Autism Spectrum Disorder gene prediction using Machine learning model and Human brain Spatiotemporal gene expression Data

Abstract

Abstract Autism spectrum disorder (ASD) is a neurodevelopmental disorder with a clear evidence of genetic basis. Although the list of ASD risk genes is growing, it is still far from complete. Co-expression analysis showed convergence across multiple ASD-associated genes during mid-fetal development in the prefrontal cortex suggesting an important connection of risk gene activity in specific places at a precise time. In this study, we used a machine learning approach to predict ASD candidate genes using the genes labeled and spatiotemporal gene expressions in the human brain. We applied six machine learning method. Ultimately, we opted for the ANN model which gave us scores that surpassed those of the other models: AUC 88.6%, AUC_PR 71.38%, F1_score 67.5%. The genes identified by our model were validated in independent datasets of risk genes. The top-ranked genes included not only those known in ASD (for example UNC13A, CHD3, GRIk3) but also novel candidates such as SNORD112, Small nucleolar RNAs that have a role in the mechanism of the epigenetic imprinting process and EVX2, transcription factor that specify the neurotransmitter fates. Our method outperformed other ASD candidate ranking system. An ontological enrichment analysis of our risk genes predicts showed biological processes evidently relative to ASD including neuron projection development, neuron differentiation, neurogenesis, synaptic signaling and also other mechanisms such as regulation of RNA metabolic process. Our study reveals that spatiotemporal gene expression patterns in human brain can distinguish ASD risk gene. Our gene ranking system is therefore a helpful resource for prioritizing candidate autism genes.