Background

Pancreatic cancers rank 12th in global incidence and are the 7th leading cause of cancer death with over 450,000 deaths in 2022 according to GLOBOCAN estimates1. More than 90% of these tumours present as pancreatic ductal adenocarcinoma (PDAC)2. While most cancers have seen improvements in 5-year survival rates, those for pancreatic tumours remain low at 12%3. This can be attributed to the cancers remaining largely asymptomatic until they reach an advanced stage resulting in only 10-20% of patients presenting with non-advanced or resectable disease4. Environmental and life-style risk factors for PDAC include cigarette smoking, obesity, alcohol use, diabetes, pancreatitis, familial history of pancreatic cancer and stress5. In the last 15 years several susceptibility genetic loci have also been identified through genome-wide association studies (GWAS) or candidate region approaches6,7,8,9.

Emerging evidence suggests that differences in the gut, oral and intratumoural microbiomes play an important role in PDAC development and progression10,11,12. Intestinal microbiota disturbance has been linked with a myriad of conditions ranging from obesity to atherosclerosis as well as tumour development at several anatomical sites13,14,15.

It is challenging to establish a clear causal relationship between gut microbiome ‘dysbiosis’ (i.e., altered composition, diversity and metabolic activity of the gut microbiome with pathogenic relevance) and pancreatic diseases through observational studies due to confounding factors and the potential for reverse causality. One potential mechanism by which the gut microbiome may influence carcinogenesis is through the alteration of circulating metabolite concentrations. Increased circulating and tumoural levels of gut derived lipopolysaccharide (LPS), a pro inflammatory bacterial membrane component, have been reported in a murine PDAC model and are correlated with reduced gut barrier functioning16. Additionally, the concentration of many bacteria-related circulating metabolites has been associated with increased or decreased risk of gastrointestinal (GI) cancers including secondary bile acids17,18,19,20,21, amino acid derivatives22, tryptophan derivatives23 and short-chain fatty acids (SCFAs)24.

Mendelian randomisation (MR) is a genetic epidemiology technique that uses genetic variants as instruments to determine the effect of an exposure on an outcome. As alleles segregate randomly during meiosis, associations with environmental confounding variables are negated. Similarly, because alleles are present before the development of PDAC, there is no possibility of reverse causation25. Therefore, MR is the optimal strategy for assessing the causal relationship between the gut microbiome and related circulating metabolites on PDAC risk.

Materials and methods

Aim

To investigate the effect of the gut microbiome and circulating microbiome-related metabolites on the development of PDAC, we first performed a comprehensive search of the literature and existing databases to identify bacteria-related metabolites. This was followed by an exhaustive search to identify SNPs associated with circulating concentrations of these metabolites. We also leveraged summary statistics from the most comprehensive exploration of genetic influences on the human gut microbiome to date26. We then performed two-sample MR using summary statistics from the studies of the Pancreatic Cancer Cohort Consortium (PanScanI-III) and the Pancreatic Cancer Case-Control Consortium (PanC4)27,28,29,30.

Metabolites data

Data on circulating gut microbiome-related metabolites (amino acids, vitamins and cofactors, fatty acids, carbohydrates, organic acids, hormones, lipids, sterols, bile acids, nucleotides and derivatives of these classes) were obtained through an extensive search in literature through Pubmed (search terms in Supplementary note 1, Additional file 1), and two metabolite databases, namely Exposome Explorer and Human Metabolome Database31,32. Specifically, Exposome Explorer is a manually curated database containing information on biomarkers that may represent risk factors for human disease31, while the Human Metabolome Database is a large database containing molecular and clinical-related data on small molecules and metabolites32. Genetic associations with the selected metabolites were downloaded from GWAS Catalog and published GWAS studies33. Only data from populations of European ancestry were used, and whenever more than one GWAS was available for a given metabolite, the largest study was selected to rely on higher statistical power. The complete list of metabolites under analysis is reported in Supplementary Table 1, Additional file 1).

Microbiome data

Genetic associations between SNPs and microbial taxa were gathered from the MiBioGen Consortium34, which represents the most significant endeavour in investigating host genetics-microbiome associations on a population scale to date. The MiBioGen Consortium comprises data from 25 cohorts for a total of 18,340 participants26. The microbiome data mainly originated from Illumina sequencing of the V4 hyper-variable region of the 16 S rRNA gene. Technical variables related to stool processing and microbial DNA extraction were taken into account (e.g., extraction kit used, mechanical lysis, enzymatic lysis, sequencing technology). The Consortium identified several associations at different taxonomical levels. However, we only focused on summary statistics from genome-wide analyses at the genus level, which was the lowest level of taxonomical resolution and comprised 119 defined genera. The complete list of microbial genera analysed is reported in Supplementary Table 1, Additional file 1.

