The recruitment and sampling were conducted from March to August 2020. Because the survey and sampling plan involved human subjects, an Institutional Review Board (IRB) approval for human studies was obtained from the University of Illinois Office for the Protection of Research Subjects (IRB protocol: 20653).

A total of 12 hydroponic farms in Illinois were identified by searching on Google Map and agriculture newsletters and the farms were contacted by phone calls or emails. Two farms participated in this study, including a vertical hydroponic microgreen farm (farm H) and an aquaponic farm that grows lettuce and tilapia together (farm A). The hydroponic system setup and operation procedures were present in Supplementary Figure 1. In farm A, the seeds were started in seed starting trays until seedlings were developed, then transferred to a flood-and-drain system until the baby lettuce rosette formed, then rosettes were transferred to a deep-water-culture (DWC) system for maturation. The water came from tilapia fish tanks, first supplied into the DWC system, then the flood-and-drain system and seed starting trays. Three lettuce cultivars were grown in the DWC system in the order of HoneyCrisp, Green Oak Leaf, and GreenCrisp from the inlet to the outlet. Farm H used two separated flood-and-drain systems and five-layer vertical structures. The first system grew five microgreen species, including garnet, radish, broccoli, cilantro, and brussels sprouts. The second system only had kale microgreens.

The two farms were asked for consent on the sampling plan, farm visit, and survey. Fresh produce samples were collected from both farms. The samples were harvested directly from the growth trays with sterile gloves and scissors and weighed into sterile sample bags (Whirl-Pak, Madison, WI, United States). From farm H, radish, garnet, broccoli, brussels sprouts, cilantro, and kale microgreens were sampled by randomly harvesting from 10 different growth trays and added up to 25 g total weight each. We also grew broccoli, radish, and kale microgreens hydroponically in the laboratory system (L1) as a control. From farm A, three lettuce cultivars at their fully grown stage were randomly sampled (three heads each cultivar) from the DW system. Four similar hydroponically grown lettuce cultivars were purchased from a grocery store in Champaign, IL, for comparison. We grew romaine and oak leaf lettuce in a flood-and-drain system and a smaller DWC hydroponic system in the laboratory (L2) as a control.

Nutrient solution samples (250 ml each) were collected from both farms directly into sterilized sample bags. From farm H, we collected two nutrient solution samples from each flood-and-drain system at the reservoir tank and the tubing system underneath the sampled microgreen trays, respectively. Three nutrient solution samples were taken from the DWC of farm A at the water inlet from the fish tank, the midpoint of the system near lettuce roots, and the end of the system near the lettuce roots, respectively.

Environmental swabs were obtained from agreed sampling items using sterile cotton swabs. The swabs were presoaked in peptone–saline water for 1 min, and 3 × 3-cm areas were carefully swabbed with the entire area of the cotton swab surface. Three swab samples were obtained from farm H, including worker’s hands, shoe soles, and sanitized growth trays. Three swab samples were obtained from farm A, including growth tray swabs from the DWC system (trayM), the flood-and-drain system (trayB), and shoe sole swab from the farmworker.

All collected samples were transferred to a microbiology laboratory at the University of Illinois in a portable cooler on the same day. The samples collected for microbiome analysis are listed in Supplementary Table 1.

A five-section survey was developed focusing on current farm food safety practices, including worker health and hygiene, food safety awareness, irrigation water treatment, equipment sanitation, and produce handling procedures (Supplementary Table 2). The questions were gathered from available farm food safety surveys with modifications to reflect the production practices of the hydroponic systems (Harrison et al., 2013; Ilic et al., 2017; FamilyFarmed, 2019). To ensure privacy, each farm was assigned a code to link back to the sampling data. The survey result was coded into numeric scores for calculating the total scores for each section.

The aerobic plate counts were performed for fresh produce and nutrient solution samples. A 25 g of the fresh produce samples were mixed with 225 mL of phosphate–buffered saline in sterilized, filtered sample bags. After homogenizing for 2 min in a stomacher and series dilution, 100 μL filtrates of the fresh produce samples or nutrient solution samples were spread onto plate count agar (BD Difco, Franklin Lakes, NJ, United States). The plates were incubated at 37 ± 2°C for 24 h and the colony-forming units (CFU) were counted. The remaining filtrates and nutrient solution samples were filtered using 0.2 μm-pored 250 mL vacuum filters (Foxx Life Sciences, Pittsburgh, PA, United States) to collect the microorganisms. The swab samples and filters were stored at –20°C for less than 2 weeks before molecular analysis.

