Intra-hospital microbiome variability is driven by accessibility and clinical activities

ABSTRACT The hospital environmental microbiome, which can affect patients’ and healthcare workers’ health, is highly variable and the drivers of this variability are not well understood. In this study, we collected 37 surface samples from the neonatal intensive care unit (NICU) in an inpatient hospital before and after the operation began. Additionally, healthcare workers collected 160 surface samples from five additional areas of the hospital. All samples were analyzed using 16S rRNA gene amplicon sequencing, and the samples collected by healthcare workers were cultured. The NICU samples exhibited similar alpha and beta diversities before and after opening, which indicated that the microbiome there was stable over time. Conversely, the diversities of samples taken after opening varied widely by area. Principal coordinate analysis (PCoA) showed the samples clustered into two distinct groups: high alpha diversity [the pediatric intensive care unit (PICU), pathology lab, and microbiology lab] and low alpha diversity [the NICU, pediatric surgery ward, and infection prevention and control (IPAC) office]. Least absolute shrinkage and selection operator (LASSO) classification models identified 156 informative amplicon sequence variants (ASVs) for predicting the sample’s area of origin. The testing accuracy ranged from 86.37% to 100%, which outperformed linear and radial support vector machine (SVM) and random forest models. ASVs of genera that contain emerging pathogens were identified in these models. Culture experiments had identified viable species among the samples, including potential antibiotic-resistant bacteria. Though area type differences were not noted in the culture data, the prevalences and relative abundances of genera detected positively correlated with 16S sequencing data. This study brings to light the microbial community temporal and spatial variation within the hospital and the importance of pathogenic and commensal bacteria to understanding dispersal patterns for infection control. IMPORTANCE We sampled surface samples from a newly built inpatient hospital in multiple areas, including areas accessed by only healthcare workers. Our analysis of the neonatal intensive care unit (NICU) showed that the microbiome was stable before and after the operation began, possibly due to access restrictions. Of the high-touch samples taken after opening, areas with high diversity had more potential external seeds (long-term patients and clinical samples), and areas with low diversity and had fewer (short-term or newborn patients). Classification models performed at high accuracy and identified biomarkers that could be used for more targeted surveillance and infection control. Though culturing data yielded viability and antibiotic-resistance information, it disproportionately detected the presence of genera relative to 16S data. This difference reinforces the utility of 16S sequencing in profiling hospital microbiomes. By examining the microbiome over time and in multiple areas, we identified potential drivers of the microbial variation within a hospital.


Revision Guidelines
To submit your modified manuscript, log into the submission site at https://spectrum.msubmit.net/cgi-bin/main.plex.Go to Author Tasks and click the appropriate manuscript title to begin.The information you entered when you first submitted the paper will be displayed; update this as necessary.Note the following requirements: • Upload point-by-point responses to the issues raised by the reviewers in a file named "Response to Reviewers," NOT IN YOUR COVER LETTER • Upload a compare copy of the manuscript (without figures) as a "Marked-Up Manuscript" file • Upload a clean .DOC/.DOCX version of the revised manuscript and remove the previous version • Each figure must be uploaded as a separate, editable, high-resolution file (TIFF or EPS preferred), and any multipanel figures must be assembled into one file • Any supplemental material intended for posting by ASM should be uploaded separate from the main manuscript; you can combine all supplemental material into one file (preferred) or split it into a maximum of 10 files, with all associated legends included For complete guidelines on revision requirements, see our Submission and Review Process webpage.Submission of a paper that does not conform to guidelines may delay acceptance of your manuscript.
Data availability: ASM policy requires that data be available to the public upon online posting of the article, so please verify all links to sequence records, if present, and make sure that each number retrieves the full record of the data.If a new accession number is not linked or a link is broken, provide Spectrum production staff with the correct URL for the record.If the accession numbers for new data are not publicly accessible before the expected online posting of the article, publication may be delayed; please contact production staff (Spectrum@asmusa.org)immediately with the expected release date.
Publication Fees: For information on publication fees and which article types are subject to charges, visit our website.If your manuscript is accepted for publication and any fees apply, you will be contacted separately about payment during the production process; please follow the instructions in that e-mail.Arrangements for payment must be made before your article is published.
ASM Membership: Corresponding authors may join or renew ASM membership to obtain discounts on publication fees.Need to upgrade your membership level?Please contact Customer Service at Service@asmusa.org.
The ASM Journals program strives for constant improvement in our submission and publication process.Please tell us how we can improve your experience by taking this quick Author Survey.
Thank you for submitting your paper to Spectrum.

