Background

A community of microbial life forms, constituting of bacteria, fungi, viruses, and parasites make up the commensal skin microbiota responsible for modulating the host immunity responses and preventing colonization and invasion by pathogens [1]. Various medical and allergic conditions are associated with alterations in one’s overall cutaneous microbial community and dysbiosis between specific skin colonizers [2,3,4], highlighting the importance of the microbiota in cutaneous health. Thanks to advances in sequencing technology, the scientific community has gained a wealth of information regarding the skin microbiota, including the biogeography of the skin microbiota, roles of host properties, activities, and the environment on cutaneous microbial communities, co-abundance and co-exclusion correlations between taxa, and changes of microbiota associated with health and disease [4,5,6,7,8,9,10,11,12,13,14,15].

Recent investigations have also reported on the temporal dynamics of the cutaneous microbiota [2, 10, 16,17,18,19,20,21], highlighting the highly personal nature of the skin microbiota across healthy individuals over weeks and months. Changes in the microbiota may also be associated with cutaneous or immunological conditions, or as responses to clinical interventions [7, 19]. Altogether, these studies reinforce the importance of understanding potential relationships between the environment, host characteristics, and microbiota variations over time, and whether microbiota changes over time is related to temporal variations in disease phenotypes and prognosis [7].

As with other ecosystems, microorganisms found on skin engage in co-abundance and exclusion relationships [22]. While correlation does not equate causation, microbial association networks can be analyzed to infer potential ecological relationships between community taxa, as well as relationships between taxa and the environment. Recent works have performed correlation network analyses to characterize potential relationships between specific taxa within the skin microbiota [5, 6, 22]. However, understanding association networks at the structural level provides additional information on how microbial associations as a whole respond to different human or temporal variables, potentially revealing how this global association framework changes, and the factors that may drive such changes over time.

Previous studies have illustrated the effects of cohabitation on the skin microbiota within individuals. Specifically, our previous work [5] recapitulates the discoveries of other studies [19, 23] that the cutaneous bacterial assemblage of cohabitants are more similar compared to non-cohabitants. While these works complement earlier reports describing the effects of lifestyles and living conditions on occupant skin microbiota [13, 15], no study has extensively and directly evaluated the potential roles of cohabitation and household occupancy in shaping the nature and extent of changes in the skin microbiota and its association networks within individuals over time. Given that cohabiting individuals and their microbiota may facilitate the transmission of microorganisms [24], a greater insight into the potential roles of household properties such as occupancy on an individual’s skin microbiota and its dynamics may be of clinical significance.

Therefore, following our previous skin microbiota studies involving a single time point [5, 6], we now provide an account of the skin microbiota (while the term microbiota commonly encompasses all microbial domains, the term is hereafter referred to as the total bacterial and archaeal community) and its dynamics, of 24 individuals within Hong Kong (HK) households over the period of 1 year. Different aspects of the nature and extent of microbiota dynamics, with a focus on the roles of cohabitation and household occupancy, are described. We show that while our observations are consistent with the individuality of the dynamics of cutaneous microbiota shown in previous works [10, 20], we extend this personalized property to the dynamics of microbial association networks. At the same time, we suggest that the extent of changes in different characteristics of the skin microbiota may be associated with the number of cohabitants an individual resides with. By focusing on the changes of multiple aspects of the cutaneous microbiota at individual and household levels, this study sets the stage for comparative analyses pertaining to microbiota dynamics between healthy individuals and those with various skin conditions, thereby potentially identifying personal, household, and related factors associated with healthy microbiota dynamics.

Results and discussion

In this study, we first characterize and discuss the nature and degree of variability of various facets of the skin microbiota within healthy individuals. This is followed by a focus of how effects of cohabitation and residential occupancy may play roles in shaping the observed differences in the degree of temporal variability in one’s skin microbiota over time.

Taxonomic and sub-genus overview of the skin microbiota over time

Top genera (i.e., those with average relative abundance of ≥ 1% across the entire dataset) are detected across all individuals and skin sites and comprise under 60% of the skin microbial community, reflecting the overall taxonomic diversity of the cutaneous microbiota (Additional file 1: Figure S1 and Additional file 2: Table S1) [25]. The top genera include Enhydrobacter, which has been detected in previous reports and is considered to be enriched on skin surfaces of Chinese individuals [5, 26,27,28], suggesting that this genus is a prevalent and stable skin colonizer in subjects of this cohort. Information is currently lacking on the ecological and physiological bases for the apparent enrichment of Enhydrobacter in the Chinese. However, an association between relative abundance of Enhydrobacter on skin and facial sebum level and hydration has been reported in a recent Asian study [28]. Further examination of genetic and/or lifestyle factors that may affect skin physiology (such as living conditions, environmental exposure, food intake, and use of cosmetic and sanitary products [11, 25, 29,30,31]) may help elucidate the physiology of this colonizer on skins of Chinese and other host populations.

