Genome wide association studies, genetic epidemiological investigations and numerous gene sequencing approaches have led to a growing appreciation of a genetic component to autism spectrum disorder (ASD). Genetic variations have consequently been linked to a broad spectrum of behavioural symptoms that fall within the classification of ASD. However, for the large part, these risk factors have not been correlated with specific symptomatology. Such a correlation might be important to dissect the heterogeneity of ASD, which is urgently needed to develop more targeted treatment possibilities. In a recent study in Molecular Autism, Patrick Bolton from King’s College London, UK, Hilgo Bruining from the Brain Centre Rudolf Magnus, the Netherlands, and colleagues, investigate the genetics of ASD with a view to determining whether specific behavioural signatures can indeed be linked to certain genetic traits. Bolton and Bruining explain how they were able to discern behavioural symptoms unique to specific genetic disorders that are known to carry an increased risk for ASD, and moreover discuss how this machine-learning approach could be applied to idiopathic ASD.
An increasing body of research has linked a variety of genetic risk factors to autism spectrum disorder (ASD). How strong do you think the genetic component of ASD is?
Twin and family studies have shown that Autism Spectrum disorder is one of the most strongly determined psychiatric disorders. Heritability estimates indicate that 60-90 percent of the liability to autism spectrum disorder is attributable to genetic factors. The known genetic risk factors included single gene disorders, chromosomal disorders and rare genetic variants. Together these are evident in approximately 15 percent of cases. However, even amongst those cases where genetic risk factors have yet to be identified, the evidence indicates that genetic factors play a major role in aetiology. We also know, however, that non-genetic factors are implicated in aetiology, although much less progress has been made in identifying these to date.
What did your study set out to investigate?
Whether specific genetic conditions that are known to cause ASD lead to different patterns of autistic symptomatology and whether cases of ASD of unknown aetiology, exhibit patterns of symptomatology that resemble those associated with specific genetic syndromes.
Data analysed in your study was generated from behavioural profiles obtained through ADI-R (Autism Diagnosis Interview – Revised). What is ADI-R and what kind of information did it provide?
The Autism Diagnostic Interview – Revised is an extensive well validated parent interview that characterises the key symptomatic manifestations of autism spectrum disorder. These include problems in communication, difficulties in engaging in reciprocal social interaction and the presence of intense preoccupations, repetitive behaviours and unusual circumscribed interests.
You employed a machine-learning approach to analyze your data. Can you explain what the Support Vector Machine is and how it works?
Machine learning concerns the construction or ‘training’ of supervised learning algorithms on labelled examples. In this study the labels were the different types of genetic disorder. The Support Vector Machine method tries to determine an optimal non-linear (flexible) combination of ADI-R items that separates the genetic labels by the largest possible margin, giving the lowest misclassification error.
What were the main results of your study, and what findings most excited/surprised you?
The most important result was that we could associate patterns of autistic symptomatology to specific genetic disorders. This is an important result implying that autistic symptom profiles might be used to designate underlying genetic aetiology. Indeed, behavioural specificity related to genetic disorders is consistent with the notion of many clinicians recognising characteristic behavioural presentations in genetic syndromes such as Down’s, Rett syndrome or Tuberous Sclerosis. Our study is the first to translate this notion into statistical evidence by machine learning pattern analysis.
Apart from its power to detect ‘hidden’ profiles, machine learning also has the advantage that it delivers a robust algorithm that can be used in other samples consisting of the same behavioural data. This offered the unprecedented opportunity to estimate the relative similarity of ‘idiopathic ASD’ to behavioural profiles designated from the selected genetic disorders, as we show in the second part of the study. Taken together, the application of support vector machine learning to autistic symptom profiles opens up a novel avenue of translational research.
What are the major strengths and limitations of the methods employed in your study?
We had ADI-R algorithm data on fairly large numbers of individuals with six different genetic disorders. However, more measures of autistic symptoms may have increased the power to detect differentiated profiles. The machine learning algorithm differentiated between some better than others. This variability might be explained by the variation in sample sizes, so in future larger samples will need to be investigated. It was also notable that the ratings of the pattern of social dysfunction were most discriminative, raising the possibility that particular styles of social impairment may be related to particular genetic risk factors.
It seems likely that the incorporation of more symptoms and other phenotypic features, such as the presence of comorbid behavioural problems like those associated with ADHD, may improve the ability to assign cases to specific classes of genetic disorder. The inclusion of other conditions such as Fragile X may also help further to improve genotype-phenotype correlations. Future studies may reveal further contrasts relating to genetic factors that are biologically meaningful.
What are the clinical implications of your findings?
Our proof of concept study indicates the existence of ‘signature’ autistic behavioral profiles related to genetic risk factors. These signatures may be helpful in disentangling the aetiological and phenotypic heterogeneity evident in ASD, but warrant replication in larger and independent samples. The approach presented in our study could hold promise as a means of stratifying patients who may benefit from treatments targeted at specific pathways and as a way of identifying those patients in whom particular interventions may have unwanted effects.