Chase Kulakowski, 3, sits on his bed surrounded by toys. He contracted AFM two years ago.

People with acute flaccid myelitis experience weakness in their limbs and paralysis.Credit: Armando L. Sanchez/Chicago Tribune/TNS via Getty

Infectious-disease researchers hunting for the cause of a mysterious illness that is paralysing children are combining machine learning with a new gene-sequencing technique to pin down the culprit.

The disease, called acute flaccid myelitis (AFM), causes limb weakness and paralysis that resembles the symptoms of polio. The US Centers for Disease Control and Prevention (CDC) in Atlanta, Georgia, has confirmed 134 cases of AFM in the United States so far this year. Many of those who develop the illness never recover.

Most of the evidence suggests that an enterovirus called EV-D681 is causing the illness, but researchers haven’t been able to find the pathogen in the spinal fluid of sick children. Scientists are trying to identify the culprit by using a combination of host-response diagnostics — which look at how the immune system responds to pathogens — and machine-learning analysis. The approach could lead to better diagnostics and provide hints about new treatments.

Host-response diagnostic tests haven’t been used in the clinic yet. But researchers are developing similar tests to help pinpoint other conditions that can be tricky to diagnose, including tuberculosis and bacterial meningitis.

Elusive target

This year’s AFM outbreak started in October, and is the third in a series of outbreaks in the United States that began in 2014. They have occurred every other year since, though researchers have yet to find a definitive explanation for the pattern. It is also taking scientists an unusually long time to determine the cause of the illness, says William Weldon, a microbiologist at the CDC.

Blood samples taken from many of the people with AFM contain the virus. But many people who never developed AFM symptoms also have the virus in their blood.

“We’ve never really had a smoking gun,” says Charles Chiu, an infectious-disease researcher at the University of California, San Francisco, who is leading the machine-learning project. He suspects that if EV-D68 causes AFM, it damages the spinal cord quickly and then drops to undetectable levels in the body.

Host-response diagnostics are useful when researchers don’t know what they’re looking for, says Purvesh Khatri, a computational systems immunologist at Stanford University in California. The composition of the immune system’s defences differs depending on which pathogens are present in the body. So instead of looking for the agent itself, Khatri says, researchers could look at what the immune system is seeing.

Searching for similarities

Most attempts to identify mystery illnesses involve searching for a pathogen’s DNA or RNA in areas of the body such as the tissues or in the blood. But the host-response technique takes a blood sample and sequences all of the 23,000 or so human genes present in the blood at any given time.

Chiu’s group is analysing these genes — collectively known as the transcriptome — using machine learning. The scientists are searching for similarities between the transcriptomes of people with the illness, and differences between the transcriptomes of those with AFM and people with other, known infections, including those caused by enteroviruses. Once the team knows which genes are relevant to AFM cases, it can test for them directly.

“We’re not relying on detecting the virus — we already know we can’t detect the virus,” says Chiu, who published some of the machine-learning methods last week2. His group hasn’t published any results yet since they're still preliminary. But their data suggest that the expressed genes common among people with AFM are those that researchers would expect to see in a person whose immune system is fighting a virus.

“I think it’s definitely promising,” says Weldon. He says that the CDC has been working with Chiu’s group, and is talking with other teams pursuing similar experimental tests based on host-immune response.

Proper training

Khatri stresses that researchers will need to train the machine-learning algorithm with data from diverse populations. Immune responses may vary depending on a person’s ethnicity or home country, which can shape which pathogens people encounter, he says.

Thorough training is especially important if researchers want to use similar host-response diagnostic techniques widely.

One group, led by infectious-disease researcher Christopher Woods at Duke University in Durham, North Carolina, has developed a transcriptomics test that can determine with 90% accuracy whether a bacterium, a virus or an autoimmune reaction is responsible for a person’s symptoms3.

The distinction is important for treatment, Woods says, and could prevent physicians from prescribing unnecessary antibiotics for viral or autoimmune diseases.

Khatri’s group has developed a test that predicts whether a person will develop active tuberculosis. About 25% of the world’s population harbours the bacterium that causes the illness, but only about 5–10% of these people develop symptoms4. The test from Khatri’s group could allow researchers to categorize and prioritize people to begin treatment before the disease becomes severe.

Chiu hopes that the host-immune response approach could also help to explain why only some people infected with EV-D68 develop AFM. His group is also sequencing the genomes of children with the condition. They hope that this information — combined with the transcriptome data — might provide hints about who could be susceptible to the illness before the next outbreak, which many researchers expect to occur in 2020. “These cases this year provide valuable data for us in evaluating how it might progress in the future if we see additional outbreaks,” Chiu says.