Mendelian randomization linking metabolites with enzymes reveals pathway regulation and therapeutic avenues

Summary

Reactions between metabolites are catalyzed by enzymes. These biochemical reactions form complex metabolic networks, which are only partially characterized in humans and whose regulation remains poorly understood. Here, we assess human biochemical reactions and regulation using Mendelian randomization (MR), a genetic observational causal inference technique, to understand the methods’ strengths and weaknesses in identifying metabolic reactions and regulation. We combine four metabolite and two protein quantitative trait locus (QTL) studies to determine how well MR recovers 945 curated canonical enzyme-substrate/product relationships. Using genetic variants from an enzyme’s transcribed (cis) region as instrumental variables, MR-inferred estimates have high precision (35%–47%) but low recall (3.2%–4.6%) for identifying the substrates and products of an enzyme. Testing reverse causality from metabolites to enzymes using genome-wide instruments yields lower precision (1.8%–8.5%) and recall (1.0%–1.9%) due to an increased multiple-testing burden. Literature review of 106 Bonferroni-significant results identifies 45 links (43%) confirmed by different degrees of evidence, including bidirectional links between linoleate and cytochrome P450 3A4 (CYP3A4) levels (p = 8.6 × 10⁻³²). Eleven enzymes in the 106 links involve drug targets, allowing for an interpretation between N-acetyl putrescine and IL1RAP (p = 2.7 × 10⁻⁷), as IL1RAP is a target of the psoriasis drug spesolimab, and putrescine levels are elevated in psoriatic tissues. This work highlights how MR can be leveraged to explore human metabolic regulation and identify both canonical reactions and previously unknown regulation.

Graphical abstract

Mendelian randomization (MR) is widely used to identify causality between human traits; however, it suffers from high error rates. Van der Graaf et al. benchmark MR using human metabolic networks to understand MR’s failure modes. Extensive biological follow-up reveals that up to 43% of MR positives have plausible literature support.

Introduction

Metabolic reactions with unfavorable activation energy are catalyzed by enzymes. These enzymes are tightly regulated to maintain optimal metabolite concentrations under fluctuating environments, preventing both the depletion of essential metabolites and their accumulation to toxic levels. Enzymes can be regulated through processes such as competitive and uncompetitive inhibition, covalent enzyme modifications, phosphorylation, and pathway inhibition or activation.¹ Yet, our knowledge of human metabolism remains incomplete: (1) novel human metabolic pathways are still being discovered,² (2) the human genome still contains “orphan” enzymes predicted to have a catalytic function that have yet to be experimentally demonstrated,^3,4 and (3) the regulation of the activation or repression of metabolic pathways is still an active field of study, for instance through metabolic flux analysis.⁵

It is possible to model metabolite pathways as a series of causal relationships, where the enzymatic activity of a protein is causal to the concentrations of the substrate and product. This approach allows the identification of new putative metabolic links and regulatory mechanisms. These conclusions are usually derived from techniques such as randomized control experiments, metabolic flux analyses, or Bayesian causal networks and have successfully identified metabolic pathways in humans and other organisms.^5,6,7,8

In this work, we explore a different causal inference technique to infer human pathway regulation: Mendelian randomization (MR). MR is an observational causal inference technique that uses genetic information to establish directional links between an exposure and an outcome.^9,10 In principle, it is possible to identify pathway regulation between metabolites and proteins using MR. Motivated by the increased availability of well-powered metabolite and protein quantitative trait locus (mQTL and pQTL) studies, as well as the ability of cis-MR methodology to meaningfully improve drug target prediction,^11,12 we expect that MR is valuable in identifying the substrates and products of an enzyme and its regulation. To this end, we explore if and how MR can be used to derive meaningful conclusions in human metabolism and its regulation.

Here, we identify regulation between enzymes and metabolites through a form of observational causal inference called MR. MR allows the identification of causal relationships between a heritable exposure and an outcome even when these traits are not measured in the same individuals, using genetic variants that influence an exposure of interest. Should there be a causal relationship between this exposure and an outcome, the genetics of the exposure should also be associated with the outcome in a proportional manner. MR operates on three main assumptions. (1) The relevance assumption: the genetic variant needs to be robustly associated with the exposure of interest. (2) The independence assumption: the genetic variant needs to be independent from any confounder of the exposure-outcome relationship. (3) The exclusion restriction: the genetic variant needs to affect the outcome only through the exposure, and there should be no other paths. Violations of the exclusion restriction assumption—otherwise known as horizontal pleiotropy—are difficult to account for, especially when there is only a limited number of independent genetic variants available.^10,13 Recent advances in MR allow for pleiotropy-robust analysis of molecular phenotypes that have a single or only a handful of associated regions.^14,15

Our goal in this study is to identify to what extent enzymatic reactions can be reidentified by MR based on available pQTL and mQTL studies. Then, we aim to study enzyme-metabolite pairs that are significantly linked by MR but are not in the ground-truth benchmark set to estimate whether these reflect true human biology or rather false positives. For this, we use two pathway references to build a ground-truth metabolic map between enzymes and metabolites in humans (Figure 1). Then, using pQTLs derived from two independent studies^16,17 and mQTLs derived from four independent studies^18,19,20,21 (Figure 1A), we estimate the causal effects that proteins have on metabolites, and vice versa, using the protein levels as a proxy for enzymatic activity, distinguishing between MR results based on instruments in the cis region of the enzyme and those based on meta-analyzing all cis and trans regions together. We find that MR methods have limited discriminative ability to determine if a metabolite is catalyzed by a specific protein. Yet, Bonferroni-significant MR estimates are enriched for true reactions. An extensive literature search of the Bonferroni-significant combinations that are not in our ground-truth metabolic map proposes possible mechanistic explanations for some of the newly identified links and points out multiple drug targets among the involved enzymes, suggesting putative metabolite downstream drug effects.

Figure 1.

Overview of the data used in this study

(A) The different metabolite quantitative trait locus (mQTL) studies and protein quantitative trait locus (pQTL) studies, combined with two pathway reaction references, Expasy + KEGG and MetaCyc. Based on the matching of proteins to metabolites, we identify 1,742 ground-truth positive enzyme metabolite links.

(B) The number of measured ground-truth positive enzyme metabolite links present in each pathway reference.

Methods

mQTL studies

We analyzed four metabolite quantitative trait studies, as described in our previous publication.¹⁵ In short, the summary statistics for four mQTL studies were downloaded from their respective resources,^18,19,20,21 and the metabolites were harmonized to Human Metabolite Database (HMDB) identifiers. This resulted in 1,109 metabolites to which an HMDB ID was assigned.

pQTL studies

We analyzed two pQTL studies: the DeCode study¹⁷ and the UK Biobank Pharma Proteomics Project (UKB-PPP).¹⁶ The DeCODE study comprises genetic associations of 4,907 plasma proteins across 35,559 individuals. The PPP study contains genetic associations for 2,923 plasma proteins across 54,219 individuals. The summary statistics and meta-data were downloaded from their respective resources and harmonized using the UniProt identifiers, which were provided by both studies.

Identification of catalytic enzymes

We used two independent resources to build our ground-truth metabolic map: KEGG + Expasy and MetaCyc. For KEGG + Expasy, we used the Expasy enzyme resource to retrieve enzymes that have catalytic effects in humans²² and the KEGG resource to identify metabolites that are involved in the reactions. Out of 8,227 enzymes from the Expasy resource, we identified 4,277 enzyme commission (EC) numbers that match human enzymes, including 3,592 unique enzymes (some enzymes catalyze multiple reactions).²³ We used KEGG to match the compounds to each enzymatic reaction. We identified 229 human KEGG pathways in which the measured compounds are present and downloaded the XML files for each human pathway. From these, we derived the 1,489 reactions for which protein and reactant information was available for at least one pQTL and mQTL study. After matching these reactions to the enzymes that catalyze them, this resulted in 1,236 measured enzyme-reactant combinations. As a secondary metabolite reference, we used the HumanCyc v.24.0 database to identify enzymes and reactants that are present in a pathway.²⁴ Matching a total of 1,925 unique reactions that can be catalyzed by 2,933 unique proteins. After matching the reactions to the measured proteins and metabolites, we ended up with 231 unique exposure and outcome combinations.

