336. Relative proportions of Kocuria were clearly correlated with time by linear regression models and showed the highest value for importance (0.3) (Supplementary Fig. S13). While crew members A and E showed a higher prevalence of Kocuria right from the beginning, built environment surfaces as well as crew members B and C showed higher proportions only later on. In contrast, Propionibacterium did not manifest itself on built environment surfaces and could only be recovered from other skin samples over time (Supplementary Fig. S14). On the other hand, signatures of Streptococcus could not be linked to a defined human source and established itself on bed and main room surfaces (Supplementary Fig. S15). Kytococcus was associated with crew members A and D at the beginning (Supplementary Fig. S16). Later on, only single events of high proportions were visible on the skin of crew member B or sampled bedroom surfaces.. Dermacoccaceae were regularly retrieved from skin and built environment surfaces with highest proportions in samples from the bedroom and from crew member D at the end of the confinement period (Supplementary Fig. S17). After Kocuria, Dermacoccaceae showed the highest importance (0.1) in linear regression models.
As a representative of the GIT/UGT, Gardnerella showed a consistent presence despite varying proportions on the surface of the toilet bowl (Supplementary Fig. S18). Likewise, signatures of Lactococcus revealed a single peak on the kitchen floor after day 50, but could not be detected on the skin of any crew member (Supplementary Fig. S19). In contrast to Faecalibacterium (Supplementary Fig. S20), Methanobrevibacter were not consistently recovered from the toilet bowl, but were clearly associated with some of the crew members (Supplementary Fig. S21). As biplot analyses identified Euryarchaeota (in particular Methanobrevibacter sp.) as the main reason for compositional changes around day 84, this genus was analyzed further. In general, signatures of Methanobrevibacter were highly associated with the human crew within the first 210 days (Fig. 3B and Supplementary Fig. S22). However, these signatures were not common on built environment surfaces and were only observed on the toilet bowl and the kitchen floor on day 115 and on day 224. This pattern was different from that of other archaeal lineages (Euryarchaeota, Thaumarchaeota and Woesarchaeota), which showed scattered peaks on built environment surfaces but not in skin samples. Enterobacteriaceae showed only single events of prevalence on the toilet bowl and were mainly associated with skin samples of crew members D, E and F between day 100 and day 250 (Supplementary Fig. S23).
Despite representing an environment-associated taxon, Enhydrobacter was present on all crew members to varying proportions and was regularly detected in samples from the toilet bowl (Supplementary Fig. S24). Signatures of Pseudomonas were observed in skin samples from crew member B and D, followed by a detectable increase on the kitchen floor and further on the skin of crew member E and F (Supplementary Fig. S25).
Shared occupancy influences the microbiome composition and function of the crew skin and abiotic surfaces
Source tracking of microbial signatures with Sourcetracker2 identified human skin as the main source of microbial dispersal. Noteworthy, the intensity of microbiome exchange was heterogeneous among the possible pairs of crewmembers. In more detail, crew member F showed the highest interactive profile (13.4%) of all crew members. This was also supported by redundancy analysis (RDA) revealing pairwise microbial exchange for two pairs of crewmembers as significant parameters on respective skin microbiome profiles (P = 0.002 RDA). Sampled locations of the built environment played only a minor role in overall microbial dispersals. Maximal microbial contributions on the crew reached only proportions of 1.2% in case of the main room. Interestingly, crew gender showed different microbial associations for bedroom and the kitchen floor samples. Nevertheless, microbial interaction profiles were highly person-specific as well as dynamic over time. Overall, trends were difficult to delineate (Fig. 4).
Further on, we were interested in whether microbial profiles and interactions between crew members and surfaces inside the HI-SEAS habitat had an impact on potential phenotypes predicted with Picrust 2 and BugBase. According to these predictions, most phenotypes showed higher proportions (e.g., potential mobile elements, potential pathogens, potential stress tolerance and especially facultative anaerobes) or recurring peaks (Gram-positive and Gram-negative phenotypes and aerobes) on samples from the crew’s skin. However, the potential to form biofilms showed a constant maximum in samples retrieved from the kitchen floor (global median 0.18, maximum 0.27, cumulative average decrease/increase -0.25/0.24) and anaerobes were increasing on the toilet bowl (minimum 0.06, maximum 0.21) while decreasing on human skin (0.02 to 0.01). Interestingly, potential pathogens showed an anti-cyclic pattern of samples from the built environment versus samples from the crew (Fig. 5).
