Evaluation of full-length nanopore 16S sequencing for detection of pathogens in microbial keratitis

Introduction

Microbial keratitis is a leading cause of preventable blindness worldwide and is the most common cause of acute medical ophthalmology admission (Centers for Disease Control and Prevention, 2014). Bacterial keratitis accounts for the majority of the microbial keratitis cases in the Western hemisphere, with a preponderance for Gram-positive organisms such as Staphylococcus aureus and Streptococcus pneumoniae, and to a lesser extent Gram-negative organisms such as Pseudomonas aeruginosa and Klebsiella pneumoniae (Ting et al., 2018; Tan et al., 2017; Ibrahim, Boase & Cree, 2009; Lichtinger et al., 2012). Prognosis is dependent on early identification of the causative organism and initiation of appropriate treatment, including antibiotic therapy (Austin, Lietman & Rose-Nussbaumer, 2017). Conventional sampling and culture technique is time-consuming, with culture results taking 48 hours and antimicrobial sensitivity results taking up to 5 days (Maurer et al., 2017). Approximately 40% of clinically suspected microbial keratitis cases are culture-negative (Sugita et al., 2013; Tananuvat et al., 2012), leading to the widespread initial implementation of empirical broad-spectrum antimicrobial therapeutic protocols, increasing the risk of antimicrobial resistance and poorer patient outcomes (Goldstein, Kowalski & Gordon, 1999), particularly in patients who do not have ready access to diagnostic laboratories, such as those in poor and low income countries and deployed military personnel (Musa et al., 2010).

Culture-independent molecular techniques such as polymerase chain reaction (PCR) has been utilised in the diagnosis of ocular infections (Kim et al., 2008), however, this requires a priori knowledge of the likely pathogenic micro-organism to determine which specific primer sets to use and is limited by the number of species that can be detected simultaneously in a single PCR assay (Bispo et al., 2018; Ung et al., 2020). Newer, high-throughput sequencing approaches can be broadly classified into two major options: targeted amplicon sequencing (selective amplification of the specific genetic region of interest such as 16S rRNA in bacteria or 18S rRNA in fungi) and metagenomic sequencing (untargeted amplification of all genomic DNA) (Ung et al., 2020). Untargeted metagenomic sequencing allows for the discovery of unexpected, novel pathogens, such as Vittaforma corneae in infectious conjunctivitis (Lalitha et al., 2019) and Torque teno virus in culture-negative endophthalmitis (Lee et al., 2015). Deep metagenomic sequencing techniques have enabled phylogenetic analysis of the temporal and geographic origin of ocular infection (Doan et al., 2016b; Kirstahler et al., 2018).

Several challenges have hindered the adoption of sequencing technologies in routine clinical practice for microbial keratitis, including sample collection methods that are frequently ineffective (Kaye et al., 2003) compounded by the low abundance of pathogens (Ung et al., 2020), high contamination from background host DNA or laboratory reagents (Gu, Miller & Chiu, 2019), lack of standardisation in sequencing and bioinformatics processing methods (Chiu & Miller, 2019). Sampling and DNA extraction methodology significantly impact upon the downstream sequencing data in samples with low biomass (Douglas et al., 2020; Sui et al., 2020). Microbial cells within the sample must be sufficiently lysed to liberate its DNA content, and this is particularly challenging for thick-walled microorganisms whereby mechanical disruption method such as bead-beating in conjunction with heat, chemical or enzymatic treatment is required in the DNA extraction protocol (De Boer et al., 2010; Ojo-Okunola et al., 2020). Clinical samples have high host DNA content, usually constituting more than 90% of the sequences, with relatively low abundance of pathogen DNA, and therefore requiring greater depth of sequencing (Ung et al., 2020; Lalitha et al., 2019). The prohibitively high running costs of large sequencing platforms mean that they are only available in select centres, and samples are generally pooled for batch processing, resulting in delayed turnaround time. Efforts have been directed to reduce the amount of host DNA present in clinical samples, especially for samples that can be obtained in abundance, such as saliva, sputum, and bronchoalveolar lavage and blood (Charalampous et al., 2019; Marotz et al., 2018; Feehery et al., 2013). However these techniques are still not compatible with a generic method that can be applied across all the different types of clinical samples.