Instrumental variant (IV) selection

To ensure adherence to the relevance, independence, and exclusion restriction core assumptions of the two-sample MR design, the model reported in Supplementary Fig. 1, Additional file 1 was strictly followed. Briefly, these three assumptions state that all genetic variants used as instrumental variables (IVs) should be associated with the exposure of interest (relevance), but not with risk factors and confounders (independence), and that the IVs should affect the outcome only through the exposure (exclusion restriction). For each exposure trait (metabolites or microbial taxa), SNPs with a p-value of association < 5 × 10−8 were selected. Whenever for this threshold there were no associations (i.e., all bacterial genera and 31 metabolites), a less stringent threshold of < 1 × 10−5 was adopted to gain sufficient IVs to perform sensitivity analyses. The SNPs were then pruned using a linkage disequilibrium (LD) threshold of 0.01 with Ldlink35. Palindromic SNPs (C/G or A/T) whose effect allele frequency ranged between 0.45 and 0.55 were removed. SNPs with a minor allele frequency (MAF) < 0.01 were also removed. To account for potential pleiotropy between the SNPs and the outcome, GWAS Catalog and Phenoscanner were screened for associations with known PDAC risk factors and potential confounders: alcohol drinking, allergy, tobacco consumption, diabetes, diet-related traits (e.g., meat, fruit, and vegetable consumption), pancreatitis, body mass index, and lipidic- and weight-related traits using a p-value threshold of 5 × 10−833,36. SNPs directly associated with PDAC risk were also removed. In addition, each SNP was removed whenever it was in LD with another SNP (r2 > 0.8) associated with one of the above-mentioned traits. The remaining SNPs were considered valid IVs.

Pancreatic cancer GWAS data

Genetic associations between SNPs and PDAC risk were obtained from four GWASs: three from the Pancreatic Cancer Cohort Consortium (PanScan I-III) and one GWAS from Pancreatic Cancer Case-Control Consortium (PanC4). All such studies are available in the Database of Genotypes and Phenotypes (dbGaP) with study accession nos. phs000206.v5.p3 and phs000648.v1.p1; project reference no. 1264427,28,29,30,37. Additional details about the original studies are reported in Supplementary note 2, Additional file 1.

After download, quality control procedures were applied using Plink 1.938 as reported previously39. Briefly, individuals with sex mismatches (genotyped sex Vs self-reported sex), heterozygosity rates higher or lower than three standard deviations from the mean, high levels of cryptic genetic relatedness (pi hat > 0.2), or call rates lower than 90% were eliminated. Genetic variants with a call rate lower than 95%, MAF < 0.5%, and not respecting Hardy-Weinberg equilibrium (p < 1 × 10−6) were removed. Imputation was carried out, separately for each dataset, using the Michigan Imputation Server, with the r1.1 Haplotype Reference Consortium HRC panel40. Imputed SNPs with an info score r2 < 0.7, a quality threshold for the selection of imputed genetic variants, and MAF < 0.5% were removed, and all four studies were merged. The final dataset comprised 7,543,430 SNPs and 15,824 individuals (among which 8,769 PDAC cases). Logistic regression adjusted for age, sex, and the top eight principal components was applied to calculate summary statistics.

Data harmonization

The datasets for each exposure and outcome were aligned and merged. For each IV that was absent from the outcome dataset, a proxy SNP was searched using the 1000 Genomes database (European population) with a r2 > 0.99 to replace the missing one. The resulting datasets were harmonized so that the beta for the association with the exposure was positive and the IVs had the same effect alleles for both the exposure and the outcome. Finally, all betas and standard errors were standardized to the standard deviation scale, to have comparable quantities for the exposure and outcome associations. The complete workflow is reported in Fig. 1.