Fluctuation of microbial quantity correlates with the presence of certain antimicrobial resistance genes and microbial phenotypes
Quantitative PCR was used to assess bacterial and archaeal abundance, and its dynamics inside the HI-SEAS habitat and the skin of its isolated crew members. As observed for microbial profiles, bacterial abundance changed to a larger extent for samples from the crew’s skin than from built environment surfaces. In general, two main phases of differing bacterial load could be determined. In the beginning (days 0 - 42) and between days 126 and 210, bacterial abundance on human skin was much lower than on selected locations of the built environment (respective mean difference for the two phases: 12.5 and 22.5%). The largest dynamics were observed around day 28 and day 182 (change in relative proportions by 43%; Fig. 6A). On the other hand, archaeal abundance peaked around day 84 (82%) but varied to much lesser extents, especially in the mid-term of the isolation period (Supplementary Fig. S26).
In addition, four markers for antimicrobial resistance (blaOXA - class A beta-lactamase, int1a - class 1 integrase, qacE∆1 - biocide resistance gene, quaternary ammonium compound and tetM - tetracycline resistance) were selected to analyze dynamics of microbial resistances in a quantitative way. TetM was most predictive for the factor time (importance = 0.2) and was constantly more abundant in skin samples between day 140 and day 294. Interestingly, beta-lactamases showed the opposite pattern, with lowest proportions between day 140 to day 294. Int1a gene abundance was highly dynamic over the whole time frame (highest global variance of 0.06) and showed peaks on built environment locations (toilet bowl and kitchen floor) on day 140, but also on human skin (especially crew member C and D) on day 308. Highest and lowest abundances of qacE∆1 regularly alternated between samples of the built environment and from human skin. Nevertheless, all four targeted resistance genes showed high dynamics and potential transfer between skin and the built environment (Fig. 6B).
89 taxa on species level could be positively correlated by Spearman rank correlations with 16S rRNA gene abundance (for instance Chryseobacterium q-value = 1.5 x 10-8, R = 0.57; Pseudomonas fragi q-value = 1.6 x 10-8, R = 0.56; Megasphaera q-value = 7.5 x 10-8, R = 0.54), while only a few taxa showed significant negative correlations (Ralstonia q-value = 1.8 x 10-4, R = -0.39; Tepidimonas q-value = 0.01, R = -0.26). On the contrary, potential significant correlations of taxa with selected antimicrobial resistance genes could not be verified by multi-hypothesis testing using FDR correction of significant p-values.
Finally, predicted phenotypes were correlated both with each other and with obtained quantitative information (16S rRNA gene copies and selected resistance genes). While all quantitative data could be significantly positively correlated with each other (especially class A beta-lactamases with biocide resistance of quaternary ammonium compounds – q-value = 1.5 x 10-13, rho = 0.66), and biocide resistance of quaternary ammonium compounds with tetracycline resistance – q-value = 2.9 x 10-7, rho = 0.5), comparisons of the qualitative information showed both positive and negative correlations. Significant positive correlations were evident between aerobes and potential biofilm formers (q-value = 8.9 x 10-11, rho = 0.60), as well as potential pathogens and stress tolerance (q-value = 4.7 x 10-12, rho = 0.63). On the contrary, significant negative correlations were observed between aerobes and facultative anaerobes (q-value = 1.0 x 10-14, rho = -0.67), as well as between potential biofilm formers and Gram-positives (q-value = 1.1 x 10-7, rho = -0.51). However, significant correlations between quantitative and qualitative measures were scarce. Only the overall bacterial load (16S rRNA gene copy numbers) showed significant positive correlations with anaerobes (q-value = 0.002, rho = 0.32) and significant negative correlations with aerobes (q-value = 0.01, rho = -0.26).