Application of Mendelian randomization to explore the causal role of the human gut microbiome in colorectal cancer

Abstract

The role of the human gut microbiome in colorectal cancer (CRC) is unclear as most studies on the topic are unable to discern correlation from causation. We apply two-sample Mendelian randomization (MR) to estimate the causal relationship between the gut microbiome and CRC. We used summary-level data from independent genome-wide association studies to estimate the causal effect of 14 microbial traits (n = 3890 individuals) on overall CRC (55,168 cases, 65,160 controls) and site-specific CRC risk, conducting several sensitivity analyses to understand the nature of results. Initial MR analysis suggested that a higher abundance of Bifidobacterium and presence of an unclassified group of bacteria within the Bacteroidales order in the gut increased overall and site-specific CRC risk. However, sensitivity analyses suggested that instruments used to estimate relationships were likely complex and involved in many potential horizontal pleiotropic pathways, demonstrating that caution is needed when interpreting MR analyses with gut microbiome exposures. In assessing reverse causality, we did not find strong evidence that CRC causally affected these microbial traits. Whilst our study initially identified potential causal roles for two microbial traits in CRC, importantly, further exploration of these relationships highlighted that these were unlikely to reflect causality.

Figure 1

Mendelian randomization framework applied to assess the causal effect of the human gut microbiome on CRC risk. CRC = colorectal cancer; MR = Mendelian randomization; SNP = single nucleotide polymorphism. MR relies on three key assumptions: (top panel) the SNPs are associated with the exposure (here, the gut microbiome); (middle panel) there are no common causes of the SNPs and the outcome (here, CRC), meaning any confounders driven by population substructure, dynastic effects or assortative mating (these may not and are unlikely to be the same confounders of the association between the exposure and outcome); and (bottom panel) the SNPs are not independently associated with the outcome (here, CRC) other than pathways through the exposure (here, the gut microbiome). Given these key assumptions, microbiome-related variants can be used to assess the causal effect of the human gut microbiome on CRC, overcoming limitations of observational epidemiological studies. In two-sample MR analyses, the causal effect of the exposure on outcome (${\mathrm{\beta }}_{\mathrm{XY}}$) is generated by a ratio of the SNP-outcome (${\beta }_{\mathit{ZY}}$) and the SNP-exposure (${\beta }_{\mathit{ZX}}$) effect estimates derived from two independent samples.

Figure 2

MR estimates of the effect of each microbial trait on overall CRC risk. AB = abundance; CI = confidence interval; CRC = colorectal cancer; MR = Mendelian randomization; OR = odds ratio; P/A = presence versus absence; SD = standard deviation. Letters in the microbial trait name represent the taxon classification level from which that microbial trait was observed, with “C”, “F”, “G”, “O” and “P” representing “class”, “family”, “genus”, “order” and “phylum”, respectively. All microbial traits that were not confidently classified at the genus level were organised into unclassified groups within higher classification levels (represented by “unclassified”). MR estimates represent the OR for CRC risk and 95% CI per SD unit change for continuous microbial traits (labelled as “AB” in brackets) or per approximate doubling of the genetic liability to presence (versus absence) of each binary microbial trait (labelled as “P/A” in brackets).

Figure 3

Colocalisation results for bacteria within the (A) Bifidobacterium AB microbial trait and (B) P/A of unclassified genera within the Bacteroidales order with overall CRC. AB = abundance; CRC = colorectal cancer; FGFP = Flemish Gut Flora Project; GECCO = Genetics and Epidemiology of Colorectal Cancer Consortium; GM = gut microbiome; GWAS = genome-wide association study; P/A = presence versus absence. Regional association plots, generated from LocusCompareR, showing the − log10(P-value) where each lead SNP is represented by a purple diamond (panel A: rs4988235 associated with the Bifidobacterium AB microbial trait and panel B: rs116135844 associated with the unclassified Bacteroidales P/A microbial trait) in relation to overall CRC. These plots were created using the FGFP and GECCO full summary-level data for microbial traits and CRC, respectively.

Figure 4

MR estimates of the effect of each microbial trait on overall CRC risk using a more lenient p-value threshold for selecting genetic instruments. AB = abundance; CI = confidence interval; CRC = colorectal cancer; mGWAS = microbiome genome-wide association study; MR = Mendelian randomization; OR = odds ratio; P/A = presence versus absence; SD = standard deviation. Letters in the microbial trait name represent the taxon classification level from which that microbial trait was observed, with "C", "F", "G", "O" and "P" representing "class", "family", "genus", "order" and "phylum", respectively. All microbial traits that were not confidently classified at the genus level were organised into unclassified groups within higher classification levels (represented by "unclassified"). MR estimates represent the OR for CRC risk and 95% CI per SD unit change for AB microbial traits or per approximate doubling of the genetic liability to P/A of each binary microbial trait. Results for the inverse variance weighted, weighted median, weighted mode and MR-Egger methods are presented when using the multi-SNP instrument for each microbial trait using a lenient p-value threshold (P < 1 × 10⁻⁰⁵) and directional consistency across cohorts included in the mGWAS and were compared to the Wald ratio estimates obtained from the main analysis.