By Deborah Borfitz
January 2, 2020 | A Duke University spin-out company is putting the last touches on a host response bacterial-viral test entering the final leg of the Antimicrobial Resistance (AMR) Diagnostic Challenge of the National Institutes of Health Division of Program Coordination, Planning, and Strategic Initiatives. Predigen, Inc. has come up with an improved host response test that fills many of the gaps in the current diagnostic armamentarium, according to company co-founder Ephraim L. Tsalik, associate professor of medicine at Duke University School of Medicine.
Predigen is one of five finalists and the sole contestant with a diagnostic test that can discriminate between bacterial and viral infection—one of the two Challenge categories, he says. The other four entrants have a test designed to rapidly detect the presence of antibiotic-resistant pathogens.
Submissions are due by Jan. 3, after which the diagnostic platforms will undergo real-world testing by a pair of facilities certified by the Clinical Laboratory Improvement Amendments of 1988 (CLIA). Up to three winners will be announced on July 31, 2020, sharing at least $19 million in prize money.
Discriminating between viral and bacterial infection is no easy task, which has contributed to the overuse of antibiotics and the rise of antibiotic-resistant organisms, says Tsalik. Currently available tests—including those based on the detection of antigens or amplification of a DNA target—miss most cases of bacterial infection. While viral pathogens are readily detected by polymerase chain reaction (PCR)-based tests and syndromic panels that include the most common pathogenic culprits, a negative result shown using these tests does not reliably exclude a viral infection.
An added complexity is that studies have shown that people who are asymptomatic are often just as likely to harbor a virus as someone experiencing symptoms, Tsalik says. So, a positive result is not particularly meaningful since it cannot distinguish an infection from colonization.
“We know the same thing is true of certain bacterial pathogens,” continues Tsalik, including group A streptococcus. “A lot of people are walking around colonized with strep who have no sore throat, but if you swab the back of their throat, you’ll detect group A strep. So, if someone comes in with a sore throat and you detect group A strep, is that what is causing their illness? We know from research we’ve done that in many cases it is not.”
Different On Purpose
The standard diagnostic approach is based on the faulty assumption that identifying a pathogen gets at the root cause of sickness, but this can be misleading, says Tsalik. Host gene expression signatures are a better alternative, but up to now have been used largely in the context of cancer because that’s where a large portion of the work characterizing transcriptomes has occurred.
Gene expression-based test MammaPrint, for example, predicts a woman’s risk of having recurrent or metastatic breast cancer and has been used to risk-stratify women to help determine if they need additional chemotherapy, Tsalik says. But that’s a decision that can be made over the course of days to weeks, not while the patient is still sitting in the office.
The MammaPrint test itself is highly complex and takes many hours to generate a result on existing instruments, he continues. Plus, the frequency with which the test is performed is relatively low.
“Infectious disease, on the other hand, is a scenario where the decision to treat or not to treat with antibiotics has to be made right then and there,” Tsalik says. However, rapid measurement of host gene expression “has never been done before.”
If time is of no concern, a few tools exist in the clinical testing world to help separate bacterial from viral infection based on host biomarkers, Tsalik says. Very high levels of the inflammatory marker C-reactive protein (CRP) tend to be more specific for bacterial infections, but people with a severe viral infection can sometimes also have elevated CRP. Moreover, a low or normal CRP does not exclude bacterial infection.
More recently, viral infections have been found to inhibit the inflammatory marker procalcitonin while levels tend to be heightened in the presence of bacterial infections, Tsalik says. But there is also considerable overlap in procalcitonin levels between patients having viral infections, bacterial infections and no infection.
Predigen’s host response test overcomes the limitations of existing biomarkers used to differentiate bacterial from viral illness using a more comprehensive assessment of RNA transcription to isolate the cause of infection.
Technical requirements for the rapid-response discriminatory test include a multiplex instrument to process blood samples. Most such diagnostic platforms don’t meet the criteria of being able to simultaneously measure dozens of messenger RNA molecules in a rapid and fully integrated manner, Tsalik says.
Point-of-care tests largely do endpoint PCR analysis to determine if a single pathogen is detectable, not measure the differential expression of a set of genes that can be found at a baseline level even in healthy people, he adds. Predigen’s host response test identifies diseases based on the upregulation of some genes and the downregulation of others.
As one of 10 semi-finalists to convince AMR Diagnostic Challenge judges it had a good idea, Predigen’s Duke team was initially awarded $50,000 to develop its concept into a prototype. The field narrowed to five contestants during the second step of the competition that required submitting data generated by the proposed diagnostic test.
For step two, the test was translated onto a platform developed by BioFire Diagnostics. However, for step three, the company pivoted to the diagnostic platform of Biomeme that up until then had been largely focused on environmental sampling for industry- and Department of Defense (DoD)-related applications.
This is Biomeme’s first foray into the human health space, Tsalik says, but it already had a CLIA-waivable platform ready for deployment in a clinical environment. The battery-operated device is the size of a Bluetooth speaker with a smartphone readout and was developed with funding from the DoD.
