We investigated ten buildings in five categories (two per each category: train station, mart (market including grocery stores), indoor parking garage, public library, and daycare center) designated by the Korean IAQCA standards for air quality in PUFs (Table S1). Selection of buildings was a convenience sampling by recruiting the buildings of which managements agreed to our study and gave us permission to access their buildings and perform environmental sampling during the coronavirus disease 2019 pandemic. All ten buildings recruited were located within the metropolitan area encompassing cities of Seoul (population: over 10 million) and Suwon (about one million) (Figure S1). Environmental surveys including air sampling at the buildings were conducted for three consecutive days in each building during both summer and winter from August 2021 through August 2022. We collected a total of 60 air samples (10 buildings × 2 seasons × 3 consecutive days) during the surveys.

Samples for airborne bacteria were collected over a period of any 8 hours between 9 a.m. and 7 p.m. every day for three days when the buildings were occupied or being utilized. Prior to sampling, the flow rate for each sampling pump was calibrated using a Primary Pump Calibrator (Model 4146, TSI, Shoreview, USA). The sampling pumps were set up at about 0.7 m above the ground at a location of high traffic area within the building, ensuring minimum mutual interference between the building users and sampling. Each sampling pump (Gillian Air Plus, Sensidyne, St. Petersburg, USA) was connected to a 37 mm closed cassette loaded with polyvinyl chloride (PVC) filters (pore size = 5 μm, SKC Inc., Eighty-Four, PA, USA) to collect air at a flow rate of 4 L/min. Cassettes and filters were sterilized for one hour inside an ultraviolet disinfection box (HU-4050, Hanshin Medical Co., Incheon, Korea) to inactivate contaminating background DNA. To obtain a sufficient amount of genomic DNA (gDNA) for 16S rRNA gene sequencing, we combined two filter samples collected side-by-side for DNA extraction in each day, resulting in an average of 3841 ± 45 L of air for the pooled daily air samples. Field blanks were manufactured and compared for data reliability. The values excluding the qPCR value of the field blanks taken to the same site were used. To prevent contamination and moisture infiltration after the sampling, the filter cassettes were sealed using parafilm and stored at −80°C until sample analysis.

Environmental conditions that may influence bacterial concentrations or community composition were also measured. These included indoor temperature, relative humidity (RH), particulate matters (PM) in particle aerodynamic diameter < 10 μm (PM₁₀), < 2.5 μm (PM_2.5), and < 1 μm (PM₁), occupant density, and carbon dioxide (CO₂) concentration (Table 2). Indoor temperature and RH were monitored using a smart probe (Testo 605i, Testo, Lenzkirch, Germany) for every two minutes during the sampling period, and daily average temperature and RH were used for data analysis. PM data were monitored using a portable aerosol spectrometer (11-d, GRIMM Aerosol Technik, Ainring, Germany) for every one minute over the sampling period every day. From these PM measurements, we calculated concentrations for PM_2.5–10 (i.e., PM₁₀ concentration – PM_2.5 concentration: coarse PM concentration) and PM_1–2.5 (PM_2.5 – PM₁). CO₂ concentrations were monitored using an IoT-based air quality monitor (Micro Watch: Indoor, Innodigital, Yongin, Korea) for every one minute during the sampling, and daily average concentrations over the sampling time were used for data analysis. Except for daycare center, the number of hourly users was determined by counting the number of people passing by the sampling station within a radius of 5 m for ten minutes for every hour and then multiplying it by 6 (Table S2). Total number of users during the sampling period for each day was then calculated by summing the number of hourly users over 8 hours. Then occupant density was calculated by dividing the total number of daily users during the sampling period by the volume of the building space that was obtained from the building managements.

