Development of a miRNA-based deep learning model for autism spectrum disorder diagnosis

JOURNAL/atin/04.03/02274269-202506000-00002/figure1/v/2025-09-16T151948Z/r/image-tiff Autism spectrum disorder is a neurodevelopmental disorder characterized by differences in social behaviors, intellectual disabilities, and various mental health conditions. It is often undiagnosed due to overlapping symptoms with other disorders and the challenging, subjective nature of behavioral analysis. However, recent studies have identified dysregulated microRNAs as potential biomarkers for autism spectrum disorder, which could enable more accurate quantitative diagnoses. This study aimed to develop a machine learning model to predict whether dysregulation of a specific miRNA is associated with autism spectrum disorder. We selected an even number of autism spectrum disorder-associated miRNAs and randomly chosen miRNAs for analysis. Data was collected on amino acid sequences, gene targets, and predicted pathway attributes to classify each microRNA. Feature selection was then performed to identify the optimal number of features for achieving the highest accuracy. Only statistically significant predictions ( P < 0.05) were included in the training dataset. The sequential model with two hidden layers emerged as the best classifier, achieving an accuracy of 95.24% for microRNA biomarkers. This model was further validated with an independent, unseen dataset, which achieved 81.67% accuracy. The study also explored the genes and pathways of significance to understand better potential causes of autism spectrum disorder, particularly those involved in regulating the pluripotency of stem cells. This study presents a rapid and efficient method for classifying microRNAs as potential biomarkers for autism spectrum disorder based on their biological characteristics. By screening for dysregulated microRNAs in patients’ blood or serum samples, this approach can enhance early diagnosis and timely intervention.