Re-analysis of metagenomic sequences from acute flaccid myelitis patients reveals alternatives to enterovirus D68 infection

Metagenomic sequence data can be used to detect the presence of infectious viruses and bacteria, but normal microbial flora make this process challenging. We re-analyzed metagenomic RNA sequence data collected during a recent outbreak of acute flaccid myelitis (AFM), caused in some cases by infection with enterovirus D68. We found that among the patients whose symptoms were previously attributed to enterovirus D68, one patient had clear evidence of infection with Haemophilus influenzae, and a second patient had a severe Staphylococcus aureus infection caused by a methicillin-resistant strain. Neither of these bacteria were identified in the original study. These observations may have relevance in cases that present with flaccid paralysis because bacterial infections, co-infections or post-infection immune responses may trigger pathogenic processes that may present as poliomyelitis-like syndromes and may mimic AFM. A separate finding was that large numbers of human sequences were present in each of the publicly released samples, although the original study reported that human sequences had been removed before deposition.

Metagenomic shotgun sequencing, in which DNA or RNA is extracted from a tissue sample and then sequenced, has the potential to detect a wide range of infections. Deep whole-genome shotgun (WGS) sequencing can detect bacteria, viruses, and eukaryotic pathogens with equal effectiveness, as long as the infectious agent is similar to a species that has been previously sequenced. Sequencing databases already contain thousands of known species, and as this number grows, the sensitivity of WGS will grow as well.
In 2014, a large outbreak of infection with enterovirus D68 was associated with both severe respiratory illness and acute paralysis, which the U.S. Centers for Disease Control and Prevention (CDC) named acute flaccid myelitis (AFM) 1 . Samples collected from 48 patients were sequenced and shown to form a novel strain, Clade B1, based on phylogenetic analysis of 180 complete enterovirus D68 sequences 2 . The same study conducted metagenomic sequencing of cerebrospinal fluid (CSF) and/or nasopharyngeal (NP) swabs from 22 of these patients and found enterovirus D68 in some NP samples that were positive based on PCR testing.
The identification of species from a WGS sample is a challenging problem that has spurred the development of multiple new computational methods [3][4][5] . Because of the large size of next-generation sequencing data sets, these methods need to be very fast, but in the context of clinical diagnosis, they also need to be accurate. We downloaded the 31 next-generation sequencing (NGS) samples from the Greninger et al. 2 study (NCBI accession SRP055445) and re-analyzed them using a computational pipeline based on the recently developed Kraken metagenomic analysis software 4 , a very fast and sensitive system that can be customized to use a database containing any species whose sequences are available.

Alternative infectious diagnoses in two subjects
Among the 22 subjects for which NGS data were available, we found at least two that had far greater numbers of sequences (reads) from a bacterial pathogen than from enterovirus D68. Neither subject had been reported in 2 as having a bacterial infection.
In one subject, US/CA/09-871, reported by Greninger et al. 2  In a second subject, US/CA/12-5837, we found a strikingly large number of reads from Staphylococcus aureus in the NP swabs. The two separate NGS files associated with this subject contained 6,858,453 and 1,343,806 reads, comprising 70% and 84% (respectively) of all non-human reads identified at the species level in each sample. The closest match was S. aureus subsp. aureus MRSA252, a methicillin-resistant strain. The coverage was deep enough, approximately 40X, that it would be possible to assemble this genome separately from the reads here ( Figure 1).

Amendments from Version 1
We have revised our manuscript to make some small text amendments, in response to Charles Chiu's comments. Greninger et al. 2 reported 2,790 reads from enterovirus D68 in this subject (our analysis found 1,641) but did not report any from S. aureus.