Sincerely, Jan Claesen Editor Microbiology Spectrum
Reviewer #1 (Comments for the Author): The paper written by Chibwe et al., uses a combination of deep-amplicon sequencing and culture-dependent approaches to identify microbial spatial patterns for prevention and control in nosocomial settings.Very few times, as clinician or researcher we have the opportunity to set a baseline for a hospital from a microbiological perspective.I consider this paper to be very well written and the results seem solid.My comments described below are small in nature: In some parts of the paper Pearson was used over Spearman and viceversa.Was there a specific reason for this?.In this context line 267, presents Pearson-ρ= 0.287 with a significant value of p-value <2.2e-16.While significant, that positive value Pearson-ρ= 0.287 seems weak.
There is a typo in line 163, it should be storage.Line 149, the fact "the volunteers sampled five surface sites of their choice" seems biased.Could you elaborate more on this experimental design?.
Line 339, besides Principal Coordinate Analysis reported here, was any other model ran?.Regarding PCoA, axis in this figure seems to indicate principal component analysis (PC1 vs PC2).I would use PCo1 vs PCo2.
Lines 500-501, antibiotic resistant organisms were identified.Was a MIC conducted during any point of the study (i.e.baseline and April-May 2018).
Reviewer #2 (Comments for the Author):

General comment:
In this manuscript, the hospital environmental microbiome was investigated by collecting surface samples across different areas in a hospital and assessed using both 16S rRNA gene sequencing and culture-based methods.The diversity and microbial composition was analyzed and machine learning models were used to identify distinct members in sample collected from different areas.The manuscript is well-written and easy to follow and the result presentation is well-organized.The topic of this manuscript should be of interest to the readers of Microbiology Spectrum.However, I would like the authors to address the following comments before I can recommend this manuscript for publication.

Sample collection:
Is there an instruction for selecting sampling locations for the healthcare workers or they can collect samples from wherever they think are high-touch areas?Considering these areas may also frequently cleaned/sanitized, it would be great to discuss what the criteria is for selecting sampling locations for microbiome monitoring.

ML:
More justification is needed for using Machine Learning in this study.ML is used to identify the distinct bacterial members in different areas, which can also be achieved via simply comparing the relative abundance/occupancy across different sampling areas using R or other statistical software.It is unclear why ML is needed here and how it is better than simply comparing the data.Also, it is unclear what the ecological and practical importance of identifying distinct members in different areas is.The authors should improve their discussion on why we want to understand this.
Several terms are defined and used in Line 350-392 and it is a bit confusing.What are the criteria for members of the distinguishing set?Are they also referred as "informative ASVs"?Are "influential ASVs" just "informative ASVs" with greater ML coefficients?
Finally, the authors compared LASSO, SVM, and forest models and concluded that "LASSO is the optical model for identifying informative ASVs" (Line 496).I think this conclusion may be biased considering only limited number of samples were used for training the models (n=120) and different classes were used as the response variable for LASSO and the other two models (binary vs. multiclass).I would recommend toning down such claim and focusing on conclusions that specifically related to the experimental setup in this study.

