A joint complex network and machine learning approach for the identification of discriminative gene communities in autistic brain

Abstract

Autism is a genetically and clinically very heterogeneous group of disorders. Gene co-expression network analysis can help unravel its complex genetic architecture through the identification of communities of genes that are dysregulated. Using a publicly available brain microarray dataset (experiment GSE28475), we performed a gene co-expression analysis based on Leiden community detection to identify stable communities of genes and used them within a robust machine learning framework with feature selection. We reached an accuracy as high as ( 98 ± 1 ) % in discriminating between autism and control subjects and validated our results on an independent microarray experiment obtaining an accuracy of ( 88 ± 3 ) % . Furthermore, we found two communities of 43 and 44 genes that were enriched for genetically associated variants and reached an accuracy of ( 78 ± 5 ) % and ( 75 ± 4 ) % on the independent set, respectively. An eXplainable Artificial Intelligence analysis on these two causal communities confirmed the pivotal role of autism specific variants thus independently validating our analysis. Further analysis on the restricted number of genes in the identified communities may reveal essential mechanisms responsible for autism spectrum disorder.