Evidence-based prioritisation and enrichment of genes interacting with metformin in type 2 diabetes

Metformin has been used for 60 years by more than 150 million people worldwide. It is the first-line monotherapy prescribed at diagnosis in people with type 2 diabetes [1], and also slows progression to type 2 diabetes in people with elevated but non-diabetic glucose levels who are unable or unwilling to adhere to lifestyle modification [2‐4].

Despite the popularity of metformin in diabetes treatment, its mechanisms of action are poorly understood; suppression of endogenous glucose production via activation of AMP-kinase (AMPK) has been hypothesised [5]. However, a preserved glucose-lowering effect has been reported in AMPK knockout mice [6]. Alternative, non-AMPK-dependent, mechanisms include inhibition of mitochondrial glycerophosphate dehydrogenase activity [7] and adenylate cyclase-mediated inhibition of the gluconeogenic pathway in favour of glycolysis [8]. In a recent study performed on Caenorhabditis elegans and extended to human cell lines, Wu et al identified two new targets of metformin action: the nuclear pore complex and the gene encoding acyl-CoA dehydrogenase 10 [9].

Diabetes treatment guidelines adopt a one-size-fits-all approach, and do not take into account interindividual variation in response. Yet there is considerable between-patient variation in treatment effects, with some responding poorly or not at all and others being highly sensitive to the drug or experiencing extreme adverse drug reactions [10]. Up to 30% of individuals treated with metformin develop nausea, bloating, abdominal pain and/or diarrhoea, and 5–10% are unable to continue with metformin treatment [11]. Heritability studies indicate that genetic variation underlies around 34% of the variability in metformin response [12].

Previous candidate gene-based pharmacogenetic studies of metformin have largely focused on loci encoding transporter proteins; little emphasis has been placed on genes in the pharmacodynamics (PD) domain, and much of the published data are inconclusive and sometimes controverted [10]. Hypothesis-free genome-wide association studies (GWASs) on metformin have identified a genome-wide significant variant, rs11212617, near the ATM gene for metformin-induced glycaemic response [13]. Given that this SNP lies in a large block of genes that are in linkage disequilibrium, the authors performed cellular work and suggested ATM to be the causal gene.

AMPK, the energy sensor, is the downstream target of metformin and is believed to be involved in the PD of metformin. Selective inhibition of ataxia telangiectasia mutated (ATM) protein by KU-55933 resulted in a marked reduction in metformin-induced AMPK activation, suggesting involvement of ATM in AMPK activation. However, cellular studies showed marked inhibition of organic cation transporter (OCT)1, an important mediator of metformin uptake by the liver, by KU-55933, suggesting that the observed attenuated AMPK phosphorylation could also be due to inhibition of OCT1 [14]. A recent GWAS study from the MetGen consortium reported an association between an intronic SLC2A2 variant, rs8192675, and the glycaemic response to metformin [15].

Owing to the vast literature on gene–metformin interactions, obtaining an unbiased overview of the evidence is extremely difficult. While meta-analysis delivers trustworthy findings if well conducted, heterogeneity in study designs, analytic strategies, population characteristics and data selection biases present challenges to such analyses [16]. Thus, to facilitate this process, automated approaches to integrate evidence from multiple sources, cataloguing the levels of evidence, validating in a real-world dataset, and using this to prioritise genes for follow-up are increasingly favoured [17, 18].

Here, we established a semi-automated text-mining pipeline to prioritise biological candidate genes that show evidence of interaction with metformin based on strength of evidence from published studies. We then evaluated the prioritised gene sets by examining their enrichment using a well-powered external dataset.

Patients vary greatly in their glycaemic response, optimal dosage and adverse drug reactions following metformin therapy [41]. Genetics accounts for about 34% of this variability [12]. Hence, there is a case for more personalised therapy in diseases such as type 2 diabetes. Understanding how genetic variation impacts the effects of glucose-lowering drugs or helps to refine the characterisation of type 2 diabetes might improve treatment effectiveness and decrease adverse drug reactions, morbidity, mortality and cost of treatment.

Although there are publications in relation to the PK, PD and clinical outcomes of metformin, there is no database that concisely summarises the mechanisms describing gene–drug interactions. In most cases, specific evidence of interactions is buried deep within the literature, making it extremely difficult to comprehend the overall weight of evidence for given interactions. This problem is not specific to gene–metformin interactions, but is one that is common to the gene–drug and gene–environment interaction literature per se. In this paper, however, we focus on interactions between metformin and genes that impact clinical outcomes, PK and/or PD of the drug using a comprehensive text-mining strategy.

Our analyses identified seven genes ranked as ‘top priority’ to predict metformin-related clinical outcomes. These genes constituted three solute carrier family genes (SLC22A1, SLC22A2 and SLC47A1) that are related to the PK of metformin, and four PD-related genes (ATM, STK11, PRKAA2 and SHBG). Fifteen genes were found to affect the PK/PD of metformin without being consistently related to clinical outcomes. A third gene set in which expression or activation is affected by the presence of metformin has also been identified from the text-mining. A GSEA using GWAS data from GoDARTS showed significant enrichment of the third category for glycaemic response.