Culture-based analysis:
It is unclear why the authors wanted to compare the sequencing based results vs. culture-based results.The culture-based analysis is highly dependent on the culture medium selection.In this manuscript, the authors selected sheep blood agar (Line 175) but no justification was provided.It would be great to provide some justification and descriptions of previous studies using similar methods here.
The results basically suggest that there is week to no significant correlation (P-values are needed) on the prevalence (need definition in the manuscript) and relative abundance between sequencing results and cell culture results (Line 419).This is expected considering only very small portions of the bacteria may grow on the agar plates and even the ones grow may have Introduction L71, "but also commensal microorganisms, which are normally non-pathogenic," suggests to be changed into "but also opportunistic microorganisms, which are normally non-pathogenic to healthy individuals".Because there are rare cases in which non-pathogenic and commensal microorganisms can cause diseases.
L76-85, suggests deleting, as the 16S amplicon sequencing is widely used, and readers in this field should have a good understanding of this technology.L117, replace "microbiome community" with "microbial community".L172, why did you use the water control?Why didn't you include an unused swab also as a control?
L183, sequence quality score above 10?To be honest, 10 is very low ... please indicate the average score of sequencing you obtained after quality filtering.
L191 suggests using greengene2 for taxonomic annotation, which was released recently and has a higher resolution at fine levels.
L198 sampling date is also one of the metadata.L219, why didn't you apply weighted and unweighted unifrac distance metrics, which were widely used in the beta diversity of 16S amplicon data analysis?L219-228, in the statistics of permanova, because you have multiple factors, how did you control the influence of others when you were testing the significance of one factor?Please illustrate clearly in the method section.It is good that you converted the categorical factor "sampling data" to a numeric factor "number of days after opening" in statistics.
L246-261 suggests moving to the method section.
L268, is there a reference for the core microbiome at a prevalence cutoff of 40%?In addition to the prevalence, the relative abundance is also important for defining the core microbiome; some previous studies have suggested using 1% as the cutoff for the core taxa, if I remember correctly.L270-291, among such many areas, which area and which sample types contain the highest relative abundance of pathogens?It suggests using a heat map showing the results.

SC4: Line 339, besides Principal Coordinate Analysis reported here, was any other model run? Regarding PCoA, the axis in this figure seems to indicate principal component analysis (PC1 vs PC2). I would use PCo1 vs PCo2.
Response: We've added an NMDS plot to assess the Bray-Curtis distances (Figure S7).In the PCoA plots, the axes titles have been changed to PCo1 vs PCo2.Response: We identified MRSA by inoculating all S. aureus isolates onto chromogenic medium for MRSA detection.We agree with the reviewer that MIC tests on AMR positive strains will be important follow-up work on this study.Unfortunately, antibiotic resistance testing was not within the scope and funding of this project.As indicated in the manuscript, we also identified Acinetobacter parvus but we did not conduct phenotypic testing for carbapenem resistance.

R2: Reviewer #2 (Comments for the Author):
General comment: GC1: In this manuscript, the hospital environmental microbiome was investigated by collecting surface samples across different areas in a hospital and assessed using both 16S rRNA gene sequencing and culture-based methods.The diversity and microbial composition was analyzed and machine learning models were used to identify distinct members in samples collected from different areas.The manuscript is well-written and easy to follow and the result presentation is well-organized.The topic of this manuscript should be of interest to the readers of Microbiology Spectrum.However, I would like the authors to address the following comments before I can recommend this manuscript for publication.
Response: We thank the reviewer for the overall positive feedback.We found the constructive comments helpful.We have carefully considered all the comments raised by the reviewer.Revisions, additional analyses, and clarifications have been made.We believe that these revisions following the reviewer's suggestions have made the manuscript more robust.

