Background
Autism spectrum disorder (ASD) is a neurodevelopmental disorder commonly complicated by gastrointestinal symptoms. Emerging studies have investigated links between ASD and gut bacteria, but little know about its relationship with microbial functions and non-bacteria microorganisms, including archaea, fungi and viruses. We aimed to identify robust multi-kingdom ASD-microbiome associations using metagenomics analysis and machine learning.
Methods
We performed metagenomic sequencing on faecal samples from 1,017 well-phenotyped children and generated 6.24 terabytes of sequence (Fig. 1A). We evaluated the impact of 227 environmental and host factors on the gut microbiome and identified 26 significant associations across diet, gastrointestinal symptoms, medication and comorbidity by permutational multivariate analysis of variance (Fig. 1B). Multi-kingdom gut microbiome, including archaea, bacteria, fungi and viruses, were profiled by Kraken 2. Microbial pathways and genes were profiled by HUMAnN 3. Associations between ASD and gut microbiome were determined by MaAsLin2 with adjustments of host confounders. A fully matched sub-cohort was constructed by a one-by-one pairing algorithm for training the machine learning model, of which the performance was further validated in other independent cohorts. 195 metagenomes from public datasets were employed as external validation cohort.
Results
Integrated analyses showed that 19 archaea, 126 bacteria, 6 fungi, 24 viruses, 909 microbial genes and 104 microbial pathways were altered in children with ASD compared with neurotypical children (Fig.1 C-D). Machine-learning analysis using a single kingdom panel showed an area under curve (AUC) of 0.61 to 0.85 in differentiating ASD from neurotypical children (Fig 2. A-B). Multi-kingdom and microbial functional markers consisting of a panel of 27 markers (2 archaea, 2 fungi, 2 viruses, 8 bacteria, 6 microbial genes and 7 pathways) showed superior diagnostic accuracy with an AUC of 0.88 at a specificity of 89.7% and a sensitivity of 90.5% for the diagnosis of ASD (Fig.2 C-E). Our model maintained an AUC of 0.81 in two independent validation cohorts of different ages, a significant (p=9.4e-07) diagnostic value in public datasets of 195 samples across different populations, and a low false positive rate in two cohorts of children with non-ASD diseases. Accuracy of the model was driven predominantly by the thiamine diphosphate biosynthesis pathway which was less abundant in children with ASD, suggesting a relationship between microbial thiamine metabolism and ASD.
Conclusion
Our findings highlighted the involvement of multi-kingdom gut microbiome in ASD-related gut-brain axis. The robust gut microbiome biomarkers identified in this study warrants further exploration for their role in the pathogenesis, diagnosis and therapeutics of ASD.

<b>Figure 1. </b><b>Associations between ASD and multi-kingdom fecal microbiome composition</b>

Figure 1. Associations between ASD and multi-kingdom fecal microbiome composition

Figure 2. Random forest classifiers trained based on multi-kingdom microbiome composition and function data for diagnosis of ASD

Yating Wan

Appearances