We sought to examine the extent of change in relative abundance of a given genus within an individual over the four seasons. Here, the extent of change in relative abundance for a particular genus is expressed as its coefficient of variation (CV). Regardless of skin site, each individual presents a distinctive signature of CV for each of the top genera (Additional file 2: Table S1 and Additional file 3: Figure S2). Previously, it was revealed that the relative abundance and the variations (as expressed by the standard deviation) of microbial taxa within the skin microbiota of a western cohort exhibit a second-order relationship, where low- and high-abundance microbial taxa vary less than those that are moderately abundant [10]. In contrast, our HK cohort shows strong linear correlative relationships across individuals (Additional file 4: Figure S3). As no genus in our dataset represents an average of over 45% of any individual’s microbiota, our correlative relationship between relative abundance and standard deviation over time is expected to be different from the previous study, where the second-order relationship observed appears to be strongly driven by taxa detected in over 50% of a subject [10].

Differential abundance (Fig. 1 and Additional file 5: Table S2), negative binomial mixed model (Additional file 5: Table S2), and oligotype (Additional file 6: Figure S4) analyses reveal that genera are comprised of multiple species/strains, and specific taxa may be enriched in different seasons, independent of household or individual factors (for example, OTU_27 and OTU_746 in Fig. 1a, b; mixed model results are shown in Additional file 5: Table S2 and also oligotypes of genera in Additional file 6: Figure S4a-d). At the same time, taxa may be enriched in specific households/individuals within a season (for example, OTU_27 and OTU_20 within WKS, and OTU_24 within TK in Fig. 1a, c, d, and also oligotypes of genera in Additional file 6: Figure S4e-g). Cohort-wide variations in taxa distribution in different seasons may suggest general temporal trends in the emergence of particular species or strains on skin [10]. However, individual- or household-specific passive exposure may contribute to certain subjects or households carrying a particular species or strain at a given time point. It is possible that particular taxa may be picked up by cohabiting members via physical contact with residential surfaces, given the overlap between occupant skin and surface microbiota within households [21, 32, 33]. However, it is not known how these individual- or household-specific taxa can stably colonize the skin, as the fluctuations in relative abundance of taxa at different time points within an individual may also represent its poor adaptability and competitiveness with co-colonizers [10]. Whether a taxon presents population-wide or individual/household-associated temporal variation in abundance may be dependent on the taxon and how it interacts with the host, the environment, and co-colonizing microbes. As different species and strains may show variations in physiologies [34, 35], future temporal analyses of taxon-level variation in subjects may prove valuable in linking individual, household, and temporal factors with detection of taxa associated with increased virulence and resistance. Compared to the marker-based analyses, multi-omics strategies may provide more detailed genomic and metabolic assessments of strain variations between individuals, households, skin sites, and over time [9, 10].

Fig. 1
figure 1

DeSeq2 differential abundance analysis of OTUs between pairwise seasonal comparison within individuals. OTUs showing cohort-wide differential abundances between seasons include a OTU 27 of Acinetobacter and b OTU 746 of Caldithrix. In addition, OTUs with household-specific differential abundance patterns include c OTU 20 of Roseomonas (household WKS) and d OTU_24 of Deinococcus (household TK). Note that despite a OTU 27 showing cohort-wide patterns, OTU 27 in occupants of WKS shows an opposite trend, showing a household-specific pattern in differential abundance. Occupants are color-coded according to households. A full list of OTUs with strong and significant differential abundances, which is defined by DeSeq2 log-fold difference of at least three and with an adjusted p value of ≤ 0.05, is presented in Additional file 5: Table S2

Full size image

Microbial diversity variations within individuals over time

Shannon diversity, taking into account both community richness and evenness, was calculated to determine the nature and range of temporal variability in microbial diversity across individuals. When all samples were included for analysis, significant differences in Shannon diversity are observed between individuals, seasons, and skin sites (FDR-adjusted p ≤ 0.02 for all, Additional file 7: Table S3). Within each season, individuals and households are the major factors explaining significant differences in diversity. Linear mixed model also suggests that each of these major factors is potential drivers for differences in Shannon diversity (p < 0.001 for the three factors of season, individuals, and households. For each test, the other two factors, as well as sequencing batches, were considered as random variables). Significant differences in forehead and forearm diversities are observed between households, while forearms also exhibit significant differences in diversity between individuals (Additional file 7: Table S3).

To understand the personal rate of change in microbial diversity over the course of 1 year, we calculated the CV of Shannon diversity at a given site within an individual. Considering all samples, the Shannon diversity CVs are different between individuals, households, and occupancy (Kruskal-Wallis (KW) FDR-adjusted p < 0.005 for all). In contrast, CVs are not significantly different between samples of different skin sites, age group, and gender (FDR-adjusted p > 0.05, KW for skin site and age group, and Mann-Whitney (MW) for gender). Communities on left forearm sites varied in CVs between individuals and households (KW FDR-adjusted p < 0.04 for both, Additional file 8: Figure S5a-b). Shannon diversity CVs in forearms are also significantly higher in adults (average within-individual CV = 0.145) compared to the elderly (CV = 0.0871, KW post hoc pairwise p < 0.05, Additional file 8: Figure S5c).