REVISED
Patient US/CA/12-5837 was sampled in 2012, two years before the outbreak of AFM, although this patient was described in Greninger et al. 2 as positive for enterovirus D68 based on clinical PCR testing and metagenomic sequencing. This patient is reported to be one of the first patients with enterovirus-D68-positive AFM 2 , but the sequence evidence indicates a severe S. aureus infection that might explain at least some of the patient's symptoms. S. aureus has been implicated in neurological complications such as myelitis 6 and meningitis 7 by mechanisms that involve not only direct invasion into the central nervous system (CNS), but also immunopathogenic responses triggered by superantigens that can target the CNS 8 . At a minimum, S. aureus infection was overlooked by the previous analysis. Although the potential role of bacterial infection in the neurological disease that affected these two subjects is difficult to assess because of the lack of clinical and CSF information, its involvement as a pathogenic co-factor should be evaluated.

Human reads included in database submission
The metagenomics data (NCBI accession SRP055445) released by Greninger et al. 2 comprise 43 files which cover 22 of the 48 subjects from their study (in their Supplementary Table 1); the study did not conduct NGS for all subjects. Our metagenomics pipeline identifies human reads at the same time that it searches for pathogens; therefore we scanned the data for human as well as microbial content. Greninger et al. 2 reported that all human sequences had been removed from these files. We found, however, that all samples contained large numbers of human reads, ranging from a low of 18,215 to a high of 6,159,868. These comprised as few as 0.5% to as many as 95.6% of the reads in each sample, as shown in Table 1.
The inclusion of human sequence data in the files deposited at NCBI was likely a result of a computational method (SURPI 5 ) that was insufficiently sensitive. Although the exact cause cannot be determined here, it is well known that sequence alignment algorithms often trade speed for sensitivity; e.g., by allowing fewer mismatches, an aligner can process reads at a much higher rate, at the cost of missing some alignments. It is less clear why the very large numbers of matches to two bacteria were missed; for both these bacteria, complete genomes from multiple strains are available in GenBank. We used both the Kraken system 4 and the Bowtie2 aligner 9 to ensure both sensitivity and speed in our analysis.
Release of sequence data is highly valuable, if not essential, for reproducibility and validation of sequencing-based studies. Failure to filter human reads from a sample is not uncommon; a recent study 10 found that Human Microbiome Project samples, from which human DNA was supposed to have been removed, contain up to 95% human sequence. This suggests that future efforts to deposit microbiome data need to employ more sensitive computational screens in order to avoid the unintentional release of human sequence data.

Methods
Sequences were extracted from SRP055445 and each file was separately run through the Kraken program version 0.10.6-beta (https:// github.com/DerrickWood/kraken) 4 , which identifies species by . We report the highest number of reads matching any one of these strains.
Author contributions SLS conceived the study. FBP ran the computational analyses. SLS, CAP, and FBP jointly analyzed the computational results and wrote the manuscript.

Competing interests
The authors declare no competing interests. to mention in the abstract that this analysis does not challenge the main findings of the original paper. As currently written, the abstract alone does not make this point.

I have read this submission. I believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.
No competing interests were disclosed. Competing Interests: data, they did not note these in the paper itself or in the table in the supplementary data that provided additional details -despite the fact that these were present in very high proportions. They also didn't report the presence of human sequence. If all these data were presented explicitly and in the main body of Greninger ., a reader would be more aware of the et al challenges of these metagenomic approaches in infectious disease. I hope this paper will encourage this awareness and stimulate discussion so I note the comments on this paper from both sets of authors and I commend Breitwieser for taking the time to respond as they have. et al.

I have read this submission. I believe that I have an appropriate level of expertise to confirm that it is of an acceptable scientific standard.
No competing interests were disclosed. Competing Interests: We would like to clarify the tables in the Supplementary Appendix of our paper: We would be happy to provide the tables corresponding to bacterial, fungal, and parasitic read counts in the NP swabs that were generated from our analysis upon request. In hindsight, it may have been better to include these tables in the original publication as part of the supplementary files. We have made two small changes to correct two minor points raised by Chiu, Greninger, and Nacacche in their comments. First we reworded a sentence about sample US/CA/09-871 to clarify that the authors did not report any reads in this sample. (In Suppl. Table 3 they list a total count of Haemophilus influenzae bacterial reads found in US/CA/09-871, but in Suppl. "They don't disagree with our finding, but say that they already knew about it, pointing to their Supplementary Table 3. Suppl. Table 3 does indeed report large numbers of "bacterial reads" for these two samples, but it does not identify them further."

