(667e) Machine Learning of Fecal Metabolites of Children with Autism Spectrum Disorder during Microbiota Transfer Therapy

Introduction: Autism spectrum disorder (ASD) encompasses a large group of early onset neurological conditions that result in impairments in social behavior and communication, which are estimated to affect 1 in 54 children under the age of 8 in the United States (Maenner et al., 2020). Although diagnosis of this disorder is only made through behavioral evaluations, many systems of the body are strongly affected in this condition. A diverse range of physiological mechanisms have been observed to be perturbed in ASD including the immune, endocrine, and gastrointestinal (GI) systems (Rossignol, D. A., and Frye, R., 2012). Notably, the prevalence of GI symptoms co-occurring with ASD (~46%) lends significant credence to investigating the relationship of ASD to the gastrointestinal system (Fry et al., 2015). Furthermore, the microbiota of individuals with ASD, both with and without the presence of this co-occurring condition, has consistently been observed to be distinct from their typically-developing (TD) peers (Hughes et al., 2018).

Thus, in recent years, there have been growing efforts related to studying the effect of the microbiome on the Gut-Brain Axis in the context of ASD etiology. The use of microbiota transfer therapy (MTT) has shown considerable potential in its capability to alleviate not only symptoms associated with GI complications, but also in some cases to reduce the severity of certain behavioral symptoms in children with ASD. For example, Kang et al. (2017) demonstrated in an open-label study that through MTT, there was an 80% reduction in GI symptoms and a 24% initial reduction in core ASD symptoms, with greater improvement in ASD symptoms at a two-year follow-up (Kang et al 2019). ASD children who underwent MTT were also observed to undergo changes in their plasma metabolite profiles to more closely resemble those of their typically developing peers (Adams et al., 2019).

This work subsequently builds on this MTT study by focusing on fecal samples taken from these participants. The purpose of this work was to examine the differences in gut metabolites between children with ASD and GI problems vs. typically developing children without GI problems, and determine the effects of gut microbiota transfer therapy on the fecal metabolites of the ASD group. Towards this purpose, both multivariate and univariate statistical analysis was applied. The development of a classification model based on metabolite panels could in turn be used to validate underlying metabolic change stemming from treatment.

Methods

The study involved 38 children, aged 7-16 years, with 18 professionally diagnosed with ASD (verified with the Autism Diagnostic Interview-Revised) and 20 determined to be TD. The participants with ASD were required to have moderate to severe GI problems, while the TD participants were required to be without GI disorders. The study consisted of 2 weeks of preparing ASD participants for MTT, 8 weeks of MTT treatment followed by 8 weeks of evaluation post treatment. TD participants did not undergo any treatment, but had fecal samples collected at the same time as ASD participants at the start of the study. Fecal samples were taken at four time points from the participants with ASD. Parents were instructed to freeze these sample immediately after collection for up to 3 days, and the samples were then transported to Arizona State University on dry ice where they were stored in a −80 °C freezer. Initial fecal samples were collected from all participants at Week 0. Samples were taken from ASD participants at the Week 3 mark from the beginning of the treatment (after about 5 days of microbiota transplant) and at the end of MTT treatment (Week 10). The ASD group was sampled again 8 weeks after all treatment ceased (Week 18).

Once the study concluded, aliquots of the fecal samples were shipped overnight on dry ice to Metabolon (Durham, NC, USA). Both the control and autism samples were blinded and randomized prior to being shipped. Metabolon utilized ultrahigh performance liquid chromatography-tandem mass spectroscopy (UHPLC-MS/MS) instruments for obtaining metabolomic information on 669 metabolites. In order to ensure continuous distribution of values across all participants, metabolites with fewer than 40% of measurements above the detection limit were removed from subsequent analysis. Univariate analysis was performed on each of the remaining 583 metabolites by finding the area under the receiver operator curve (AUROC) for classifying between the ASD and TD cohorts at Week 0. Depending upon what type of distribution the metabolites followed, a Wilcoxon signed-rank test or a paired t-test was performed on to determine if the ASD cohort significantly changed from Week 0 to Week 18.

