Charna Albert
August 2024—The number of laboratories performing clinical metagenomic next-generation sequencing is limited, as is the number of sample types for which it’s available, but the range of pathogens mNGS detects is wide open.
At the University of California San Francisco clinical microbiology laboratory, where validation and automation are marching forward, clinical use cases and impact for mNGS have come into focus, and movement into machine learning is taking place. Charles Chiu, MD, PhD, who is the laboratory’s director and professor of laboratory medicine and infectious diseases, paints a picture of where mNGS stands.
In his talk at the AMP annual meeting last year, Dr. Chiu opened with the case of a 60-year-old woman with a history of hypertension and spinal stenosis who presented with a four-day history of fever, cough, nausea, myalgias, fatigue, dizziness, and diarrhea. The patient’s urinalysis and chest X-rays were negative, and clinical laboratory test results were unremarkable aside from mild hyponatremia. Cerebrospinal fluid analysis by lumbar puncture showed a mixed pleocytosis with a white blood cell count of 517 (43 percent neutrophils, 46 percent lymphocytes), a low glucose of 17, and a high protein of 200. An extensive microbiological workup, including CSF bacterial and fungal cultures, cryptococcal Ag, HSV/VZV/enterovirus PCR, BioFire meningitis/encephalitis panel, and West Nile virus IgM, was negative. An MRI/MRA of the brain revealed layering debris within the occipital horns and meningeal and ependymal enhancement consistent with meningoencephalitis.
“Based on the standard clinical workup, this case didn’t have specific findings that were distinct for any type of infection. In fact, her symptoms could be entirely consistent with a noninfectious disease syndrome,” Dr. Chiu said. “But certainly all potential pathogens that may cause this particular illness must be considered on a differential diagnosis.”
The case highlights the diagnostic challenges infectious disease presents, he said. “In hospitalized patients, roughly 40 to 60 percent of the time we cannot diagnose cases of meningitis and encephalitis. Twenty to 40 percent of the time we’re unable to conclusively determine the cause of a patient’s sepsis syndrome. With respiratory infections the percentage can vary, but up to 60 percent of the time the cause remains unknown, and even for organ-specific syndromes such as hepatitis, the cause remains unknown in 30 to 50 percent of cases.” A large fraction of these may be noninfectious, but infection is almost always on the differential, he added.
Bacterial, viral, fungal, and parasitic pathogens can be detected with a single mNGS test. “Notably we’re missing prions, which are protein-based,” Dr. Chiu said. Like multiplex PCR, mNGS is a direct detection method, making it necessary for nucleic acid to be present from the pathogen at the time of sampling. Dr. Chiu and colleagues are working through the challenges of introducing mNGS into the clinical laboratory, where “the turnaround time for this is absolutely critical,” he said.
The laboratory offers a validated test for the diagnosis of meningitis/encephalitis from cerebrospinal fluid, and it has validated the assay for plasma, viral respiratory fluids, and body fluids. “To some extent the protocols overlap,” he said, “but we wanted to develop a module of protocols where we can introduce mNGS from various sample types.” The Food and Drug Administration has granted breakthrough device designation to the CSF and viral respiratory mNGS assays, and applications to the FDA for regulatory approval are pending.
For bioinformatics, Dr. Chiu and his colleagues developed a customized pipeline known as SURPI+, which identifies microbial sequences by alignment to publicly available bioinformatics databases (Naccache SN, et al. Genome Res. 2014;24[7]:1180–1192; Miller S, et al. Genome Res. 2019;29[5]:831–842). “Of course, many of these databases are incomplete and there may be misannotations,” he said. “So we had to tailor the pipeline to handle misannotations and accurately identify the presence or absence of a given organism.”
Dr. Chiu and others in 2023 cofounded Delve Bio to commercialize UCSF’s CSF mNGS assay. Among the other U.S.-based laboratories offering clinical mNGS testing are Karius (plasma mNGS), Day Zero Diagnostics (plasma and respiratory mNGS for outbreak investigation), and MicroGenDX (microbial mNGS for multiple sample types, including joint fluid). “Despite clinical mNGS being available for roughly five years,” Dr. Chiu said, “we still have a limited number of sample types for which clinical mNGS is available.”
Dr. Chiu and others investigated mNGS of CSF for the diagnosis of infectious meningitis and encephalitis in a study of 204 pediatric and adult hospitalized patients (Wilson MR, et al. N Engl J Med. 2019;380[24]:2327–2340). Fifty-eight infections of the nervous system were diagnosed in 57 patients. Of the 58, mNGS identified 13 infections (22 percent) that were not identified by clinical testing at the source hospital. Among the remaining 45 infections (78 percent), mNGS made concurrent diagnoses in 19 infections. Of the 26 infections not identified by mNGS, 11 were diagnosed by serologic testing only, seven were diagnosed from tissue samples other than CSF, and eight were negative by mNGS owing to low titers of pathogens in CSF. Eight of 13 diagnoses made solely by mNGS had a likely clinical effect, with seven of 13 guiding treatment.