Marker genes using conserved, housekeeping regions of the genome interspersed with variable regions have been utilised to infer phylogenetic links and microbial taxonomy (Clarridge, 2004; Raina et al., 2019). The universal 16S ribosomal RNA (rRNA) gene sequence in bacteria, which is approximately 1,550 base pairs long, composed of a highly conserved region interspersed with nine variable regions (V1-9), is the most commonly used marker gene for assessing bacterial profiles (Clarridge, 2004; Akram et al., 2017; Achtman & Wagner, 2008). Amplification of the 16S rRNA region by PCR reduces the background host contamination and by only sequencing a smaller part of the genome, this drastically reduces the sequencing depth requirements and cost. The vast majority of clinical studies have only sequenced part of the 16S gene, ranging from the single variable region of V4, V6 to three variable regions V1-3 or V3-5, because the widely used Illumina sequencing platform is only capable of producing short reads of less than 500 bases (Johnson et al., 2019; Holm et al., 2019). This short-read sequencing further compounds the taxonomic resolution of 16S rRNA sequencing, which is typically limited to genus-level resolution (Achtman & Wagner, 2008). Choice of the hypervariable region affects the taxonomic assignment, for example the V4 hypervariable region providing better whole bacterial diversity in human gut microbiome studies whereas the V1-V2 hypervariable region is more specific for skin microbiota profiling (Santos et al., 2020). Several studies have shown that full-length 16S rRNA reads provides better taxonomic resolution compared to reads that only target a certain region of the 16S rRNA gene (Winand et al., 2020; Nygaard et al., 2020). The choice of 16S reference database also impacts upon the taxonomic assignment (Nygaard et al., 2020; Park & Won, 2018; Szabó et al., 2016). For example, the expanded Human Oral Microbiome Database contains references to microbes specifically from the aerodigestive tract whilst the National Center for Biotechnology Information (NCBI) 16S RefSeq, the Ribosomal Database Project (RDP) and the SILVA rRNA databases include taxa from all sources (human and non-human hosts) and the environment (RefSeq, 2020; Cole et al., 2014; Quast et al., 2013; Escapa et al., 2018).

With the introduction of long-read sequencing technologies, such as the Pacific Biosciences (PacBio) and Oxford Nanopore Technologies (ONT) sequencing platforms, we are now able to sequence in real-time the full-length of the 16S rRNA gene (V1–V9 regions) (Johnson et al., 2019; Winand et al., 2020). Major advantages of using nanopore sequencing include portability (pocket-sized ONT MinION sequencer vs large PacBio sequencer), relatively low cost (~£140 per sample by multiplexing 12 samples per run and £820 for the ONT MinION vs ~£280 per sample and £390,000 for the PacBio Sequel II System) and rapid workflow (less than an hour from sample preparation to data analysis with nanopore sequencing vs 3.5 hours with PacBio) (De Maio et al., 2019; 16S Sequencing and Analysis, 2020; Ashton et al., 2015; PacBio Sequel Systems, 2020; Workflow-PacBio, 2020). Nanopore sequencing has been utilised for on-the-field, real-time genomic surveillance of Ebola (Quick et al., 2016) and Zika (Quick et al., 2017; Faria et al., 2017) viruses. In addition, nanopore sequencing has also been used for pathogen detection in proof-of-concept clinical studies of sepsis (Leggett et al., 2020), lower respiratory tract (Charalampous et al., 2019; Yang et al., 2019), urinary tract (Schmidt et al., 2017) and prosthetic joint infections (Sanderson et al., 2018). However, there is relatively high error rates in nanopore sequencing (~95% raw read accuracy for nanopore) (Rang, Kloosterman & De Ridder, 2018), which affects the discriminatory power of 16S rRNA gene for species level classification (Winand et al., 2020). Strategies to reduce nanopore sequencing error rate are constantly evolving with improvements to the pore chemistry and basecalling software (Rang, Kloosterman & De Ridder, 2018).

