Progress has been made in the search for endophenotypes – biological traits of a condition linked to its genetic causes – of autism spectrum conditions (ASC). A new paper by scientists in the ARC, published in Frontiers in Computational Neuroscience, describes how analaysis of the relationships between data points in MRI data, using advanced machine-learning algorithms, can identify differences in the size of brain structures characteristic of autism.
A useful tool in the hunt of endophenotypes is the study of healthy individuals that are closely related to individuals affected by the condition. These individuals would be expected to express endophenotypes to lesser degrees, as they share a certain percentage of genes with affected individuals. The current study achieved this by studying 132 participants, of which 52 had a diagnosis of ASC, 40 were unaffected siblings, and 40 were unaffected control participants with no family history of autism. Each participant underwent a structural MRI scan.
Then, the authors used a machine-learning algorithm and trained it using the imaging data. This means that the algorithm gets “fed” the data from each participant, and progressively learns to distinguish the experimental groups from each other based on the features of the group. In particular, the algorithm was trained to distinguish sibling subjects from control subjects, and ASC subjects from control subjects.
This produced two lists of areas of the brain in which the control group differed from the ASC and the sibling group. The overlap between those areas are, therefore, areas in which both siblings and affected individuals show differences in brain structure – and therefore possible endophenotypes.
Professor John Suckling, senior author of the study, commented that “machine Learning is a relatively new way of looking at imaging data that may give us an edge in terms of sensitivity over more standard statistical techniques. Neuroendophenotypes – heritable brain structures or functions that are a marker for familial risk – have been elusive in autism, but offer a real opportunity to get a clearer picture of this highly complex condition.”
To test these possible endophenotypes, the experiment then tried to use each area of overlap as a diagnosis criterion on the subjects. This means taking each individual’s data, and, blind to whether they are siblings, cases or controls, putting them into a category purely based on possible endophenotypes. A true endophenotype would be expected to falsely categorise some siblings as having a diagnosis of autism, as the endophenotype would also be expressed in the group.
Indeed, the experimenters found that 20% of the markers they found occurred in both subjects with autism and sibling subjects. Those 20% of areas are therefore strong candidates for being endophenotypes of autism.
Professor Suckling added that “machine learning identified as endophenotypes some of the brain regions that were previously known to be consistently different in individuals with autism, and may help us understand why unaffected brothers and sisters sometimes go on to develop the condition in later life.”
This work was partly supported by the University of Granada and the University of Liège. The study was also funded by a Clinical Scientist Fellowship from the UK Medical Research Council to Michael Spencer and by the UK National Institute for Health Research Cambridge Biomedical Research Centre.Adapted from Department of Psychiatry News