Summary statistics harmonization

We jointly harmonized summary statistics from all pQTL and mQTL studies in the same way: first, if necessary, the files were lifted over to the human genome build GRCh37 to match our allele reference using the UCSC liftover tool (https://genome.ucsc.edu/cgi-bin/hgLiftOver). Second, variants were matched to the UK10K linkage disequilibrium (LD) reference panel,²⁵ matching variants on their genomic positions and alleles, while removing palindromic variants and variants that have a minor-allele frequency of less than 0.5% in the panel.

We then harmonize the effect sizes of the summary statistics to the standard deviation per standard deviation scale:

$${\beta }_{standardized}=\frac{z}{\sqrt{n+z^2}},\text{and}$$

$$se\left({\beta }_{standardized}\right)=\frac{1}{\sqrt{n+z^2}}\text{,}$$

where n is the variant sample size and z is the Z score of the association. We set n to be the maximum sample size if n was not available for a given variant. The p value and effect size sign of the variant-trait association were converted into a Z score if the Z score was not available.¹⁵ This normalization step converts all genome-wide association study (GWAS) summary statistics into a magnitude of standard deviation change per allele change. Therefore, all MR estimates reported in this study are on a standard-deviation-per-standard-deviation scale. If the per-variant sample size was available, we removed variants that were missing more than 5% of the maximum sample size in a region.

The genotype information underlying the LD matrices for the UK10K data resource was downloaded from the EGA under accession IDs EGA: EGAD00001000740 and EGA: EGAD00001000741. For data access, see https://www.uk10k.org/data_access.html.

Identification of associated regions

We identified associated exposure regions used for identification of MR instrumental variables (IVs) through p value clumping of genetic variants at a p value threshold of smaller than 5 × 10⁻⁸, using the plink (v.1.90b7)²⁶ –clump command with a clumping window of 250 kb and an LD threshold of 0.01 r². Overlapping clumped regions were combined. Non-overlapping regions were analyzed by all four cis-MR methods.¹⁵ If a method required IVs, these were clumped from the full region using the same p value and LD threshold. To reduce the chance of estimating reverse causality, we removed the protein cis region if the protein is the outcome in the meta-analysis.

As all the instruments selected are univariable and have p < 5 × 10⁻⁸ (Z score = 5.32), we estimate that the F statistic for any IV must be > 28, ensuring that there are no weak instruments used in the study.

MR and colocalization methods used in this study

We used four MR methods: MR-IVW, MR-IVW-LD, MR-PCA, and MR-link-2. These methods have been chosen because they are appropriate to use on one or multiple associated regions. We have also performed statistical colocalization on all the loci using the “coloc” package.²⁷ Code implementations for all the methods can be found at https://github.com/adriaan-vd-graaf/mrlink2.

We performed three styles of MR analyses: (1) cis-MR, (2) cis + trans MR, and (3) meta-analyzed bidirectional meta-analyzed MR. We performed exposure-region-based MR for every exposure-associated region in both protein-to-metabolite and metabolite-to-protein directions and meta-analyzed the region-based MR estimates.

In cis-MR, we performed regional MR when any clumped variants in the gene region (defined in the section identification of associated regions) are within 250 kb of the start or end of the gene that transcribes the enzyme according to Ensembl BioMart (https://www.ensembl.org/info/data/biomart/index.html).

In the cis + trans analysis, we analyzed all associated regions of an enzyme and treated each individual exposure-associated region as an independent MR analysis. Here, we only test the direction from the enzyme to the metabolite.

In the meta-analyzed MR, we performed a bidirectional MR identifying causality between enzymes and metabolites and between metabolites and enzymes. This can identify canonical metabolic reactions (from the enzyme onto the metabolite) but also their regulation (both the enzyme onto the metabolite and the metabolite onto the enzyme). We performed an MR for all exposure genetic associations of an enzyme or a metabolite, and we meta-analyzed these estimates. The enzyme-to-metabolite direction is the only direction that we consider truly causative, but we include both directions in the final approach, as we perform biological follow-up in both directions.

Discriminative ability analysis using AUC

We used the area under the receiver operator characteristic curve (AUC) to assess the discriminative ability of different MR methods. True links were defined as enzymes and their reactants, whereas the true non-links are all other protein-metabolite combinations. The AUC was calculated with the scikit-learn library.²⁸

Manual literature search and classification of evidence

For all MR-link-2-significant estimates in the meta-analyzed MR analysis, we performed a literature review. Specifically, we searched for both the protein-metabolite combination and the metabolite in combination with the products and substrates of the reaction that the protein catalyzes in PubMed, OMIM, and the BRENDA enzyme database.^29,30 We classified our findings into five categories of evidence.

• (1)Canonical reaction: significant combinations where the exposure is the enzyme that directly catalyzes the metabolite outcome.
• (2)Reverse canonical evidence: causal relationships where the metabolite acts on its catalyzing enzyme.
• (3)Pathway evidence: causal relationships where the metabolite and the enzyme are in the same pathway.
• (4)Homologous evidence: causal relationships where a paralog (similar gene in the same organism) or ortholog (similar gene in another organism) is related to the metabolite.
• (5)Supporting evidence: these combinations are cryptic and may represent novel and unknown regulations; it is, however, also possible that these combinations are false positives.

OpenTargets matching drugs to their consequences

We downloaded the “molecule” and “diseases” tables from the OpenTargets platform (https://platform.opentargets.org/downloads; June 2024) and matched 1,550 OpenTargets drug targets to 358 of the unique enzyme measurements. We conducted a literature search of all drug targets that are Bonferroni significant in our original analysis.

Network degree of metabolites

Using the igraph (v.0.11.6) package in Python, we created a full network between all the available metabolites in KEGG + Expasy and in MetaCyc. We overlapped the measured metabolites in both graphs and identified the number of incoming nodes and the number of outgoing nodes for each metabolite and enzyme. We used Spearman correlation between the degree of the created graphs and the estimated absolute MR effect sizes and p values.

Tissue-specific eQTL studies

To identify the merit of plasma proteomics compared to tissue-specific gene expression, we analyzed all tissues of the GTEx consortium v.8 eQTL dataset, which were downloaded from the eQTL catalog.^31,32 Briefly, the GTEx consortium v.8 eQTL dataset contains (cis) eQTLs for up to 670 individuals across 49 different tissues. We matched the gene expression Ensembl IDs to UniProt identifiers in our comparisons with the proteomics dataset.

Enrichment analysis of metabolite types

We performed an enrichment analysis using a Fisher’s exact test to identify enrichment of HMDB metabolite classes (at three resolutions: super class, class, and subclass of metabolites) across the ground-truth positives of the MR analysis, comparing the number of findings outside the class to the number of findings inside the class. Across all ground-truth positive causal combinations in the cis + trans analysis, we determined how often metabolites from a certain class were found among the Bonferroni-significant results compared to all metabolites outside the selected class.

GWAS overlap with Sakaue et al.