Discuss this Article
Yes, we knew about the detection of in sample US/CA/09-871 and Haemophilus influenzae MRSA (methicillin-resistant in sample US/CA/12-5837, both of which are Staphylococcus aureus) nasopharyngeal / oropharyngeal (NP/OP) swabs. We would be happy to provide the actual output from the SURPI pipeline (Naccache, et al., Genome Research, 2014) that shows the presence of these two bacteria, which were easily detected. As we previously stated, we reported the bacterial reads from the SURPI pipeline (5,614,487 reads from US/CA/09-871 and 28,676,383 reads from US/CA/12-5837); a "re-analysis" using a different algorithm was not necessary.
"In contrast, Supplementary Table 5 in their paper includes specific read counts for 75 bacterial species that they found in most of their samples. These species include two species of Staphylococcus and one species of Haemophilus, but and are not listed. S. aureus H. influenzae Thus from the published paper, it is not possible to conclude that the authors were aware of either of these two species." We respectfully request that the authors read our paper and tables more carefully. In Supplementary Table 5, we report the bacterial species found in only the cerebrospinal fluid (CSF) samples, not in nasopharyngeal / oropharyngeal (NP/OP) swabs. Again, we specifically state in our paper: ". As this protocol reduces sensitivity of detection and speciation for non-viral microbes (i.e. bacteria, fungi, and parasites), only viral sequences are shown for the NP/OP samples. " "Greninger, Naccache and Chiu also write that "the purpose of metagenomic NGS on NP/OP samples was to 'aid in the recovery of enterovirus D68 genome sequences and detect potential co-infections from other viruses." We agree that NGS can be valuable for this purpose, but we do not agree that findings of bacterial DNA -particularly when the bacteria dominate the sample, as in these two cases -should be ignored." We did not "ignore" the findings of bacterial RNA/DNA. First, we emphasize that the sample library preparation on the NP/OP samples was a RNA, not a DNA preparation. It was post-treated with post-DNase to enrich for RNA viral sequences. As such, interpretation of bacterial sequence data from this library is problematic. It is entirely possible that other bacteria were present but their DNA was selectively degraded, for instance. Also, such nuclease treatment will affect both the number and distribution of bacterial reads in the metagenomics sample, so attempting to derive quantitative information ("overwhelming presence of bacterial sequences") and, even worse, attempting to attribute clinical significance from the metagenomics data is not valid.
Greninger et al. also make the point that they believe that "bacterial reads in NP/OP swabs from children most often reflect colonization/carriage and not infection." We agree: we would expect any metagenomics sample from the nasopharyngeal tract to contain many bacterial species. However, as we explain in our paper, the two samples in question have an overwhelming number of reads from just one species: in US/CA/09-871, over 93% of the reads were -2.4 million H. influenzae reads -and in the two files for the other patient, 70% and 84% of the reads were -over 8 S. aureus million reads. It is possible that these merely represent colonization and were not causing disease. Even if so, we still feel it would have been important to acknowledge the presence of these pathogens in these samples (assuming the authors were aware of them) and to discuss each of pathogens in these samples (assuming the authors were aware of them) and to discuss each of them, at least briefly.
We re-iterate that it is dangerous to over-interpret the data as it appears as if Salzberg and colleagues are doing here. Please see the previous paragraph on why the number and distribution of bacterial metagenomic reads in these cases are not reliable. As mentioned before, we do report the number of bacterial reads found in NP/OP samples in Supplementary Table 3. We did not discuss them given that (1) the focus for this study was on looking for clinical associations with AFM, (2) analysis of the NP/OP samples were problematic because the preparation was not suitable / reliable for bacterial metagenomics analysis and the difficulty in discriminating colonization from genuine infection, and (3) the presence of reads to many and multiple bacterial species across all of the NP/OP samples, not just the.two containing MRSA or .