Microbial compositional variations within individuals over time

We analyzed variations in microbial community compositions between subjects, households, and within individuals over time using weighted UniFrac distances, taking into account both the presence and abundance of OTUs. Skin microbiota significantly cluster by season, individual, household, and age groups overall, as well as within separate skin sites, as shown by both analysis of similarities (ANOSIM) and permutational multivariate analysis of variance (PERMANOVA) analyses (Additional file 9: Table S4). Microbiota do not appear to cluster strongly by skin sites within each season in our cohort (and only marginally in summer).

Within an individual, the microbiota of forearm and palms are more similar within than between seasons (Additional file 9: Table S4 and Additional file 10: Figure S6. Forehead sites were not analyzed, as there is only one forehead sample per individual per season). Also, within an individual, the extent of microbiota dissimilarity on a skin site between any two time points is significantly and positively correlated with that of another skin site (Table 1), consistent with the work of Flores et al. [20]. Furthermore, extremity sites within individuals with greater extent of temporal change in microbial diversity also vary greater in the change in community composition over the year (Additional file 11: Figure S7). While correlations between temporal community variation and average community diversity (i.e., beta-diversity versus alpha-diversity) within an individual over time have been reported previously [10, 20], here, we demonstrate that the extent of changes in the two community properties (beta-diversity versus CV of alpha-diversity) are also correlated.

Table 1 Spearman’s correlation based on pairwise weighted UniFrac distance within individual over time
Full size table

As with previous investigations [10, 17], the microbiota of the same site within some individuals between seasons closer together are more similar, as suggested by the significantly negative time-decay relationships (Additional file 9: Table S4). As samples for this study were collected within a single year, it is currently not known whether the decay relationships observed will persist for longer time frames as in other ecosystems [36], or whether an annual or seasonal cyclical pattern will emerge over longer periods of time, as seen in the gut microbiota of different mammals [37] and potentially in humans [38]. A greater understanding of time-decay relationships of cutaneous microbiota has forensic implications, where the identifiability of a person’s microbiota on a surface may decrease over time [21]. Given that time-decay relationships are more likely to be a personal trait rather than one shared by members of a cohort or household level (based on our results here and those reported previously [20]), additional works will be required to understand why some subjects present time-decay relationships and some do not, and how inter-individual time-decay variability affects the feasibility of using microbiota data in forensic applications.

Overall, our observations are congruent with previous western studies looking at the temporal dynamics of the skin microbiota [10, 17, 20]: (1) individuals experience ranges in the extents of temporal changes in relative abundance of genera and overall microbial diversity and community composition, (2) time-decay relationships in microbiota are present but not a ubiquitous property on skin, and (3) within the same individual, multiple skin sites present significant correlations in the extent of microbiota variations at any two time points, and (4) there exists significant correlations between the extent of temporal variations between different skin sites. Therefore, despite known differences in microbial community compositions across continental populations [5, 11, 13,14,15, 30], some temporal properties of cutaneous microbiota may be more universal. Such knowledge regarding the behaviors of microbial communities within and between individuals over time may provide fundamental basis for subsequent and more focused microbiota analyses in special populations around the globe, such as those adopting different lifestyles and are living in drastically different environments [13].

Dynamics of microbial networks are individualized in associating taxa and overall network structure

While the dynamics of microbial diversity and community composition within individuals have been examined above and in previous works [7, 10, 17, 20], little information is currently available on the temporal properties of microbial correlative association networks. As correlative association may represent ecologically significant relationships between microbial constituents, tracking correlations over time in each individual will allow the assessment on the factors governing network structure property at a personal level. Here, insights into SParse InversE Covariance Estimation for Ecological Association Inference (SPIEC-EASI) [39] were used to (1) identify co-abundance and co-exclusion relationships between specific taxa within individuals and (2) determine whether within-individual network structure dynamics can potentially be explained by individual, household, or temporal factors.

Within an association network for each individual over the course of 1 year (networks of individuals in WKS are presented as examples in Fig. 2; networks of remaining individuals in cohort are presented in Additional file 12: Figure S8, and Additional file 13: Table S5), positive correlations represent an average of 82.2% (71.0–92.2% for 24 individuals) of all associations, consistent with previous works demonstrating that the majority of significant correlations between skin sites are positive [22, 40]. Intra-genus correlations OTU pairs are more likely to be positive, whereas OTUs of different genera may be involved in both positive and negative correlative associations (Fig. 2, Additional file 12: Figure S8, Additional file 13: Table S5, and Additional file 14: Figure S9). A taxonomically diverse collection of OTUs are involved in significantly strong associations, suggested by the large number of OTU nodes classified as “minor/unclassified.” Moreover, most of the highly connected nodes (i.e., key hub OTUs) are not necessarily the most abundant taxa (Additional file 13: Table S5). Relatively minor and rare microbial members within the microbiota have been appreciated as possibly important drivers of community composition dynamics [41]. Here, we complement our previous works demonstrating that these rare taxa are not merely present on skin [5], but may engage in correlative relationships with other microbial members, and are potentially central members in the maintenance of association network structure [42]. Individuals within a residential unit do not share similar hub OTUs, suggesting that microbial members central to association structures are not strongly influenced by household factors.