The aim of this study was to evaluate the potential of full-length 16S nanopore sequencing directly from clinical samples of bacterial microbial keratitis, focussing on the optimisation of the DNA extraction and bioinformatics pipeline to make 16S nanopore sequencing feasible. Firstly, using a defined mock microbial community, we evaluated and benchmarked our bioinformatics analysis pipeline for taxonomic assignment. Then, we optimised the sample collection using an ex vivo porcine model of microbial keratitis. Next, we performed a proof-of-concept study in which we applied the sampling technique that provided the highest DNA recovery and the optimised bioinformatics pipeline to prospectively collected samples from patients with suspected microbial keratitis, comparing the nanopore sequencing results to standard microbiology culture methods (Fig. 1).

Materials and Methods

Results

Optimising sampling technique for collection of microbial keratitis on ex-vivo porcine model

To optimise the sampling technique on ex-vivo porcine model of microbial keratitis, comparisons were made between the different collection materials, both pre- and post-inoculation. Pre-inoculation, the average C_T of 16S was not found to differ significantly between the collection materials (p = 0.909) (Table S3). Post-inoculation, the DNA concentration was found to differ significantly across the collection materials (p < 0.001) (Table S4; Fig. S3). The highest concentrations were observed for the Sugi^® Eyespear and Isohelix^™ SK-2S, with medians of 82.2 and 96.6 ng/ul respectively, whilst the lowest concentrations were observed for the PTFE 4mm, at a median of 0.19 ng/ul. For the average C_T, no significant difference across the collection methods was detected for 16S (p = 0.242). The 16S: β-Actin ratio was consistent across all of the collection materials (p = 1.000).

Six out of the twelve collection methods provided sufficient DNA yield for 16S Nanopore sequencing and 16S copy number quantification—Sugi^® Eyespear, Isohelix^™ SK-2S, MW1021D Dryswab^™ Polyester, MW130 Hospiswab^™ Albumin, MW1041 Cotton and MW1021 Dryswab^™ Rayon (Fig. 2). All three species of the mock microbial community inoculated onto the ex-vivo porcine eyes were detectable on nanopore sequencing. The Sugi^® Eyespear swab was chosen for patient sample collection as it provided the highest mean rank DNA concentration that was sufficient for nanopore sequencing.

Discussion

We have developed a proof-of-concept study optimising the sample collection method and downstream bioinformatics pipeline for full-length 16S rRNA gene identification by nanopore sequencing in the setting of microbial keratitis.

The use of different swabs and collection methods for microbial keratitis has a direct effect on the DNA yield—the Sugi^® Eyespear and Isohelix^™ SK-2S swabs provided the highest DNA concentration. However, the ratio of host to microbial DNA recovery is similar across all collection methods. Differences in the DNA yield could be explained by the absorption efficacy of the swabs (Bruijns, Tiggelaar & Gardeniers, 2018). The absorption capacity of the swab materials is dependent upon the swab tip dimensions and morphology of the sorbent material—how tightly wound the sorbent material is to the shaft (Bruijns, Tiggelaar & Gardeniers, 2018). The Sugi^® Eyespear swab, primarily designed for use in ophthalmic theatres due to its high tensile strength in absorbing fluids, has been proven to be effective in recovering DNA from corneal tissue in our study. Previous studies have also demonstrated that DNA recovery is inversely proportional to the fiber density on the swabs (Brownlow, Dagnall & Ames, 2012; Verdon, Mitchell & van Oorschot, 2014). Both Sugi^® Eyespear and Isohelix^™ SK-2S have swab tips made of cellulose fiber, which have high DNA-binding capacity (Su & Comeau, 1999).