We assessed whether the pleiotropy of an exposure-associated locus had any impact on the findings of the MR methods. Using 214 case-control GWAS traits analyzed by Sakaue et al.,³³ we counted for each exposure-associated region how many complex traits the lead exposure SNP is associated with (at p < 5 × 10⁻⁸). If this SNP was not available in a GWAS, we took the most significant proxy SNP in a 100 kb window at a minimum LD of 0.9 r² using plink with the UK10K as the LD reference.

Results

Matching pathway references with pQTLs and mQTLs

To build a ground-truth positive dataset of enzymatic reactions that can be matched to in pQTL and mQTL studies, we first use two human pathway reference datasets to identify enzymes and their reactants in our metabolic map: KEGG + Expasy and MetaCyc to build a metabolic map^24,34,35 (methods). KEGG + Expasy contains 1,489 reactions, while MetaCyc contains 1,925 reactions (Figure 1A). Each reaction consists of one enzyme, one or more substrates, and one or more products (Table S1; methods). We matched enzymes from our metabolic map to two pQTL studies: the DeCODE study, comprising 4,907 proteins across 35,559 individuals,¹⁷ and the UKB-PPP, comprising 2,923 proteins across 54,219 individuals (Figure 1A). After matching proteins to our metabolic map, 2,696 and 2,180 protein measurements had catalytic activity according to KEGG + Expasy and MetaCyc, respectively (Figure 1A). Across pQTL studies, 1,624 unique proteins had a catalytic function according to at least one pathway reference (Table S2). We matched the metabolites from our metabolic map to those in four mQTL studies: Shin et al., Lotta et al., Chen et al., and Borgess et al.^18,19,20,21 This resulted in 330 unique enzymes across 437 measurements in reactions with 127 unique metabolites across 214 measurements. Overall, our ground-truth positive metabolic map encompasses 1,310 protein-metabolite pairs, including 540 and 770 pairs in which the metabolite acts as substrate and product, respectively (Figures 1A and 1B; Table S3).

When considering these ground-truth positive reactions from the metabolic map, enzyme abundances to reactant concentration are the ground-truth positives, as we expect that enzyme abundance causally influences the concentrations of substrates and products. We make a distinction between the enzyme-onto-metabolite direction vs. the metabolite-onto-enzyme direction, as the former can be understood as directly causative: an enzyme converts a metabolite in a well-understood physical process. As a reaction occurs, we expect substrate depletion and an increase in product concentration. Under the assumption that enzymes catalyze a reaction in a single direction, we expect to see negative causal estimates from the enzyme to the substrate and positive causal estimates from the enzyme to the product. In contrast, we do not consider the metabolite-onto-enzyme direction directly causative, as there is no obvious physical process through which the reactant influences the abundance of the catalyzing enzyme. Instead, we consider such influences as indirect regulatory effects, such as, for instance, the regulation between cholesterol concentrations and the degradation of high-mobility group (HMG)-coenzyme A (CoA) reductase,³⁶ as well as between xenobiotics and cytochrome P450 3A4 abundances.³⁷

MR analyses

We perform MR using four cis-MR methods (MR-IVW, MR-PCA, MR-IVW-LD, and MR-link-2) to test how well MR can identify the causal relationships between an enzyme and its substrates and products (methods). We consider three forms of MR analysis: (1) cis-MR, where only the genetic variants proximal to the gene encoding for the enzyme of interest are used as IVs in the MR; (2) cis + trans MR, which performs MR independently for each exposure-associated locus; and (3) bidirectional cis + trans meta-MR, which meta-analyzes exposure-region-based MR estimates (methods).

First, for the cis-MR analysis, we identify enzyme-to-metabolite effects using only the cis region (where the enzyme is transcribed) as a source of MR instruments.

Secondly, in cis + trans MR, we perform MR in each exposure-associated region separately, yielding multiple (region-specific) causal effect estimates. These MR effect estimates are tested against the ground-truth set independently.

Thirdly, for meta-analyzed MR, we combine the individual region-based MR estimates from the cis + trans MR for each enzyme-to-metabolite MR, as well as for each metabolite-to-enzyme MR, using inverse-variance weighting meta-analysis (methods).

We only test the enzyme-to-metabolite direction when performing cis and cis + trans MR, as we are testing the merit of these methods for identifying enzymatic reactions, and a cis region is not available for a metabolite. In contrast, for the meta-analyzed MR, we also perform the metabolite-to-protein direction, as here, we broaden the research question and assess any potential indirect regulation through a literature search.

In all these estimates, we consider the protein-to-metabolite direction to be the only truly causative direction, and we perform literature follow-up on significant findings only for the most conservative meta-analysis MR approach.

Cis-MR to identify reactants and their enzymes

We applied MR to cis regions that associate (SNP p ≤ 5 × 10⁻⁸) with 159 enzymes (197 measurements) that are in the ground-truth positive metabolic map. These enzymes catalyze reactions involving 88 metabolites (139 measurements), leading to 499 protein-metabolite pairs that are the ground-truth positives in this study (Table S4). All other 25,420 possible pairwise comparisons between enzymes and metabolites are considered ground-truth null links. We provide an example cis-MR result in Figure 2. Here, arginase 1 converts arginine into ornithine and urea (EC 3.5.3.1) (Figure 2A) along with the MR-link-2 results, where the product causal direction is strictly positive (Figures 2B–2D).

Figure 2.

Example of MR effect estimation

(A) Considering the reaction catalyzed by arginase 1 (ARG1), with arginine as a substrate, and ornithine and urea as the products, one can estimate a causal relationship between the enzyme and the reactants, provided they are measured in each study.

(B) The causal estimates between ARG1 and the arginine substrate; since arginine is a substrate, a negative causal effect is expected.

(C) The causal estimates between ARG1 and the reaction product ornithine.

(D) The causal estimates between ARG1 and the reaction product urea. It is not possible to estimate a causal effect if either the enzyme or the metabolite is unmeasured in a single study. p values shown are from the MR-link-2 likelihood ratio test.

As cis regions are only available for enzymes, this methodology will only be able to identify causality from enzyme concentrations to metabolite levels. When testing for the directionality of MR methods on substrates and products, MR-link-2 is the only method that identifies a significantly larger causal effect for products than for substrates (Kruskal-Wallis p = 0.011) (median product causal effect estimate $\hat{\alpha }$ = 0.008, median substrate $\hat{\alpha }$ = −0.013) across 219 substrates and 320 products from our metabolic map (Figure S1).

Using the ground-truth positive enzyme-metabolite combinations, it is possible to determine the discriminative ability of different MR methods, which is limited, with an AUC ranging between 0.543 and 0.554 (Table S5; Figure S2). Here, we identify causality based on p values irrespective of the effect sign, as this may be difficult to discern, as enzymes often have bidirectional catalytic effects. We found that precision (proportion of ground-truth positives over all detected) for all MR methods was highest in the low p value range; we show precision and recall (proportion of ground-truth positives over all positives) on a log scale (Figure S3A and S3B). Analyzing causal estimates below the cis-specific Bonferroni threshold ($p<\frac{0.05}{25,919}=1.93\times {10}^{-6}$), all methods had a low recall of 3.2%–4.6% combined with a high precision of 35%–47% (Figure S3). Each MR method had similar precision and recall at any specific p value rank, with MR-IVW-LD having the highest recall (with lower precision) and MR-link-2 the highest precision (with lower recall), in line with our previous work.¹⁵ Considering only unique enzyme-to-metabolite combinations, MR-PCA identified the largest number (n = 28) of unique enzyme-to-metabolite links, including 12 ground-truth positive controls (Figure S3C) and 16 likely false positives (Figure S3D). In contrast, MR-link-2 identified eight ground-truth positive controls (Figure S3C) and five likely false positives (Figure S3D).