Fig. 2
figure 2

Microbial association network for individuals of household WKS over the period of 1 year. Significant correlative associations between OTUs (represented by nodes) within each individual determined based on the SPIEC-EASI pipeline. Correlations between OTUs can be positive (represented by blue edges) or negative (represented by red edges). SPIEC-EASI correlations with magnitude of < 0.05 are not represented in figure. OTUs belonging to one of top 20 taxonomic families are color-coded, whereas OTUs of other families are grouped into “ minor/unclassified” group, represented by dark gray nodes. Networks of individuals within WKS presented here as a visual example of the differences in network structures. Network representations of remaining individuals are included in Additional file 12: Figure S8, and a full list of significant correlative associations is presented in Additional file 13: Table S5

Full size image

While strong positive associations are detected within OTUs of Enhydrobacter across individuals of different households (QB3z and SW3z), this genus is also involved in positive associations with OTUs of other genera (OTU_4 of Enhydrobacter with OTUs of Chryseobacterium and Acinetobacter within households of STW and TMb, respectively Additional file 14: Figure S9). Consistent with our previous work [5], Enhydrobacter is negatively associated with OTUs of Streptococcus (HFC3z, Additional file 14: Figure S9). OTUs of Enhydrobacter and Corynebacterium can be involved in both positive and negative associations (within individual TMb3x), demonstrating that species and strains of Enhydrobacter and other genera deserve further analyses in not only its physiology on skin but also their potential interactions within the skin microbiota.

With the exception of individual MOS3x, most of the associations between OTUs of Propionibacterium and Staphylococcus are positive correlations. Propionibacterium acnes and Staphylococcus epidermidis (the dominant Staphylococcal species on skin) can be mutually inhibitory, but genetic variations between strains influence the extents to which one species affect the physiology and survival of the other [43]. Previous species-level analysis of Staphylococcus in our cohort reveals that as detected in some healthy individuals in the western world [18], most OTUs of this genus are in fact Staphylococcus aureus [5]. Intriguingly, S. aureus has been shown to respond to molecules secreted by P. acnes at particular conditions [44]. While correlation analyses act as gateways for identifying interactive relationships between microbial members, it provides no information on the environmental and strain physiological states necessary for such interactions to occur. Coupled with the recent identification of separate evolutionary patterns between continental strains of S. aureus [45], the results here again beckon the need for shotgun metagenomics sequencing to examine microbial associations at higher resolution, not only to understand species- and strain-level patterns in microbial associative relationships but also the metabolic and physiological bases for such relationships [10].

Following the description of taxa within association networks, we assessed the dynamics of association network structures within individuals over the four seasons by generating a network for each season within every individual. Graphlet correlation distance information between network pairs was used to visualize similarities in graphlet properties between association networks in a multidimensional scaling (MDS) plot (Fig. 3). Graphlet MDS ordination suggests that the dynamics of the association networks are highly personal and do not appear to be governed by seasonal or household factors considered in the study. Similarly, the dynamics of the association network structure, as measured by node degree distribution (Additional file 15: Figure S10a) and natural connectivity (Additional file 15: Figure S10b-c), also appear to be individualized. Microbial association network structures have been analyzed and compared between different natural and human ecosystems [5, 22, 39, 40, 46], but not between individuals, nor over multiple time points within individuals. Here, we surmise that the temporally personal nature of the human skin microbiota is not merely limited to its community diversity and composition [10, 20] but extends to the potential ecological interactions within its members within this microbiota. Whether these personal differences in association network dynamics play roles in differences in susceptibilities to conditions associated with skin microbiota dysbiosis is unclear, but our association network findings here necessitate a more individual and egocentric approach in the future by linking personal attributes to microbial associations, and whether the nature of changes in association networks over time can act as indicators or predictors for cutaneous health and disease. Our observations here also raise the question of, despite the diversity observed across individuals and over time, whether a core microbial association network for the skin microbiota exist across healthy individuals at a time and across time within an individual [22], and how this core network is affected by the overgrowth of opportunistic pathogens associated with disease onsets [18]. These valuable questions should be addressed in future large-scale skin microbiota investigations.

Fig. 3
figure 3

Multidimensional scaling plot (MDS) of graphlet correlation distances between microbial association networks of individuals over four seasons. Network structures (each network represented by a data point) in each season that are closer together along the MDS axes have more similar association graphlet structures. The MDS result is separated into four plots for visual ease, and MDS axes are identical in the four plots, as ordination was performed with data points from all four plots included for analysis. Points are color-coded by household and shaped according to individuals within the households