Choice of primers affects the amplification efficiency of the 16S rRNA gene region, with primers that amplify the entire 16S rRNA gene spanning the V1-9 variable regions (27F and 1492R) providing better classification of reads compared to primers that only amplify a portion of the 16S rRNA region, as shown in studies by Winand et al. (2020) and Nygaard et al. (2020). Therefore, we have used the 27F and 1492R primer pairs provided in the commercially available ONT Rapid 16S barcoding kit. Other preliminary studies have also shown that sequencing the whole rrn operon (~4,300 bp), which includes the 16S rRNA gene, internal transcribed spacer (ITS) region, and the 23S rRNA gene, may provide better taxonomic resolution (Cuscó et al., 2019). However, as there is a lack of updated and curated rrn operon databases, users will need to retrieve and compile whole ribosomal operon reference database for their specific usage (Benítez-Páez & Sanz, 2017), whereas, curated or updated 16S rRNA reference databases are more readily available (RefSeq, 2020; Quast et al., 2013).

Another factor influencing the quality of nanopore sequencing is the choice of basecalling software (Rang, Kloosterman & De Ridder, 2018). We used the Guppy ‘flip-flop’ high accuracy model for basecalling, based on Wick et al’s study showing that the model, which utilises the recurrent neural network algorithm, performed better than the other basecalling programs (Albacore, Scrappie and Flappie) (Wick, Judd & Holt, 2019). Using Minimap2 (Li, 2018), we aligned our nanopore sequencing reads against the three different publicly available databases, NCBI 16S RefSeq, RDP and SILVA. Performance matrix comparison of thirteen different classification tools by Urban and colleagues revealed that Minimap2 provided robust alignments that were closely aligned to their mock community taxa (Urban et al., 2020). However, similar to the challenges encountered by Urban et al. (2020), we have had issues of high memory usage on Minimap2, which necessitated a reduction in the number bases loaded into memory to process in the query batch (command -K 100M). By comparing against a defined mock community, we observed differences in the taxonomic assignments between the databases—with the NCBI 16S RefSeq database clustered at 100% providing the most accurate assignments, which could be attributed to the differences in the database size and sequence validation steps (Balvočiūtė & Huson, 2017; Park & Won, 2018). The NCBI 16S RefSeq database is manually curated and near full length 16S sequences are preferentially selected (RefSeq, 2020). In our benchmarking steps, we have used the CD-HIT program (Li & Godzik, 2006), which employs heuristic greedy incremental clustering algorithm to cluster the reference sequences of the databases into 97%, 99% and 100% similarity to approximate taxonomic assignments. Schmidt, Matias Rodrigues & Von Mering (2015) assessed different clustering programs (hierarchical and heuristic algorithms) and observed that the CD-HIT program was robust, computationally efficient and provided reproducible clusterings. In our study, we have demonstrated that the 100% identity threshold provided a more optimal OTU assignment compared to 97% or 99% identity thresholds, which is consistent with previously published studies (Edgar, 2018; Mysara et al., 2017).

Nanopore sequencing reads were in concordance with current gold standard clinical microbiology culture techniques in Patient A. In the case of Patient B where cultures were negative, in the setting of previous antibiotic use, the nanopore sequencing result was in agreement with Illumina short-read sequencing suggesting an identification of a putative organism in the context of a false negative culture. The human ocular surface is paucibacterial. The conjunctival microbiome predominantly consists of Corynebacteria, Propionibacteria, Staphylococcus, and Streptococcus—with ‘approximately 1 bacterium for every 20 human conjunctival epithelial cell collected on conjunctival swab’ (Doan et al., 2016a). It is likely that the presence of Serratia marcescens and Bacillus subtilis at such high relative abundance in the patient samples, with correlating clinical signs and symptoms of infective keratitis, would constitute a positive test. A major challenge in applying high-throughput sequencing in clinical practice is distinguishing between true polymicrobial keratitis, ocular commensal or contaminant (Ung et al., 2020). Hence, we have taken meticulous steps to account for any potential contamination in our study—swabs of the unaffected contralateral eye and negative control swabs at the same time point and clinical environment of the patient had been taken, processed and sequenced in the same manner as the swabs of the affected eye. These negative control swabs had significantly fewer reads, less than 1% of the reads from the affected eye.

