For some time, there has been a consensus that autism spectrum disorders (ASD) are not a unitary disease entity. They represent a collage of symptoms that tend to co-occur in groups of people. The symptoms can present quite differently from person to person and are likely driven by different biological processes.
Conor Liston, MD, PhD
Conor Liston, MD, PhD, a psychiatrist at NewYork-Presbyterian/Weill Cornell Medicine, and neuroscientists from Weill Cornell Medicine leveraged machine learning to analyze newly available neuroimaging data and classified people with autism spectrum disorder into four distinct subtypes. The subtypes differ based on brain activity and behavior. The study was published in the March 9 issue of Nature Neuroscience.
"We sought to understand the biological basis explaining why people with autism may present with different kinds of symptoms but have the same diagnosis," explains Dr. Liston, who is co-senior author of the study. "Our work highlights a new approach to discovering subtypes of autism that might one day lead to new diagnostic methods and treatments."
A previous study published by Dr. Liston and colleagues in Nature Medicine in 2017 used similar machine-learning methods to identify four biologically separate subtypes of depression. Subsequent work has shown that those subgroups respond differently to various depression therapies.
“Our work highlights a new approach to discovering subtypes of autism that might one day lead to new diagnostic methods and treatments.” — Dr. Conor Liston
Building on that success, the team set out to determine if similar subgroups exist among individuals with autism and whether different gene pathways underlie them. ASD is a highly heritable condition associated with hundreds of genes and has a diverse presentation and limited therapeutic options. Lead author and neuroscientist Amanda Buch, PhD, pioneered new analyses for integrating functional MRI (fMRI) neuroimaging data with gene expression data and proteomics.
A 3D prism cube represents the machine learning of the three brain-behavior dimensions, etched onto the prism's glass. White light or “data” passes into the prism or "machine learning algorithm," splitting into four colored light paths that represent the spectrum of autistic people in the four autism subtypes. The painted background of a sequencing array represents the molecular associations of the autism subtypes. Image courtesy of Dr. Amanda Buch.
“One of the barriers to developing therapies for autism is that the diagnostic criteria are broad," she says. “To personalize autism therapies, we need to understand and target this biological diversity.”
Key to this study was a large national dataset of neuroimaging data from people with autism that had not been available until recently. Using machine learning, Drs. Liston, Buch, and neuroscientist and statistician Logan Grosenick, PhD, correlated fMRI data with symptoms from 299 people with ASD and 907 neurotypical individuals. “New methods of machine learning that can deal with thousands of genes, brain activity differences, and multiple behavioral variations made this study possible,” says Dr. Grosenick, co-senior author, who pioneered the machine-learning strategy used for biological subtyping in both the autism and depression studies.
They linked brain connection patterns with three dimensions of symptoms associated with ASD:
- Social skills and communication
- Restricted interests and repetitive behaviors
- Intellectual disability or heightened intellectual skills
Using hierarchical clustering, they grouped people in the study based on their scores in the three dimensions and uncovered at least four subtypes with different patterns of atypical brain connectivity.
Two of the groups had above-average verbal intelligence. One group had severe deficits in social communication but fewer repetitive behaviors, while the other had more repetitive behaviors and less social impairment. The connections between the parts of the brain that process visual information and help the brain identify the most salient incoming information were hyperactive in the subgroup with more social impairment. These same connections, however, were weak in the group with more repetitive behaviors.
Two other groups had severe social impairments and repetitive behaviors but verbal abilities at opposite ends of the spectrum. Despite some behavioral similarities, the investigators unearthed completely distinct brain connection patterns in these two subgroups.
“The relationship between brain connectivity and autism symptoms is complex and surprising.” — Dr. Conor Liston
The team then looked at whether the findings in each subgroup were correlated with regional differences in gene expression. "The idea is that if Region A, Region B, and Region C all have atypical connectivity in one subtype and Regions A, B, and C also have elevated expression of a particular set of genes, some of those genes could be involved in regulating the atypical connectivity we see in autism," Dr. Liston postulates. "We found that this was indeed the case. There was a strong correspondence between gene expression and atypical connectivity patterns. The genes that were important for explaining atypical connectivity in brain networks were the same genes that other groups had previously shown to be involved in autism."
The team replicated their findings in a different dataset, identifying the same four autism subtypes. Dr. Buch conducted an unbiased text-mining analysis of biomedical literature. This showed that other studies had independently connected autism-linked genes with the same behavioral features associated with the subtypes identified in the new study.
"These findings told us that we were onto something here and that this was real. That was interesting and encouraging, and perhaps a little surprising," notes Dr. Liston. There were other surprises, too. "Many of the connectivity features that explained individual differences in different symptom domains were actually not atypical. They fell within a normal range — meaning if we measured them in the neurotypical brain, they would probably be within approximately the same range as they are in these people with autism," Dr. Liston adds.
In one subtype, having elevated connectivity might be associated with problems of social communication, for example. In another subtype, however, having weakened connectivity could be associated with social communication difficulties. "What that's telling us is that the relationship between connectivity and symptoms is complex and surprising — and sometimes not intuitive," notes Dr. Liston.
Finally, the researchers also found evidence of protein-protein interactions that helped explain brain and behavioral differences between people with and without ASD. They analyzed network interactions between proteins associated with atypical brain connections, searching for proteins that might serve as a "hub." Oxytocin, a protein previously linked with positive social interactions, was a hub protein in the subgroup of individuals with more social impairment but relatively limited repetitive behaviors. Studies have looked at the use of intranasal oxytocin as a therapy for people with autism with mixed results, though Dr. Buch noted it would be interesting to see whether oxytocin treatment is more effective in this subgroup.
"Whenever we're talking about molecular targets, we're always very careful and cautious not to overstate our findings," Dr. Liston cautions. "But if we find that atypical connectivity patterns are associated with particular molecular signaling modules that might be therapeutic targets down the road, that certainly will warrant further study." Other molecular targets include G-protein coupled receptors and signaling pathways involving synapse function and the immune system. "I would hesitate to call them therapeutic targets at this point, but definitely molecular signaling modules that should be further explored and might one day lead to therapies," adds Dr. Liston.
Next, the team will study the subgroups and potential subgroup-targeted treatments in animal models. They are also collaborating with other researchers who have large human datasets and further refining their machine-learning techniques.
Might there be more than four subgroups of ASD? Absolutely, Dr. Liston notes. "This is just one way of subgrouping people," he says. But other features of people with ASD could lead to more subgroups — such as differences in their microbiomes, comorbidities such as epilepsy or digestive disorders, and genomic variations. "Ideally, we'd like to have in-depth information on the genome and all of the risk variants for autism that we know can influence biology," he asserts. "With more data, we could probably come up with better solutions to subgrouping. This is just one way of doing it. But I think it's a step forward."