Full size image

Roles of cohabitation on microbiota changes within individuals over time

Microbiota within individuals are more similar than those between non-cohabitants regardless of season (Fig. 4 and Additional file 16: Table S6), which is similar to our previous study [5] based on a single time point and conglomerating all subjects as a unit. However, depending on the individual and/or time point, the similarity in microbiota between a resident and his/her cohabitant may not be significantly greater than that between that resident and non-cohabiting individuals (by post hoc test indicating no significant differences in average pairwise weighted UniFrac distances between cohabitants and between households for a particular individual, Additional file 16: Table S6). Conversely, in other instances, the microbial assemblage within a resident is not significantly more similar than that between his/her cohabitants (by post hoc test indicating no significant differences in average pairwise weighted UniFrac distances between within individual and between cohabitants, Additional file 16: Table S6). Therefore, the effects of cohabitation on microbiota similarities between individuals may be more personal and complicated than previously appreciated. Song et al. [23] report that cohabiting couples share more similar microbiota compared to between parents and their children and that dogs may be a factor in homogenizing skin microbiota between cohabitants. In contrast, a more recent study shows lack of microbiota differences between cohabitants of different relationships [19]. Our cohort do not present enough individuals to assess how particular relationships between household members affect individual microbiota over time, and how domesticated animals contribute to the overlapping of microbiota between cohabitants over time. Nonetheless, given the lack of consistency over seasons in how more or less similar cohabitants’ microbiota are between different individuals and households in our cohort, we hypothesize that similarity between microbiota of cohabitants is more complex than the cohabitants’ relationships alone. Alternatively, specific interactions between cohabitants, differences in skin physiologies, and lifestyle [23, 25] between different occupants within a household likely play more prominent roles in determining each occupant’s microbiota, how variable the microbiota is within the occupant, how similar the microbiota of his/her cohabitants are, and whether such similarities change over time.

Fig. 4
figure 4

Density plots of pairwise UniFrac distances between samples of the same individuals (green), between cohabitants (orange), and between non-cohabitants (purple) across seasons. Density plots are faceted by household and season. Households with data from only one occupant (ADMa, FH, HFC, HHa) do not have cohabitant curves

Full size image

The contribution of cohabitants’ skin microbiomes as potential sources of an individual’s microbial assemblage in each season was assessed using Bayesian SourceTracker analysis [47]. For individuals with cohabitants (i.e., excluding households ADMa, FH, HFC, and HHa where no cohabitant microbiota data is available), the proportions of each subject’s microbiome coming from cohabitants range from 7.0% (SW3y sink in winter) to 94.2% (WKS3x sink in summer) (Fig. 5). No one season appears to be associated with an increased sourcing of individuals’ microbiomes from cohabitants across households, suggesting that the potential effects of cohabitants in sourcing one’s skin microbiome has minimal temporal effect within the year of sampling. In contrast, the overall proportions and dynamics of cohabitants’ microbiome as potential sources of each individual’s skin microbial community are personal in nature. In addition, there is an overall significant and positive correlation between the proportion of one’s skin microbiome potentially sourced from cohabitants and the number of cohabitants that sink subject is living with (Kendall’s τ = 0.210, p = 0.01). However, this can merely be due to the increased number of cohabitant samples as sources available for analysis. Nonetheless, in all households, the proportion of unknown sources is low (average proportion of 3.3% within an individual at a given time point). This suggests that the microbial assemblages across our subjects overall are relatively similar, and any individual- and household-based differences may be due to microbial taxa that are low in abundances. Given that this cohort is a relatively small collection of cohorts within a single city, a similar analysis can be performed in cohorts across continents to test whether global skin microbiota, which are known to present community differences [5, 11, 13,14,15, 30], present similar extents of source overlap over time, so as to potentially infer the size of the global skin pan-microbiota [5, 48].

Fig. 5
figure 5

Estimated proportions of cohabitants’ skin microbiota as potential sources of an individual’s skin microbiota. Source prediction are grouped by individual and season. For a given individual, a microbiota source can be from cohabitants (red) or from non-cohabitants (green). Individuals from households ADMa, FH, HFC, and HHa either live alone or do not have cohabitant microbiota data. Microbiota that cannot be sourced with confidence are indicated as “Unknown Sources” (blue). Source proportion estimated using the Bayesian SourceTracker analytic tool [47]

Full size image

Roles of occupancy on degree of microbiota changes within individuals over time

Members within a household do not appear to share similar signatures of CV in relative abundance of specific genera, depending on the genus and skin site. However, significant correlations can be observed between CVs of relative abundance and household occupancy. Specifically, individuals within households of higher occupancy are associated with greater CVs of Propionibacterium and Acinetobacter over the period of the year (Additional file 17: Figure S11a-c). Conversely, a significantly negative correlation, as shown for Bacillus on palms, suggests that individuals with fewer cohabitants tend to have greater rates of changes in relative abundance of this genus (Additional file 17: Figure S11d). In addition to changes in relative abundance of specific genera, individuals residing in households with higher occupancy also present higher CVs of Shannon diversity on left forearm but not for the right forearm and other sites (Additional file 17: Figure S11e). The significance on the left forearm is intriguing and may perhaps be related to the combined effects of handedness (since all but one individual in this study are right-handed) and environmental exposure. However, this is speculative, and additional investigations in controlled settings will be required to assess the roles of these effects, and with a focus on how cohabitation, handedness, and environmental exposure affect the temporal dynamics of skin microbiota. Thus, it is currently not known whether the different correlation trends in different skin sites are ecologically significant. Metagenomic shotgun sequencing may prove beneficial in explaining any potential ecological and physiological rationale for the correlative observations here.

To examine whether the number of cohabitants within a household is also associated with variations in community composition of the skin microbiota of each occupant and site, Kendall τ correlation was performed between the mean UniFrac distances for a particular site within an individual between any two seasons, and the number of occupants present in his/her household. Forehead and left forearm sites from individuals living in residences of higher occupancy show more temporally variable microbial communities over the course of the year (Additional file 17: Figure S11f).

