Introduction

The increasing connectivity of the modern world, changing demographics, and climate change increase the potential for novel and known viral pathogens to emerge and rapidly spread in new and unexpected areas, as could be seen during the emergence and global threat of Ebola virus in recent outbreaks1. Early detection or ruling out of high impact (emerging) infections as causes of disease is a hallmark of preparedness, but research in response to recent outbreaks of Ebola, Zika and yellow fever has shown that these pathogens circulated for extended periods of time before being recognized, leading to costly delays in public health response2,3,4,5. One of the key challenges is how to prioritize local investments in detection capacity, given the diversity of emerging diseases, the unpredictable nature of outbreaks, and the limited resources available for outbreak preparedness. Understandably, surveillance of infectious diseases mainly targets common conditions and is scaled up in response to the emergence of pathogens and in particular disease outbreaks, rather than the costlier approach of broad range testing for any relevant infectious disease. The changing dynamics of infectious diseases related to global change, however, require rethinking of this model for public health preparedness, as incidence-based surveillance provides a fragmented and limited scope of which pathogens are circulating in the general population, particularly in low resource settings where access to healthcare and laboratory diagnostics is restricted6,7. Therefore, in its reorganization in response to the West African Ebola outbreak, the World Health Organization has launched the term “Disease X” to call for novel ideas for preparedness to unpredictable disease outbreaks8. Thus, there is a need for novel approaches to viral surveillance providing a broader and less biased insight into the circulation of viral pathogens to supplement the more targeted surveillance. Genomic epidemiology using real-time pathogen sequencing has become part of the routine toolbox for outbreak tracking once the cause of the outbreak is known9,10. In addition, metagenomic sequencing has been put forward as a potential catch-all surveillance tool, but the step from research to routine implementation is extremely challenging11,12, and thus, careful validation is needed to avoid overpromise and wasting of resources.

Here, we set out to explore the potential use of metagenomic sequencing of urban sewage as an add-on strategy for global disease preparedness. One key driver of emergence is the amplification of rare zoonotic and vector-borne diseases in densely populated regions where infrastructure needs are outpaced by rapid urban developments. This leads to the formation of slums, favorable conditions for viral disease vectors, disparity in access to clean water, sanitation and healthcare, and an increase in close human-animal interaction due to deforestation13,14. The advantage of using sewage-based surveillance is that it represents the entire population of the catchment area, sample collection is straightforward, and the anonymization by default makes it less challenging to use than patient-based surveillance regarding privacy laws. Using sewage to detect viruses with low case fatality rate but overall high population level impact has been tested successfully to monitor the progress of the global polio-elimination program, particularly in regions where non-replicating polio-virus vaccines are used15,16. The huge potential of environmental surveillance was illustrated when a silent epidemic of wild-poliovirus type 1 in Israel was detected, which led to a mop-up vaccination campaign and resolution of the epidemic, without a single case of paralytic poliomyelitis17. In addition, small-scale studies have already shown the potential for using metagenomic sequencing of sewage extracts for the detection of a range of virus families18,19,20 (Table 1 in Appendix). While these studies have largely focused on viruses with a replication phase in the gastro-intestinal tract, the fecal and/or urinary shedding of, for instance, measles virus, yellow fever virus, Zika virus, West Nile virus, Ebola virus, SARS coronavirus, and MERS coronavirus suggests the potential utility of sewage testing to capture circulation of these pathogens as well21,22,23,24,25. Moreover, metagenomic sequencing has the potential to detect any viral genomic material in the sample, without targeting a specific viral pathogen or limiting for only known viral pathogens. In this study, we pilot the use of metagenomics to describe a comparative snapshot of the virome from sewage samples of high-density urban areas across all continents. We provide a critical appraisal of technical and analytical biases and discuss the potential utility for human and animal disease monitoring and surveillance, as well as the additional steps needed to go towards routine implementation.