The High-Pure PCR Template Preparation Kit (Roche, Basel, Switzerland) was used to extract gDNA from the sampled air filters (Rittenour et al., 2012). In a 2 ml screw-cap microcentrifuge tube, 300 mg of sterilized glass beads (Sigma, St. Louis, MO, USA) with a size range of 212–300 μm were added. Subsequently, two filters that were cut into smaller pieces and 600 μL of CelLytic^™ B Cell Lysis solution (Sigma, St. Louis, MO, USA) were added into the tube. Samples were pulverized five times for 30s at 1484 × g using a homogenizer (Allsheng, Hangzhou, China). After centrifugation at 8000 × g for 1 min, the supernatant was collected and centrifuged again at 8000 × g for 1 min. When the amount of the resulting substance exceeded 200 μl, it was divided into aliquots of 200 μl each. To each aliquot, 200 μl binding buffer and 40 μl proteinase K were added, mixed, and incubated for 10 min at 70°C. The sample solution consisting of 100 μL of isopropyl alcohol was transferred to a composite of a filter tube and a collection tube, followed by centrifugation for 30 s at 8000× g. When dividing an equal volume of the same sample, the previous step was repeated to place it in the same filter tube collection tube assembly. The samples were subjected to a washing procedure using an inhibitor removal buffer and wash buffer as per the method prescribed by the kit manufacturer. Finally, 200 μl of elution buffer was added to a new tube to elute gDNA from the sample. The gDNA samples obtained were stored at a temperature of −20°C until they were analyzed.

Targeting the V3 and V4 regions, all sample sequencing libraries were prepared according to Illumina 16S Metagenomic Sequencing Library protocols at a commercial laboratory (Macrogen, Seoul, Korea). The 2 ng of input gDNA was amplified in PCR with 5x reaction buffer, 1 mM of dNTP mix, 500nM each of the universal F/R PCR primer, and Herculase II fusion DNA polymerase (Agilent Technologies, Santa Clara, CA). The thermal conditions for the first PCR were as follows: 3 min of the initial denaturation at 95°C, and 25 cycles 30 sec of denaturation at 95°C, 30 sec of annealing at 55°C, and 30 sec of elongation at 72°C, followed by a 5-min final extension at 72°C. The first PCR products were purified using AMPure beads (Agencourt Bioscience, Beverly, MA). For gene amplification with the primer sequences in amplicon sequencing, 16S V3–V4 region of the bacterial DNA was used as shown in Table 1. The second PCR was the same as the first PCR condition, except for the number of cycles (10 cycles). The concentration and purity of amplified DNA were assessed using VICTOR Nivo (PerkinElmer, Waltham, MA, USA) and DNA size verified by TapeStation D1000 ScreenTape (Agilent Technologies, Waldbronn, Germany). After sequencing, the Illumina MiSeq raw data were processed using the Cutadapt program (v3.2) (Martin, 2011) to generate paired-end FASTQ files for each sample utilizing the index sequences. Errors in the amplicon sequencing process were corrected using the DADA2 (v1.18.0) package in R (v4.0.3) (Callahan et al., 2016). Sequences with an anticipated error rate of 2 or above were excluded, focusing on paired-end reads. After correcting for sequencing errors, the paired-end sequences were combined, and DADA2 was employed to form Amplicon Sequence Variants (ASVs). ASVs less than 350bp were excluded from the generated set. To compare and analyze the microbial communities, normalization was performed using QIIME (v1.9) (Caporaso et al., 2010). Subsampling was performed based on the read count of the sample with the minimum number of reads among all samples. Each ASV sequence was aligned to the reference DB (NCBI 16S Microbial DB, NCBI_16S - 2022–05-26) using BLAST+ (v2.9.0) (Camacho et al., 2009) to assign taxonomic information to organisms. If the query coverage of the best matching result in the database or the identity of the matched region was less than 85%, taxonomy information was not assigned.