Statistical analysis

The inverse-variance weighted (IVW) method with multiplicative random effects was used as the reference model for causal estimation. The potential heterogeneity in the estimate was calculated using Cochran’s Q and the I2 statistics41, and weak instrument bias was assessed using the approximated first-stage F statistic42 and applying a debiased IVW estimator that is advocated to be higher than 2043. The weighted median, MR-Egger, Lasso, RAPS, and contamination mixture MR methods were applied as additional sensitivity methods to verify the consistency of the IVW result under more flexible and varying assumptions44,45,46. To identify potential “outlier” SNPs driving the associations, forest plots for single estimates and leave-one-out method calculations were evaluated, together with data from funnel plots, and the PRESSO outlier test which tests for the presence of horizontal pleiotropy47. Finally, to account for both potential pleiotropy and outliers in a single easy-to-interpret model, the radial regression-based approach of the IVW and Egger methods was applied, and the relative radial plots were inspected48. Specifically, the radial methods identify outliers based on their contribution to the total heterogeneity as estimated by the Cochran’s Q and Rucker’s Q statistics for the radial IVW and Egger, respectively48. In this way, we were able to identify possible outliers based on multiple non-redundant tests. A summary is reported within the workflow in Fig. 1.

The estimates obtained by each MR method were converted to odds ratios (ORs). An estimate was considered statistically significant if the IVW method and at least two additional methods among the weighted median, contamination mixture, RAPS, and lasso were statistically significant, conditional to the absence of heterogeneity and pleiotropy.

An adjusted threshold for statistical significance was calculated based on Bonferroni correction, by dividing the nominal threshold of statistical significance by the number of traits analysed: 0.05/(109 + 119) = 2.19 × 10−4.

All statistical analyses were performed using the packages MendelianRandomization (version 0.9.0), MRPRESSO (version 1.0), mr.raps (version 0.2), and RadialMR (version 1.1) in RStudio version 4.3.

Fig. 1
figure 1

Study Workflow.

Full size image

In this study, the exposure is the abundance of circulating metabolites and gut microbial taxa at the genus level of taxonomical classification. The outcome is PDAC risk. The workflow reported above shows the criteria for IV selection and the relative thresholds, and the methods applied to estimate the causal effect of each exposure on PDAC risk.

Results

Circulating metabolites

Among the 109 metabolites that were analysed, mannitol, methionine, and stearic acid were estimated as causally related with decreased PDAC risk, whereas carnitine, hippuric acid and 3-methylhistidine were causally associated with increased risk (Table 1). The results of the IVW method for mannitol and methionine (odds ratio per standard deviation [ORSD] = 0.97; 95% CI: 0.95–0.99, p = 0.006) and (ORSD=0.97; 95% CI: 0.94-1.00 p = 0.031 respectively) were supported by all other MR- and non-MR-based sensitivity analyses (Supplementary Figs. 2–3, Additional file 1), and no pleiotropy, outliers and heterogeneity were detected using Egger, radial Egger methods, PRESSO, Cochran’s Q (p > 0.05), and I2. One outlier was identified according to the radial IVW method for mannitol, but after its removal, the results were not significantly different (Supplementary Fig. 4, Additional file 1). Moreover, the potential issue of weak instrumental bias was excluded based on the approximated F statistic, which was 21.2 and 21.3 for mannitol and methionine, respectively, and based on the debiased IVW estimator (66.9 and 93.2, respectively).

Since only two IVs were used as proxy for stearic acid levels, only the debiased and radial IVW methods could be applied as sensitivity analyses (Supplementary Fig. 5, Additional file 1). Both alternative IVW methods supported the result of the standard IVW model (ORSD=0.93; 95%CI: 0.87–0.99, p = 0.027), and no evidence of weak instrument bias (approximated F statistic = 55.7, debiased IVW estimator = 77.4) was observed.

The causal effects of carnitine, hippuric acid and 3-methylhistidine on PDAC risk (ORSD=1.01; 95% CI: 1.00-1.03, p = 0.027, ORSD=1.02; 95% CI: 1.00-1.04, p = 0.038 and ORSD=1.05; 95% CI: 1.01–1.10, p = 0.02 respectively) were supported by most MR methods, and no specific issue related to heterogeneity in the estimates, weak instruments, pleiotropy, or the presence of outliers affecting the results was detected (Supplementary Figs. 6–9, Additional file 1). When considering the Bonferroni threshold, however, none of the reported causal associations between metabolites and PDAC risk remained statistically significant.

Summary statistics for mannitol, methionine, stearic acid, carnitine, hippuric acid and 3-methylhistidine are reported in Supplementary Table 3, Additional file 1.