Haemphilus influenza
Our focus was also on examination of "sterile sites" (e.g. cerebrospinal fluid and blood) for sequences to pathogens, since it is much easier to demonstrate that an infectious agent is associated or even causal if it is detected in a sterile site. We did not think that specifically mentioning Haemphilus e from a positive control NP/OP sample would be relevant to the study, nor MRSA from the influenza nose/throat of an AFM patient in part because it would be very difficult to distinguish colonization from infection. Also note that infectious disease diagnosis of MRSA infection is primarily made by positive cultures or detection of the bacteria from sterile or invasive sites such as blood, cerebrospinal fluid, and bronchoalveolar lavage fluid and from nasal swabs which are mainly used to screen for MRSA not colonization.
"Greninger et al. express concern that our analysis suggests that MRSA or infections H. influenzae are involved in the pathogenesis of AFM. We do not suggest this, but we suggest that the role of bacterial infection as a co-factor should be evaluated in patients suspected of having AFM. Although not the point of our study or of the Greninger et al. study, this is an important issue in clinical practice and for future studies of patients suspected to have AFM, due to the potential role of bacteria in triggering inflammatory myelopathies, encephalomyelitis syndromes such as Acute Disseminated Encephalomyelitis (ADEM), or superantigen-induced disorders, which may present clinically as longitudinal extensive myelopathies that may resemble AFM." We agree that bacterial infection can be a co-factor. However, as previously mentioned, the library preparation from NP/OP was not suitable for bacterial metagenomics analysis. To evaluate this properly, the samples can and should be prepared as bacterial metagenomics libraries, and a complete analysis of the results, not just mentioning the e and MRSA -as done here --but Haemophilus influenza also all of the other bacteria in the other samples, performed. Even after this analysis is done, much more study would need to be done to investigate the role of bacterial infection as a co-factor. The inadequate sample preparation for bacterial metagenomics analysis and non-sterile sample type (NP/OP) make our NP/OP data not suitable for investigating "the role of bacterial infection as a co-factor".
The authors do point out one error in our paper, where we stated that patient US/CA/09-871 was reported (by them) as having "encephalitis and severe respiratory illness." This was an error on our part, and we will submit a revised manuscript where we correct this to read "severe respiratory illness." We feel that this is a major error in the paper, as one of the two patients did not even have acute flaccid myelitis. Also, it is incorrect to say "severe respiratory illness"; this was a patient with an upper respiratory infection that may not have been clinically severe. We feel that it is inappropriate in general to report the results of metagenomics analyses in this fashion without understanding the clinical context and how the sample was collected and prepared. Note also that given this error, the title becomes even more misleading: "re-analysis of metagenomics sequences from acute flaccid myelitis patients reveals misleading: "re-analysis of metagenomics sequences from acute flaccid myelitis patients reveals alternatives to enterovirus D68 infection." "Regarding our second main criticism -that the authors inadvertently left large numbers of human reads in their data, which were supposed to have been filtered to remove human DNA -the authors agree. As we mentioned in our paper, this problem has occurred elsewhere too: J. Allen and colleagues recently showed (Ames et al 2015, cited in our paper) that some of the Human Microbiome Project samples, all of which were supposed to have been cleaned of human reads, contain up to 95% human reads. We believe it is vitally important that the biomedical community be aware that the choice of computational methods is a very critical one, particularly when analyzing the large data sets that are becoming ever more common." We agree and have taken steps to ensure patient privacy (requesting that SRA data be removed from the database). However, we also believe that it is unfair to target our paper specifically, as many other earlier published studies including the Human Microbiome Project, as the authors point out, have the same issues. We will be erforming a bioinformatics analysis comparing the different computational methods for their ability to remove human reads.
We believe it is vitally important that the biomedical community be aware that the choice of computational methods is a very critical one, particularly when analyzing the large data sets that are becoming ever more common.