Results

Data quality evaluation

Urban sewage samples and associated metadata (Supp. File 1) were obtained from 62 countries across all continents between January and April 2016 from the influent of wastewater treatment plants prior to treatment or from open sewage systems in low- and middle-income countries. All samples were previously processed for the detection of bacterial antimicrobial resistance genes using DNA metagenomics26. Here we focus solely on viral DNA and RNA metagenomics (methods) and the analysis of the viral data. Sewage samples are highly variable in terms of composition and DNA abundance and therefore potential biases that might impact the final read abundance and diversity of the sewage virome were evaluated. Initially, an extensive evaluation of the technical factors that may impact the resulting data to gain a deeper understanding of potential pitfalls was performed. First, read abundance was evaluated as a proxy for viral abundance. Sequencing protocols for virome analysis in sewage typically require an amplification step to provide enough DNA input for sequencing, which can result in artificial duplication of sequence reads and thereby impact the quantitative interpretation of the data substantially (Fig. 1a). Indeed, the observed viral species richness was negatively correlated with the number of amplification cycles needed to obtain enough DNA as input for sequencing (Fig. 1b), while the average fold replication of a read was positively correlated (Fig. 1c). The impact of dereplication on the individual species level read counts varied greatly within a sample. Especially in samples with a low number of reads after dereplication (Fig. 1d) the decrease in read counts for a species ranges from 600 to fivefold . These differences have a profound effect on the species distribution within the sample, and thus the interpretation thereof. The effect of dereplication is much less variable between species in samples with a high number of reads after dereplication (Fig. 1e). Therefore, the optimal use of virome sequencing depends on the initial abundance of viral sequences in the sample and extra amplification may only increase the coverage of the same viruses, but does not increase the richness of the virome, which needs to be carefully considered when designing and interpreting sewage metagenomics studies.

Figure 1
figure 1

Effect of read preprocessing on data interpretation. (a) Number of reads before preprocessing (blue bars) after quality control (red bars) and read dereplication (green bars). The x axis shows sample identifiers ordered by number of dereplicated reads. (b, c) Effect of number of PCR replication cycles on library concentration (color), species diversity (b) and read replication rate (c). (d, e) Fold replication of raw reads by species level annotation (points). X axis separates superkingdom or “Unknown” annotations. (d) shows sample LVA_31 with a high replication rate and panel e shows sample MLT_63 with a low replication rate.

Full size image

Besides the influence of read replication on read abundance, the richness of the virome can be impacted by the presence of non-viral sequences. Typically, the metagenomic data contain a large fraction of unknown reads, and, despite the virus specific sample preparation, non-viral reads, including archaeal, bacterial, and eukaryote DNA.

While the overall proportion of reads for the different domains was comparable in most samples, multidimensional scaling of the non-viral read counts showed that some samples were very divergent from the central cluster and were manually marked as outliers (Fig. 2a, dashed line). Viral read abundance was low in these outlier samples (Fig. 2b, right panel). There was no significant correlation between the concentration of human or bacterial read fractions with any of the measured sample characteristics, such as pH, conductivity, and type of sewer system.

Figure 2
figure 2

Effect of non-viral background read abundances on viral read abundance and the chosen outlier samples in the sewage metagenome data. (a) A multidimensional scaling of Bray–Curtis dissimilarity between samples based on the normalized read counts of bacterial, archaeal, eukaryote (human), and viral content. “Unknown” indicates reads that could not be assigned any annotation. The red labels indicate the effect of the different annotations on the position of a sample in the plot. Gray circle indicates the samples that were manually assigned to be outliers. (b) A scaled bar chart of relative read abundance showing the outliers in a separate facet to the right.

Full size image

Exploration of the sewage virome

Based on the data quality assessment, we analyzed viral diversity in the samples after dereplication and following annotation by both Kajiu and Centrifuge as described. Between 0.09% and 22% of the reads could be annotated as viral (median of 6%), with high abundances of bacteriophages, plant- and insect viruses (Fig. 3). Most abundant were bacteriophages, representing on average 77% (ranging from 9 to 94%) of the annotated viral reads in the sewage. In particular Microviridae (median of 18%, range 0.5–51% of reads), Siphoviridae (median of 17%, range 0.22–67% of reads), Myoviridae (median of 9%, range 0.08–41% of reads), and Podoviridae (median of 4%, range 0.02–25% of reads), were highly abundant. These bacteriophage families could be found around the globe without obvious regional differences when using read annotations at this taxonomic level. Although specific bacteriophages have been studied extensively as potential indicators of human fecal pollution, bacteriophage taxonomy is relatively poorly defined, making accurate classification challenging at genus and species level27,28. Hence, geographical patterns at a more fine-grained level of annotation may be lost in our analysis. Moreover, interpretation of patterns of bacteriophage abundance could be obscured by the fact that bacteriophages can encounter bacterial hosts in the sewage in which they can multiply. As described elsewhere, the analysis of the bacterial resistomes of the same samples showed clear segregation of sequences from Africa and Asia versus those from Europe and the US26. A more detailed analysis is needed to assess if there is a relation between specific bacteriophages and the resistomes, as environmental viromes have been shown to be a potential reservoir for antimicrobial resistance genes29.