Table 1 Causal effects of circulating metabolites on PDAC risk. The table reports the results of the MR analyses estimating the causal effect of circulating metabolites on PDAC risk. The number of IVs used as a proxy of each exposure is also reported for each metabolite. Nominally significant p-values ≤ 0.05 are marked in bold.
Full size table

Microbial taxa

Among the 119 genera analysed, a causal relationship was observed between increased abundance of the Romboutsia genus and decreased risk of PDAC, with an OR of 0.87 (95% CI 0.80–0.96) and p = 0.004 (IVW method). Notably, the approximated first-stage F statistic was 21.2, excluding potential bias due to weak instruments. Moreover, no heterogeneity was detected using Cochran’s Q (p = 0.565) and I2 (0%). The weighted median and contamination mixture methods supported the result, with ORs of 0.85 (95% CI: 0.74–0.97) and 0.80 (95% CI: 0.69–0.89), and p-values of 0.019 and 0.002, respectively. Additionally, both the radial IVW and the debiased IVW confirmed the direction and effect size of the IVW method (Table 2). Moreover, the reliability of the asymptotic approximation of the IVW estimate was checked, yielding a value of 75.65. The radial IVW and leave-one-out forest plots and PRESSO method did not show the presence of influential or outlier IVs (Supplementary Fig. 10, Additional file 1), and neither the Egger nor the radial Egger methods detected pleiotropy (pintercept > 0.05).

Supportive evidence of a causal effect between Clostridium sensu stricto 1 genus and lower PDAC risk was also observed, with an OR of 0.88 (95% CI: 0.78–0.99) and p = 0.027. No heterogeneity was observed based on Cochran’s Q (p = 0.625) and I2 (0%), and no weak instrument bias was suggested (approximated F statistic of 20.3). Lasso and RAPS methods supported the result with similar effect sizes: ORs of 0.88 (95% CI: 0.78–0.99) and 0.87 (95% CI: 0.79–0.97), respectively. The weighted median and contamination mixture approaches did not support the result, however, there were no outliers or influential IVs detected by either forest and funnel plots or the PRESSO method. The standard and radial Egger methods suggested an inverse direction of the effect, even though not statistically significant. The results of scatter, forest, and funnel plots are reported in Supplementary Fig. 11, Additional file 1. However, none of these two causal associations between microbial taxa and PDAC risk remained statistically significant after Bonferroni correction.

Summary statistics for Romboutsia and Clostridium sensu stricto 1 are reported in Supplementary Table 3, Additional file 1.

Table 2 Causal effects of gut microbial taxa on PDAC risk.
Full size table

Discussion

To our knowledge, this MR study is the largest to date to investigate causal associations between gut microbiota abundances using genetic associations reported by the MiBioGen consortium, having leveraged summary statistics from the PanScanI-III and PanC4 studies, including a total of 8,769 PDAC cases and 7,055 controls. Previous studies assessing the impact of gut microbiota abundances on PDAC risk have been conducted using the more modestly sized United Kingdom Biobank (case n = 587) and FinnGen (case n = 1416) GWAS summary statistics49,50. Additionally, we conducted thorough searches of the literature and existing databases to identify bacteria associated metabolites and related SNPs.

Metabolites

Higher circulating mannitol was found to be associated with reduced PDAC risk. Mannitol is a sugar alcohol produced by lactic acid bacteria and used as a sweetener as it does not induce hyperglycaemia due to its slow absorption by the body51,52. Dietary mannitol has been suggested as a prebiotic and has been seen to significantly increase concentrations of gut SCFAs butyrate and propionate in animal models53. There is a paucity of reported associations of circulating mannitol with PDAC risk, however metabolomic analysis of fecal samples from CRC cases and controls revealed mannitol was present exclusively in samples from controls54. A similar analysis performed on CRC tumour and adjacent tissue found mannitol was found exclusively in adjacent mucosa55. Additionally, dietary supplementation of mannitol has been demonstrated to promote lipid metabolism and may prevent obesity in mouse models56. Mechanistically, mannitol may exert indirect effects through stimulation of SCFA production and obesity prevention or through direct anticancer effects such as free radical scavenging and antiproliferative activities57,58.