To extend the true ground-truth positive set, we added additional links that were reported in the BRENDA database³⁰ (methods). We determined for each enzyme-metabolite combination whether any of the metabolites had been found as a substrate, product, or inhibitor in the enzymes that were otherwise missed by our original ground-truth positive dataset. Two out of five unconfirmed links found by MR-link-2 were inhibitors or substrates according to BRENDA. In the case of MR-PCA, 4 out of 16 links represented a substrate, product, or inhibitor relationship (Table S6; Figure S4; methods).³⁰

Some enzymes catalyze multiple reactions, suggesting that their change in abundance may have a smaller influence on a single metabolite, leading to smaller MR estimates. Indeed, we identified only negative correlations (Spearman r range = −0.08 to −0.06) between the number of reactions that a metabolite is involved in and the absolute effect sizes of the MR analyses when including that metabolite as an outcome (Figure S5; Tables S3 and S7). This implies that highly connected metabolites in the metabolic map are less amenable to univariable MR analysis, as their concentration is determined by multiple factors (Table S7; Figure S5; methods). Similarly, we identify negative correlations when comparing the MR effect size when the enzyme catalyzes more reactions (Spearman r range: −0.057 to −0.067).

Using the node degree information to define a more reliable set of null links allowed us to reduce the case-control imbalance in our analysis, which could impact receiver operator characteristic curves from which AUC is derived. For this, we limit the true null links in our analysis to the enzymes that catalyze more than 11 reactions in the KEGG + Expasy network and the number of reactions the molecule is involved in (Table S3; Figure S5). Such a restriction may favor enzymes that are more widely researched, and hence, the lack of evidence more likely translates to evidence of absence. In this analysis, the AUC values are comparable (0.545 compared to 0.554 for MR-PCA) to the original analysis considering all null links in our analysis (Figures S6A–S6C), and no large differences between MR methods were observed (AUC range: 0.529–0.545). The largely unchanged results may imply that the high-confidence null set may not contain more true null links.

The protein measurements used in this study are based on blood plasma, which is unlikely to be the most relevant tissue for the metabolic reactions considered in our metabolomic ground-truth dataset. Therefore, we used cis eQTLs from all tissues (in the GTEx dataset) for the different MR analyses instead of blood pQTLs.³¹ Across all tissues, we identify up to 12 ground-truth positives at Bonferroni-significant thresholds. One of these true positives (glycine amidinotransferase-to-guanidinoacetate) was not found when we used blood pQTLs, while the amylase alpha 1C-to-maltose true positive link was unique to the results based on pQTLs (Table S8). Interestingly, when analyzing tissues independently, whole blood led to the highest recall (8 ground-truth positives), while liver exhibited only one ground-truth positive. However, it is likely that the number of true positive findings is more driven by sample size (liver n = 208 vs. blood n = 670) than the relevance of the tissue source of the eQTL (Table S8).

Cis + trans MR to extend causality

Considering that the cis-MR results had high precision with relatively low recall, we considered that including trans-associated loci could increase the number of causal links that may be biologically informative. We set out to investigate the ability of MR methods to identify further protein-to-metabolite causality, using trans-associated loci as regions that may individually contribute to identifying ground-truth positives (cis + trans).

After analyzing a total of 451,502 combinations of exposure region-enzyme-metabolite triplets, of which 4,987 are considered ground-truth positives, at a Bonferroni significance threshold ($\frac{0.05}{451,502}=1.1\times {10}^{-7}$), recall ranged between 0.4% and 1% depending on the MR method (Figure S7A), while precision ranged between 11.6% and 1.8% (Figure S7B; see data and code availability). Even though we expected to increase recall in the trans analysis over the cis analysis, at the cost of precision, the inclusion of separate trans-region-based MR estimates identified fewer unique true positives for MR-link-2 compared to the cis-MR analysis. This is likely due to the increased multiple testing burden and an increase in the set of positives, as not all ground-truth positives can be tested in the cis analysis. At each method-specific Bonferroni p value threshold (either cis only or cis + trans), MR-link-2 identified eight unique ground-truth positives in the cis analysis (p < 1.93 × 10⁻⁶), compared to seven in the cis + trans analysis (p < 1.1 × 10⁻⁷). Applying the more lenient cis-only p value threshold (p < 1.93 × 10⁻⁶) to the cis + trans analysis, MR-link-2 identifies 9 significant ground-truth positives, which is one more than obtained in the original cis analysis (Figures S3C, S7C, and S7D; Table S5). Nonetheless, including trans regions in an analysis is not without merit: some trans regions were individually significant at the stringent multiple testing threshold; for instance, ARG1 abundances were regulated in trans on chromosome 19, providing a concordant Bonferroni-significant true positive causal estimate on ornithine concentrations ($\hat{\alpha }=0.67,p=4.0{\times 10}^{-8}$) (see data and code availability).

Similar to the cis-only analysis, we tested if MR effect sizes produced by the cis + trans MR were correlated with the number of reactions in which a metabolite or an enzyme might be involved. When considering the number of reactions an enzyme is involved in, we identify exclusively negative correlations (Spearman correlation range: −0.016 to −0.047). In the metabolite case, all but one comparison has a negative Spearman correlation (Spearman correlation range: −0.011 to 0.005) (Table S7; methods). These correlation values are smaller than the ones obtained for the cis-MR effects, which hints that trans results may be noisier.

Using the large numbers of combinations in the cis + trans analysis allowed us to estimate which metabolite classes were overrepresented in the ground-truth positives of the MR analysis. Across 324 metabolite class combinations (leading to an enrichment-specific Bonferroni p threshold of 1.5 × 10⁻⁴), we identified significant positive enrichments for the HMDB super class “organic acids and derivatives” (max p = 1.98 × 10⁻⁸), the HMDB metabolite class “carboxylic acids and derivatives” (max p = 1.1 × 10⁻⁴), and the HMDB metabolite subclass “amino acids, peptides, and analogs” (max p = 3.6 × 10⁻⁶) (Table S9). This suggests that MR can more often identify causality when the outcome metabolite is an organic acid.

Including trans loci in molecular traits can lead to false positives due to pleiotropy. Some of these assumption-violating genetic regions may be picked up by looking up pleiotropic signals of the top SNPs for complex diseases. For this, we combined 214 case-control GWASs from Sakaue et al.³⁸ and identified how often the most highly associated SNP from the exposure region was associated with complex diseases at the genome-wide significance level (methods). Each exposure locus overlaps a median of 5 case-control traits, with a maximum of 38. We found that the number of Bonferroni-significant causal effects increased as the exposure locus became more pleiotropic (Figures S8A–S8C; Table S10), while the positive predictive value (PPV) (at each method’s Bonferroni-significant level) started to drop as the exposure SNP was associated with more than 10 diseases (Figure S8C). When only looking at the top 1,000 most significant MR results for each method (making the methods comparable in terms of rank), we found that the highest PPV was found for exposures whose SNPs were associated with 4–15 complex traits (Figure S8D).

Sometimes, investigators use colocalization methodologies to identify violations of MR assumptions. After applying coloc to all loci in the cis + trans analysis and selecting for the inclusion of loci at a PP4 > 0.9, we identify that there is an increase in precision for MR-IVW, MR-IVW-R, and MR-PCA, which comes at the cost of recall (Figure S9; Table S11). However, since MR-link-2 implicitly assesses colocalization in its model, it did not benefit from an increase in precision or recall upon filtering for coloc PP4 > 0.9 (Figure S9; Table S11).

