The investigation of neural substrates of autism spectrum disorder using neuroimaging has been the focus of recent literature. In addition, machine-learning approaches have also been used to extract relevant information from neuroimaging data. There are only few studies directly exploring the inter-regional structural relationships to identify and characterize neuropsychiatric disorders. In this study, we concentrate on addressing two issues: (i) a novel approach to extract individual subject features from inter-regional thickness correlations based on structural magnetic resonance imaging (MRI); (ii) using these features in a machine-learning framework to obtain individual subject prediction of a severity scores based on neurobiological criteria rather than behavioral information. In a sample of 82 autistic patients, we have shown that structural covariances among several brain regions are associated with the presence of the autistic symptoms. In addition, we also demonstrated that structural relationships from the left hemisphere are more relevant than the ones from the right. Finally, we identified several brain areas containing relevant information, such as frontal and temporal regions. This study provides evidence for the usefulness of this new tool to characterize neuropsychiatric disorders.