To profile the microbiome, bacterial DNA was extracted from the swabs and filters using the DNeasy PowerSoil Pro kit (QIAGEN, Hilden, Germany), following the manufacturer’s instructions. Positive control with 12 known bacterial strains was extracted together with the regular samples. The DNA quality was evaluated by gel imaging, and concentration was determined using the Qubit dsDNA HS Assay Kit and Qubit 2.0 Fluorometer (Invitrogen, Thermo Fisher Scientific, Oregon, United States). The extracted DNA samples were stored at –20°C before further analysis.

The Wave StrainID kit (Shoreline Biome, Farmington, CT, United States) was used to sequence the amplicon that spans the full-length 16S, ITS, and partial 23S rRNA genes. The procedures were previously described by Graf et al. (2021). Amplicon libraries were created using the SMRTbell express template prep kit 2.0 (catalog number 100-938-900; Pacific Biosciences, Menlo Park, CA, United States) according to the manufacturer’s instructions. The library was sequenced on 1 SMRTBell 8M on a PacBio Sequel II system (Pacific Biosciences) using the circular consensus (CCS) sequencing mode at the University of Illinois Roy J. Carver Biotechnology Center, Urbana, IL, United States. The circular consensus reads (ccs) were determined with a minimum predicted accuracy of 0.999 and the minimum number of passes set to three in the SMRT Link software package 5.1 (Pacific Biosciences). A total of 826,974 ccs reads with a mean read length of 2,421 bp were produced using default settings.

The ccs were further processed using SBanalyzer 2.4 (Shoreline Biome) following the workflow described by Graf et al. (2021). Briefly, all reads were sorted into FASTQ files by sample with no trimming and classified by mapping to the Athena database (Shoreline Biome). The taxonomic assignments were made at >97% identity for strain level and >95% for species-level matched with reference 16S-ITS-23S rRNA sequence in the Athena database. After demultiplexing, the ccs were further processed with DADA2 (version 1.9.1) to obtain amplicons with single-nucleotide resolution (Callahan et al., 2019). ASVs were assigned eight taxonomy levels: kingdom, phylum, class, order, family, genus, species, and subspecies. The taxonomic ID and the corresponding read count for all samples were created at the end resulting in a total of 1,939 ASVs. Additionally, contaminant ASVs were detected and removed with the R package “decontam” using a prevalence-based contaminant identification with a threshold value cutoff of 0.5 (Davis et al., 2018). After decontamination, 1,096 ASVs remained.

Microbial community analysis was performed using the “Bioconductor,” “microbiome,” “phyloseq,” and “vegan” packages and visualized with ggplot2 in R (Gentleman et al., 2004; McMurdie and Holmes, 2013; Oksanen et al., 2020; Wickham; Lahti and Shetty). The dataset was transformed compositionally to visualize the phylum-level compositions and the 50 most abundant ASV genera. An untransformed dataset was used for community diversity analysis. Alpha-diversity indexes “Chao1,” “Shannon,” and “Simpson” were calculated. The normal distribution of individual alpha-diversity indexes was tested with the Shapiro–Wilk normality test. The data were evaluated for homogeneity of variance using Levene’s test. The statistical significance of alpha-diversity was evaluated using a repeated-measures ANOVA followed by Tukey’s test. The alpha-diversities were also compared by sample types or sampling location using a pairwise t-test with Benjamini–Hochberg procedure to reduce the false discovery rate. Beta diversity was visualized using principal coordinate analysis (PCoA) plots with phylogenetic-based weighted and unweighted UniFrac distance, as well as count-based Bray–Curtis distance. Permutational Multivariate Analysis of Variance (PERMANOVA) analysis was performed using the “adonis” function to measure the differences in beta diversity indexes by sampling location or type (Anderson, 2017).