Meta-analyzing MR between enzymes and metabolites

Based on the information that cis as well as trans regions can be informative, we decided to combine all associated regions into a single causal estimate per enzyme-to-metabolite pair, as well as in the reverse direction, additionally testing the effect of metabolites on enzyme abundances using the standard inverse-variance weighting meta-analysis (methods). This is the first analysis where we include the reverse metabolite-to-enzyme regulation direction. We still only consider the enzyme-to-metabolite direction causative and attribute any significant results in the metabolite-to-enzyme direction to indirect regulation, which we later follow-up in our literature search. Combining all enzyme-metabolite pairs resulted in 187,036 combinations (437 enzymes × 214 metabolites × 2 directions), of which 146,062 could be instrumented by MR, setting our meta-analysis Bonferroni-corrected threshold for significance at p$<\frac{0.05}{146,062}={3.42\times 10}^{-7}$ (methods).

Of these, 945 enzyme-to-metabolite combinations represent ground-truth positive controls in our metabolic map (Table S12). We find that all tested MR methods in meta-analysis have modest discriminative ability (AUC: 0.512–0.531) (Table S13). When considering Bonferroni-significant results, there is strong enrichment for true metabolite reactions across all methods (odds ratio: 2.9–13.2, p = 3.0 × 10⁻⁴–7.4 × 10⁻⁸) (Table S14). The differences in precision and recall between MR methods were more pronounced compared to cis-MR (Figures 3A and 3B); MR-link-2, MR-PCA, MR-IVW, and MR-IVW-LD identified 106, 660, 870, and 912 Bonferroni-significant effects, respectively. Methods differ strongly in their precision (Figure 3A; Table S12), with the best precision obtained by MR-link-2 (8.5%) compared to up to 2.3% for other methods. This comes at the cost of lower recall: nine out of 945 true comparisons pass Bonferroni significance with MR-link-2, compared to 19 for MR-IVW-LD (Figure 3B; Table S12).

Figure 3.

Precision and recall curves of different methods for bidirectional MR between enzymes and metabolites when meta-analyzing results across exposure loci

(A and B) Recall (A) and precision (B) as a function of MR p value ranking.

(C) Set memberships for the true positives identified by each method at Bonferroni-significant threshold. The plot visualizes how the true, significant protein-metabolite link overlaps across methods: black dots indicate the set membership of a method. The vertical bars show the number of links in each of the presented sets of true positives, while the horizontal bars denote how many true positives were found by each method.

(D) Set memberships of ground-truth null findings (false positives) at each method’s Bonferroni threshold. The plot is the same as that in (C) but for ground-truth null findings (identified at the Bonferroni-significance threshold). TP, true positive; FP, false positive; FN, false negative.

These precision and recall curves can also be used to compare MR methods. Generally, all methods have similar precision and recall, but there are some differences, especially when considering the highest-ranked predictions. Here, we see that MR-link-2 shows higher recall (more true positives) than less pleiotropy-robust methods (MR-IVW, MR-IVW LD, and MR-PCA) (Figure 3A; Table S12). A similar trend can be observed for precision in these critical ranges, where MR-link-2 has higher precision than the other tested methods, while the other methods follow roughly the same pattern (Figure 3B; Table S12).

Significant meta-analyzed MR-link-2 results show different types of evidence for pathway regulation

Since our ground-truth dataset of reactions represents only a fraction of the full literature evidence linking metabolites and enzymes, we performed a follow-up literature review to better annotate an MR-derived positive set of candidate reactions that were not reported in our imperfect metabolic map. For each protein-metabolite pair, we meta-analyzed all region-based MR estimates to yield the most robust links. Considering the labor-intensive process of a manual literature follow-up, we chose to analyze only the Bonferroni-significant MR-link-2 results, as such a set showed the highest precision among methods. We manually annotated all 84 unique (across 106 pairs of measurements) MR-link-2 Bonferroni-significant meta-analyzed results (Figure 4) and classified the identified links into five classes: (1) “evidence of a canonical reaction” (n = 12, 11%), (2) “reverse canonical evidence” (i.e., the metabolite influences the enzyme, n = 8, 7.5%), (3) “pathway evidence” (i.e., the metabolite and enzyme share a pathway, n = 6, 5.6%), (4) “homologous reaction evidence” (i.e., the enzyme was shown to interact with the metabolite based on homologous enzymes, n = 4, 3.8%), and (5) “supporting evidence” (i.e., biological evidence present that the enzyme and metabolite are related in model organisms or shared biological pathways, n = 15, 14%). The manual annotation of the Bonferroni-significant results, excluding “supporting evidence,” is shown in Table 1, and all annotations are visualized as a graph in Figure 4. The manual annotations of the significant MR-link-2 meta-analyzed results that include “supporting evidence” are listed in Table S15.

Figure 4.

Bonferroni-significant MR-link-2 results with their biological evidence

The pie chart depicts the numbers found for each class of interpretation and serves as a legend. The metabolites are depicted as diamonds, and proteins are depicted as circles. A blunted arrow indicates a negative causal relationship, and a pointy arrow indicates a positive causal relationship.

Table 1.