“A quarter of all the diagnoses made in this very curated cohort of patients with undiagnosed meningoencephalitis were made using mNGS, the majority of which had direct clinical impact, meaning they directly affected a patient’s treatment and/or management,” Dr. Chiu said.
At UCSF, which has offered mNGS testing of CSF for the past five years, test volumes have risen over time. “We’ve run the test for patients from nearly every state in the continental U.S.,” Dr. Chiu said (doi.org/10.1101/2024.03.14.24304139). They’ve done a look-back study of the clinical use cases for which mNGS has maximum clinical impact and found high-yield diagnostic scenarios to be when an infectious disease consultant or neurologist suspects infection, when there are imaging abnormalities such as a brain abscess, when the patient is immunocompromised or has pleocytosis, and when the patient has had an arboviral exposure. “This is a good test for arboviral infections of the central nervous system,” he noted.
To quantify the clinical impact of the testing, one outcome Dr. Chiu and colleagues have studied is mortality at 60 days. Among patients who were mNGS-negative, 887 (92 percent) were alive at 60 days. Among those who were mNGS-positive, 74 (78) percent were alive at 60 days. “I don’t want to give the impression that sending off mNGS increases mortality in patients,” Dr. Chiu said. “The conclusion here is that patients with positive mNGS tend to be the patients who have infections, and they tend not to do as well. In contrast, the patients who are mNGS negative are more likely to have a noninfectious etiology such as autoimmune, autoantibody syndrome, or other causes like multiple sclerosis, lupus, or amyloidosis,” for which the clinical prognosis is typically better (de Lorenzi-Tognon M, et al., in preparation, 2024).
In the case he opened with, of the 60-year-old woman who presented with symptoms of meningoencephalitis, mNGS testing detected reads to Mycobacterium tuberculosis (MTB). “The infectious disease clinicians had not entertained Mycobacterium tuberculosis as a potential diagnosis here,” Dr. Chiu said. “She did not have any suggestive clinical or exposure history that would suggest MTB.” The result was read as being at subthreshold levels in the patient CSF, “meaning the number of reads did not exceed our previously established limits for calling it positive. So strictly speaking, this is a negative test by mNGS,” he said, with the detection of MTB reads reported as a comment. The mNGS result, however, was subsequently confirmed by another microbiologic test. “So this was a case of Mycobacterium tuberculosis infection of the CNS.”
Dr. Chiu and others evaluated the accuracy of mNGS testing in a study of 87 body fluid samples from 77 patients (abscess, joint, peritoneal, pleural, cerebrospinal, urine, and bronchoalveolar lavage fluids) collected as residual samples after routine clinical testing in the microbiology laboratory.
Those 87 samples were used to evaluate the accuracy of mNGS testing by either Illumina or Oxford Nanopore Technologies nanopore sequencing. PCR and mNGS testing were concordant for six of eight culture-negative body fluids, but mNGS detected two additional occult pathogens. Body fluids from an additional five patients with high clinical suspicion of infection but negative microbiological testing were also analyzed with mNGS, which detected a pathogen in all five patients (Gu W, et al. Nat Med. 2021;27[1]:115–124).
“We showed that you can run the same protocol on a wide variety of body fluids,” Dr. Chiu said. Clinical validation, however, remains a challenge, as does streamlining the test for use in the routine clinical laboratory. “We managed to take a [CSF] test that has a two- to three-day turnaround time down to 12 to 24 hours, and a test which used to have 150 steps down to 15 steps,” he said. Many of these steps are now fully automated.
The test’s sensitivity is comparable to that of virus-specific PCR, Dr. Chiu said. “It might be slightly worse than the reported limits of detection, but we’re achieving limits of detection of about 500 copies per mL by 95 percent probit analysis.” It’s also possible to calculate viral load information from mNGS data, he said, by spiking ERCC (External RNA Controls Consortium) reads with a known concentration in each sample, calculating the standard curve for each sample, and comparing the relative ERCC normalized counts between the sample and a quantified virus positive control. “This is how we potentially would be able to obtain quantitative information from mNGS data,” he said.
SARS-CoV-2 variant identification for public health purposes is one clinical use case for viral respiratory mNGS, as is identifying uncommon and unexpected respiratory infections associated with critical illness. More precise subtyping, too, is another indication, he said, “in particular of more pathogenic rhinovirus and enterovirus strains” (doi.org/10.1101/2024.02.28.24303521).
Plasma mNGS is the more widely available clinical mNGS test, Dr. Chiu said, and its indications include culture-negative endocarditis, fever of unknown etiology, lung nodules or unexplained pneumonia, difficult-to-biopsy end-organ disease, immunocompromised and transplant patients with suspected infections, and ICU patients on empiric antibiotics.