Overall, the number of residents within a household appears to be associated with changes in the dynamics of the personal skin microbiota over 1 year at multiple levels: (1) the extent of variation in the relative abundance of specific colonizers (CVs of Propionibacterium, Acinetobacter, and Bacillus relative abundance across multiple skin sites), (2) the variation in community diversity (as determined by CVs of Shannon diversity within each individual on forearm sites), and (3) the variation in community structure (through average weighted UniFrac distances between forehead and left forearm sites of the same individuals over seasons). Our households are limited to a maximum number of six occupants, and it is currently not known how residences with more individuals, or other built environment settings (such as workplaces and other public spaces of higher occupancy) would affect one’s skin microbiota and its dynamics.

However, it is unlikely that occupancy directly shapes microbiota changes within occupants, as individuals living in different households of the same occupancy still have different microbiota and variable CVs. Instead, we propose that the number of residents within a household indirectly affects an occupant’s microbiota dynamics by how an occupant interacts with the cohabitants and their microbiota. A household with more people may facilitate increased contact, which may facilitate direct transfer of personal skin taxa and potentially pathogenic microorganisms [12, 24]. Alternatively, occupants may release personal microbiota into the residential air and onto surfaces, and microbial members may subsequently be picked up by its cohabitants through passive exposure or active surface touching [21, 32]. We suggest that over single or multiple time points, the relationships between one’s own microbiota and that of its cohabitants and others (that is, how similar or different their microbiota are, Fig. 4 and Additional file 16: Table S6) depend on the biogeographical (across skin sites within a time point) and temporal (within a site over multiple time points) variability of a person’s microbiota (which is in itself now known to be highly personal [10, 20]) and the nature of interactions between the person and his/her cohabitants. Individual factors, in addition to household occupancy and cohabitation, may as a result influence the temporal stability of one’s baseline microbiota over time, which is consistent with our pairwise UniFrac (Additional file 16: Table S6) and SourceTracker (Fig. 5) observations. Therefore, occupancy and cohabitation represent part of a collection of household properties, along with living conditions and lifestyles [15] that may influence the nature and the extent of change in one’s skin microbiota over time. In addition, given that population density in HK is among the highest in the world, it is currently not known how household properties that take into account the crowdedness of households (such as occupant density or privacy indexes as defined previously for microbiota of residences [49]) will influence skin microbiota of residents. We believe that separating these household properties, perhaps via controlled chamber and inoculation experiments [7, 17], will be crucial in specifying the relative proportion of contribution of each household factor in shaping an individual’s microbiota dynamics.

Conclusion

While our findings corroborate previous western studies in identifying the personalized properties of skin microbiota dynamics, we have highlighted in this study possible associations between occupancy to multiple aspects of microbiota dynamics within individuals, including for the first time, the changes of microbial association network structures within every individual over 1 year. Our findings underscore the need for close examination on individual physiological and lifestyle properties in explaining changes in cutaneous microbiota but also an in-depth assessment of the person’s living environment, including number of cohabitants and their interactions over time. Also, while a daunting task, a more individualistic approach may therefore be beneficial to determine factors governing changes in microbial association network structures within an individual, thereby gaining understanding on properties associated with network dynamics associated with adverse skin conditions.

Finally, there is little information regarding whether individuals suffering from cutaneous ailments present different temporal patterns on their skin microbiota from that of healthy individuals, in ways similar to how gut microbiomes of individuals suffering from various gut conditions have distinctive microbiota temporal traits [50]. Data presented here therefore can act as community-level baseline in understanding how the microbiota dynamics of healthy individuals compare with those with various skin conditions. Future longitudinal and comparative microbiota works will potentially provide perspectives on whether disease onset, progression, and recovery are related to changes in microbiota-host relationships over time and identify host and other factors that may be associated with these changes.

Methods

Cohort characteristics and sampling

A total of 480 skin samples from this study are part of a larger seasonal analysis of skin, air, and surface microbiomes of Hong Kong residences [21], and a continuation of previous single-season skin microbiota works [5, 6]. Ethics approval for subject sampling was granted by the City University of Hong Kong Ethics Committee (reference number 3-2-201,312 (H000334)). After being informed about the nature of the study, as well as their roles and responsibilities as subjects, written informed consent was given by all individuals.