The presence of pathogens in all samples was screened against the NCBI pathogen database (NCBI, 2021). The spoilage bacterial species associated with fresh produce were screened (Kaczmarek et al., 2019). The spoilage ASVs were grouped by species or genus. The reads were transformed compositionally to percentage values and visualized as heatmaps using the ‘‘pheatmap’’ function¹.

ASV1628 was marked as “E. coli unclassified” and it contains 38 sequence reads. We used it as an example to further examine potential pathogenicity. The 38 seed sequences were mapped against the SILVA databases with EMBL-EBI/ENA, GTDB, RDP, LTP taxonomy, using the SILVA Alignment, Classification, and Tree Services (ACT) to small subunit (16S) and large subunit (23S), respectively (Leinonen et al., 2011; Pruesse et al., 2012; Quast et al., 2013; Cole et al., 2014; Yilmaz et al., 2014; Chaumeil et al., 2020). The 38 seed sequences were also mapped against the bacterial whole-genome sequences in NCBI nucleotide collection (nt) using the Basic Local Alignment Search Tool (BLASTn; Zhang et al., 2000; Morgulis et al., 2008). For each seed sequence, the aligned sequences of the 10 most closely related strains with the highest identities were fetched and fast minimum evolution trees were constructed (Rzhetsky and Nei, 1993). The closely related pathogenic strains were screened and aligned with seed sequences using ClustalW in Jalview 2 (Version 2.11.1.4; Larkin et al., 2007; Waterhouse et al., 2009). After alignment, the Jukes–Cantor distances between seed sequence and fetched sequences were calculated using MegaX and with gamma parameter = 1 (Jukes and Cantor, 1969; Stecher et al., 2020). Potential pathogens are those within a 0.03 Jukes–Cantor distance of a known pathogen sequence (Ibekwe et al., 2013).

Microbial source tracking was achieved using the SourceTracker (version 1.0.1) R package with default parameters (Knights et al., 2011). The sequencing depth at 1,000 sequences/sample was chosen according to Zwirzitz et al. (2020) that 1,000 is adequate to provide a comparable result to deeper sequenced datasets. The nutrient solution and environmental samples were assigned as sources, and the fresh produce samples were assigned as sinks. The outputs were visualized using the Sankey flow diagram². The shared ASVs between different system samples were visualized using the Venn diagram³.

The overall microbial loads were quantified using aerobic plate counts. The microgreens from the hydroponic farm (farm H) and the laboratory system (L1) showed similar bacterial loads, ranging from 7.3 to 8.6 log CFU/g (Figure 1A). The lettuce samples from the aquaponic farm (farm A) had significantly lower bacterial counts (3.7–4.0 log CFU/g) than the lettuce samples from the grocery store (G, Figure 1B).

The farm H nutrient solution showed significantly higher bacterial loads than that of the L1. The bacterial loads of farm A nutrient solution were significantly lower than that of the laboratory hydroponic systems (L2). In farm A water system, the highest bacteria count was detected at the inlet from the fish tank (3.7 log CFU/mL).

The food safety awareness and practice of two farms were surveyed, and the scores are shown in Supplementary Table 3. Farm H implemented overall better food safety practices than farm A and received higher scores in equipment and environment sanitation and food safety awareness, such as sanitizer application and usage of personal protection equipment. Both farms paid minimum attention to the treatment of circulating water, which may result in the accumulation of waste and microbial hazards in the system. In fresh produce handling, farm A used tap water to rinse the seeds before starting germination, while farm H did not implement any seed treatment.

The microbiome compositions and the genera of the 50 most abundant ASVs are displayed in Figure 2 and Supplementary Table 4. Proteobacteria was the most abundant group in microgreens systems (Figure 2A). Pseudomonas was the dominant genus in the farm H, and P. alcaligenes (ASV1746), P. fluorescens (ASV1756 and ASV1758), and P. lutea (ASV1767) were among the 50 most abundant ASVs. Actinobacteria and firmicutes are resistant to sanitation treatments (Ramlal et al., 2021). They were both presented in the blank microgreen growth medium, sanitized trays, and on workers’ shoes. Cyanobacteria are the common photosynthetic bacteria in wastewater and are known as “blue-green algae.” They were commonly observed in lettuce systems (farm A and L2) as green substances on the growth trays and near plant roots. Five cyanobacteria ASVs were identified within the 50 most abundant ASVs. Cyanobacteria over-grow would produce hazardous toxins and deplete oxygen, and cause aquatic animal death (Ezenarro et al., 2021).