During the final phase of the competition, two anonymous CLIA-certified labs will identify clinical samples that will be run on the instrument, says Tsalik. In addition to confirming the test works, the evaluation criteria being used by the prize committee includes the device’s simplicity of operation and speed of delivering results—i.e., if it would be easily implemented in a clinical lab environment.
For Predigen’s entry, the plan is for the labs to enroll patients presenting with an acute respiratory illness in the emergency department, Tsalik says. Clinical staff affiliated with the competition will ultimately determine if those patients had a bacterial infection, a viral infection or no infection and compare those findings to results provided by the host response test. “It’s effectively a miniature version of a clinical trial… to evaluate the performance as well as the user-friendliness of the diagnostic test.”
Role Of Machine Learning
The process of discovering a host gene expression signature requires agnostically measuring tens of thousands of genes to come up with dozens to hundreds of meaningful combinations, depending on the question, says Tsalik. But it’s easy to “overfit” and find patterns that do a good job of distinguishing different groups of patients but have no association with the biology.
So, machine learning algorithms are deployed that have been trained on known data to make predictions for new cases coming in, helping to identify signatures. Machine learning technology also aids in the development of software associated with the test. “It’s an important part of the process we’ve developed over the past 15 years and have been applying to a variety of different questions,” Tsalik says.
Ultimately, machine learning is being asked to make a binary decision about whether an individual belongs in one group (viral) or another (bacterial), he explains. The problem is that genes in the model “tend to be very specific for viral infection because the host response to viral infection is generally driven by the interferon pathway… a strong component of the biology that drowns out a lot of the other things that a machine learning algorithm might identify.
“Effectively, what you end up with is a signature that is characteristic of viral infection,” Tsalik continues. “If patients don’t have that signature, the model will label them as having a bacterial infection,” including people who are healthy. Therefore, it was critically important to develop a test that assesses more than just bacterial or viral infection.
In many cases, healthy people are used as a control group to develop these types of models. However, people getting a diagnostic test are probably not healthy to begin with. Instead, the culprit may be an exacerbation of chronic obstructive pulmonary disease, the lingering effects of the common cold, asthma, allergies or heart failure—"the list of conditions that can present with exactly the same symptoms as a bacterial or viral infection is really long,” says Tsalik. This points to the importance of Predigen’s clinical validation studies that included a control group comprised of people who are sick but without infection.
A further clinical study will start in early 2020 using the MASTERMIND (MASTER protocol for evaluating Multiple INfection Diagnostics) strategy of the Antibacterial Resistance Leadership Group (ARLG), a key supporter of the clinical research undertaken by Predigen. This is in addition to a recently completed clinical validation study, called RADICAL-2, which enrolled 1,200 subjects at ten U.S. sites presenting with acute respiratory illness. Data from these studies are expected to be submitted to the U.S. Food and Drug Administration in the latter half of 2022, says Tsalik.
In an initial study looking at the performance of the host response signatures in discriminating between bacterial, viral, and non-infectious illness, the test proved to be highly reliable, he says. “We saw AUCs [areas under the curve] on the scale of troponin for myocardial infarction. In the context of infectious disease, we have not seen a test with performance characteristics that good.”
Results of that study were presented at ID Week 2018. Additional experiments have since been conducted and a manuscript is being prepared for publication, Tsalik says.
Predigen is developing the host response signatures using gene expression data pulled largely from clinical samples that have been continuously collected by Duke over the past two decades, says Tsalik. That has allowed it to develop signatures for fungal infection, sepsis diagnosis and sepsis risk assessment as well as evaluate the performance of these signatures in challenging populations, such as immunocompromised patients.
Predigen is actively working on a variety of host gene expression signatures in the sepsis domain. One such test measures an 18-gene combination at the time of admission to predict if they are at high, moderate or low risk for a bad outcome as defined by mortality, says Tsalik.
The company also has a sepsis diagnostic signature to detect the presence of the life-threatening condition and is working on a sepsis predictive signature for at-risk patients that can indicate if sepsis is “brewing” before symptoms become obvious, he says. It is collaborating with Qvella on both the sepsis diagnostic and predictive tests, with support from the Biomedical Advanced Research and Development Authority (BARDA).
In the cardiovascular arena, Predigen founders and their colleagues at Duke have developed a host response signature that helps identify patients who would benefit from more aggressive medical management, such as the addition of more potent anti-thrombotic therapies, Tsalik says. The company is presently seeking strategic partnerships to evaluate this test in clinical studies.
Further out, Predigen is interested in developing a fully disposable, paper-based test it could market to consumers in partnership with the Johns Hopkins Applied Physics Laboratory, says Tsalik. “Acute respiratory illnesses are such a common problem, and most of them are viral and do not require or benefit from a physician visit. If we can empower patients to make some of their own diagnoses, that could really change the paradigm.”
There is already a movement underway to bring influenza testing into the home, he notes, with financial backing from BARDA. As currently envisioned, specimen collection would be via finger stick or nose swab with results readable by a smartphone and transmitted to a physician for a telemedicine consult. Any needed medications might even be drone-delivered directly to patients’ doorstep.
“All of the technical components to do this have been developed in the past for other applications,” says Tsalik. “They have just not been assembled [yet] for this particular application.”