A novel approach to find Biomarkers affecting Autism Spectrum Disorder (ASD) Using Machine Learning

Abstract

Autism Spectrum Disorder (ASD) is a neurodevelopmental condition that presents with challenges in social interaction, communication, and repetitive behaviors. Early detection is essential but remains difficult due to the variability of symptoms and their overlap with other developmental disorders. Advances in diagnostic techniques increasingly employ machine learning (ML) models to analyze biological and behavioral data, aiming to improve early detection and supplement traditional diagnostic methods with efficient, cost-effective tools. In this study, we analyzed gene expression data (GSE18123) comprising samples obtained from two distinct platforms, which were used to construct two separate datasets. To minimize the risk of data leakage, one dataset was designated for feature reduction, while the other was used for training and testing the machine learning model. From the original gene expression data, we derived a pathway expression dataset. To address the common issue of an excessive number of pathways compared to the available samples, we developed a secondary dataset by creating paired and concatenated samples. The ML model was trained on this secondary dataset. During testing, each test sample was paired with all training samples, and predictions were made for each pairing. The predicted labels were then aggregated to derive the final label for each test sample. The primary ML method used in this study was Random Forest, although the approach is adaptable to other machine learning techniques. For feature importance analysis, we identified pathways with high importance scores in both of their appearances in the secondary dataset. Pathways with consistently high scores were selected, and genes frequently appearing within these pathways were prioritized as potential biomarkers. Our model demonstrated strong performance, achieving an accuracy of 0.85, precision of 0.81, recall (sensitivity) of 1.00, and an F1-score of 0.90. These results suggest that the model is effective at distinguishing between ASD and non-ASD cases. Additionally, we identified over 50 candidate genes for ASD, several of which have been reported in previous studies. Notably, we also discovered a novel gene, ATP6V1F, which scored highly and may represent a new potential biomarker associated with ASD.

Milad Besharatifard

Ph.D. student in Computer Science specializing in Soft Computing and Artificial Intelligence

Fatemeh Zare-Mirakabad

Associate Professor

My research interests include bioinformatics, computational biology and artificial intelligence.