From 24 healthy individuals (coded one of 3u to 3z) of 11 households (ADMa, FH, HFC, HHa, MOS, QB, STW, SW, TK, TMb, WKS), samples were collected in January (winter), April/May (spring), August (summer), and November/December (autumn) of 2014 (Additional file 18: Table S7). All samples were collected between 19:00 and 20:00, as individuals were most likely to be at home during those times and to minimize the disruption caused to the individuals due to sampling. For each individual, five skin samples (forehead, left/right forearm, left/right palm) were collected at his/her residence by samplers wearing gloves sterilized with 70% ethanol to minimize contamination or skin taxa transfer between sampler and the subject. All subjects included this study are ethnically Chinese and are long-term residents of Hong Kong. Subjects of this study had not taken antibiotics and antifungals at least 3 months prior to each sampling episode. The individuals in this study were living in households throughout rural and urban areas of Hong Kong to cover a broad local geographical scope. Individuals and households were selected to cover a range of age and lifestyle choices such as smoking, pet ownership, and allergies. All households involved in this study did not use pesticide or have purchased new furniture during the course of the sampling periods. While antibiotic ingestion may or may not have an effect on skin microbiota dynamics [19, 20], we note that all subjects had not taken antibiotics 1 month prior and after each sampling period. Subjects were instructed to not wash their hands or shower immediately prior to sampling. Samples were collected as previously described [5]. Briefly, skin samples were collected using sterile cotton swabs moistened with 100 μL swab solution (0.15 M NaCl and 0.1% Tween 20) [25], and each sample was collected using a back-and-forth swabbing motion for 15 s. Samples were stored in − 80 C within 1 h of sampling and were stored until genomic DNA (gDNA) extraction.

Genomic DNA extraction, 16S rRNA gene amplification and sequencing

Following sampling, gDNA was extracted using the PowerSoil DNA Isolation Kit (MO BIO Laboratories, Inc., Carlsbad, CA, USA), with modifications as described previously [5]. Sterile cotton swabs not exposed to skin surfaces were included in the extraction process as negative controls to account for possible contamination. Purified gDNA samples, including negative controls, were sent to Health Genetech Corporation (Taoyuan City, Taiwan) for PCR, library preparation, and sequencing. PCR, library preparation, and Illumina MiSeq sequencing were prepared as described [5, 51]. Briefly, the 515f/806r primer pair was used to target and amplify the V4 hypervariable region of the 16S rRNA gene [29]. PCR amplification was performed in a 20-μL reaction volume containing 10 μl 2× Phusion HF master mix (New England BioLabs, Ipswich, MA, USA), 0.5 μM each forward and reverse primer, consisting of customized barcodes present on both primers for multiplex sequencing, and 50 to 150 ng DNA template. The PCR conditions consisted of an initial of 98 °C for 30 s, followed by 30 cycles of 98 °C for 10 s, 54 °C for 30 s, and 72 °C for 30 s, as well as a final extension of 72 °C for 5 min. Positive amplification was verified by agarose gel electrophoresis. Amplicons in triplicates were pooled and purified using AMPure XP beads (Agencourt, Brea, CA, USA) and quantified using a Qubit double-stranded DNA HS assay kit on a Qubit fluorometer (Invitrogen, Carlsbad, CA, USA), all according to the respective manufacturers’ instructions. For library preparation, Illumina adapters were attached to amplicons using the Illumina TruSeq DNA sample preparation kit v2. Purified libraries were applied for cluster generation and sequencing on the MiSeq platform using paired-end 150-bp reads.

Sequence analysis

A total of 12,212,096 forward sequences in .fastq format underwent quality filtering using the “fastq_filter” command in USEARCH [52], based on a maximum expected error of 1 error/read, reads trimmed to a length of 139 bp and shorter reads removed. Eighty percent of reads (9,792,397 reads) were of acceptable quality. Filtered reads were then de-multiplexed and binned into OTUs using UPARSE with 97% sequence identity threshold [53]. OTUs were provided with taxonomic information using the “assign_taxonomy.py” command in QIIME [54] based on the Greengenes reference database (May 2013 version, 99,322 sequences). Chimeras were detected in UCHIME2 (high-confidence mode) against the Greengenes database to minimize risk of identifying false-positive chimera OTUs [55]. OTU lineages present in an average of > 5% of reads in negative controls (each control averages 20,000 reads and 600 OTUs) were considered contaminants and were removed from all samples. OTUs classified as chloroplasts and mitochondria were also removed. Following read filtering and OTU removal, a total of 8,093,026 reads are retained for community analyses described below. Following quality control and OTU taxonomic classification, archaeal OTUs represent ~ 0.3% of reads detected.

Bioinformatics analyses

The coefficient of variation (CV) of relative abundance for top genera for each individual over the four seasons is calculated by the standard deviation of the genera within that individual, divided by the mean of the relative abundance of that genera in the individual. A low CV represents a relatively stable relative abundance for that given genus on a given individual, and high CV represents less stable relative abundance. Total Shannon (alpha-) diversity for each sample was estimated using the “breakaway” package (version 4) in R available from CRAN [56]. Shannon diversity values presented in Additional file 7: Table S3 are average values following 999 reiterations. Similar to previous skin microbiota works [10, 20], the CV for Shannon diversity within each individual, representing the change of within-sample microbial diversity of an individual over the four seasons, is calculated by the standard deviation of the Shannon diversity values for that individual, divided by the mean of the Shannon diversity of the same individual. Weighted UniFrac distances were computed using the QIIME script “beta_diversity.py.” Analyses of similarities (ANOSIM) were computed on based on β-diversity data using “vegan” package in R. The PERMANOVA pseudo-F statistic and significance were calculated using QIIME’s “compare_categories.py” script (with 999 permutations) based on the weighted UniFrac distance matrix generated from QIIME. Differential abundance of OTUs between sample groups were performed with DeSeq2 using the QIIME script “differential_abundance.py,” performed separately for each individual and each season pair. OTUs that showed significant DeSeq2 results in more than half of the individuals for a given season pair were included for negative binomial and zero-inflated negative binomial mixed model analyses (“glmmADMB” package in R) to confirm significant differences in abundance for these OTUs, using season as a fixed effect and count data as response variable. Variations between individuals, skin sites, and sequencing batches were considered as random effects. OTUs that appear in less than 20% of samples, and OTUs with less than 100 reads were not included for the analysis. OTUs of top genera that were detected at an average of ≥ 1% relative abundance and were present across all samples were subjected to oligotyping (version 2.1 from http://merenlab.org/software/oligotyping/) [57]. To remove oligotypes with low read counts, a minimum substantive abundance (M) of 10 was adopted. A linear mixed effects model using the R package “lme4” was used to determine whether Shannon diversity differences can possibly be driven by seasonal, individual, and household factors, with individual, site, household, and/or sequencing batch differences as random effects when not considered as a predictor effect. The analysis of variance (ANOVA) on the hypothesized and null models was performed to determine statistical significance in Shannon diversity differences for each predictor factor.