Figure 3
figure 3

Heatmap of the viral diversity at viral family level (when available) and non-viral fraction. The read abundance after quality control and dereplication is shown ordered by total read abundance after preprocessing and facetted by continent. The heatmap follows the same ordering. Color gradient represents log-transformed relative abundance of reads belonging to the taxonomic groups indicated. The top four rows of the heatmap show read abundances of non-viral annotations, the other rows show read abundance by viral family, or “no family” if only genus or species level annotation was available. Vertical facets represent subdivision of the viral families based on their inferred host. Black arrows indicate outlier samples based on an overabundance of background sequences.

Full size image

Global patterns of viruses related to vegetable consumption and to insects detected in urban sewage

The second largest fraction of the virome (0.02–69%, median 3.4%) consisted of plant-related viruses. On average, more than 84% of these reads belonged to the Virgaviridae family. Especially viral species related to infections of cucumber, tomato, tobacco and pepper plants could be detected in sewage, as indicated by species level taxonomy (Fig. 4b). Apart from a sample from Kenya, the abundance of vegetable-consumption-related viruses was higher in samples from Europe and North America compared to samples from the rest of the world (Fig. 4a) (Welch’s t-test, p-value = 0.06). The global presence and high abundance of plant viruses has led to the proposal that they may be good indicators for human fecal contamination alike specific bacteriophage populations30. However, this remains to be validated given the geographic variation observed in our dataset, which could reflect differences in diet and/or agricultural practices in these countries.

Figure 4
figure 4

Overview of the global distribution and abundance of plant viruses and insect viruses in urban sewage (a) Global distribution of all plant viruses (b) The four most abundant plant virus species and their global spread. (c) Global distribution of all insect related viruses. (d) Top 5 most abundant insect virus genera. Datapoints represent absolute read numbers and read fraction by varying size and color respectively. Viral species are ordered by summed read abundance across samples and samples are ordered by total read abundance from left to right. Facets represent continent of sample origin.

Full size image

A median of 1.4% (ranging 0.1–74%) of the sewage virome consisted of viruses associated with insects, comprising mainly species from the genera Ambidensovirus, Cripavirus, and Brevidensovirus (Fig. 4d), known to infect a range of crickets, cockroaches, fruit flies, and mosquitos31. In the global distribution there was an increased abundance of insect viruses in samples from around the equator, mainly in samples from Africa (Fig. 4c) (Welch's t-test, p-value = 0.0004). One exception was the sample from Finland, which had a high abundance of insect virus reads (13.7%) in comparison with samples from other European countries (1.5%). Several reads were found to be annotated as “Aedes albopictus densovirus 2”, “Aedes aegypti Thai densovirus”, and “Anopheles gambiae densonucleosis virus”. There is some evidence that these densoviruses may be associated with Aedes aegypti, Aedes albopictus and Culex mosquitos32,33. Current data are not sufficient to meet the requirements for sewage surveillance, but these findings show the potential to track mosquitos by looking for mosquito specific viruses.

Detection of vertebrate viruses and investigation of known human pathogens

About 1.7% (ranging 0.01–11%) of the virome consisted of vertebrate viruses. Most abundant were small ssDNA viruses from the families Circoviridae and Parvoviridae, and members of the Picornaviridae, Astroviridae and Adenoviridae families (Fig. 5a). Vertebrate viruses were detected widely across the samples, but did not show distinct geographical patterns of abundance. Circoviruses were especially highly abundant across most sewage samples and, as novel variants of circoviruses have been associated with several diseases in pigs34. Further longitudinal sewage surveillance could potentially be used to detect epidemiological patterns of emerging circovirus variants.

Figure 5
figure 5

Overview of the most abundant vertebrate viruses and specific human viruses and their distribution worldwide in urban sewage. (a) Distribution of the top ten most abundant vertebrate viral families. (b) Relative abundance of viruses encountered in clinical surveillance (c) World maps showing distribution of viruses encountered in clinical surveillance. Coloring of the maps delineates differences in climate by geographical location. Datapoints represent absolute read numbers and read fraction by varying size and color respectively. Viral families are ordered by summed read abundance across samples and samples are ordered by total read abundance from left to right. Facets represent continent of sample origin.