Sample collection:
GC2: Is there an instruction for selecting sampling locations for the healthcare workers or they can collect samples from wherever they think are high-touch areas?Considering these areas may also be frequently cleaned/sanitized, it would be great to discuss what the criteria is for selecting sampling locations for microbiome monitoring.
Response: This point was also raised by Reviewer 1.The reviewer's comment helped us see the need to better clarify the justifications of our approach.To encourage awareness and compliance with hospital infection control policies, volunteers were given an orientation to the study, examples of high-touch surfaces in the hospital, and training in sample collection.We agree that there may be bias in the choice of surface sampled, but we feel that this was balanced by the benefit that the volunteers would be better at identifying high-touch surfaces in their own workplace, and that we had serial samples from the same sites at the same time of day over the study period.To bring this point to light we've added to the Discussion: Line 430-434:"Healthcare workers chose which high-touch surfaces to sample which increased their awareness of potential microbial reservoirs in their workspace, evidenced by the diverse types of surfaces (Table S3).Potential issues of sampling bias were mitigated by repeated sampling over time and analyzing the most frequently sampled surfaces."

ML:
GC3: More justification is needed for using Machine Learning in this study.ML is used to identify the distinct bacterial members in different areas, which can also be achieved via simply comparing the relative abundance/occupancy across different sampling areas using R or other statistical software.It is unclear why ML is needed here and how it is better than simply comparing the data.
Response: While statistical methods can be used to identify differential bacteria, machine learning models better account for variability introduced by the complex interdependent relationships between different bacteria in the microbiome and identify a robust set of features through cross-validation.The reviewer's comment helped us see the need to better clarify the justifications for our approach.We have added the following to the introduction: Line 90-96:"The application of machine learning to 16S rRNA sequencing data can identify biomarkers through feature selection.Machine learning models account for inherent complexity and variation within the microbiome which would be otherwise overlooked by traditional statistical methods.The identified biomarkers can be potential pathogen-associated sensors, such as the association between Rothia sp. and SARS-CoV-2 virus found in another hospital study.These types of cross-validated models can be integrated with infection control and should be further explored." GC4: Also, it is unclear what the ecological and practical importance of identifying distinct members in different areas is.The authors should improve their discussion on why we want to understand this.
Response: In our study, the models are used for their explanatory value and identify key ASVs which are different among the areas.There is also future potential for similar models to be implemented in infection control and surveillance.The reviewer's comment helped us see the need to better clarify the potential applications of the biomarker results.
The following has been added to the Discussion: Line 487-496: "Using LASSO classifications in an explanatory analysis, we identified several ASVs that are characteristic for each area.These statistical results point even more strongly to the unique microbiome signatures created in each area, likely by the inherent differences in the occupants and activities each area houses.For example, ASVs of Escherichia-Shigella, Leptotrichia, and Johnsonella were biomarkers of the pediatric surgery ward which indicates higher shedding of these gut and oral microbes.These microbes may originate from patients that have recently undergone surgery and shed them, such as from incisions or incubation tubes.Identification of these biomarkers and dispersal patterns could aid in infection control monitoring by specifying which bacteria to target in culturing methods and which areas to inspect."

SC1:
Several terms are defined and used in Line 350-392 and it is a bit confusing.What are the criteria for members of the distinguishing set?Are they also referred to as "informative ASVs"?Are "influential ASVs" just "informative ASVs" with greater ML coefficients?
Response: A few sentences were changed to clarify informative and influential ASVs.They read, Line 346-357: "LASSO classifications were performed to identify informative ASVs.These are the smallest set of ASVs that distinguishes the area types from each other.After repeating 10-fold cross-validation ten times on each model, 156 ASVs were identified as informative.
[…] Among the informative ASVs, we defined influential ASVs as those with a LASSO coefficient absolute value greater than one, which indicates those ASVs as particularly informative.We identified 18 influential ASVs."

SC2:
Finally, the authors compared LASSO, SVM, and forest models and concluded that "LASSO is the optical model for identifying informative ASVs" (Line 496).I think this conclusion may be biased considering only limited number of samples were used for training the models (n=120) and different classes were used as the response variable for LASSO and the other two models (binary vs. multiclass).I would recommend toning down such claim and focusing on conclusions that specifically related to the experimental setup in this study.