Another challenge in using full-length 16S rRNA sequencing is the difficulty in differentiating species and subspecies strains in certain bacterial genera with high sequence homology and similarity, notably the Bacillus subtilis species complex (Public Health England, 2018). This is illustrated in Patient B, whereby different taxonomic species within the Bacillus subtilis group such as Bacillus amyloliquefacians, Bacillus licheniformis, Bacillus velenzensis and Bacillus halotolerans have been assigned. Classical phenotypic tests by colonial appearance on culture, presence of ß-haemolysis, or biochemical tests to discriminate between these subspecies are unreliable. Other laboratory identification methods such as Matrix-Assisted Laser Desorption Ionisation–Time of Flight mass spectrometry to detect microbial protein composition is highly variable dependent upon endospore formation (Shu & Yang, 2017), whilst Multilocus Sequence Typing, which relies on PCR amplification and sequencing of six or seven well-conserved, housekeeping genes within the bacterial genome, is unable to provide distinct phylogenetic typing of Bacillus owing to the difficulty in designing primers for genetic sequences with high similarity. Although Pulse Field Gel Electrophoresis, which utilises endonuclease restriction enzymes, subsequent separation of DNA fragments by gel electrophoresis and staining under ultraviolet light for bands, is highly discriminatory, it is time consuming (over 30 h) and requires specialist laboratory equipment (Public Health England, 2018). Hence, Public Health England advocates initial clustering and genus identification by 16S rDNA for Bacillus, which could then be followed by more in-depth whole genome sequencing for more accurate strain characterisation (Public Health England, 2018).

Nanopore sequencing technology has the potential to provide rapid, real-time diagnosis of causative pathogens in a healthcare setting, with relatively low cost. Although in this proof of concept study, we have used 16S primers which specifically amplify bacterial DNA, non-biased deep metagenomic detection of pathogens and its antimicrobial resistance genomes from cultured clinical isolates using nanopore sequencing have previously been demonstrated (Schmidt et al., 2017; Szabó et al., 2016). Direct RNA sequencing on the nanopore platform would also enable identification of ‘live’ pathogens and host gene profiling for transcriptome signatures related to the infection (Lalitha et al., 2019). However, challenges still remain in terms of limiting background contamination (Glassing et al., 2016), reducing the error rates of sequencing and improving base-calling algorithms (Wick, Judd & Holt, 2019). Performing molecular diagnostics on ocular samples is inherently more difficult compared to other types of clinical samples, such as blood, urine or cerebrospinal fluid, as ocular samples are many magnitudes smaller in volume (Doan et al., 2016b) and are highly abundant in human cells (Lalitha et al., 2019).

This is the first study to evaluate the use of full-length 16S nanopore sequencing for detection of pathogens in microbial keratitis. We have optimised collection methods and demonstrated the bioinformatics pipeline for bacterial microbial keratitis. Our study is limited by the small sample size of patient cohort and the use of 16S rRNA primers, which specifically amplifies bacterial genome. To resolve this, larger clinical sample studies involving unbiased metagenomic sequencing are required to determine the sensitivity and specificity, as well as the cost effectiveness of nanopore sequencing.

Conclusion

We have optimised collection methods and demonstrate a novel workflow for identification of bacterial microbial keratitis using nanopore sequencing.

Supplemental Information