New study uses machine learning to find biomarkers for an autism subtype linked to maternal immune reaction

Using a simple blood test, a new study is the first to use machine learning to predict a subgroup of autism, known as MAR-ASD, with 100 percent accuracy.  

Researchers at University of California-Davis used blood samples to find certain proteins in mothers of children with MAR-ASD, or maternal autoantibody-related autism spectrum disorder.  

In previous studies, these proteins were found in 23 percent of children with autism, but only 1 percent of typically developing children. Researchers estimate that MAR-ASD may affect up to 18 percent of people with autism. 

Using a new machine learning technique to look for patterns in the blood samples, the research team's method correctly predicted autism with 100 percent accuracy in a subset of children with MAR-ASD.  

This development, funded in part by Autism Speaks to study author Judy Van de Water, represents an advancement that might allow for a simple blood test to become a new screening tool for early childhood diagnosis of autism in infancy, before traditional behavioral diagnostic tools can be effectively used. 

“This study paves the way for the future development of real-world clinical tools to screen for autism using a simple test,” said Thomas W. Frazier, chief science officer at Autism Speaks. “Improving early-childhood screening and getting interventions to children early in life are critical to their development as they grow. Machine learning techniques are a fast, reliable, and effective tool to help in achieving this goal.”

In the current study, researchers suggest that the target proteins affect how brain cells grow as well as other biological processes critical to development. These processes can influence pathways in the brain that may lead to autism.  

One particular protein, CRMP1, had the strongest correlation to a child’s autism severity as measured by the Autism Diagnostic Observation Schedule, the standard tool for diagnosing autism. 

This study is the first to explore the ability of machine learning to predict the risk of having a child with autism with 100 percent accuracy. 

Researchers collected biological samples from mothers enrolled in the Childhood Autism Risks for Genetics and Environment (CHARGE) study. The current study, partially funded by Autism Speaks, included 450 mothers of children diagnosed with autism and 342 mothers of typically developing children.  

The researchers used a subset of the samples to train the machine and a separate subset to test its accuracy. While no singular autoantibody can predict autism, the researchers identified 12 autoantibody pattern combinations, of two or more maternal autoantibodies, that can.  

“This study paves the way for the future development of machine learning technology in screening for autism,” said Thomas W. Frazier, chief science officer at Autism Speaks. “Increasing early-childhood screening and timely interventions is a critical objective of Autism Speaks. Machine learning methods coupled with biological information could become a fast, reliable and effective tool to help in achieving this goal.”  

Future studies aim to further explore these autoantibody patterns as predictors of autism by using larger data sets and looking at other autism metrics and subcategories.