Many cancer cells including tumour initiating cells are dependent on the essential amino acid methionine for progression in a process known as the Hoffman effect59,60. Further, a methionine restriction diet has been seen to inhibit cancer growth and enhance the efficacy of existing therapies in cell culture and animal models61,62. In contrast, we found that increased circulating methionine was protective against PDAC risk. Methionine is essential for T-cell proliferation and function indicating that methionine may be a double-edged sword63,64. Ming and colleagues found low methionine intake reduces T cell abundance, exacerbates intestinal tumour growth and impairs tumour response to immunotherapy in mice. Low methionine intake was seen to reduce the conversion of methionine to hydrogen sulfide by the gut microbiome, a mechanism critical for immune cell activation and survival65. In a sub-study (n = 31,626), they also found that UK Biobank participants with low protein/methionine intake had significantly higher overall cancer risk compared to participants in the high-intake group65. In vivo, circulating methionine has previously been noted to be higher in healthy volunteers relative to pancreatic cancer patients in two Japanese case-control studies66,67. In experimental studies it was observed that proliferation was reduced and apoptosis increased in pancreatic cancer cells following methionine treatment in vitro68. Methionine intake was also significantly inversely associated with pancreatic cancer risk in a large prospective cohort from Sweden (n = 81,922) and nested case-control samples within prospective cohorts from China (n = 387) and Singapore (n = 162)69,70. Therefore, our result expands the current knowledge of an existing association between circulating methionine levels and PDAC risk to a plausible causal association.

The long-chain saturated fatty acid stearic acid was associated with reduced PDAC risk. This finding should be interpreted with caution as only two IVs were available for our analysis; therefore, sensitivity MR analyses could not be run. Dietary stearic acid has been reported to promote the relative abundance of Akkermansia and Lactobacillus in the gut and bolster gut barrier integrity in mouse models71. Circulating stearic acid has been seen to be lower in PDAC patients relative to controls72,73. PDAC tissue was seen to have lower stearic acid concentration relative to healthy adjacent tissue in two separate cohorts74. Stearic acid has also been seen to significantly induce expression of the TNF-related apoptosis-inducing ligand (TRAIL), trigger apoptosis, and inhibit proliferation in pancreatic cancer cells in vitro74.

Higher carnitine was also found to be associated with an increased PDAC risk. Circulating carnitine has previously been associated with a reduced risk of multiple GI cancers and has established anti-inflammatory and antioxidant activities75,76,77. However, high circulating levels of carnitine have also been reported in patients with colon cancer78. Carnitine plays a significant role in shuttling fatty acids into the mitochondria for fatty acid oxidation, one of the most common ATP-generating processes of PDAC cells79,80. Expression of multiple carnitine transporters is altered in many cancer types, conferring a survival advantage by supplying carnitine essential for fatty acid oxidation81. In the case of PDAC, expression of the carnitine transporter SLC22A5 is associated with tumour progression82. Finally, knockdown of carnitine palmitoyltransferase 1 C (CPT1C), an enzyme that catalyses fatty acid carnitinylation, inhibited the tumourigenesis of PANC-1 cells in vivo and suppressed xenograft tumour growth in situ, illustrating the role of the carnitine system in PDAC83. Taken together, these studies suggest increased circulating carnitine may facilitate increased fatty acid oxidation and consequent cancer cell survival and progression.

Higher concentrations of hippuric acid, a glycine conjugate of benzoic acid, were found to be associated with increased risk of PDAC. Hippuric acid is a microbial–host co-metabolite produced in the liver, following the metabolism of phenylalanine to benzoic acid by the gut microbiome. Levels of circulating hippuric acid rise with the consumption of phenolic compounds (such as fruits and whole grains) and are associated with increased gut bacterial Shannon diversity and improved metabolic health84. However, circulating levels have been seen to be higher in colorectal cancer (CRC), pancreatitis and PDAC relative to controls85,86. Additionally, β-cell proliferation is increased by infusions of hippuric acid in mouse models87. A tissue metabolomic study also revealed an increase in hippuric acid between pancreatic tumour tissues and adjacent pancreatic tissues as well as its utility as prognostic marker in PDAC88. In mouse CRC models, circulatory hippuric acid levels positively associated with tumour weight and the expression of oncogenes, including ROBO3, JAK3 and BEST489. Additionally, at high concentrations, hippuric acid can disrupt the redox balance by inducing mitochondrial ROS production in vitro90,91. Therefore, our result is supported by the current knowledge and adds further evidence to a link between circulating hippuric acid and PDAC risk.

Increase in circulating concentrations of the post-translationally modified amino acid 3-methylhistidine was found to be associated with increased PDAC risk. A potential relationship between 3-methylhistidine and carcinogenesis has not been well characterised. Circulating levels have been observed to be increased in both case-control and prospective studies of prostate cancer but few other associations have been reported92,93.

Overall, considering the relatively small effect sizes, alterations in the individual assessed metabolite concentrations may not be clinically relevant in stratifying the population for PDAC risk. However, future studies may assess their utility as components of multifactorial risk assessments and may lead focused studies to explore the correlation between the metabolome and PDAC risk.