Metagenomic NGS is limited in that detection relies on the presence of nucleic acid or protein from the pathogen.
“We do have some ways of diagnosing disease based on host response,” Dr. Chiu said, “for instance, procalcitonin as a single-protein biomarker for distinguishing between bacterial and viral infection, reviewing differential cell counts to distinguish viral or bacterial or fungal infection, or serology, looking for antibodies.” But the applicability of these methods is limited, he said. “So we’ve been interested in doing host response profiling for diagnosis of infections,” using a combination of mNGS testing and machine learning methods.
In mNGS testing, RNA sequencing is used to identify RNA viruses. “We’re now leveraging RNA sequencing data to analyze human RNA reads, which we would normally throw away. The idea is to be able to assay the host response.” To do so, he and his colleagues have generated a CSF host response classifier for diagnosis of neurologic illness (Omura C, et al., in preparation, 2024). To generate the classifier, they split 464 samples (91 bacterial, 43 fungal, 175 viral, and 155 autoimmune/noninfectious) into training and test sets at an 80:20 ratio. With the training data, they generated predictive models that could distinguish between the different categories, evaluating the performance of the model with the 20 percent test set. “We’ve been able to develop a model that has an overall accuracy of about 83 percent in distinguishing between these various infectious and noninfectious syndromes.”
Also interesting, he continued, is that in using a “leave-one-out” approach with machine learning, they can identify not only broad categories such as bacterial and viral infection but also potentially specific types of infection such as dimorphic fungal infections from cocci or histoplasma, or acute flaccid myelitis associated with enterovirus D68 or A71. They have developed individual classifiers based on RNA gene host response biomarkers that can distinguish, for example, dimorphic fungal infections and parasitic syndromes, he said. “And we could even try to provide insight into noninfectious syndromes,” because many such syndromes can mimic infection in patients presenting with meningoencephalitis, among them malignancy, lupus, neurosarcoidosis, and amyloidosis. “Essentially, we develop biomarkers and combine all of this into a result,” he said. “In short, what we’re doing is repurposing the RNA data from a metagenomic assay for pathogens into a host-response-based assay, without doing additional wet lab work.”
Diagnosing enterovirus D68/A71-associated acute flaccid myelitis is one potential use case for mNGS-based host response profiling. “This is an interesting syndrome because in acute flaccid myelitis, you almost never detect the enterovirus that appears to be the causative pathogen from spinal fluid,” Dr. Chiu said. “In fact, the diagnosis is usually made by finding the enterovirus in other samples, such as respiratory secretions or blood.”
Dr. Chiu shared an example of a representative clinical CSF mNGS run for which host response analysis was performed. In Fig. 1 are the positive results. In the first sample (row one), where the host response classifier predicted a bacterial infection, the diagnosis was Acinetobacter baumanii meningitis. In row two, the classifier prediction was bacterial, and the diagnosis was Klebsiella aerogenes meningitis. “The third row is interesting, where the patient had both a bacterial and mycobacterial host response. It turns out this was a patient [with HIV] who had a subthreshold result and had Mycobacterium avium meningitis.” Another patient with HIV (row four), where the classifier predicted a viral and parasitic host response, had subthreshold Toxoplasma gondii meningoencephalitis. And the host response classifier predicted a viral infection in a patient with HSV-2 meningoencephalitis (row five).
In the future, Dr. Chiu said, host response profiling might provide information complementary to mNGS. One criticism of mNGS is the low diagnostic yields. “We typically get only a five to 10 percent positivity rate in our testing.” One reason for the low yields is that mNGS would not be done unless the case is very difficult to diagnose. “Typically these patients have already undergone a very extensive workup.” The other reason is that the cause of disease may be noninfectious. “And being able to generate host response analysis could provide complementary information.”
Host response analyses could be complementary in outbreak investigations, too, Dr. Chiu noted, citing publication of an investigation that used mNGS and traditional diagnostic testing to confirm the infection source, and highlighted the potential for host-response-based profiling to complement mNGS methods for pathogen detection (Gould CV, et al. Lancet Microbe. 2023;4[9]:e711–e721). “Although it [host response analysis] doesn’t give you the actual cause of the virus, it at least may point you in the right direction.”
Dr. Chiu returned to the case of the patient with Mycobacterium tuberculosis. The patient’s host response profile had a main category prediction of likely bacterial infection and a subcategory prediction of likely atypical bacterial infection but was negative for MTB or other specific etiology subcategory. “We don’t have a lot of MTB cases in that classifier so I’m uncertain how robust it may be,” he said. “But it at least appears that the host response would have pointed the clinician in the right direction had the mNGS test itself been negative. And the patient was treated for MTB with clinical response.”
Charna Albert is CAP TODAY associate contributing editor.