Doctors have observed for centuries that patients on multiple medications can have varied and unexpected adverse events. Some of these can be linked to the recognized adverse effects of an individual drug. Others can be traced to pairs of drugs that share a common pathway of metabolism (e.g., both are metabolized by a particular isoform of cytochrome P450), thereby rendering the “effective dose” very different from the intended one. Still others may result from two drugs interacting with redundant biological pathways, such that the effects are not additive but nonlinear or synergistic. This may happen, for example, when there are two parallel pathways that perform the same function. If one is blocked, the other can still perform the function and maintain homeostasis. However, if both are blocked, there is a sudden and dramatic decrease in the ability of the cell to perform the function. These kinds of interactions have been demonstrated clearly through synthetic genetic knockout experiments.8
Our results are consistent with the hypothesis that pravastatin and paroxetine may interact in this manner with biological pathways that are critical for glucose metabolism. Our results highlight not only the possibility of such interactions but the potential clinical significance of these interactions.
Paroxetine and pravastatin show evidence of a synergistic interaction
An analysis of the prescriptions at the three sites from which we drew data revealed that ~6% of the patients on pravastatin were also on paroxetine and ~4% of patients on paroxetine were also on pravastatin. Given that there were ~18 million prescriptions for pravastatin and ~15 million for paroxetine in 2009,3
we estimate that between 550,000 and 715,000 individuals received prescriptions for combined therapy with these two drugs in the United States that year (see Supplementary Table S2
online). In this retrospective observational study, a majority of patients showed increases in random blood glucose levels when paroxetine was added to a drug regimen that included pravastatin (and vice versa). We stress that the clinical significance of our finding is not clear. However, it is important to evaluate whether this effect extends to fasting blood glucose levels and whether it can potentially push glucose-intolerant patients into frank T2DM. Patients on pravastatin are already at increased risk for T2DM (as part of the metabolic syndrome9
), and therefore changes in blood glucose could complicate both the management of these patients and decisions regarding the diagnosis of T2DM.10
The mechanism of the interaction between pravastatin and paroxetine is not clear. The literature is contentious regarding the effect of pravastatin on glucose and on diabetes,11,13-19
but hyperglycemia is not a well-recognized clinical side effect. Paroxetine has played a role in two reports of hyperglycemia involving multiple drug interactions.20,21
Paroxetine is known to be associated with diabetes,22
which could explain an associated increase in blood glucose. We controlled for this in several ways. First, we found no significant change in blood glucose in patients receiving paroxetine alone; second, we included the sequence in which the drugs were taken as a covariate and did not find a significant association. There are molecular connections linking pravastatin and paroxetine to diabetes-related pathways. The pleiotropic effects of pravastatin are mediated by inhibition of Rac1,23
which also plays a crucial role in the translocation of a glucose transporter (GLUT4) to the plasma membrane.24
Paroxetine targets the serotonin transporter protein SLC6A4, and this interaction leads to inhibition of serotonin reuptake.25-27
In β-islet cells, serotonin is involved in collocating insulin to the plasma membrane before it is secreted. Interestingly, both Rac1 (ref. 28
) and SLC6A4 (ref. 29
) interact with syntaxin 1A, which is associated with impaired glucose regulation30
and promotes insulin secretion.31
These proteins provide a clear set of pathways that require careful study in order to elucidate a possible mechanism for synergy between these two drugs.
No evidence of class effects between SSRIs and statins
Our data do not suggest that the interaction of pravastatin and paroxetine is a general class effect between all statins and all SSRIs. Although we do see a small effect on glucose levels when pravastatin is taken in combination with any of the SSRIs, no significant interaction effect was observed in our statistical model. Of note, the FDA-AERS data mining procedure did not identify other specific statin–SSRI pairs as being disruptive of glucose hemostasis. In fact, only atorvastatin plus fluoxetine and rosuvastatin plus sertraline demonstrated a significant increase in glucose values relative to baseline in the paired t-test analysis. However, after correcting for covariates in the ANCOVA, the effects on glucose levels were no longer significant with respect to either of these drug combinations.