Response:
We thank the reviewer for the thoughtful comment.Additional clarifications have been made to describe the approach and refine the interpretation of the results.In fact, though the models vary in being binary or multiclass, we examined the accuracy on a class-by-class basis.To clarify this, the sentence was changed to: Line 497-498: "From a methodological perspective, SVM and random forest models yielded lower accuracies than LASSO models for each area type, with the exception of the Pathology Lab." We also intend to emphasize that LASSO performed best for our particular case, not generally.In order to focus on this aspect in our conclusion, we specify: Line 501-503: "Our findings demonstrate that the simpler LASSO is the optimal model for our purpose of identifying informative ASVs from our specific dataset." Last but not least, Table 3 has also been revised to more clearly distinguish binary and multiclass model accuracies.
Culture-based analysis: GC5: It is unclear why the authors wanted to compare the sequencing based results vs. culturebased results.The culture-based analysis is highly dependent on the culture medium selection.In this manuscript, the authors selected sheep blood agar (Line 175) but no justification was provided.It would be great to provide some justification and descriptions of previous studies using similar methods here.

Response:
We agree that culture-based results are highly biased based on the medium used.We chose sheep blood agar as that was the most cost-effective medium to identify the majority of clinically-relevant bacteria in the hospital.One of the primary reasons for including bacterial culture in our study was to prove that there were viable bacteria on these surfaces as one of the questions from clinician stakeholders is whether the 16S results have any clinical relevance for patients or for hospital infection control.The comparison between sequencing and culture was to highlight the additional value of sequencing over culture for microbial community structure characterization.We have added to the sentence to the Methods to add appropriate justification: Line 169-170: "Sheep blood agar was the cost-effective medium that would identify most clinically-relevant bacteria."

SC3:
The results basically suggest that there is weak to no significant correlation (Pvalues are needed) on the prevalence (need definition in the manuscript) and relative abundance between sequencing results and cell culture results (Line 419).This is expected considering only very small portions of the bacteria may grow on the agar plates and even the ones that grow may have very different growth rates, which may affect the relative abundance (assuming the relative abundance here is after growth on agar plates).
Response: We agree that the result is as expected.We've added the following to the Discussion to highlight this point: Line 513-515:"Even among the genera identified by both methods, culturing detected them disproportionately.This result is expected considering the selective agar medium and differential bacterial growth rates." We have also clarified the definition of prevalence: Line 412-414: "The prevalences of the genera, as in the proportion of samples the genera was detected in by culturing and 16S sequencing, had a Spearman correlation of 0.239 (pvalue = 0.012, Figure 5B)." Introduction SC1: L71, "but also commensal microorganisms, which are normally non-pathogenic," suggests to be changed into "but also opportunistic microorganisms, which are normally non-pathogenic to healthy individuals".Because there are rare cases in which nonpathogenic and commensal microorganisms can cause diseases.
Response: Thank you for the suggestion.The sentence now reads: Line 68-71: "Hospitalized patients often have comorbidities or immunocompromising conditions that make them more susceptible to serious infections from not only pathogenic microorganisms, but also opportunistic microorganisms, which are normally non-pathogenic to healthy individuals." SC2: L76-85, suggests deleting, as the 16S amplicon sequencing is widely used, and readers in this field should have a good understanding of this technology.
Response: The reviewer's comment helped us see the need to provide more specific information here.This paragraph and the citations used are meant to specifically anchor the use of 16S amplicon sequencing in hospital microbiomes.This part of the literature review has been succinctly summarized in the following way: Line 76-78: "Temporal and spatial dynamics in the hospital environment have been previously investigated using culture-dependent and culture-independent methods in different hospital settings globally."SC3: L117, replace "microbiome community" with "microbial community".
Response: Thank you for the suggestion.The sentence now reads: Line 113-116: "Our main hypotheses were that (i) the bacterial microbial community could change before and after the patient occupancy and vary with hospital areas and (ii) several microbes can be associated with specific hospital areas." SC4: L172, why did you use the water control?Why didn't you include an unused swab also as a control?
Response: In our experience, the most significant contamination arises from the library preparation steps and carryover from prior sequencing runs.We try to minimize this by rotating our indices between sequencing runs.The water control allows us to detect these types of contamination.We agree that an unused swab control would have the added benefit of detecting potential contamination from the swab manufacturer.In our experience, the swabs and other sterile sample collection devices have not been a source of contamination in our laboratory.These devices are gamma irradiated by the manufacturer to minimize the amount of amplifiable bacterial DNA.
SC5: L183, sequence quality score above 10?To be honest, 10 is very low ... please indicate the average score of sequencing you obtained after quality filtering.