Full size image

A selection of viral taxa was analyzed containing human pathogenic viruses from the Astro-, Entero- Noro-, Sapo-, Adeno- and Rotaviridae families that are known to be abundant across the world as causes of diarrheal disease (Fig. 5b). Most abundant and widespread were the astroviruses. Enteroviruses were present to a lesser extent but could be detected in sewage samples from across the globe as well. Members of the noro-, sapo-, adeno-, and rotaviruses were only sporadically detected. Further investigation of samples with high human astrovirus content showed mostly evidence of the classic Human Astrovirus 1, 2 and 4 that are common causes of diarrheal disease, and sporadic detection of other clades such as Human Astrovirus MLB and Human Astrovirus VA for which less is known regarding clinical impact35. Mapping of human enterovirus reads resulted in 102 small contiguous sequences which were typed using the enterovirus typing tool36. Mainly Enterovirus C (46%) and B (9%) were detected. Further subtyping of for instance poliovirus was not possible because of a lack of coverage of the standardized genotyping region VP1. The same mapping was done for norovirus, resulting in 13 contigs of 84 to 962 nucleotides in length. Most norovirus sequences were typed as either GII, with capsid type 6, 10 and 17, and GIV, all viruses that are commonly found in outbreak based surveillance37. Sapovirus sequences, all belonging to type GI, were found in seven of the samples. Adenovirus and rotavirus hits were sporadically detected across all sampling sites and upon further investigation showed mainly adenovirus C and rotavirus A hits.

It is known that noroviruses, astroviruses and rotaviruses follow a winter seasonality and enteroviruses follows a summer seasonality pattern38,39,40. The time of sampling of the sewage was in a 3-month timeframe between January and March, which corresponds to the winter period in the northern hemisphere, therefore a higher prevalence of winter seasonal viruses was expected in those. When looking at the global distribution of viruses, the average abundance of astro- and noroviruses was higher in the northern hemisphere, and the reverse pattern was observed for enteroviruses, with higher average abundance in the southern hemisphere during the sampling period (Fig. 5c). Given the cross-sectional nature of our study we acknowledge that these seasonal patterns will have to be confirmed using longitudinal sampling which would allow for meaningful statistical analysis, but our first observations align with what is generally expected at that time of the year.

Discussion

This global sewage study gives, for the first time, a catch-all metagenomic comparison of the urban sewage virome of major cities across the world. We show that it is possible to detect a wide diversity of viruses in sewage samples and we identify geographical and seasonal differences in abundance for specific viral groups, including those that are currently targeted by surveillance for diarrheal and neurological disease, as well as viruses that could be used as indicators for presence of specific mosquito species. In addition, we provide the global scientific community with a geographically very broad resource for searching for novel virus sequences as novel pathogens continue to emerge. The pilot study also highlights some important challenges that need to be addressed to take the technology forward, such as how to deal with low input samples and the overabundance of phages, plant, and insect viruses in the sample. Metagenomic sequencing of viruses is a complex and evolving technology which is currently far from being standardized. Differences in sample preprocessing, sequencing technology, and data analysis can have a major impact on the viral read abundance, diversity, and the proportion of sequences that are annotated41,42. In our study, we eliminated lab-to-lab variability by performing all sample preparation, sequencing and analysis at the same location, which, apart from the analysis, is obviously not feasible for global surveillance. Further work is ongoing, including the development of fieldable sample treatment and sequencing protocols, comparison of effects of sample preparation on viral richness and further exploration of applicability, by longitudinal sampling and sampling in the presence of known ongoing outbreaks.

A critical challenge of using metagenomic sequencing for surveillance purposes remains the interpretation of sequence annotations. With the development of high-speed k-mer based annotation tools such as the ones used in this study, annotation can be performed rapidly and with few false negatives. However, erroneous and mis-annotated entries in public databases, together with inconsistency in the sequence-based taxonomic classification of viruses, make annotation to the species level challenging. Major steps have been taken to create a more consistent sequence based viral taxonomy27,43, but these approaches have not yet been integrated in fast viral annotation tools. Also, deposits of large volumes of virus sequences without a clear host association or pathogenicity data in public databases44 make it difficult to interpret the relevance of such findings. In our data, many of these “environmental viruses” could be identified. Given the increase in virus diversity in reference databases, it is striking how many sequence reads can remain unclassified with the currently used methods. This is in line with previous observations, where 40–90% of the sequence reads could not be classified45. It can very well be that the currently unclassified sequence reads represent potential new viruses, including novel pathogens.

In conclusion, we show the potential of global viral surveillance using metagenomic sequencing of sewage without ignoring the complexity of the approach. However, with improvements in sample preprocessing, sequencing methods and interpretability of viral sequence annotation this potential will increase.

Methods

Urban sewage sample and metadata collection

Samples were obtained from 62 countries from all continents as previously described26. All samples were taken before wastewater treatment. A questionnaire was filled in with information on sampling site, sample consistency and sample temperature, including transport time, storage time, and temperature before shipping. All samples were taken in a timeframe of 3 months from January until March 2016. In addition to sample specific data, additional metadata (Supp. File 1) was collected such as demographics, type of industry in the surrounding area, weather conditions and catchment area of the sewer. Upon arrival, samples were thawed at room temperature and 250 ml of the raw sewage was taken and centrifuged at 10,000 g for 10 min. The pellet was removed for bacterial content determination and DNA metagenomic sequencing26 and the supernatant was used to perform the virus specific sample pretreatment and sequencing.

