A recent study published in the journal Scientific advancesUS researchers have used 3D transport-based morphometry (TBM) to identify and visualize brain changes associated with 16p11.2 gene copy number variants (CNVs), improving prediction accuracy and advancing precision medicine for autism.
Study: Generative machine learning uncovers genes-brain-behavior links in autism. Image credit: jittawit21 / Shutterstock
background
Characterized by social, communication, and behavioral impairments, autism is influenced by genetic and environmental factors, with an estimated heritability of up to 90%. Despite this, diagnosis is primarily based on behavior, and genetic testing is rare. Over 200 CNVs associated with autism have been identified, with particular focus on the 16p11.2 region. Endophenotypes can link genetics and behavior. New machine learning techniques, such as 3D TBM, may uncover the relationship between genes, brain, and behavior, advancing precision medicine. Further research is essential to improve understanding and develop better diagnostic and therapeutic approaches.
About the Research
In this study, subjects were recruited from the Simmons VIP Project and were reviewed by the Johns Hopkins University Institutional Review Board, which found them exempt as they were de-identified from existing databases. Participants were referred from clinical genetic centers, laboratories, web-based networks, and by self-referral. Screening and medical record review were performed by Geisinger and Emory University, and 16p11.2 CNVs were tested by fluorescence in situ hybridization. Inclusion criteria included recurrent breakpoints at 16p11.2 without other pathogenic CNVs or unrelated syndromes. Exclusion criteria included environmental neurocognitive influences, severe birth asphyxia, prematurity, and lack of English fluency.
Behavioral testing used the Autism Diagnostic Observation Schedule, Autism Diagnostic Interview, and Social Responsiveness Scale. Primary phenotyping sites included the University of Washington Medical Center, Baylor University Medical Center, and Boston Children's Hospital, and Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, Text Revision (DSM-IV-TR) criteria were used. Cognitive measurements used standardized tests to assess full intelligence quotient (IQ). High-resolution brain imaging was performed at the University of California and Children's Hospital of Philadelphia.
Control subjects were recruited locally near the imaging sites and were matched for age, sex, handedness, and non-verbal IQ, and were excluded from the study for major DSM-IV diagnoses, family history of autism spectrum disorder (ASD), other developmental disorders, dysmorphic features, or genetic abnormalities.The study cohort included brain images of 206 individuals in the control (N = 118), deletion (N = 48), and duplication (N = 40) groups.
T1-weighted magnetization-adjusted gradient echo images (MPRAGE) were collected using a standardized protocol. Preprocessing included exclusion of non-brain tissues, segmentation of gray and white matter, and normalization for brain size. Optimal mass transport-based 3D TBM technology transformed images to identify and visualize tissue patterns linked to 16p11.2 CNVs and was combined with machine learning for automated detection and visualization.
research result
Duplication and deletion carriers had a range of diagnoses, often multiple diagnoses per individual. Analysis of variance (ANOVA) revealed significant differences in brain tissue volume between groups, but volume alone was insufficient to distinguish the cohorts. Deletion carriers were generally younger, likely due to earlier medical procedures. Despite efforts to age-match the cohorts, this difference persisted.
Age and sex could not accurately distinguish 16p11.2 CNVs, and the addition of brain parenchymal volume did not significantly improve classification accuracy.
This study used T1-weighted MPRAGE images (n = 206) from the Simons VIP dataset. Images were co-registered and segmented into gray and white matter tissues using statistical parametric mapping software. After tissue mass normalization, TBM transformed each image into the transport domain relative to a reference image and transport maps were generated and analyzed.
TBM allowed for efficient data representation, capturing 96% of the white matter variance with 132 components and 96% of the gray matter variance with 46 components compared to 184 and 182 components, respectively, in the image domain.
Canonical correlation analysis revealed a significant relationship between gray and white matter distribution (Pearson correlation coefficient = 0.56, P < 0.01), justifying separate analyses. After adjusting for covariates, no significant correlations were found between brain parenchymal volume and gray or white matter tissue distribution.
The genetic cohorts were highly separable in white and grey matter transport domains using penalized linear discriminant analysis (pLDA). The genetic cohorts were further separable based on white matter distribution, indicating that 16p11.2 CNVs have a dosage-dependent effect on brain structure in direction 1. Classification performance on the test set using 10-fold cross-validation showed an accuracy of 94.6% for white matter and 88.5% for grey matter.
3D TBM enabled direct visualization of brain endophenotypes that drive CNV classification. Visualization demonstrated that 16p11.2 CNVs affect brain regions diffusely rather than focally, with characteristic tissue shifts highlighted by inverse TBM transformation. These shifts showed reciprocal patterns of tissue expansion/contraction between deletion and duplication carriers.
We found a significant association between TBM scores and speech impairments, and the Direction 1 score was found to be highly sensitive and specific in detecting these impairments in deletion carriers. TBM scores showed a strong relationship with IQ, highlighting the potential of TBM to link brain endophenotypes with behavioral outcomes. This technology will improve our understanding of gene-brain-behavior relationships and support the development of targeted therapies.
Conclusion
In summary, this study reveals new details about brain structural patterns associated with genetic CNVs in autism. These patterns can accurately predict CNVs solely from brain images of new individuals. Moreover, the discovered patterns are sensitive to speech disorders and explain some variation in IQ. This result was made possible by 3D TBM, a generative machine learning approach that directly interrogates the biological mechanisms that influence brain mass distribution. By revealing the structural networks underlying CNV-associated endophenotypes, this study advances our understanding of the biological basis of autism.
