Predicting Autism Spectrum Disorder using Gene Expression Signatures and Machine Learning

Abstract

Objective: Genomic data may be a source that aids the identification of individuals with autism spectrum disorder (ASD) posing high heritability. Here we applied a genomic approach to detect a biological signature from peripheral blood with promising performance in the prediction of ASD individuals. Methods: We utilized the published microarray data GSE26415 from Gene Expression Omnibus (GEO) database, including 21 ASD individuals and 21 controls. Thirty differentially expressed probes were identified by LIMMA package in R language (corrected p-value < 0.05) and were further analyzed using machine learning methods. Results: The hierarchical cluster analysis was found to categorize 1 ASD individual and 17 controls into one group, and 20 ASD individuals and 4 controls into the other, respectively. For robustness of classification, we adopted supervised machine learning models. Support vector machine showed both sensitivity and specificity of 100% for classifying ASD individuals from controls. Both linear discriminant analysis and K-nearest neighbor indicated sensitivity of 100% and specificity of 88.9%, respectively. Conclusion: Our findings demonstrate that gene expression profiles identified in peripheral blood from ASD individuals can be utilized for a biological signature of ASD.