qPCR analysis was also performed using the primers listed in Table 1 (331F and 797R for the 16S rRNA region). We prepared the recommended qPCR reaction mix in 0.2 ml PCR tube (GmBH, Switzerland). The total volume was made up to 20 μl as follows: 4 μl of 5× HOT FIREPol EvaGreen qPCR Supermix (Solis BioDyne, Tartu, Estonia), 1 μl of DNA template, 14.6 μl of nuclease-free water (Qiagen, Hilden, Germany), and 0.2 μl of each forward and reverse primers (0.1 μM of each). For the qPCR system, LineGen 9600 (BIOER, Hangzhou, China) was used with the following thermal conditions: 95°C for 12 min of initial denaturation, 40 cycles of 95°C for 15 s of denaturation followed by 50°C for 20 s of annealing, and then 72°C extension for 30 s after the cycles. Diluted Escherichia coli (E. coli) was used to generate a standard curve for qPCR analysis (Lee et al., 2021; J. I. L. Yang et al., 2022). The OD₆₀₀ value before the dilution of the E. coli sample used to produce the standard curve was 0.7, and the corresponding DNA concentration measured after extraction of DNA in the E. coli standard solution was 55.0 ng/μL. Standard sample of E. coli concentration was 1.08×10⁷ (Number of copies of DNA template per/μl) that was serially diluted to 10⁻⁵ times by a factor of 10. PCR efficiency (R² = 0.998, PCR efficiency = 98.4%) was verified using a standard curve, and accuracy was confirmed using a melting curve. The total bacterial DNA copy number was calculated from the linear regression equation of the standard curve and the cycle threshold (Ct) values of each sample.

The data analysis of the 60 air samples at the genus level was performed using Statistical Package for the Social Sciences (SPSS, version 28) and the vegan package in R (v4.2.2). Unverified genus data were excluded from the analyses. Pearson’s correlation coefficients were computed among the environmental factors surveyed. Differences in DNA copy number concentration and alpha diversity indices (Shannon-Weaver and Pielou evenness index) of bacteria among the five building categories and between two seasons were analyzed with analysis of variance (ANOVA) and t-tests (J. I. L. Yang et al., 2022). Tukey’s honestly significant difference test was used for multiple comparisons among five building categories (Brown, 2005). The Bray-Curtis dissimilarity index was calculated using relative abundance of each bacterial genus in each sample. To compare similarity in bacterial community composition within and between groups, we used nonmetric multidimensional scaling (NMDS) and analysis of similarities (ANOSIM) (Pereira et al., 2021). Multivariate linear regression of log-transformed bacterial copy number/m³ was performed to examine the adjusted effects of environmental conditions on the concentrations of DNA copy number. The final multivariate model included building category, season, sampled city, occupant density, PM₁ and PM_2.5–10 concentration, and CO₂ concentration. Temperature, RH, and PM_1–2.5 were excluded from the final multivariate model because of high correlations with season and PM_2.5–10 concentration, which resolved multicollinearity issues in the model. We used the Permutational Multivariate Analysis of Variance (PERMANOVA) method to examine the adjusted effects of environmental conditions on the similarity in bacterial community composition using the Bray-Curtis dissimilarity index (Anderson, 2017). We determined that the p-value < 0.05 was statistically significant.

The average PM₁₀ concentration in the library (11.40 ± 4.35 μg/m³) was significantly (multiple comparison p-values < 0.05) lower than in the daycare center (27.05 ± 7.10 μg/m³) and the parking garage (29.29 ± 18.84 μg/m³) (Table 2). Parking garages had significantly lower occupant density than train stations, marts, and daycare centers (multiple comparison p-values < 0.01) (Table 2). Similarly, the library had significantly lower occupant density than the marts and daycare centers (p-value < 0.05). The average concentration of CO₂ in the daycare center was significantly (p-value < 0.001) higher than the train station, parking garage, and library. The occupant density in three buildings in Seoul (station #1, mart #2, daycare center #2) was significantly (p-value < 0.001) higher than those in Suwon. The average indoor temperature and RH were significantly (p-value < 0.001) higher in summer (26.50 ± 2.51°C and 62.14 ± 5.52%, respectively) than winter (15.77 ± 5.93°C and 29.46 ± 10.77%, respectively). Figure S2 presents the correlations among environmental parameters. Indoor temperature was positively and highly correlated with RH (R = 0.75, p-value < 0.001). Conversely, PM₁ showed negative correlation with indoor temperature (R = −0.39, p-value < 0.01). RH was weakly but positively correlated with PM_2.5–10 (R = 0.32, p-value < 0.05). The PM_2.5–10 had a strong positive correlation with PM_1–2.5 (R = 0.78, p-value < 0.001), while PM_1–2.5 showed a moderate positive correlation with PM₁ (R = 0.50, p-value < 0.001). Only PM_2.5–10 concentration was weakly correlated with CO₂ concentration (R = 0.30, p-value < 0.05).