Response:
We thank the reviewer for the insightful comment.Additional analyses and visualizations were performed to demonstrate the quality filter process.We would like to reassure the reviewer that the average read quality after filtering are 36.5 for the forward reads and 35.2 for the reverse reads, respectively, which indicates successful quality filtering.
Quality filtering and denoising was completed in a single step using the DADA2 pipeline in QIIME 2 which does not output the resulting quality scores.The denoising statistics provided from QIIME 2 were added for transparency.Additionally we simulated the same quality filtering and denoising using DADA2 in R so that the quality scores could be examined (Table S1).The average read quality is 36.5 for the forward reads and 35.2 for the reverse reads which indicates successful quality filtering (Table S2, Figure S1).SC6: L191 suggests using greengene2 for taxonomic annotation, which was released recently and has a higher resolution at fine levels.
Response: We thank the reviewer for the thoughtful suggestion.Additional taxonomic annotations were performed using Greengene2.We have referenced the Greengenes2 database in our taxonomic annotation.A comparison of the classification by SILVA and Greengenes2 for key ASVs discussed in the study is available in Table S10.Due to the increased potential for misclassification at the species level, we discuss phylogeny at the genera a family levels which are largely consistent between both methods of classification.
SC7: L198 sampling date is also one of the metadata.
Response:Thank you for bringing up this oversight.The sentence now reads: Line 196-197: "The metadata included locations (towers, levels, and room numbers), area types (six areas as described earlier), room types, surfaces, and sampling date." SC8: L219, why didn't you apply weighted and unweighted unifrac distance metrics, which were widely used in the beta diversity of 16S amplicon data analysis?
Response: We have added PCoA plots based on weighted and unweighted UniFrac and Jaccard distances in the Supplementary (Figures S6).These diversity metrics lead us to the same conclusion that there are two distinct clusters: the high diversity group, and the low diversity group.
method.The annotation bar indicates the area type of the sample.B) The prevalences and C) relative abundances of each genera in each sample as measured by 16S rRNA gene sequencing and as measured by culture analysis.The weak correlation demonstrates both the compatibility of the methods and added value of 16S sequencing for wider data capture.
SC13: Table 1 and table 2 suggest to be put as supplementary files.
Response: Table 1 and Table 2 have been revised to be more concise, and the pairwise statistical analyses have been moved to the Supplementary (Table S6 and S7)

FIG
FIG S7 (R1.SC4) NMDS plots based on Bray-Curtis distance of all door handle, keyboard, and office electronic samples taken after the hospital opened for inpatient care (n = 120) and color coded by the area type of where the sample was taken including multivariate t distribution ellipses of the low and high diversity groups.Sample sizes: Pathology lab: n = 20, IPAC offices: n = 42, Microbiology lab: n = 8, PICU: n = 19, NICU: n = 15, Pediatric surgery ward: n = 16.PERMANOVA test: Ho: The centroids of the groups are the same.Ha: The centroids of the groups are not the same.p = 0.001.Ho can be rejected.Solid symbols are area types in the low diversity group, and open symbols are in the high diversity group.

FIG
FIG S1 (R3.SC5) Average read quality across all reads at each base during quality steps.The shaded region represents the 25th to 75th quartile range.

Table 1
and table 2 suggest to be put as supplementary files.
DiscussionL448, too absolute, the healthcare workers can bring in external microbes in the NICU.