Bonferroni-significant MR-link-2 results with supporting evidence

Evidence	Causal relationship (no. of significant estimates)	Causal estimate	p value	Interpretation
Canonical evidence	ARG1 → arginine (3×)	−0.66	6.8 × 10−47	arginase 1 converts arginine into ornithine and urea (KEGG: R00551)
Canonical evidence	ARG1 → ornithine (2×)	0.39	1.4 × 10−21	arginase 1 converts arginine into ornithine and urea (KEGG: R00551)
Canonical evidence	NPL → N-acetylneuraminate	−0.37	2.3 × 10−17	N-acetylneuraminate pyruvate lyase converts N-acetylneuraminate into pyruvate and N-acetyl-D-mannosamine (KEGG: R01811)
Canonical evidence	PKLR → pyruvate (2×)	0.21	1.7 × 10−15	pyruvate kinase converts phosphoenolpyruvate into pyruvate and ATP (KEGG: R00200)
Canonical evidence	AMY1(A|B|C)→ maltose	0.17	2.8 × 10−13	the genes AMY1A, AMY1B, and AMY1C encode for glycosylases that convert starch and related oligosaccharides into maltose (KEGG: R11262)
Canonical evidence	TYMP → 2′-deoxyuridine	−1.18	3.1 × 10−18	thymidine phosphorylase converts deoxyuridine and orthophosphate into uracil and 2-deoxy-D-ribose 1-phosphate (KEGG: R02484), missed in our reaction reference
Canonical evidence	ACY1 → N-acetylglutamate	−0.13	3.9 × 10−8	aminoacylase converts an N-acyl-amino acid (e.g., N-acetyl-glycine) into a carboxyl and amino acid (e.g., glycine) (KEGG: R01263), missed in our reaction reference
Canonical evidence	ACY1 → glycine	−0.10	3.0 × 10−7	aminoacylase converts an N-acyl-amino acid (e.g., N-acetyl-glycine) into a carboxyl and amino acid (e.g., glycine) (KEGG: R01263), missed in our reaction reference
Reverse canonical evidence	N-acetyl-aspartyl-glutamate → FOLH1 (2×)	−0.10	6.4 × 10−40	glutamate carboxypeptidase II (FOLH1) converts N-acetyl-aspartyl-glutamate into N-acetyl-l-aspartyl-aspartate and glutamate (KEGG: R10687)
Reverse canonical evidence	aspartate → DARS1	0.39	3.2 × 10−7	aspartyl-tRNA synthetase (DARS1) converts tRNA and aspartate into aspartyl-tRNA (KEGG: R05577)
Reverse canonical evidence	alanine → ACY1	0.30	3.2 × 10−13	aminoacylase converts N-acyl-amino acid (e.g., N-acetyl-alanine) into a carboxyl and an amino acid (e.g., alanine) (KEGG: R01263), missed in our reaction reference
Reverse canonical evidence	N-acetylneuraminate → NPL	−0.64	2.5 × 10−8	N-acetylneuraminate pyruvate lyase converts N-acetylneuraminate into pyruvate and N-acetyl-D-mannosamine (KEGG: R01811)
Reverse canonical evidence	2′-deoxyuridine → TYMP	−0.30	1.6 × 10−7	thymidine phosphorylase converts deoxyuridine and orthophosphate into uracil and 2-deoxy-D-ribose 1-phosphate (KEGG: R02484), missed in our reaction reference
Reverse canonical evidence	arginine→ ARG1 (2×)	−0.32	1.2 × 10−14	arginase 1 converts arginine into ornithine and urea (KEGG: R00551)
Pathway evidence	TYMP → uridine (2×)	−0.57	3.3 × 10−11	thymidine phosphorylase converts uracil into 2′-deoxyuridine; uridine and 2′-deoxyuridine are separated by two reactions: uridine → uracil → 2′-deoxyuridine
Pathway evidence	CYP3A4 → linoleate	0.36	7.9 × 10−21	cytochrome P450 3A4 is an oxidizing enzyme with a wide range of substrates, including fatty acids; linoleate is a substrate of human recombinant CYP3A4 and human and rat liver microsomes39
Pathway evidence	linoleate → CYP3A4	0.32	8.6 × 10−32	cytochrome P450 3A4 is an oxidizing enzyme with a wide range of substrates, including fatty acids; linoleate is a substrate of human recombinant CYP3A4 and human and rat liver microsomes39
Pathway evidence	1-stearoyl-2-oleoyl-gps → PCYT2	0.57	9.8 × 10−8	phosphate cytidylyltransferase 2, ethanolamine (PCYT2) is involved in phospholipid biosynthesis40; 1-stearoyl-2-oleoyl-gps (18:0/18:1) is a phospholipid intermediate, suggesting a feedback mechanism
Pathway evidence	CKB → glycine	0.17	8.2 × 10−8	creatine kinase B phosphorylates creatine; creatine is produced in a two-step reaction from glycine and L-arginine catalyzed by arginine:glycine amidinotransferase (AGAT) and guanidinoacetate methyltransferase (GAMT)41
Homologous evidence	ALPI → glucose	−0.03	2.6 × 10−9	intestinal alkaline phosphatase (ALPI) is a wide-specificity phosphoric-monoester hydrolase; yeast orthologs can metabolize glucose42
Homologous evidence	ALPI → histidine	0.04	2.9 × 10−9	alkaline phosphatase intestinal (ALPI) is a wide-specificity phosphoric-monoester hydrolase; histidine was found to inhibit alkaline phosphatase in human adipocytes43
Homologous evidence	ALPI → pyruvate	−0.03	2.0 × 10−7	alkaline phosphatase intestinal (ALPI) is a wide-specificity phosphoric-monoester hydrolase; yeast orthologs can metabolize pyruvate44
Homologous evidence	CKB → valine	−0.17	1.9 × 10−7	in mice, valine inhibits creatine kinase B (CKB) activity45; CKB, which is primarily expressed in the brain, buffers ATP, while valine is a source of ATP in the brain,46 suggesting a feedback mechanism between the two sources of energy

Causal relationships are classified according to whether there is evidence of a canonical reaction, reverse canonical evidence (the metabolite influences the enzyme), pathway evidence (the metabolite and enzyme share a pathway), or homologous reaction evidence (the enzyme has been shown to interact with the metabolite from homologous evidence). Enzymes or metabolites might be measured in multiple studies, leading to multiple estimates (indicated by a 2× or 3× in the “causal relationship” column). We only report the most significant result and the number of Bonferroni-significant (p ≤ 3.4 × 10⁻⁷) combinations. The full results are provided in Table S15.

Canonical reactions

Twelve Bonferroni-significant protein-to-metabolite effects represent canonical reactions where the enzymes are estimated to causally influence the substrate or product of the reaction (Table 1; Figure 4). Three of these direct reactions were missed by our metabolic map due to differences in harmonization of compounds and reactions and retrieved through manual curation: TYMP (thymidine phosphorylase), causal to 2′-deoxyuridine levels ($\hat{\alpha }=-1.2,p=3.1\times {10}^{-18}$); ACY1 (aminoacylase 1), causal to N-acetyl-glutamate levels ($\hat{\alpha }=-0.13,p=3.9\times {10}^{-8}$); and ACY1, causal to glycine levels ($\hat{\alpha }=-0.10,p=3.0\times {10}^{-7}$).

Reverse canonical reactions

MR-link-2 identifies eight metabolites that influence their metabolizing enzyme (Tables 1 and S15; Figure 4). These reactions can be further divided depending on whether the enzyme also causally influences the metabolite (i.e., bidirectional effects; n = 4; arginine → ARG1 [2×], N-acetyl-neuraminate → NPL, and 2′-deoxyuridine → TYMP [Table 1]) or not (i.e., unidirectional effects; n = 4; [2×] N-acetyl-aspartyl-glutamate [NAAG] → FOLH1, aspartate → DARS, and alanine → ACY1). In the latter scenario, all tested enzymes had an IV, and the MR estimate from enzyme to metabolite was non-significant (Table S12). This suggests the existence of feedback mechanisms wherein metabolites sense and regulate enzyme levels. In such examples, NAAG has a negative effect on its catalyzing enzyme, folate hydrolase (FOLH1, also known as glutamate carboxypeptidase 2; measured in two studies: $\hat{\alpha }=-0.10,p=6.4\times {10}^{-40}$ and $\hat{\alpha }=-0.07,p=1.6\times {10}^{-25}$). The enzyme converts NAAG into N-acetylaspartate and glutamate, suggesting a potential feedback mechanism that would prevent the accumulation of glutamate, which has neurotoxic properties at high concentrations^47,48 (Table 1).

Pathway regulation

We identify five enzyme-metabolite pairs that are closely related in the same pathway but not connected through a canonical reaction (Table 1; Figure 4). For instance, TYMP → uridine (uridine was measured twice; two causal estimates: $\hat{\alpha }=-0.57,p=3.2\times {10}^{-11}$ and $\hat{\alpha }=-0.55,p=5.6\times {10}^{-8}$) is two reactions away from 2′-deoxyuridine, a TYMP substrate (uridine → uracil → 2′-deoxyuridine) (KEGG: hsa01232).^34,35 We also found that creatine kinase B (CKB) concentration causally affects glycine levels ($\hat{\alpha }=0.16,p=8.2\times {10}^{-8}$). Creatine is a substrate of CKB and is two reactions away from glycine (glycine → guanidinoacetate → creatine).⁴¹ Interestingly, no MR method identified a causal relationship between CKB and creatine (p > 0.09) (the reverse direction) (Table S12), suggesting that the causal link between glycine and creatine may represent a regulatory mechanism.