Although we agree with this statement, we believe that a "re-analysis" was not necessary for our paper as we detected both bacteria described by Salzberg and colleagues. We would have been happy to provide this data upon request. In addition, the re-analysis has numerous errors and issues of over-interpretation that did not take into account the clinical context and not understanding important details related to the study (e.g. upper respiratory infection does not imply "severe respiratory illness"; the NP/OP sample preparation was sub-optimal and problematic for bacterial metagenomics analysis; our patient with e did not have acute flaccid myelitis). We believe that the re-analysis by Salzberg Haemphilus influenza and colleagues simply confirms the results from our published analysis, although there are obvious differences in clinical interpretation. Alexander Greninger, MD/PhD Samia Nacacche, PhD Charles Chiu, MD/PhD We are authors from the paper by Greninger, et al., (2015)  We thank Drs. Greninger, Naccache, and Chiu for their response, which makes some valid points that we will comment on further here. But first we wish to acknowledge that our re-analysis does not affect the main finding of their paper; i.e., that the 2014 outbreak of enterovirus D68 represented a novel strain, as shown by their phylogenetic analysis. We don't question that finding, and we think it is an important contribution to the understanding of acute flaccid myelitis.
In their comment, they respond to our first main criticism that they appear to have missed the fact that two of their samples were dominated by the presence of (sample US/CA/09-871) and Haemophilus influenza (sample US/CA/12-5837). They don't disagree with our finding, but say that they Staphylococcus aureus already knew about it, pointing to their Supplementary Table 3. Suppl. Table 3 does indeed report large numbers of "bacterial reads" for these two samples, but it does not identify them further.
In contrast, Supplementary Table 5 in their paper includes specific read counts for 75 bacterial species that they found in most of their samples. These species include two species of Staphylococcus and one species of Haemophilus, but and are not listed. Thus from the published paper, it S. aureus H. influenzae is not possible to conclude that the authors were aware of either of these two species.
Greninger, Naccache and Chiu also write that "the purpose of metagenomic NGS on NP/OP samples was to 'aid in the recovery of enterovirus D68 genome sequences and detect potential co-infections from other ." We agree that NGS can be valuable for this purpose, but we do not agree that findings of viruses bacterial DNA -particularly when the bacteria dominate the sample, as in these two cases -should be ignored.
Greninger et al. also make the point that they believe that "bacterial reads in NP/OP swabs from children most often reflect colonization/carriage and not infection." We agree: we would expect any metagenomics sample from the nasopharyngeal tract to contain many bacterial species. However, as we explain in our paper, the two samples in question have an overwhelming number of reads from just one species: in US/CA/09-871, over 93% of the reads were -2.4 million reads -and in the two files for the H. influenzae other patient, 70% and 84% of the reads were -over 8 million reads. It is possible that these S. aureus merely represent colonization and were not causing disease. Even if so, we still feel it would have been important to acknowledge the presence of these pathogens in these samples (assuming the authors were aware of them) and to discuss each of them, at least briefly.
Greninger et al. express concern that our analysis suggests that MRSA or infections are H. influenzae involved in the pathogenesis of AFM. We do not suggest this, but we suggest that the role of bacterial infection as a co-factor should be evaluated in patients suspected of having AFM. Although not the point of our study or of the Greninger et al. study, this is an important issue in clinical practice and for future studies of patients suspected to have AFM, due to the potential role of bacteria in triggering inflammatory myelopathies, encephalomyelitis syndromes such as Acute Disseminated Encephalomyelitis (ADEM), or superantigen-induced disorders, which may present clinically as longitudinal extensive myelopathies that may resemble AFM.
The authors do point out one error in our paper, where we stated that patient US/CA/09-871 was reported (by them) as having "encephalitis and severe respiratory illness." This was an error on our part, and we will submit a revised manuscript where we correct this to read "severe respiratory illness." Regarding our second main criticism -that the authors inadvertently left large numbers of human reads in their data, which were supposed to have been filtered to remove human DNA -the authors agree. As we mentioned in our paper, this problem has occurred elsewhere too: J. Allen and colleagues recently showed (Ames et al 2015, cited in our paper) that some of the Human Microbiome Project samples, all of which were supposed to have been cleaned of human reads, contain up to 95% human reads. We believe it is vitally important that the biomedical community be aware that the choice of computational methods is a very critical one, particularly when analyzing the large data sets that are becoming ever more common. Steven Salzberg, Ph.D.