Gut microbiome

Clostridium sensu stricto 1 of the Firmicutes phylum, a genus that includes beneficial and pathogenic strains, was also found to be protective. Previously, Clostridium sensu stricto 1 has been shown to have lower abundance in fecal samples of patients with pancreatic cancer and pre-cancerous lesions relative to healthy and non-alcoholic fatty liver disease controls94. Higher gut abundances of Clostridium sensu stricto 1 have also been associated with a reduced risk of hepatocellular cancer in a study on 142 individuals95.

The butyrate producing Romboutsia genus of the phylum Firmicutes was found to be associated with reduced PDAC risk. Low abundances of Romboutsia have previously been associated with other GI tract cancers. Markedly lower Romboutsia levels were measured in colorectal polyp and tumour tissue relative to healthy colonic tissue96. Fecal abundances of Romboutsia have also been shown to be significantly lower in patients with hepatocellular carcinoma when compared to their healthy first-degree relatives97. Higher abundances of Romboutsia in stool samples have been associated with a decreased risk of esophageal cancer in a previous MR analysis98. Fecal abundances of Romboutsia have also been inversely correlated with serum markers of inflammation TNF-α and IL-1β, and markers of reduced gut barrier integrity D-lactic and diamine oxidase in patients with severe pancreatitis, an established PDAC precursor disease99,100. Members of the Romboustia genus, namely Romboutsia timonensis and Romboutsia ilealis were, respectively, observed to be depleted in the faeces of Spanish and Chinese patients with PDAC relative to controls through shotgun and 16 S sequencing approaches101,102.

There is little in the literature on interactions between Clostridium Sensu Stricto 1, Romboutsia and the reported metabolites. Circulating concentrations of methionine have previously been positively associated with abundances of Romboutsia in rodent models and abundances of Romboutsia and Clostridium were both negatively correlated with urine concentrations of hippuric acid in individuals treated with metformin103,104. However, these associations are tenuous and require functional studies for verification.

To our knowledge, only two studies assessing the contribution of genetically predicted variance in the microbiome or metabolites to pancreatic cancer risk using the PanScan and PanC4 genetic data from dbGaP have been published, both by Zhong and colleagues105,106. They identified 5 metabolites and 1 bacterium, as well as 44 metabolites, to be causally associated with PDAC risk, respectively; however, none of these were replicated in the present study. These discrepancies can likely be attributed to the lower info score of 0.3 used in the previous studies, whereas we included only those with an info score > 0.7, increasing reliability. Additionally, in the first study by Zhong and colleagues, they used the TWAS/FUSION framework based on four methods, specifically, best linear unbiased predictor (BLUP), least absolute shrinkage and selection operator (LASSO), Elastic Net (enet) and top SNPs to develop prediction models to identify associated metabolites and applied these models to the PanScan and PanC4 genetic data. The microbiome data used in their MR analysis was obtained from the Finnish FINRISK study, which comprised 5,959 individuals, making comparisons with the current study challenging105.

This study has several advantages. Firstly, we leveraged the most comprehensive summary statistics available for both gut microbiome populations and pancreatic cancer from the MiBioGen Consortium (18,340 individuals) and PanScan-PC4 studies (8,769 cases and 7055 controls), respectively. Moreover, being an MR-based study, any potential issue of reverse causation and confounding was eliminated. Finally, we have applied the most current MR methods and sensitivity analysis including stringent tests for invalid instruments as well as adhering to STROBE-MR guidelines107.

Our study is not without limitations. As SNPs at the more stringent genome-wide significant threshold 5 × 10−8 were not available for all exposures analysed, we used SNPs at the more relaxed threshold of 1 × 10−5. Our findings rely on genetic data exclusively from individuals of European ancestry. As such, they cannot be generalised to other populations. Additionally, for the gut microbiome taxa analysis, no species-level data are available. We therefore assessed features at a higher rank which may not be as biologically relevant due to the broad range of functions by members within these genera. Finally, many of the assessed circulating microbiome-associated metabolites have sources such as diet and endogenous metabolism. As such it is difficult to assess relative contributions.

In conclusion, we used two-sample MR to assess the impact of gut microbiome characteristics and circulating concentrations of gut microbiota-associated metabolites on PDAC risk. The genera Romboutsia and Clostridium sensu stricto 1 and metabolites mannitol, methionine and stearic acid were estimated to be causally related to decreased PDAC risk, whereas carnitine, hippuric acid, and 3-methylhistidine were implicated as causally related to increased PDAC risk.