Homologous evidence

Sometimes, a homologous enzyme has activity that can be related to the metabolite that MR-link-2 implicates. This can either be through activity in another organism (orthologous enzyme) or through activity in a copy of the gene (paralogous enzyme). We identify that alkaline phosphatase, intestinal (ALPI) has a causal effect on three metabolites: glucose ($\hat{\alpha }=-0.03,p=2.6\times {10}^{-9}$), pyruvate ($\hat{\alpha }=-0.03,p=2.0{\times 10}^{-7}$), and histidine ($\hat{\alpha }=0.04,p=2.9\times {10}^{-9}$). ALPI is a wide-acting phosphatase, and studies of the yeast ortholog have shown that pyruvate and glucose can be reactants of the enzyme^42,43 (Table 1). In line with this, ALPI levels have been found to be risk factors for diabetes.^49,50 We also estimated that CKB causally lowers valine concentration ($\hat{\alpha }:-0.17,p=1.9\times {10}^{-7}$), which can be related to energy pathways, as CKB stores adenosine triphosphate (ATP) in creatine in neuronal cells and muscles, whereas valine can be converted into ATP in neuronal cells.^46,51 In mice, valine has been shown to inhibit CKB, providing orthologous evidence of a causal relationship.⁴⁵

Regulation that has supporting evidence

Fifteen other enzyme-metabolite pairs have additional emerging evidence, either from animal studies or their presence in related pathways, or there is a reasoning as to why the causal relationship may be valid but no definitive human experimental result (Table S15). Three combinations (PLA2G7 onto linoleate [18:2n6], linoleate [18:2n6] onto PLA2G7, and glycine onto ornithine transcarboxylase [OTC]) were recurrently identified in 2 independent studies, representing independent lines of evidence for the causality of these relationships. Furthermore, we identified multiple bidirectional causal relationships between lipoprotein-associated phospholipase A2 (PLA2G7) and the essential fatty acid linoleate, as well as with cholesterol. MR-link-2 identified two linoleate-to-PLA2G7 causal relationships (PLA2G7 measured twice, two causal estimates: $\hat{\alpha }=0.19,p=9.7\times {10}^{-14}$ and $\hat{\alpha }=0.36,p=1.3\times {10}^{-47}$) and two causal relationships from the enzyme to the metabolite (PLA2G7 measured twice, two causal estimates: $\hat{\alpha }=0.05,p=4.7\times {10}^{-8}$ and $\hat{\alpha }=0.24,p=2.2\times {10}^{-61}$). There is emerging evidence linking PLA2G7 and linoleate: PLA2G7 releases linoleate from dorsal root ganglia⁵² and PLA2G7 is protective of the peroxidation of conjugated linoleic acids.⁵³ In line with this, we identify a causal relationship between cholesterol levels and PLA2G7, which has been associated with low-density-lipoprotein particles, the primary cholesterol transporter in blood.⁵⁴

Regulation without supporting evidence

For 62 pairs, our literature search did not find any supporting evidence for causal relationships. Some patterns suggest that they may be false positives, while others could reflect novel biology, which currently has limited evidence (Table S15). One example involves eight links derived from two pQTL studies and two mQTL studies that link bilirubin and biliverdin to butyrylcholinesterase (BCHE) (range of $\hat{\alpha }=-0.1\text{to}-0.15$, range of p = 2 × 10⁻⁷ to 9 × 10⁻¹⁵), which might represent confounding by liver damage, which causes increases in both BCHE and bilirubin.⁵⁵ Alternatively, there is weak evidence that bilirubin and biliverdin are related to choline esterase from in vitro experiments and in model organisms.^56,57

Integrating drugs and their effects into biological understanding

Using the OpenTargets resource, we matched drugs to their target proteins and assessed the downstream consequences of these proteins on metabolites.⁵⁸ Such an analysis provides insights into whether a drug has the intended effect on molecular phenotypes and can help identify if the perturbation of the gene that the drug targets has unintended side effects on metabolite levels.¹¹ Out of the 50 unique proteins that MR-link-2 identifies as having at least one causal effect, 11 are targets of compounds that are either registered as medicines (n = 20), are under investigation, or have failed trials (n = 19) (Table S16). One notable example is that MR-link-2 identifies a causal relationship between N-acetyl putrescine and IL1RAP ($\hat{\alpha }=0.10,p=2.7\times {10}^{-7}$), which is targeted by the drug spesolimab, used to treat psoriasis. Putrescine and other polyamines are present in psoriatic tissues, supporting the biological relevance of the estimated causal relationship.^59,60 It is worth noting that putrescine was not measured in the mQTL studies included in this work, and N-acetyl putrescine is one reaction away from putrescine (KEGG: R01154).

Discussion

In this study, we performed bidirectional MR between 214 metabolites and 437 enzymes to assess whether MR is suitable to identify known biochemical reactions and discover putative enzymatic regulation of metabolites. In cis-MR, all MR methods perform similarly and have high precision (35%–47%) and low recall (3.2%–4.6%), resulting in limited discriminative ability (AUC < 0.554). Extending our analyses to include genome-wide IVs, we performed bidirectional MR between enzymes and metabolites. Discriminative ability remains low (AUC < 0.531), but we found a strong enrichment for direct pathway regulation among Bonferroni-significant meta-analyzed MR results. This low discriminative ability can be attributed to the difficulty in identifying causality but perhaps also to the fact that the ground-truth null sets may contain under-researched true links. This is why we followed up with all seemingly true null links and created a high-confidence subset (based on high node degree). We then reran the analysis using this stricter set of null links and obtained similar conclusions in terms of precision and recall ranks of the different methods.

In line with characterizing the ground-truth null links that may be unknown, we further focused on metabolite-protein links identified by MR-link-2, a method that we developed, which showed high precision in the Bonferroni significant results and an overall lower type 1 error rate and a better recall/precision ratio compared to other methods.¹⁵ We justify this focus as our aim was to understand how well MR methods can recover bona fide metabolite-enzyme relationships rather than method comparison. Importantly, the tested MR methods boast similar precision and recall when comparing their performance at a fixed number of findings. The major differentiation between MR-link-2 and the other methods is that MR-link-2 has better type I error control. Manual review of 106 Bonferroni-significant enzyme-metabolite pairs identified 12 canonical reactions, eight reverse canonical reactions, four combinations with homologous evidence, six with pathway evidence, and 23 with supporting evidence. For the remaining combinations, we could not identify literature support. Finally, when matching 11 enzymes to therapeutics using OpenTargets, we identify the causal relationship between N-acetyl putrescine and IL1RAP that aligns with the therapeutic effect of spesolimab, used to treat psoriasis.

One important benefit of MR analysis is that the exposure and outcome do not need to be measured in the same individual.¹⁰ This allows for a phenotype measured in one cohort to be tested for causality with a phenotype measured in another cohort. This provides the opportunity to study a much broader set of potential causal relations, such as bidirectional regulation between metabolites and enzymes or the metabolic downstream consequences of drugs. Still, we warrant caution in the interpretation of these analyses, as MR results still suffer from large type I error rates, and robust validation of regulatory pathway mechanisms would require targeted experimentation and case-control trials.

Indeed, there are substantial challenges to overcome before MR can be widely applied to study enzyme-metabolite links. An inherent limitation of large-scale metabolomics and proteomics studies, coupled with the genotype information of participants, is that not all molecular species are measured. Even among those that are, not all are instrumentable in MR. A promising avenue is to use protein or metabolite ratios to understand intermediate pathway phenotypes.^61,62 These composite phenotypes are still understudied but could provide indications of unmeasured phenotypes or mechanisms in known pathways, even though analyzing ratio phenotypes comes with its own challenges.⁶³ A second major challenge is tissue specificity. Measurements underlying this study are derived from blood. Yet, most human metabolism is tissue specific, with certain tissues being more important for certain pathways.⁶⁴ It is still unclear how much our results will translate to other tissues, although the limited specificity of our analysis already provides evidence that we will not be able to recover most elements of pathway regulation in ground-truth datasets.

There are still unknowns about how to conduct an appropriate MR analysis; for instance, it is unclear how much MR performance improvement can be expected from further increased sample size to derive molecular QTLs. Power analyses for MR studies usually do not consider locus selection mechanism and are hard to extend to pleiotropy-robust methods,^65,66 making it difficult to distinguish true null links from links where MR is simply underpowered.

