Model predicts drug toxicities, may accelerate approval process

A model created by Weill Cornell Medicine researchers has the potential to impact the preclinical drug development pipeline by computing how likely a given compound is to result in manageable toxicity in clinical trials, according to study results published in Cell Chemical Biology.

The data-driven approach, called the PrOCTOR score, integrates the structure and properties of a novel drug, gathers data from past clinical trials in which a novel drug failed trial due to toxicity events, and provides a new score of measure that may help shape future therapeutic design.

Olivier Elemento

Olivier Elemento, PhD , associate director of the Institute for Computational Biomedicine at Weill Cornell Medicine and head of the computational biology group at the Caryl and Israel Englander Institute for Precision Medicine, spoke with HemOnc Today about the predictive model’s potential to reshape the future of drug development.

Question: How did this model come about?

Answer: It came about from a realization that the way we try to develop new drugs in general, whether it is academia or the pharmaceutical industry, is inefficient. It takes almost a decade for new drugs to be approved, and we knew that there was room for improvement. We think the main reason is that we are not learning well enough from past successes and failures. We realized there was an opportunity to try to apply some of our big-data techniques to the drug discovery process. It is very interesting that people have not really looked at large datasets of clinical trials and all of the aspects of small molecules. Research is showing that, if you look at many aspects of what a drug does, what it tries to target [and] what its chemical structure looks like, several of these features — including unexpected ones — are very predictive of whether a drug will fail human trials due to toxicity reasons. We have created a new predictive model called the PrOCTOR score that integrates all these features. We think we can dramatically speed up the process of developing drugs by using this model.

Q: How was the model evaluated?

A: This is a predictive model, which means that this is a model that we train an initial set of drugs with — drugs that failed other trials due to toxicity events. We analyzed a large number of trials, looking at the features of the drug, and built this model using a form of artificial intelligence, which is at the heart of the model. Obviously, in order to integrate this model, we have to test it using multiple datasets and drugs that were not originally included in these datasets. A set of molecules were identified in this category, and one set of molecules was actually drugs that were approved not in the United States, but in Japan and Europe. We also tested this approach on many more drugs and small molecules that were not in our training sets. Some of these drugs had been looked at by the FDA and were found to have different side effects. We have shown we can predict the number of side effects that a drug has by using this model.

Q: What are the potential long-term implications?

A: There are a variety of implications, one of which is that we have a much clearer idea of what a toxic drug looks like. This really has important implications for the pharmaceutical industry, because all of the drugs in the pipeline can now be looked at using this approach. The industry can now get a prediction for drug toxicities, and maybe researchers can tweak the drug so that it can be less toxic. We also have a web app. People can input their drug into the app and see which features are responsible for drug toxicities. We are potentially providing the recipe for how to make a drug less toxic. We realize that people have been talking about how a lot of our data are incomplete because people are not reporting their results. For our approach to work, we need to know whether a trial failed and why it failed. This information needs to be available so we can use this information to build better predictive models.

Q: Do you plan to conduct more research on the model?

A: What we are doing now is trying to predict specific types of toxicity in a variety of different settings. We are really taking a broad approach to predicting toxicity, but we will need more data because there are different ways in which a drug can be toxic. We think we can use a very similar approach to predict different toxicities. One idea is that we can use specific toxicity predictions to better design clinical trials by enrolling individuals that are unlikely to experience specific severe toxic events. We also may be able to use such predictive models to implement an improved form of ‘precision medicine’ through which individuals likely to experience toxic side effects would be detected early and treated with alternative drugs. We also are trying to go beyond toxicity and predict efficacy, which is a little harder to do. However, we do think we can predict if a drug can do what it is supposed to do.

Q: Is there anything else you would like to mention ?

A: Drug toxicity in humans is a major problem. Some promising small molecules that would efficiently kill tumor cells are frequently not able to get FDA approval or must be withdrawn from the market after approval due to severe toxicities. Oncologists frequently have to discontinue otherwise efficacious treatments due to severe drug side effects occurring in a small number of patients. Our results show that clinical toxicity can be addressed using a data-driven approach that crunches data from past clinical trials and overlooked drug features to accurately predict toxicity before any human receives the drug. We expect that this applied research will translate into fast drug development and eventually will lead to more less-toxic drugs available to oncologists to treat their patients. – by Jennifer Southall

For more information:

Olivier Elemento , MD, can be reached at Weill Cornell Medicine, 1300 York Ave., New York, NY 10065; email: jeg2034@med.cornell.edu.

Reference:

Gayvert KM, et al. Cell Chem Biology. 2016; doi:10.1016/j.chembiol.2016.07.023.

Disclosure: Elemento reports no relevant financial disclosures.