Note: Most Internet Explorer 8 users encounter issues playing the presentation videos. Please update your browser or use a different one if available.

Clasification of Autistic Specturm Disorders Using Blood-Based Gene Expression Profiles

Thursday, 2 May 2013: 09:00-13:00
Banquet Hall (Kursaal Centre)


Background:  We previously constructed and cross-validated a support vector machine (SVM) that successfully distinguished children with an autistic spectrum disorder (ASD) from typically developing (TD) children based on the expression levels of just 48 genes in peripheral blood.  Subsequently, we have collected and assayed gene expression in peripheral blood samples from approximately 200 additional subjects.  Here we report on the classification accuracy of our previously derived gene-expression signature of autism in this new independent sample.  We also performed a series of contrasts between ASD and TD children and other groups of affected children, such as those with global developmental delay (DD) or language delay (LD).

Objectives: Our objectives were to evaluate the generalizability of a previously derived blood-based biomarker of ASDs, and to build better biomarker profiles with greater disorder-specificity and generalizability.

Methods: Children were ascertained through a network of pediatricians in San Diego County, USA, based on whether (ASD, DD, LD) or not (TD) any developmental red flags were identified at or after 12 months based on Wetherby and Prizant's Communication and Symbolic Behavior Scales.  Gene expression levels in peripheral blood mononuclear cells were measured using Illumina WG microarrays.  Differentially expressed genes were identified between age- and sex-matched diagnostic-group subsamples by ANCOVAs, and diagnostic classifiers were constructed and optimized by SVM.

Results:  The identical SVM that obtained approximately 70-90% accuracy in our initial study attained an accuracy of 58% in the newly collected sample, with a corresponding sensitivity of 55%, specificity of 62%, positive predictive value of 65%, and an area under the receiver operating characteristic curve (AUC) of 0.59.  This model performance, while far from perfect, was significantly better than chance expectation (p=0.012).  Newly derived models contrasting ASDs vs. TD children or those with DD or LD attained externally cross-validated AUCs between 0.6 and 0.85.

Conclusions:  These results suggest that the continued pursuit of blood-based biomarkers of ASD and other developmental disorders or delays is warranted.  Additional analyses—including further cross-validation, re-optimization of classifier parameters, and more precise quantification of distinct mRNA isoforms—should yield even more accurate, stable, and generalizable classifiers of these conditions, which may pave the way for molecular diagnostic testing.

| More