Sample processing for sequencing

Viral extraction was performed on 40 ml of sewage supernatant as previously described46. In short, the conductivity was measured to exceed 2000 µs and the pH of the samples was adjusted to pH 4. Afterwards 10 ml PEG 6,000 was added and the samples were incubated overnight at 4˚C under agitation.

After incubation the samples were centrifuged a 13,500 g for 1.5 h at 4 °C. The supernatant was removed, the pellet was dissolved in warm glycine buffer and 1 mL of chloroform-butanol (50/50) was added. After mixing, the sample was centrifuged for 5 min at 13,000 g at 4 °C. The filtrate was collected through a series of filters with 5 µm, 1.2 µm, 0.45 µm and 0.22 µm pore size.

Unprotected free DNA was removed by incubation with Ambion Turbo DNase for 30 min at 37 °C. Total nucleic acid content was extracted using Roche NA isolation kit and cDNA was made using superscript III (Invitrogen) using random hexamers that avoids amplification of human rRNA47. dsDNA was made using Klenow (NEB) and samples were sheared using Ion Shear Plus Enzyme Mix II. Libraries were amplified for 15 cycles using High Fidelity Platinum PCR reaction. The library concentration was determined using Ion Torrent quantification kit (Thermo Fisher). If the concentration was below 20 nM, extra amplification cycles were performed. Sequencing was performed on the Ion Torrent S5XL platform to generate around 10 million sequence reads per sample.

Data preprocessing

Raw fastq files were quality trimmed using FastP48. Read ends were trimmed to mean quality 25 with a sliding window of 5. Reads were trimmed to 400 nucleotides by default because the chemistry of Ion Torrent sequencing technology allows for reads of maximally 400 nucleotides long and longer reads were observed to contain high Phred score non-sense repetitive patterns in the tail region. Reads shorter than 50 nucleotides were discarded as well as reads with an average Phred score below 25. Duplicate reads were removed using CD-HIT49 by clustering reads that start at the exact same position in the genome and have over 90% sequence identity in the first 50 nucleotides of the read, because of variable read length and observed insertion and deletion errors in the beginning of the reads.

Read based analysis

Due to the expected high diversity of viruses present in the sewage samples, a read based annotation of the data was chosen, contrary to an assembly-based approach. Annotation was performed using two taxonomic annotation tools: Kaiju50 and Centrifuge51. Kaiju performs taxonomic annotation based on an amino acid (AA) level which provides a higher sensitivity. This is especially important for the annotation of viral sequences given the high mutation rate of viruses52 compared to other organisms. In parallel with Kaiju, Centrifuge was run, which uses nucleotide (nt) identity for taxonomic annotation. Combining a nucleotide and an amino acid based matching approach ensures that both coding and non-coding read sequences can be annotated. In addition, the combination of two read annotation tools with different annotation strategies was chosen to give more robust mapping results.

The databases used for taxonomic annotation consisted of archaeal, bacterial and human RefSeq sequences and were extended with all viral and phage entries in GenBank version 23053 because of the limited viral and phage sequence diversity in the RefSeq database.

Recommended quality thresholds and parameters for metagenomic data were used for both Kaiju and Centrifuge. Kaiju was run in greedy mode with a score cutoff of 70 and an error of 5. Centrifuge was run with a score threshold of 300 and a hit length cutoff of 50. If neither method produced a hit the read was annotated as “Unknown”. BASTA54 was used to determine the last common ancestor (LCA) of each hit given by both methods without restrictions on hit quality.

The final read counts passing QC were determined by the sum of read annotations at a certain taxonomic level and were normalized by total dereplicated read count to adjust for differences in sequencing depth and data quality55,56,57. The LCA taxon was used if the annotation at a certain taxonomic level was absent. Manual regrouping of taxonomic levels was performed to calculate read counts of human pathogenic viruses and read counts by host group. For sample comparison, read counts were normalized by Hellinger transformation58. Sample-wise comparison was done by calculating the Bray–Curtis dissimilarity between the normalized read counts using the R package Vegan59. Further investigation of the annotation of specific viral species was performed by mapping the reads against a redundant set of reference genomes using KMA with default parameters60. The maps of global read distribution were created using the continent subdivision from the “rnaturalearthdata” R package and the Köppen-Geiger climate classification61.