Machine Learning Finds Potential Biomarkers Associated with Autism

Machine learning tools identified patterns of maternal autoantibodies indicating the likelihood and severity of autism in children, according to a study published in Molecular Psychiatry.

The team noted that while the incidence of autism spectrum disorder (ASD) has been rising, ASD-risk biomarkers are still lacking. According to the researchers, in 2018 the CDC estimated that one in 59 children are affected with autism in the US, making ASD a top health concern and a substantial socioeconomic burden for affected families and the healthcare system.

Autoantibodies are immune proteins that attack a person’s own tissues. Previously, the research team found that a pregnant mother’s autoantibodies can react with her growing fetus’s brain after its development.

In the current study, researchers analyzed plasma samples from 450 mothers of children with autism and 342 mothers of typically developing children to detect reactivity to eight different proteins that are abundant in the fetal brain. The team then used a machine learning algorithm to determine which autoantibody patterns were specifically associated with a diagnosis of ASD.

The group developed and validated a test to identify ASD-specific maternal autoantibody patterns of reactivity against eight proteins highly expressed in the developing brain. This maternal blood test uses an Enzyme-Linked-ImmunoSorbent Assay (ELISA) platform, which is both quick and accurate.

“The big deal about this particular study is that we created a new, very translatable test for future clinical use,” said Judy Van der Water, a professor of rheumatology, allergy and clinical immunology at UC Davis and the lead author of the study.

The machine learning algorithm analyzed approximately 10,000 patterns and identified the top three patterns associated with maternal autoantibody-related autism spectrum disorder (MAR ASD), a condition that accounts for about 20 percent of all autism cases: CRMP1+GDA, CRMP1+CRMP2 and NSE+STIP1.

“For example, if the mother has autoantibodies to CRIMP1 and GDA (the most common pattern), her odds of having a child with autism is 31 times greater than the general population, based on this current dataset. That’s huge,” said Van de Water. “There’s very little out there that is going to give you that type of risk assessment.”

The team also found that reactivity to CRMP1 in any of the top patterns significantly increases the odds of a child having more severe autism.

With these maternal biomarkers, there is the potential for very early diagnosis of MAR autism and more effective behavioral intervention.

“The implications from this study are tremendous,” said Van der Water. “It’s the first time that machine learning has been used to identify with 100 percent accuracy MAR ASD-specific patterns as potential biomarkers of ASD risk.”

The study paves the way for more research on potential pre-conception testing, which could be a viable option for high-risk women over 35 or who have already given birth to a child with autism.

“We can envision that a woman could have a blood test for these antibodies prior to getting pregnant. If she had them, she’d know she would be at very high risk of having a child with autism. If not, she has a 43 percent lower chance of having a child with autism as MAR autism is ruled out,” Van de Water said.

The team is currently researching the pathologic effects of maternal autoantibodies using animal models.

“We will also use these animal models to develop therapeutic strategies to block the maternal autoantibodies from the fetus,” said Van de Water. “This study is a big deal in terms of early risk assessment for autism, and we’re hoping that this technology will become something that will be clinically useful in the future.”