Decoupling diet from microbiome dynamics results in model mis-specification that implicitly annuls potential associations between the microbiome and disease phenotypes-ruling out any role of the microbiome in autism (Yap et al. 2021) likely a premature conclusion

Abstract

Recently, and in a tour de force effort, Yap et al. performed a comprehensive association analysis of factors such as demographics, psychometrics, diet, stool metagenomics, stool consistency, and genome-wide SNP genotypes with autism in 247 Australian children. Surprisingly, the authors suggest their data show a strong correlation between diet and autism spectrum disorder (ASD) but only negligible, if any, ASD-specific microbiome signals. While the first conclusion comes as no surprise, we were rather puzzled by the second conclusion given the growing evidence of strong associations between the microbiome and ASD phenotype and the wide consensus on a close connection between diet and microbiome composition and function. The causal model proposed by Yap et al. seemed to imply that diet and the microbiome were two independent variables. A careful review of the approach used by the authors confirmed our suspicions that the statistical models were mis-specified, i.e. they had a questionable biological assumption--the independence of diet and microbiome--embedded in them. We have run side-by-side simulations of the causal linear model proposed by Yap et al. and of an analogous model in which diet and the microbiome are treated as co-dependent variables. We show how the Yap et al. model can preemptively exclude any potential host-microbe interactions if the diet-microbiome independence assumption is violated. We believe large-scale efforts such as the one described by Yap et al. are essential to advance our understanding of the potential role of the microbiome in ASD and other diseases. But these are highly complex systems to analyze and thus ensuring that the statistical methods used are accurate is essential to avoid drawing any potentially misleading conclusions due to subtle causal assumptions propagated by the statistical models themselves.