Using Data Science to Reduce Animal Testing - Crowdsourcing for Drug Repurposing

I have been working with teams to study virtual control groups to reduce the need for animals in study control groups, however data science can be used to entirely circumvent the need for animal studies in some cases. This can be done repurposing already studied drugs and re-using all of their associated animal and human studies. Be sure to see the webinar on reducing animal testing in January.

Drug repurposing is of course not a new idea. David Fajgenbaum, MD, MBA, MSc has been leading recent notable efforts to identify treatments for rare diseases by repurposing approved therapies.

Given Dr. Fajgenbaum's work I will use Castleman's Disease as an example although the analysis can be carried out for any indication.

The FDA FAERS database is a significant resource to study drug-adverse event relationships. However a less appreciated aspect is that when an event is reported the physician reports all drugs administered to the patient, and the reason that particular drug was administered. These concomitant drugs are often not related to the reported adverse event. In the example below shows the drugs administered in a case report. The indications are MedDRA preferred terms reported by the prescribing physician.

Currently there are a about 26 million such drug-indication reports in FAERS. It is well known that the FAERS is 'noisy' with errors and extraneous reports. The FDA's Ana Szarfman worked with William DuMouchel to develop and deploy the "multi-item gamma Poisson shrinker (MGPS)" Bayesian algorithm that is now the gold standard for assessing statistical relationships in pharmacovigilance. The FDA uses this analysis to confirm early signals of adverse events caused by drugs. This algorithm is implemented in the Oracle Empirica product and is also available in R. This is only one of many ways that this data in FAERS might be analyzed to find repurposing indications.

In this case I turned this algorithm to look at the 26 million drug-indication records for concomitant drugs in FAERS. The statistical analysis was used to identify statistically interesting uses of drugs on and off-label - thus uncovering the insights that treating physicians had. The analysis was carried out and the empirical Bayes geometric mean (EBGM) value computed for each drug-indication pair. The EBGM value is a ratio of how much more often a given drug was used for a particular indication over any other drug. An EBGM value greater than 2 of the lower bound of the 95% confidence interval was considered to be significant. This filtering reduced the 26 million reports to about 150,000 significant drug-indication pairs. The calculation was carried for the data integrated up to each year shown in the graph below.

A report of drug use unfortunately does not give direct information about efficacy. However, continuous use of a drug over a period of time may suggest that physicians have seen efficacy.

The chart above shows the change in statistics for treatments over time for drugs prescribed for Castleman's Disease from 2005 to 2023. Siltuximab was approved to treat Castleman's in 2014; its use is on-label. Tocilizumab is approved for Castleman's in Japan, but not in the United States. Cidofovir failed a clinical trial for Castleman's in 2004. It has not recently been reported in connection with Castleman's so it's signal has been dropping. One can see that the statistics of many treatments are declining, possibly because they have been found not to be efficacious for many patients. Once a statistical signal appears it takes time for other reports to accumulate to dilute it. In this case the approval of siltuximab and high number of reports (the Y axis is a log scale) may reduce the computed statistical significance of other treatments in this analysis.

The FAERS statistics may lag the use and disuse in treatments, but it can identify other diseases treated by the same therapeutics. The network below shows the other diseases that those drugs used for Castleman's treat; and the commonalities among them. The most obvious mechanism in common is immune suppression, consistent with Dr. Fajgenbaum's experience that Sirolimus, an immunosuppresive, treats the disease. It is notable that Sirolimus was not identified as statistically significant for Castleman's until very recently.

In this context one might suggest that other immunosuppressants that do not appear here could be efficacious for Castleman's disease. The case studies also suggest that drug combinations may be used to treat the disease.

The same analysis was used to analyze therapeutics reported to be used for weight control, weight loss, obesity and similar indications. The analysis is below, filtered to show the most interesting trends.

One can see that semaglutide (Wegovy?) had significant use for weight control before 2019. It was first approved in the US in 2017, and for weight loss in 2021. It's use for weight control became statistically significant before approval for that indication in 2019.

The leader in weight control reports in the graph, statistically, tesamorelin, is not approved for weight loss; it is approved for HIV associated lipodystrophy. It may be that some physicians see efficacy for weight loss. The recent rise of cascara, a dietary supplement made from the skin of coffee fruit, is interesting. It jumped up by a factor of 50x in 2018. (note Y axis is logarithmic) Tirzepatide was approved to control blood sugar and as an adjunct to weight loss in 2022; it has accumulated enough reports to become statistically signifiant. Bupropion, approved for depression in 1985, has enough reports to be connected to weight loss in just the last FAERS data release. It is interesting that chlordiazepoxide (Librium) has also been prescribed for weight loss for the past few years although it has been used for anxiety since 1959.

In conclusion, this evidence-based approach can support new indications for therapies based on physician experience. It has implicit biases and weaknesses from requiring indications to appear in adverse event reports, but provides data to support repurposing and generates ideas for exploration.