Using Fisher discriminant analysis (FDA), the 165 metabolites that had been identified as having an AUROC value above 0.60 were used to develop a multivariate model for distinguishing between the ASD and TD cohorts. An exhaustive search was performed through all possible combinations of 2, 3, and 4 of the remaining metabolites to determine the models which best separate the ASD and TD groups at Week 0, as determined by AUROC. A 5-metabolite model was developed by iterating through all possible 1000 top 4-metabolite models augmented by each of the 161 remaining metabolites. An optimized model for each metabolite number was selected based upon the accuracy attained via leave-one-out cross validation. The optimized models were subsequently applied to ASD measurements taken at Week 3, Week 10, and Week 18 to examine the effects of MTT. Using kernel density estimation, the probability density function of each model was computed.

Results and Discussion

Preliminary analysis using univariate methods revealed that none of the individual fecal metabolites achieved an AUROC greater than 0.8, which generally indicates poor applicability for classification. The highest univariate AUROC value was 0.77, corresponding to carnitine. It was observed that 10.9% of the 165 metabolites with AUROC greater than 0.60 significantly changed following the MTT therapy when comparing the ASD group before and after treatment. Approximately 68% of the variance observed in the fecal metabolome can be explained by the gut microbiome, which underscores the potential impact MTT can have on reshaping metabolite concentrations observed (Zierer et al., 2018).

All optimized multivariate models using 3 or more elements were able to achieve an AUROC greater than 0.9, highlighting that a multivariate analysis can provide better classification than what can be determined using univariate analysis alone. There were two distinct models that were identified using five separate metabolites as having achieved the same accuracy (94.6%) after cross-validation. The constituent metabolites in these panels were largely identical (imidazole propionate, theobromine, hydroxyproline, 2-hydroxy-3-methylvalerate) apart from one containing indole and the other adenosine. The effectiveness of the 5-metabolite fecal models for classification changed significantly before and after MTT. The type II error rate was initially observed to be 5% for both models but was observed to be 56% 8 weeks after MTT was completed, thereby indicating that distinguishing between the ASD and TD cohort is not reliably possible after MTT.

Conclusion

This study investigated differences in fecal metabolites between a group of children diagnosed with ASD and GI symptoms and their typically developing peers with no history of GI symptoms in the context of MTT. Prior to beginning treatment, the univariate analysis demonstrated that individual fecal metabolites had limited potential to distinguish between ASD+GI and TD cohorts. However, multivariate metabolite models showed the potential of fecal metabolite panels to effectively classify ASD and TD children. Following MTT, 10.9% of metabolites that most greatly differed between the ASD and TD groups significantly changed. The multivariate models that had been developed prior to treatment were also unable to effectively classify between both groups. The metabolites in question that shifted prior/post treatment may point to possible metabolite/cellular mechanisms that may be implicated in ASD, and the classification panel could potentially serve to validate the efficacy of a course of MTT treatment.

Adams, J. B et al., “Multivariate Analysis of Plasma Metabolites in Children with Autism Spectrum Disorder and Gastrointestinal Symptoms Before and After Microbiota Transfer Therapy”. Processes, 7(11), (2019)

Frye, R. E et al., “Gastrointestinal dysfunction in autism spectrum disorder: the role of the mitochondria and the enteric microbiome”. Microbial Ecology in Health & Disease, 26(0) (2015)

Hughes, H. K et al. “The Gut Microbiota and Dysbiosis in Autism Spectrum Disorders”. Current Neurology and Neuroscience Reports, 18(11), 81 (2018)

Rossignol, D. A., and Frye, R. E., “A review of research trends in physiological abnormalities in autism spectrum disorders”. Molecular Psychiatry, 17(4), 389–401 (2012).

Kang, D.-W et al., “Long-term benefit of Microbiota Transfer Therapy on autism symptoms and gut microbiota”. Scientific Reports, 9(1), (2019)

Kang, D-W et al., “Microbiota Transfer Therapy alters gut ecosystem and improves gastrointestinal and autism symptoms: an open-label study”. Microbiome, 5(1), 10 (2017)

Maenner MJ, et al., “Prevalence of Autism Spectrum Disorder Among Children Aged 8 Years — Autism and Developmental Disabilities Monitoring Network — Autism and Developmental Disabilities Monitoring Network, 11 Sites, United States, 2016”. Morbidity and Mortality Weekly Report, Surveill Summ 2020;69 (2020)

Zierer, J et al., “The fecal metabolome as a functional readout of the gut microbiome”. Nature Genetics, 50(6), 790–795. (2018)