Interpretable Machine Learning Reveals Dissimilarities Between Subtypes of Autism Spectrum Disorder

Area of Science:

• Bioinformatics and Computational Biology
• Neuroscience and Genetics
• Machine Learning in Healthcare

Background:

• Autism spectrum disorder (ASD) is a complex neurodevelopmental condition with a heterogeneous genetic basis.
• Analyzing molecular alterations in ASD requires methods that yield interpretable biological insights.
• Interpretable machine learning offers legible models crucial for understanding biological mechanisms and clinical subgroups.

Purpose of the Study:

• To apply interpretable machine learning to gene expression data from ASD individuals.
• To construct and visualize a nonlinear gene-gene co-predictive network for ASD.
• To differentiate between ASD subtypes using network topology and centrality measures.

Main Methods:

• Utilized rule-based learning models constructed from three independent gene expression datasets of ASD patients.
• Visualized the model as a nonlinear gene-gene co-predictive network.
• Analyzed network topology and estimated centrality distance to identify dissimilarities among ASD subtypes.

Main Results:

• Identified autism as the most severe ASD subtype.
• Pervasive developmental disorder-not otherwise specified and Asperger syndrome were found to be closely related, milder subtypes.
• Discovered a significant co-predictive relationship between EMC4 and TMEM30A genes, suggesting potential co-regulation.

Conclusions:

• Interpretable machine learning effectively elucidates complex genetic patterns in ASD.
• The methodology provides a framework for distinguishing ASD subtypes and identifying key gene interactions.
• This approach holds promise for transcriptomics and other omics data analyses in bioinformatics.