Clinical implications for diabetic patients receiving combination therapy
We designed the initial clinical analysis to exclude diabetic patients to mitigate confounding factors. For example, a diabetic patient may alter his or her medications based on self-assessment of glucose levels and thereby negate the putative effect of the paroxetine-pravastatin interaction. On the other hand, diabetics who observe large changes in their glucose levels may be more likely to come in for treatment at the hospital and have their blood glucose measured, thereby introducing selection bias into our patient cohorts. Nevertheless, an analysis of the effect of combined therapy with paroxetine and pravastatin in patients with diabetes revealed a large increase in random blood glucose levels after the start of therapy (48 mg/dl, 2.7 mmol/l, ). This finding indicates that these drugs may complicate the treatment and management of diabetes and could lead to a high incidence of adverse events. Further study is required to establish the clinical significance of these interactions and to indicate whether physicians should consider alternatives to combination treatment involving paroxetine and pravastatin in patients with diabetes.
Limitations of retrospective observation study
Our observations have several limitations revolving around potential covariates. A traditional covariate analysis was not possible in this study. Our retrospective observational data do not allow for controlling many potential covariates that a traditional prospective study offers. It is therefore impractical to enumerate, much less measure, all the covariates from the EMR data. Previous work in surveillance studies has identified these limitations and highlighted the benefit and utility of simple associations to identify avenues of further study.32
Nonetheless, we identified potentially significant variables and examined their individual effects.
First, random blood glucose measurements are known to be highly variable within individuals from hour to hour. However, we found no significant differences in the time of day at which glucose levels were measured in patients, both before and after combined treatment with pravastatin and paroxetine. Second, we assumed that patients who were prescribed pravastatin and/or paroxetine were taking the drugs; however, our databases do not have verified data regarding patient compliance (however, this limitation would reduce, not increase, the signal). Third, many of the patients were receiving not only pravastatin and paroxetine but also many other medications. However, we found no significant correlation between the presence of other concurrent medications and the change in blood glucose levels. Finally, we do not know why glucose measurements were ordered for the patients in our cohorts; however, we controlled for this uncertainty by also assessing glucose level changes in patients on either drug alone and in a set of control analyses of patients receiving similar treatment regimens. The vast majority of glucose measurements were random and not carried out with the patient in a fasting condition.
We are not aware of any literature, guidelines, or conventional wisdom suggesting that pravastatin and paroxetine should be used preferentially in pre-diabetes patients, and our evaluation of other combinations of statins and SSRIs suggests that there are no differences in the baseline characteristics of patients on these two drugs as compared with other statin/SSRI combinations. If there had been a preexisting clinical concern about a prediabetic state in patients who were prescribed combination treatment with these two drugs, we should have seen more frequent measurements of fasting glucose in these patients; however, we observed that the blood glucose measurements were almost exclusively random, suggesting that these measurements were taken as part of a routine screening panel and not out of specific concern about glucose levels. We cannot rule out other potential correlated variables that we did not include in the statistical model, but the adverse-event data mining, the triple-replicated observation in three hospitals, and the initial analysis of the molecular links between the two drugs point to a causative role for the combination treatment. Many of these limitations could be addressed in a prospective clinical trial in which fasting blood glucose, insulin secretion, and insulin action are evaluated before and after the start of pravastatin and paroxetine combination therapy. As a first step, we are following this up with in vivo studies in mice. shows preliminary data supporting our observations.
Figure 4 Preliminary analysis of mean and standard error of fasting blood glucose concentrations in five groups of mice (n = 10/group). Asterisks indicate significance in a multivariate linear model with covariates (***P < 0.001, **P < 0.01, * (more ...)
Data mining and hypothesis generation
The use of data mining to generate the hypothesis motivating our investigation provides an excellent example of the power of sharing large data sets. The noise in these large data sets is justifiably used as a reason to be suspicious of results obtained from them. However, the results reported here suggest that these data sets allow detection of potentially important clinical signals. The increasing availability of clinical data repositories, suitably consented and de-identified to support translational research, is a key element in the ability to generate hypotheses and test them via discovery-oriented data mining.33
Indeed, the mature EMR systems at three institutions enabled us to discover this putative drug interaction very quickly.
In conclusion, we present a novel method for discovering putative drug interactions from the FDA’s AERS. We validated the generated hypothesis with a retrospective observational study of patients on combined therapy with paroxetine and pravastatin. To determine the significance of an interaction effect, we analyzed three independent data sets and also performed a combined analysis. Across the data sets from all three sites, we found that patients on pravastatin and paroxetine showed a surprisingly large increase in random blood glucose levels relative to baseline values (19 mg/dl, 1.0 mmol/l).