A case for screening real‐world data for collateral drug benefits: Glucagon‐like peptide 1 receptor agonists and bile acid diarrhea

Collateral drug benefits are hitherto unknown beneficial effects that might lead to repurposing of already marketed drugs. A randomized controlled trial has found liraglutide to be non‐inferior to colesevelam in reducing bile acid diarrhea. We hypothesized that this collateral drug benefit of liraglutide could have been detected using observational data.

• We present results compatible with trial findings in both magnitude and direction obtained in observational data using a symmetry analysis design.
• Hitherto unknown beneficial effects, collateral drug benefits, have been difficult to detect in clinical practice.
• Such collateral drug benefits might be captured by systematic screening for inverse associations in large real-world datasets.

| PURPOSE
Collateral drug benefits (CDBs) are beneficial drug effects that are unrelated to the reason for prescribing the drugs.Discovery of CDBs can ultimately lead to repurposing of already registered drugs.Repurposing has many advantages over developing new drugs from scratch, since the repurposed drug already has a known safety profile, has lower costs in development and shorter time in development. 1In addition, repurposed drugs are less expensive, which has value in ensuring equity.Discovery of CDBs has largely been serendipitous, based on mechanistic thinking or clinical observations.A substantial amount of registered drugs do not have completely known modes of action and posses affinity for multiple ligands and targets, 2,3 which suggests that we cannot rely on mechanistic thinking to discover CDBs.Since there is no infrastructure in place to process clinical observations suggestive of CDBs, as there is for adverse drug reactions, we cannot rely on clinical observations either.An efficient approach to detect CDBs could be systematic analyses of large health care registries.
As an illustrative case, we have chosen the recently published discovery of the glucagon-like peptide 1 receptor agonist (GLP-1RA) liraglutide as an effective drug against bile acid diarrhea (BAD). 4In the randomized controlled trial the main outcome was the proportion of participants experiencing a reduction in daily stool frequency of 25% or greater after 6 weeks.The risk difference was À27% in favor of liraglutide (one-sided 95% CI À100 to À6) compared to colesevelam.This corresponded to liraglutide had decreased stool frequency by À1.83 stools per day (95% CI À2.32 to À1.35), corresponding to a reduction of 54.0% compared to baseline.BAD is a debilitating condition, and the discovery of this collateral drug benefit could lead to repurposing of GLP-1Ras for treatment of BAD.
The aim of the study was to evaluate in whether this CBD would have been detectable in observational data from before this finding was published.

| METHODS
The current mainstay of treatment for BAD is bile acid sequestrants (BAS). 5In addition to treatment of bile acid associated diarrhea, BAS are also used as cholesterol-lowering treatment, but owing to high cost and low palatability, they are not used much for this indication. 5nce patients who incidentally are treated with GLP1-RA are already effectively treated for BAD, we would expect a lower-than-expected rate of BAS initiations among its users, that is, an inverse association between GLP1-RA and BAS initiation.
Using the Danish national prescription registry, 6 we identified all individuals who redeemed a prescription of GLP-1RA and of BAS during the period 1995-2022.In the main analysis, all GLP-1RAs were included.In the secondary analysis, only liraglutide and semaglutide were analyzed since the remaining marketed GLP-1Ras (exenatide, lixisenatide, and dulaglutide) have very limited usage in Denmark.The exposure of interest was use of GLP-1RA, and initiation of BAS was the outcome of interest (Supporting information S1).In the analysis, use of BAS served as a surrogate of BAD.
Data were analyzed according to a sequence symmetry analysis (SSA) design. 7The SSA is a self-controlled analysis, and all comparisons are made within individuals initiating both the exposure and the outcome drug.This makes the SSA robust toward time-invariant confounding as well as efficiently processed. 8We indexed individuals on the date of the first ever prescription of any GLP-1RA and restricted the analysis to all individuals who also had a first ever prescription of BAS between 365 days before and until 365 days after the index date.In the sensitivity analyses, we have used the first prescription of liraglutide and of semaglutide, respectively, in two separate analyses.Individuals were then classified according to the order of initiations (GLP-1RA !BAS or BAS !GLP-1RA).The effect measure of interest is the sequence ratio (SR) which was calculated as the number of GLP-1RA !BAS sequences divided by the number of BAS !GLP-1RA sequences.The SR estimates an incidence rate ratio, and ratios below one can be interpreted as protective effects.We obtained 95% confidence intervals based on Jeffrey's interval for binomial proportions. 9As the exposure-anchored SSA is susceptible to temporal prescribing trends for the outcome drug, we adjusted for such trends using null-effect SR adjustment. 10As a sensitivity analysis, and to evaluate the robustness of our results, we repeated the analysis with shortened time windows of 90 and 180 days.The negative control exposures, dipeptidyl peptidase 4 inhibitors and sodium glucose cotransporters, were drugs with similar indications as GLP-1RA, but without an assumed effect on BAD.A study protocol was registered prior to the commencement of any statistical analyses. 11For codes used to define exposures, negative controls and outcomes, see Table S1.

| RESULTS
We identified 158 individuals who initiated both GLP-1RA and BAS.
Their median age was 58 years and 70% were female.The crude SR was 1.00 (95% CI 0.73-1.37)and the trend-adjusted SR was 0.96 (95% confidence interval 0.70-1.31).When stratifying on the type of GLP-1RA, we found results compatible in direction, with the previous trial (SR liraglutide 0.75, 0.51-1.10 and SR semagltuide 1.23, 0.80-1.89).Despite the non-significant confidence intervals for liraglutide, the performed sensitivity analyses support the direction of the association, indicating a reduced rate of initiation of BAS following initiation of liraglutide 4 but not of semaglutide. 12When using observation windows of 180 and 90 days, we found sequence ratios of 0.61 (95% CI: 0.40-0.93)and 0.60 (95% CI: 0.34-1.04).Since BAS also can be used as a cholesterol lowering drug, we repeated the main analysis while excluding individuals who had redeemed a prescription for statins prior to BAS initiation, which yielded a sequence ratio of 0.56 (95% CI: 0.33-0.96).Post hoc, all analyses were repeated using negative control exposures and both dipeptidyl peptidase 4 inhibitors (SR 1.05, 95% CI: 0.73-1.52)and sodium glucose cotransporter 2 inhibitors (SR 1.00, 95% CI: 0.71-1.40)yielded null-results (Table 1).

| CONCLUSION
The results from this sequence symmetry analysis are compatible in both magnitude and direction with the results from the previous randomized controlled trial. 4Compared to the existing case report and randomized trial, the strength of our study is the use of real-world data, showing that the trial results may be generalizable to a realworld population.The main limitation of our study is that we cannot rule out time-varying confounding that may have influenced our results.Still, the concordance of our findings with trial evidence, the elimination of time-invariant confounders and the compatible null associations found for drugs with similar indication increase our confidence in the results.The finding of stronger inverse association in the subgroup with no previous use of statins points to a possible misclassification bias in the main analysis.We assumed that BAS generally would reflect BAD, but it is to some extent used as a third line drug for hypercholesterolemia. Which implies that in the subset with no past use of statins, BAS would have higher positive predictive value as marker of BAD, thereby conferring less bias.The fact that the inverse association was detectable in data originating from before the trial's publication renders it unlikely that we would have any appreciable confounding by indication, and it points to the added value in being able to identify the drug effect earlier than by other channels, for example, trial evidence.Thus, our findings support the case for using epidemiological analyses of real-world health care databases as an instrument to discover collateral drug benefits that eventually might result in drug repurposing.
T A B L E 1 Number of sequences and sequence ratios for main and sensitivity analyses using a sequence symmetry analysis design.