Metformin to Treat Cancer

Metformin is a drug that has been used since the 1950s as first-line treatment for type 2 diabetes and that is currently being studied as a cancer treatment. The quest into metformin as a potential treatment for cancer began in 2005 with an observational case–control study from Scotland that reported a reduction in the incidence of cancer with metformin use (odds ratio 0.77; 95% confidence interval [CI]: 0.64, 0.92).1 This “Research Pointers” publication, which advanced the hypothesis that metformin could lower the incidence of cancer in patients with diabetes, was particularly notable because of the well-documented elevated risk of several cancers in patients with diabetes.2 The study generated considerable excitement in the oncology community and spawned a keen interest in investigating metformin as an agent for cancer prevention and treatment in patients with or without diabetes.3,4 As a result, in the 10-year period following the initial hypothesis-generating publication, several observational studies have been conducted in various healthcare databases to verify this hypothesis. This efficient and rapid pharmacoepidemiologic approach of exploiting existing healthcare databases to assess the real-world effects of medications, including identifying new indications for older drugs, is now popular and widespread. Most of these observational studies reported beneficial effects of metformin, showing reductions in cancer incidence and improved prognosis associated with metformin use, thus “confirming” the 2005 study results. In addition, several meta-analyses of these observational studies have contributed to reinforce these findings.5–9 These confirmatory observational studies were thus rapidly followed by the launch of randomized controlled trials to assess the efficacy of metformin as a treatment for several cancers. In this commentary, I submit that this quest into metformin as a treatment for cancer is a likely instance of serious missteps in translational research, with currently running randomized trials motivated by observational studies that have sufficiently serious biases to question their conclusions. This metformin–cancer example should not be construed as a questioning of the importance of otherwise rigorous observational studies in the armamentarium of clinical research. With this example, I also raise the issue of challenges in interpreting data from multiple randomized trials prompted by flawed observational data. THE FIRST RANDOMIZED TRIAL The first randomized controlled trial of the efficacy of the antidiabetic drug metformin in the treatment of cancer was recently published, focusing on pancreatic cancer.10 This double-blind, randomized, placebo-controlled trial evaluated the addition of metformin to standard systemic therapy with gemcitabine and erlotinib in treating patients with advanced pancreatic cancer. Conducted over the period 2010–2014, it randomly assigned 121 patients to placebo (n = 61) or metformin (n = 60), followed up for 6 months. The sample size was determined to provide sufficient power to detect a reduction of 25% in the 6-month mortality, from an expected 50% to 25%, with the addition of metformin compared with placebo. Mortality at 6 months turned out to be 36% in the placebo group and 43% in the metformin group (hazard ratio 1.06; 95% CI: 0.72, 1.55). In response to this negative finding, the authors advocated that future research should use more potent biguanides and focus on patients with specific tumor markers. The accompanying editorial stated that “despite this disappointing outcome, attempts to repurpose metformin for treating cancer should not be abandoned. More than 100 studies assessing metformin in various stages and types of cancer are currently underway.”11 This enigmatic trial raises several questions: How solid is the evidence for studying metformin, a first-line drug of choice for the management of type 2 diabetes mellitus, as a treatment for cancer? Why are there over 100 oncology trials assessing the effectiveness of this drug currently underway? Why dare venture with confidence into such an ambitious goal of reducing all-cause mortality by one half with metformin? OBSERVATIONAL STUDIES OF CANCER INCIDENCE The evidence for studying metformin as a treatment for cancer is based primarily on the numerous observational studies conducted on the metformin–cancer issue, which generally reported similar beneficial results on reducing cancer incidence with metformin, thus “confirming” the findings of the 2005 hypothesis-generating study. The magnitude of these beneficial effects of metformin on cancer was in most studies quite spectacular, if not simply incredible. For example, one study reported that the use of metformin was associated with remarkably lower incidence of many cancers, such as rate ratios of 0.12 for any cancer, 0.36 for colorectal cancer, 0.06 for liver cancer, and 0.15 for pancreatic cancer.12 Another study reported a rate ratio of 0.55 for lung cancer incidence with metformin use within a short follow-up span, an impact not seen even with the most effective smoking cessation programs.13 These studies were subsequently shown to be affected by immortal time bias.14 This bias, which results from misclassifying the time between cohort entry and the first metformin prescription during follow-up as exposed when in fact it is unexposed, has been described in several contexts in studies of drug effects.15–20 It systematically results in risk reductions that can be greatly exaggerated in favor of the exposure. Several of the more recent observational studies, that used the correct study designs and statistical methods of data analysis to avoid immortal time bias, did not find an association between metformin and the incidence of several cancers.21–30 The most recent study, published in this issue of Epidemiology, also finds no evidence for a reduced risk of breast cancer with metformin use.31 STUDIES OF CANCER PROGNOSIS A number of observational studies also investigated the effects of metformin not as prevention, but as a treatment for cancer, by studying patients diagnosed with cancer, including prostate, colorectal, ovarian, and human epidermal growth factor receptor 2 positive (HER2+) breast cancer.32–35 These studies, comparing patients with cancer using metformin with those not using it, all reported major risk reductions on all-cause mortality, cancer-specific mortality, and cancer progression, with the use of metformin. Here again, the spectacular magnitude in the reduction in cancer outcomes with metformin use, with hazard ratios ranging from 0.38 to 0.66, were all shown to be the result of immortal time bias introduced in these studies.14 For example, the cohort study of 1983 consecutive patients diagnosed with HER2+ breast cancer in one center over the period 1998–2010 followed women for all-cause and cancer-specific mortality.32 The use of metformin was measured at any time during the median 4 years of follow-up. Among the diabetics, the hazard ratio of all-cause death associated with metformin use compared with nonuse