Deep learning causal study between the gut microbiome composition and autism spectrum disorder manifestation

Abstract

Autism Spectrum Disorder (ASD) encompasses a range of neurodevelopmental conditions characterized by early impairments in communication and social interaction, often accompanied by repetitive behaviors. Although its etiology remains unclear, both genetic and environmental factors—including gastrointestinal disturbances—have been implicated. Recent research has highlighted a potential link between ASD and alterations in gut microbiota composition (GMC), with some studies reporting microbial imbalances associated with symptom severity. However, inconsistent methodologies, non-reproducible results, and demographic biases hinder the generalizability of current findings. This thesis investigates the use of machine learning (ML) techniques to model and explore the relationship between gut microbial profiles and ASD. ML offers powerful tools for analyzing complex, nonlinear data across heterogeneous populations, addressing methodological inconsistencies and uncovering patterns that traditional statistical approaches may miss. The objectives of this work are to: (1) identify key microbial predictors of ASD across diverse cohorts, (2) quantify the relative importance of specific bacterial taxa, and (3) simulate simplified microbiota dynamics relevant to ASD. The research was carried out in three stages. First, classical ML algorithms were applied to uncover hidden relationships between microbial profiles and ASD diagnosis, revealing how different bacterial genera may contribute to ASD manifestation in cohorts with diverse GMCs. Second, in silico simulations were performed to visualize the impact of diet on gut microbiota structure and to observe clustering behaviors among bacteria under different dietary regimes. Finally, a semi-supervised model was developed using synthetic data and engineered features, grouping bacteria according to their primary metabolic functions and incorporating these functional categories as novel predictors. In conclusion, focusing on bacterial metabolic functions rather than isolated taxa provides a more robust and interpretable framework for understanding the GMC-ASD relationship, potentially supporting earlier diagnosis and improved insights into the environmental dimensions of ASD.