Our multivariate linear regression model explained 76.9% (adjusted R²=0.50) of the total variance in log-transformed DNA copy number concentration. From the model, we found that season and CO₂ were significantly (p-values < 0.015) associated with airborne bacterial concentration (Table 3). The bacterial concentration in winter was significantly lower than summer after adjusting for other environmental conditions. CO₂ concentration was positively associated with airborne bacterial concentration. The indoor CO₂ concentration can be influenced by multiple factors such as ventilation rate, occupancy rate, and occupants’ activity level (Lazovic et al., 2016; Settimo et al., 2020) and may be a surrogate measure of a combination of these factors in our study. Kwan et al. reported that human occupancy and resuspension are significant sources of the increase in indoor microbes in residential buildings (Kwan et al., 2020). The human occupancy of rooms increased bacterial abundance by approximately 81 times, compared to non-occupied environments (Hospodsky et al., 2015). In a chamber study, the fluorescent biological aerosol particles emission rate increased by 5–6 times from human walking, with more than 60–70% of the emissions originating from the floor (Bhangar et al., 2016). The coarse PM concentration in our study was positively associated with the bacterial concentration but with no statistical significance. A previous study demonstrated significant positive association of coarse particles with airborne bacterial DNA copy number concentration in school classrooms (J. I. L. Yang et al., 2022) and with airborne cultivable bacterial concentrations in coastal cities (Montero et al., 2016). Altogether, the positive but weak association of CO₂ concentration with airborne bacterial levels that we found in this study might be associated with human occupants who are one of the strong bacterial sources in indoor environments.

From PERMANOVA including the same variables in the multivariate regression model of bacterial concentration, we found that building category, season, and CO₂ concentration were significantly (both p-values < 0.04) associated with the bacterial community composition (Table 4). Season had the most significant impact, accounting for 44% of the total variation in the microbial dissimilarity, followed by building category (13%), and CO₂ concentration (2%). These findings from the PERMANOVA model also supported the results of NMDS plots and ANOSIM. The effect of the building category on community composition in our study is most likely explained by the different purpose and usage of the building. For example, unique bacterial compositions in daycare center, train station, and mart (Figure 3) could be driven by building environments and users that are quite different from one another. A positive correlation between CO₂ levels and the abundance of commuter-related bacterial genera such as Enhydrobacter and Micrococcus has been also reported by a study of Hong Kong subway network using DNA sequencing data (Leung et al., 2014). The authors also suggested this association might be driven by the level of commuter density using the subway system.

One of the limitations of our study is that we had only two buildings for each category in two geographical regions. It was extremely challenging to obtain permission from management to access the buildings for our sampling because of concerns for viral transmission during the COVID-19 pandemic period. Even with this challenge, we were able to recruit ten buildings and collect 60 air samples in summer and winter. However, because of the limited sample size per category in limited regions, caution should be exercised when generalizing the study findings. Even with this limitation, considering that data on indoor microbiota from non-residential and non-school buildings using culture-independent method are scarce, our study findings provide very useful and basic information on indoor microbiota and associated environmental factors in institutional and commercial buildings. In future studies, a longitudinal sampling for a longer period of time in many institutional and commercial buildings in different regions and collection of more detailed information on building design and characteristics would be the most useful to better elucidate associations of those building environmental factors with microbial compositions. This research could provide valuable information on designing healthier buildings with healthier indoor microbes.