was 0.52 (95% CI: 0.28, 0.97), whereas for breast cancer-specific mortality it was 0.47 (95% CI: 0.24, 0.90). These considerable mortality reductions are the direct result of immortal time bias caused by misclassifying the immortal time between breast cancer diagnosis (cohort entry) and first prescription of metformin in follow-up as exposed to metformin, when in fact it is still unexposed.14 This misclassification provides an artificial survival advantage to the users of metformin by incorrectly adding the immortal time to their survival (Figure).FIGURE: Illustration of immortal time bias in study of breast cancer patients entering cohort at the time of diagnosis: The top patient initiated metformin after diagnosis yet is classified as a metformin user for the entire follow-up time. The time between breast cancer diagnosis and the first metformin prescription is thus immortal (thick grey line), because the subject must survive to receive this prescription, and is also misclassified as exposed to metformin when in fact it is unexposed, leading to immortal time bias. Figure is available in color online.Here again, several of the more recent observational studies that used correct approaches to avoid immortal time bias did not find an association between metformin and the prognosis of several cancers.36–44 ONGOING RANDOMIZED TRIALS These generally highly biased observational studies, caused by the incorrect way in defining metformin exposure, along with corresponding meta-analyses that combined their results, provided the impetus to initiate the many ongoing randomized controlled trials of metformin for cancer treatment. It is estimated that there are currently over 100 such trials, while a search of Clinicaltrials.gov with the terms “metformin” and “cancer” report 300 studies registered in its website as of 26 February 2017. The largest of these randomized controlled trials is the ongoing MA.32 study conducted by the National Cancer Institute of Canada Clinical Trials Group, a multicenter phase III placebo-controlled trial involving over 3,600 women with early breast cancer (https://clinicaltrials.gov/ct2/show/NCT01101438). The trial aims to compare metformin to placebo as adjuvant therapy on the primary endpoint of invasive disease-free survival over a 5-year follow-up. This trial started recruitment in July 2010, with an expected study completion date of December 2017. The trial is powered to detect a hazard ratio of 0.76 for the rate of recurrence with metformin. While one can hope for positive results, the evidence from observational studies on which this trial is founded is weak and mainly incorrect. THE WAY FORWARD Thus, to date, over 100 ongoing trials of metformin as a treatment for various cancers have been initiated on the basis of biased observational studies. What should be the way forward? First, the major bias in the observational studies, namely immortal time bias, could have been avoided with the appropriate study design and statistical analysis. The authors of these observational studies still have the data and could be compelled to reanalyze their studies with the proper approaches and report these results accordingly. As well, new observational studies using more rigorous designs and analyses that avoid these biases could be undertaken before more randomized trials are initiated. For example, recent observational studies that used correct statistical methods to avoid this bias did not find a benefit of metformin in women with breast cancer on all-cause and cancer-specific mortality, as well as in lung, colorectal and prostate cancers.36–44 Certainly, one cannot overlook other methodologic challenges beyond immortal time bias that afflict observational studies in this field, including confounding resulting from incomplete information in the databases and protopathic bias from changes in medications as the patient approaches death. Some of these may also explain in part a beneficial effect of metformin on cancer-specific mortality in women with breast cancer, even after the proper time-dependent analyses.45 Perhaps a consensus statement would be useful at some point on the most efficient way to conduct observational studies in this particular population of patients with cancer. Second, the results of the over 100 trials of metformin already underway will certainly be published in the medical literature. Under the pessimistic assumption that metformin does not improve survival in patients with cancer, one can still expect that chance will yield around five trials that will “show” a mortality advantage with metformin, at the (one sided) 5% statistical level of significance. Most likely, these few significant studies could be hailed as major findings for the (yet unknown) cancer under study. Thus, the oncology and statistical communities should be prepared to properly appraise the value of such findings not as single positive studies, but rather as a few among over 100 such trials. Finally, an important issue is whether this metformin paradigm precludes the use of observational studies to identify and investigate new indications for old drugs. I certainly do not think so. On the contrary, observational studies are an extremely efficient tool to study novel pleiotropic drug effects in large real-world populations, quickly and inexpensively as they generally exploit already available data from claims databases or electronic medical records. The metformin–cancer example, however, warns against the uncritical acceptance of all observational studies at face value, simply because they are published in the peer-reviewed literature. This would avoid repeating similar futile quests based on erroneous data in the treatments of menopause and chronic obstructive pulmonary disease.46–48 The biases identified in the observational studies of metformin and cancer are well known in epidemiology. In fact, the most recent observational studies that used proper epidemiologic and statistical time-dependent methods to study the question, thus avoiding these vexing biases, did not find an association between metformin and cancer incidence or mortality. More rigorously conducted and reviewed observational studies will provide a more accurate path to follow in the quest of novel usages for existing drugs. ABOUT THE AUTHOR SAMY SUISSA is a James McGill Professor of epidemiology and biostatistics at McGill University in Montreal, specializing in pharmacoepidemiology. He has conducted studies on the effects of medications used for various chronic diseases and authored over 440 peer-reviewed papers. He heads the Canadian Network for Observational Drug Effect Studies (CNODES), serves on the editorial boards of various scientific journals, including Pharmacoepidemiology and Drug Safety, and is Section Head editor for F-1000 Medicine. He is a fellow of the Canadian Academy of Health Science, has received the Distinguished Investigator award from the Canadian Institute of Health Research (CIHR), and the FC Donders Professorship Award from Universiteit Utrecht, The Netherlands.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.