A Critical Synthesis of Machine Learning in Autism Spectrum Disorder Genomic Research: From Transcriptomics to Microbiome.

Autism spectrum disorder (ASD) is a complex neurodevelopmental condition characterized by persistent impairments in social communication, restricted interests, and repetitive behaviors. This narrative review synthesizes advances in machine learning applications to ASD genomic research through May 2026, spanning gene expression analysis, whole-exome sequencing (WES), non-coding variant interpretation, multi-omics integration, single-cell transcriptomics, epigenetic profiling, and gut microbiome analysis. A purposive, thematic literature synthesis approach was employed, allowing broad coverage of emerging methodological innovations and biological insights. We critically evaluate state-of-the-art deep learning architectures-including the Separate Translated Autism Research Neural Network and SHapley Additive exPlanations-based explainable artificial intelligence frameworks. Reported discrimination across the field varies widely, from receiver operating characteristic-area under the curve (ROC-AUC) values near 0.66 to implausibly perfect values of 1.00; the best-validated specialized genomic architecture achieves only modest discrimination (ROC-AUC≈0.73). We emphasize that interpretability and predictive performance are orthogonal properties: specialized architectures yield biologically interpretable feature attributions despite modest discriminative power; and several extreme AUC values in the literature are, in our assessment, more consistent with overfitting or data leakage than with genuine signal, although the primary reports did not always provide the information needed to definitively attribute them. Key themes include: (1) identification of differentially expressed genes through meta-analysis of transcriptomic data; (2) validation of predictive gene features from large-scale WES; (3) detection of non-coding regulatory mutations affecting synaptic transmission pathways; (4) discovery of gut microbiome signatures associated with ASD classification; and (5) discovery of data-driven subtypes enabling precision medicine stratification. Critical challenges include population bias toward European ancestry, socioeconomic ascertainment bias, modest predictive effect sizes, conflation of association with causation, and gaps between computational prediction and clinical utility. Future directions emphasize multi-modal data integration, diverse cohort expansion, engagement with neurodiversity perspectives, and regulatory science development.