Microbial source-tracking analysis

To examine the roles of cohabitants as potential microbial sources of an occupant’s skin in each residence and time point, Bayesian SourceTracker [47] analysis was used to estimate the proportional contribution of the cohabitants’ skin microbiome to the individual microbiome at the given season. OTUs present in less than 10% of the samples were excluded for the analysis. SourceTracker predictions were generated for each individual across four seasons, with the skin samples of each individual in one season as the sinks, and the skin samples of the remaining individuals in the same season as potential sources. The results were presented as the source proportions from cohabitants, non-cohabitants, and unknown sources based on whether the sink and source communities can be sourced from the same household.

Microbial association network analysis

SParse InversE Covariance Estimation for Ecological Association Inference (SPIEC-EASI) was used to assess potential ecological associations between microbial taxa. SPIEC-EASI has been recently applied to other microbiota investigations [58, 59] and is currently one of the preferred methods for association network analysis, as it is robust to issues of compositional bias, conditional independence, and dimensionality, all properties commonly encountered in microbiota data [39]. Two sets of association network analyses were performed, where (1) microbiota data for all seasons within an individual was combined (to assess for overall microbial associations within an individual, set of 24 networks), and (2) where microbiota data for each season within each individual was analyzed separately (to assess for dynamics of association networks within individuals, set of 96 networks). For each group, abundance data over different samples underwent centered-log ratio transformation. OTUs with less than 150 and 10 reads were not included in the first and second analysis, respectively. Based on transformed data, the Stability Approach to Regularization Selection (StARS) method [60], suitable for high-dimensional data, was used to infer the network structure, employing the node-based neighborhood selection procedure, with a minimum lambda ratio of 0.01 and reiteration of 50 times as recommended [39]. Networks as shown in Fig. 2 were constructed with Cytoscape (version 3.4.0) [61].

Using SPIEC-EASI, general network structural attributes, including degree distribution (distribution of the number of significant associations (i.e., edges) per each OTU (i.e., node), and natural connectivity (a proxy for network stability and robustness to targeted node removal, by assessing the extent of alternative paths present between any two nodes) for each network were also determined. Natural connectivity for network within each individual and season was assessed based on node removal of either decreasing node betweenness centrality, which is a measure of the paths between a node to all other nodes, or decrease node degree, which is a measure of the number of edges a particular node contains. Natural connectivity is presented as the proportion of the total network size, and a higher natural connectivity represents a more stable and robust association network upon node removal. Graphlet correlation distance approach as described by Yaveroğlu and colleagues [62] was also performed on the set of 96 networks to compare the roles of season, household, and individuals in shaping association network structure. Briefly, each network was broken down into subgraphs (graphlets), each consisting of up to four nodes, and performed the graphlet frequency distribution of each node across networks. A graphlet correlation matrix was generated for each of the 96 networks based on pairwise Spearman correlations of the 11 non-redundant orbits of all nodes. This matrix is then used to calculate the graphlet correlation distances, and relationships between networks which were then visualized into multidimensional scaling plots constructed using R. Each network is represented by a position on the plot, and networks that are closer to each other in the MDS plot space are considered to be more similar in network structure.

Statistical analyses

The nonparametric Mann-Whitney (MW) and Kruskal-Wallis (KW) tests were employed to determine significance when comparing between two or more comparison groups, respectively. p values were adjusted for multiple comparisons using the false-discovery rate (FDR) algorithm in R. Potential batch effect between sequencing runs were considered as a possible confounding factor when performing multiple comparisons throughout the study (Additional file 19: Table S8) and have been treated as an extra sample variable along with other sample variables for FDR-multiple comparison using “p.adjust” in R. Where indicated in the main text, post-hoc KW pairwise comparison tests for significance between individual groups were performed using the kruskalmc function in R package pgirmess (http:// cran.r-project.org/web/packages/pgirmess/index.html), following significant KW observations (adjusted p ≤ 0.05). Spearman’s and Kendall’s τ ranked correlations were computed using the “cor.test” function in R.