A multi-university research team co-led by University of Virginia engineering professor Gustavo K. Rohde has developed a system that may spot genetic markers of autism in brain images with 89 to 95% accuracy.
Their findings suggest doctors may someday see, classify and treat autism and related neurological conditions with this method, without having to depend on, or wait for, behavioral cues. And meaning this truly personalized medicine could lead to earlier interventions.
“Autism is traditionally diagnosed behaviorally but has a robust genetic basis. A genetics-first approach could transform understanding and treatment of autism,” the researchers wrote in a paper published June 12 within the journal Science Advances.
Rohde, a professor of biomedical and electrical and computer engineering, collaborated with researchers from the University of California San Franscisco and the Johns Hopkins University School of Medicine, including Shinjini Kundu, Rohde’s former Ph.D. student and first writer of the paper.
While working in Rohde’s lab, Kundu — now a physician on the Johns Hopkins Hospital — helped develop a generative computer modeling technique called transport-based morphometry, or TBM, which is at the guts of the team’s approach.
Using a novel mathematical modeling technique, their system reveals brain structure patterns that predict variations in certain regions of the person’s genetic code — a phenomenon called “copy number variations,” during which segments of the code are deleted or duplicated. These variations are linked to autism.
TBM allows the researchers to tell apart normal biological variations in brain structure from those related to the deletions or duplications.
“Some copy number variations are known to be related to autism, but their link to brain morphology — in other words, how several types of brain tissues reminiscent of gray or white matter, are arranged in our brain — shouldn’t be well-known,” Rohde said. “Checking out how CNV pertains to brain tissue morphology is a vital first step in understanding autism’s biological basis.”
How TBM Cracks the Code
Transport-based morphometry is different from other machine learning image evaluation models since the mathematical models are based on mass transport — the movement of molecules reminiscent of proteins, nutrients and gases out and in of cells and tissues. “Morphometry” refers to measuring and quantifying the biological forms created by these processes.
Most machine learning methods, Rohde said, have little or no relation to the biophysical processes that generated the information. They rely as a substitute on recognizing patterns to discover anomalies.
But Rohde’s approach uses mathematical equations to extract the mass transport information from medical images, creating recent images for visualization and further evaluation.
Then, using a unique set of mathematical methods, the system parses information related to autism-linked CNV variations from other “normal” genetic variations that don’t result in disease or neurological disorders — what the researchers call “confounding sources of variability.”
These sources previously prevented researchers from understanding the “gene-brain-behavior” relationship, effectively limiting care providers to behavior-based diagnoses and coverings.
In response to Forbes magazine, 90% of medical data is in the shape of imaging, which we do not have the means to unlock. Rohde believes TBM is the skeleton key.
“As such, major discoveries from such vast amounts of information may lie ahead if we utilize more appropriate mathematical models to extract such information.”
The researchers used data from participants within the Simons Variation in Individuals Project, a bunch of subjects with the autism-linked genetic variation.
Control-set subjects were recruited from other clinical settings and matched for age, sex, handedness and non-verbal IQ while excluding those with related neurological disorders or family histories.
“We hope that the findings, the flexibility to discover localized changes in brain morphology linked to repeat number variations, could point to brain regions and eventually mechanisms that might be leveraged for therapies,” Rohde said.
Publication
Discovering the gene-brain-behavior link in autism via generative machine learning was published June 12, 2024, in Science Advances.
Additional co-authors are Haris Sair of the Johns Hopkins School of Medicine and Elliott H. Sherr and Pratik Mukherjee of the University of California San Francisco’s Department of Radiology.
The research received funding from the National Science Foundation, National Institutes of Health, Radiological Society of North America and the Simons Variation in Individuals Foundation.