As the study sample size increases, it is likely that more causal links from other tissues can be proxied by blood, albeit with weaker effects. Here, it is important to consider that more subtle QTLs need to be analyzed with pleiotropy-robust methods, as the mixture of different tissue effects can violate the exclusion restriction in MR. This study also highlights that the selection of suitable genetic regions for MR can be improved, as illustrated by the cis-MR analysis, which has the highest precision and a high recall, with the caveat that not all phenotype combinations can be tested. When extending to the cis + trans analysis, the recall (in terms of unique true positive findings) does not necessarily increase, even though some loci do provide correct inference. This indicates that putting biological priors on certain genomic regions may be the correct course of action to reduce the type 1 error rate.

Taken together, the results of our study indicate that there is strong potential for the analysis of molecular pathways using observational causal inference techniques, which is illustrated by our biologically informed findings. Such approaches are not only cost-effective, but their top-ranking metabolite-protein links are many-fold enriched for true associations, warranting experimental follow-up.

Data and code availability

• •This study did not analyze newly generated data.
• •The individual MR results for each associated region in the dataset have been published on Zenodo under the Zenodo doi: https://doi.org/10.5281/zenodo.17965632. This table contains all cis + trans-regional MR results per method (MR method); associated regions; exposure (exposure accession and protein name) and outcome (outcome accession and metabolite name) combination, with their respective MR estimate (alpha, SE, and p(alpha)) and meta-data on the origin of the exposure and outcome (protein dataset, UniProt ID, metabolite dataset, metabolite HMDB ID, and Ensembl ID); and finally, if the combination is considered causal based on our ground-truth dataset (causality).
• •Other data generated during this study are available in Tables S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15, and S16.

Acknowledgments

We would like to thank all study participants for their altruistic donations of biological materials. Z.K. was funded by the Swiss National Science Foundation (SNSF; 315230-219587).

Author contributions

A.v.d.G. and Z.K. conceptualized the study. A.v.d.G. and S.R. performed data curation. A.v.d.G. and Z.K. performed formal analysis, including statistical and computational techniques to synthesize and analyze data. Z.K. acquired funding. A.v.d.G., S.R., C.A., and Z.K. developed and designed the methodology. Z.K. provided computational resources. A.v.d.G. and S.R. performed software development. A.v.d.G. created visualizations and wrote the original manuscript draft. A.v.d.G., C.A., S.R., and Z.K. reviewed and edited the manuscript.

Declaration of interests

The authors declare no competing interests.

Published: February 5, 2026

Supplemental information

Table S1. Reference of reactions used

All retrieved reactions from both KEGG+expasy and Metacyc databases (database), with their respective reaction identifier (reaction_id), enzyme commission enzyme number (enzymes) and their enzyme definition (definition).

Table S2. Reference of catalyzing proteins

All combined catalyzing proteins (Gene name, Ensembl ID, UNIPROT ID) measured in the two protein quantitative trait locus studies (study, accession) harmonized in this study, as well as where the enzyme comes from.

Table S3. Combined reaction table

All enzyme (Exposure accession, Exposure name) metabolite (Outcome accession, Outcome name) combinations that we can consider ground truth positives in this study after harmonizing the ground truth datasets with the protein quantitative trait locus (QTL) and metabolite QTL studies, considering the enzyme (reaction reference, dataset) and if the metabolite is a product / substrate. The KEGG + Expasy and MetaCyc network degree is also shown per metabolite and enzyme.

Table S4. Cis Mendelian randomization estimates

Cis-Mendelian randomization estimates per method (MR method), exposure (exposure accession, exposure name) and outcome (outcome accession) combination, with their respective MR estimate (alpha, se, p(alpha)), and meta-data on the origin of the exposure and outcome (pQTL dataset, uniprot ID, Metabolite dataset, metabolite HMDB ID), and finally if the combination is considered causal based on our ground truth dataset (causality).

Table S5. AUC of cis Mendelian randomization results

Area under the receiver operator characteristic curve (AUC) for cis-Mendelian randomization estimates per MR-Method, combined with the number of ground truth positives in the datasets and the number of ground truth null links (negatives).

Table S6. BRENDA annotation analysis

Brenda annotations (BRENDA Annotations) of Bonferroni significant (P < 1.93 ⋅ 10) (MR method) cis MR results (exposure, outcome, alpha, p(alpha)), their annotation in the ground truth dataset (causality), their metabolite and protein names (Metabolite name and Protein name).

Table S7. Metabolite network node degree and effect size correlation

The Spearman correlation values and P value between the MR estimate and the amount of reactions a molecule is involved in (Network degree type). Correlations are calculated for the absolute effect size of the MR method and the P value of the MR estimate. Both for the cis only and cis + trans analysis.

Table S8. GTEx Mendelian randomization results

cis-Mendelian randomization estimates per method (MR method) on the positives in our ground truth using GTEx as the exposures, exposure (exposure ENSG ID, Gene name) and outcome (outcome accession) combination, with their respective MR estimate (alpha, se, p(alpha)), and meta-data on the origin of the outcome (Metabolite dataset, metabolite HMDB ID), and finally if the combination is considered causal based on our ground truth dataset (causality).

Table S9. Metabolite class enrichment

Fisher exact test enrichment results of HMDB metabolite super class, class and subclass. Enrichment is calculated by comparing the class membership number in the ground truth dataset with the class membership amount in the Bonferroni significant cis + trans results, relative to all other classes in the ground truth and other classes significant and in the ground truth

Table S10. PheWAS results

For all exposures (accession) the associated regions and the number of overlapping GWAS hits at P < 5 10−8 based on the case control GWAS’es from Sakaue et al.

Table S11. Coloc precision and recall

Precision and recall of the cis + trans comparisons per MR method at the cis + trans Bonferroni significance threshold (P < 1.1 10−7), when filtering for coloc PP4 > 0.9 (Filtered) in the locus or not (Base), separated by if the locus is found in cis, in trans, or combined.

Table S12. Mendelian randomization meta-analysis estimates

All meta-analyzed combined Mendelian randomization results per method (MR method), associated region, exposure (exposure accession, Protein name) and outcome (outcome accession, Metabolite name) combination, with the number of meta-analyzed regions (n associated regions) and their respective meta-analyzed MR estimate (weighted alpha, weighted se, weighted p(alpha)), and meta-data on the origin of the exposure and outcome (protein dataset, UniProt ID, Metabolite dataset, metabolite HMDB ID, Ensembl ID), and if the combination is considered causal based on our ground truth dataset (causality).

Table S13. AUC of Mendelian randomization meta-analysis results

Area under the receiver operator characteristic curve (AUC) for meta-analyzed Mendelian randomization estimates (MR) per MR-Method, combined with the number of ground truth positives in the datasets and the number of ground truth null links.

Table S14. Reaction enrichment results

Enrichment results when comparing Bonferroni significant meta-analyzed results to all other results, with per method (MR method) the baseline probability of identifying a metabolic reactions (baseline probability), at Bonferroni significance (probability in Bonferroni significant), the resulting odds ratio and the fisher exact test P value.

Table S15. Biological interpretation of MR-link-2 results

Bonferroni significant MR-link-2 results with all supporting evidence. Causal relationships are classified according to whether there is evidence of a canonical reaction, reverse canonical evidence (the metabolite influences the enzyme), pathway evidence (the metabolite and enzyme share a pathway), homologous reaction evidence (the enzyme has been shown to interact with the metabolite from homologous evidence) or Supporting evidence, when there is some evidence of a relationship existing, but it is not as strong as the other types. Enzymes or metabolites might be measured in multiple studies, leading to multiple estimates (indicated by a ‘2x’ or ‘3x’ in the “Causal relationship” column).

Table S16. Drugs and targets

Bonferroni significant MR-link-2 estimates that overlap with the genes listed as drug targets in the OpenTargets database. For each overlapping gene we list their relationship with the metabolite, the drug name, the disease indications and the highest trial the drug has passed.