Genes that showed consistent changes in cellular and molecular functions in the presence of metformin may have potential clinical relevance in the search for biomarkers that predict the therapeutic outcome of metformin. This includes SLC2A4 and G6PC. SLC2A4 encodes GLUT4, which plays a crucial role in regulating blood glucose homeostasis by facilitating insulin-stimulated glucose transport into skeletal muscles [42]. Metformin is shown to modulate translocation of GLUT4 in skeletal muscle and adipocytes [43]. G6PC gene encodes G6PC, a rate-limiting enzyme in hepatic glucose production. It is involved in glucose production via the gluconeogenesis and glycogenolysis pathways. Cellular studies have shown metformin to suppress glucose phosphatase enzyme expression independent of AMPK [44]. While the top genes according to the GSEA were SLC2A4 and G6PC, there is a possibility that the true signal could be from neighbouring genes. ESM Fig. 4 shows regional association plots around SLC2A4 and G6PC.

The stringent significance threshold for a GWAS could overlook moderate-association signals that may have detrimental collective effects in certain pathways. Therefore, integrating GSEA guided by carefully curated evidence from the literature, as we did here, is likely to identify signals that are overlooked in a GWAS. In this study, we tested enrichment of text-mining-based prioritised gene sets on a GWAS of metformin response. Gene sets that were given high priority (those related to clinical outcome, PK or PD) showed no significant enrichment for multiple modest associations. This is probably due to the fact that our scoring system is subject to the publication bias that is well known to affect candidate gene association studies. For example, there are many small studies reporting the association between variants in metformin transporter genes and glycaemic response, resulting in a higher priority of PK genes in our gene set ranking. However, a meta-analysis of most published studies, in up to 8000 metformin-treated individuals with type 2 diabetes from the MetGen consortium, showed the putative pharmacogenetics variants in five transporter genes had no significant impact on the glycaemic response to metformin [45], which is in line with our GSEA. The ranking system prioritises published evidence dominated by PK studies. Yet PK variants do not seem to alter metformin response [45]. Therefore, this is largely due to candidate gene-based approaches. The molecular and cellular-based approaches looked at mostly PD studies and hence cast the net wider, so being less focused on the transporters.

The enrichment of association signals within the low-priority gene set highlights the need to follow up association signals for loci with multiple modest effects for glycaemic response that have been previously overlooked, such as SLC2A4 and G6PC, which rank high in the GWAS, with p values of 3.24 × 10⁻⁰⁴ and 4.77 × 10⁻⁰⁴, respectively.

Although the pipeline described here has generated novel evidence of gene–metformin interactions, it by no means renders a complete overview of all relevant literature. For example, only 60% of the index articles were accessible as full text, and conversion of the native PDF file into text format might have caused relevant data to be lost, as some articles used formats that are incompatible with the file templates used in our pipeline. Furthermore, data contained in tables and figures in the index papers were extracted in a semi-structured format, making data retrieved from tables and figures challenging to analyse and interpret. Thus, for papers in which important results were included in tables and figures, but not articulated in the title, abstract or main text of the article, evidence of interactions may have been overlooked.

Some genes are reported in papers in a non-standard form that may not be captured by the dictionary-based named entity recognition used in this study, and in some articles, non-standard abbreviations for metformin, instead of its original generic or brand name (e.g. the term ‘MET’), are occasionally used. Such abbreviations are not found in any standard drug abbreviation protocol. Sentences with such abbreviations are likely to be incompletely characterised using the text retrieval strategy adopted here. Other barriers to data assimilation include the extent to which anaphora (‘use of grammatical substitute [as pronoun or a pro-verb] to refer to the denotation of a preceding word or group of words’ [46]) and epiphora (‘the repetition of a word or words at the end of two or more successive verses, clauses, or sentences’ [46]) were resolved, which may also have resulted in loss of information; this might have occurred, for example, when a gene and a drug were found in separate sentences (i.e. not described in the same sentence).

Larger studies with regular iterations that use extensive literature coverage, SNP-level annotation and more intensively automated machine learning approaches are likely to extend the observations reported here. GSEA using a larger GWAS dataset for metformin response is likely to increase the power and produce convincing results. Replication using other GWAS datasets including SNP–SNP interaction analyses is also likely to extend the number of validated interactions, as some of those prioritised using our algorithm may have ethnically specific effects that are undetectable in the ethnically homogeneous GoDARTS cohort.

We have developed a semi-automated text-mining and evidence-scoring algorithm that could help to organise and extract useful information from the literature on gene–drug and gene–environment interactions. According to our analysis, genes that encode transporter proteins such as OCTs and MATEs yield the strongest evidence of modifying the therapeutic outcomes of metformin. In addition, genes in the LKB1–AMPK pathway are also found to be related to the therapeutic outcomes of the glucose-lowering drug. However, we did not detect enrichment for the highly prioritised gene sets using GWAS data from the GoDARTS cohort; instead it was the gene set that were ranked with lower evidence scores that showed statistically significant enrichment in the GoDARTS.

Given that the genomic architecture of drug response is complex and the mechanism of metformin action is still not clearly known, candidate gene studies investigating drug response to metformin have had limited success. In an alternative approach, here we have identified novel genes potentially associated with metformin action. Using a text-mining approach of the published literature, we have identified a gene set derived from cellular and functional studies associated with metformin response. The association of genetic variation in these genes, including SLC2A4 and G6PC, needs further replication and follow-up.