The microbiomes from the same sample type and sampling location were arranged closely on the tree (Figure 2A). Samples of similar types from different locations were also closely clustered with each other, such as the lettuce sample from the grocery store and the L2, and the kale microgreens from the farm H and L1.

The microbiome diversities were examined using multiple alpha-diversity (Figure 3A) and beta-diversity indexes (Figure 3B). The microgreens from farms H and L1 had similar community alpha-diversities. However, the nutrient solutions from farm H had higher community richness than that of L1. Farm H nutrient solution was shared by multiple microgreens, potentially adding to the community richness. The lettuce samples and environmental samples (shoe and tray swabs) from farm A showed the highest community richness among all samples (Figure 3A).

Permutational Multivariate Analysis of Variance test revealed significant compositional differences between different sample types and different sampling locations (Supplementary Table 5). According to the unweighted UniFrac distance, farm H microgreens had similar ASVs with worker’s hand, while its nutrient solutions had similar ASVs with shoes and tray; two lettuce samples from farm A had similar ASVs with worker’s shoes instead of other lettuce (Figure 3B left). Microgreen and lettuce samples from different locations clustered together on the weighted UniFrac plot, indicating plant microbiomes from different sampling sites had shared phylogenetic patterns (Figure 3B middle).

We screened the microbiomes for the presence of plant, human, and fish pathogens as well as spoilage organisms (Figure 4). The zero-tolerance human pathogen E. coli O157:H7, Salmonella spp., and Listeria monocytogenes were not identified in either farm. An unclassified E. coli (ASV1628) was presented on farm A worker’s shoes (Figure 4A). E. coli is used as an indicator for fecal contamination in field farming (Kaczmarek et al., 2019). According to the survey, farm A had an outdoor farm adjunct to the greenhouse, and the workers did not switch shoes between farms. Even though a sanitizer sink was presented at the entrance, it failed to eliminate the risk of bringing in contaminants from the open environment. Pseudomonas aeruginosa is a human pathogen and was abundant in farm H kale and broccoli microgreens, L2 lettuce, and on farm A worker’s shoes. Plant pathogen P. syringae was identified in low abundance on the farm H worker’s shoes. P. syringae is known to cause bacterial leaf spots on multiple cruciferous microgreen species (UMass Extension, 2018). P. putida is a beneficial organism for promoting plant growth but also an opportunistic human pathogen, and it was abundant in farm H nutrient solution. The fish pathogen Aeromonas hydrophila (ASV1571 and ASV1572) was identified in low abundance in farm A GreenCrisp lettuce but not in other samples (Figure 4). If water was circulated between the plant system and the fish tank without pretreatments, there would be a potential risk of pathogen transmission via water to fish.

Spoilage bacteria would deteriorate foods and generate unpleasant odor, taste, and texture. The major fresh produce spoilage is by lactic acid bacteria, and this group includes Lactobacillus, Leuconostoc, Pedicoccus, Lactococcus, and Enterococcus (Kaczmarek et al., 2019). Lactobacillus, Lactococcus, and Enterococcus were identified in multiple samples of farm A and nutrient solution of farm H (Figure 4B). The Pseudomonas spp. were abundant in farm H microgreens and on the sanitized trays, farm A shoes and lettuce, and G lettuce, including spoilage species of P. fluorescens, P. viridiflava, and P. tolaasii. Farm H had two systems and according to the survey, system 1 nutrient solution (H1_NS_tank1 and H1_NS_cyc1) had been circulated for almost 1 month, while system 2 nutrient solution (H1_NS_tank2 and H1_NS_cyc2) was recently changed. The spoilage Pseudomonas spp. was relatively lower in system 2 nutrient solution and kale microgreens compared to system 1 and the other five microgreens. The genus Clostridium was identified in farm A samples, and it included multiple spoilage species. However, the foodborne outbreak causing species C. botulinum and C. perfringens were not detected. The farm A worker’s shoes reserved multiple spoilage bacteria, again, indicating inadequate sanitation.

