Autism spectrum disorder (ASD) is characterized by substantial phenotypic and genetic heterogeneity, which greatly complicates the identification of genetic factors that contribute to the disease. Study designs have mainly focused on group differences between cases and controls. The problem is that, by their nature, group difference-based methods (e.g., differential expression analysis) blur or collapse the heterogeneity within groups. By ignoring genes with variable within-group expression, an important axis of genetic heterogeneity contributing to expression variability among affected individuals has been overlooked. To this end, we develop a new gene expression analysis method—aberrant gene expression analysis, based on the multivariate distance commonly used for outlier detection. Our method detects the discrepancies in gene expression dispersion between groups and identifies genes with significantly different expression variability. Using this new method, we re-visited RNA sequencing data generated from post-mortem brain tissues of 47 ASD and 57 control samples. We identified 54 functional gene sets whose expression dispersion in ASD samples is more pronounced than that in controls, as well as 76 co-expression modules present in controls but absent in ASD samples due to ASD-specific aberrant gene expression. We also exploited aberrantly expressed genes as biomarkers for ASD diagnosis. With a whole blood expression data set, we identified three aberrantly expressed gene sets whose expression levels serve as discriminating variables achieving >70% classification accuracy. In summary, our method represents a novel discovery and diagnostic strategy for ASD. Our findings may help open an expression variability-centered research avenue for other genetically heterogeneous disorders.