The pathogenicity of ASV1628 (“E. coli_unclassified”) was uncertain. Thus, we further mapped the 38 sequences in ASV1628 against multiple rRNA sequence databases (Supplementary Table 6). The large subunit (23S) databases (mean identity 93.2%) classified all 38 sequences as Escherichia–Shigella with 84 neighbor strains having >97% identity and 9 neighbor E. coli strains aligned 100% with the seed 23S partial sequence. Better identities were achieved using small subunits (16S) databases (mean identity 99.1%) with 130 neighbor strains having >97% identity, and some seed sequences were classified as Enterobactor, Pantoea, or Salmonella. We were not able to increase taxonomy resolution with small subunit or large subunit rRNA databases.

To further confirm the pathogenicity, we mapped 38 seed sequences against the whole genome data in NCBI and calculated the Jukes–Cantor distance between seed sequences and selected pathogenic strains (Supplementary Figure 3). Jukes–Cantor distance measures species level similarity and a distance < 0.03 was used for potential pathogen classification (Schloss and Handelsman, 2006; Bibby et al., 2010; Chidamba and Korsten, 2015). Among the 38 seed sequences, 9 sequences might come from multiple serotypes of pathogenic E. coli, Shigella flexneri, Klebsiella oxytoca, or Klebsiella pneumoniae (Table 1). One sequence (E. coli_85263602) was closely related to several strains of E. coli O157:H7.

The bacteria transmission route in the hydroponic/aquaponic systems was investigated using SourceTracker and visualized using Sankey flow diagrams (Figure 5). Environmental samples were designated as sources for testing against the bacterial communities from the fresh produce samples. In farm H (Figure 5A left), the worker’s shoes and hands contributed 21.1–89.7% and 1.5–55.1% of the microgreen microbiome, respectively. For farm A (Figure 5B left), worker’s shoe and the baby plant growth tray contributed 8.6–49.8 and 29.9–66.6% lettuce microbiome, respectively. There were 15.2–32.0% of ASVs in lettuce samples, which were not shared with any environmental samples.

When looking at the laboratory systems, L1 microgreens and L2 lettuce had 85.24–99.47% and 47–78% of the ASVs from unknown sources (Figures 5A,B, right). To better understand the source of bacteria, we sampled the microgreen growth medium separately in L1 and the root and edible part in L2. The growth medium or nutrient solution contributed a minor amount of ASVs to microgreens. The nutrient solution reservoir tanks shared large portions (66–96%) of ASVs with the lettuce root and smaller portions (0–46%) with the lettuce leaf. The ASVs from unknown sources could be part of the plant’s background microbiome.

To take a better look at the plant background microbiome, we constructed Venn diagrams to investigate the microbiome overlaps (Figure 6 and Supplementary Table 7). Farm H microgreens shared 53 ASVs with L1 microgreens and 27 unique ASVs with L1 system environment, while L1 microgreens shared 39 unique ASVs with farm H environment (Figure 6A). In the lettuce systems, there were 141 and 193 overlapping ASVs among all the lettuces (G, farm A, and L2) and between farm A and L2 (Figure 6B). Besides, the L2 lettuce shared 29 unique ASVs with farm A environment, and 46 unique ASVs from farm A lettuce also appeared in the L2 nutrient solution. This result agrees with the observation in beta-diversity that similar fresh produce samples from different systems had shared microbiome patterns. The share ASVs between lettuce samples as well as microgreens mainly belong to genera Pseudomonas and Pantoea. Pseudomonas and Pantoea were reported as part of the background microbiome of lettuce leaf (Rastogi et al., 2012). The lettuce samples also shared ASVs from Cyanobacteria that are associated with the water environment.

Thus, we further proposed a bi-direction bacteria transmission route inside the HCS (Figure 7). With good personnel hygiene, fresh produce would be the major source of bacteria. With poor personnel hygiene, contaminants would enter the system via human activities. Tools, water, and workers are vehicles for bacteria transmission, and poor facility hygiene would result in cross-contamination and accumulation of hazards.