Computational Methods Used to Predict Drug Side Effects

Computer programs that simulate biological events in virtual reality are playing an increasingly important role in many areas of pharmaceutical research. Adrian Elcock, D.Phil., University of Iowa assistant professor of biochemistry, is studying whether purely computational approaches can be used to identify unintended drug interactions that might cause side effects.

In general, a drug works by binding to a specific receptor molecule in the body to produce a therapeutic effect. If the same drug cross-reacts with a different receptor molecule, a side effect can occur that can range from a mildly unpleasant physical effect to a life-threatening reaction. Because side effects often are not discovered until late in the drug development process, side effects that limit or even prohibit a drug's use can represent a huge waste of time and money.

"Basically every drug has side effects," Elcock said. "We are trying to develop computational methods that would allow us to predict side effects at the outset of the drug development process before a drug is even in clinical tests."

Elcock and Bill Rockey, an M.D./Ph.D. student in the UI Medical Scientist Training Program, tested the ability of a common computational method known as a drug-docking program to successfully predict drug-receptor side interactions. The UI researchers used a series of docking simulations to compare the program's computational results, which predict the likelihood that a drug will bind to a particular receptor, with experimental results that measure real receptor-drug interactions. The researchers found that, given certain constraints, the docking program could successfully predict which receptor molecules would interact with a particular drug and which would not. The study, which was funded by the UI, was published in a recent issue of Proteins: Structure, Function and Genetics.

The UI study examined a series of clinically important drugs known as kinase inhibitors, including Gleevec, a recently approved treatment for acute myeloid leukemia. Kinases are a large family of proteins, which control many important biological pathways.

"The whole point of this study was to compare the computational results from our approach with real experimental data for these drugs," Elcock said. "We were able to predict with a good degree of confidence which kinases would be inhibited by these drugs and which ones wouldn't."

Importantly, the method did not produce any false negatives, situations where the docking program predicts a weak receptor-drug interaction, but the real experiment shows a strong interaction.

"If we had missed any strong interaction between a drug and a receptor, that would be a very bad sign for the usefulness of this method because a side effect could be caused by just a single interaction," Elcock explained.

The method did produce several false positives. That is, it predicted strong interactions that were not found in the real experiments. However, getting false positives should not be a major problem.

"This method is not a replacement for the actual experimental testing," Elcock said. "But it may be a very good guide. So long as this method allows us to throw out a lot of the chaff, which we were able to do, having a few false positives left over to be eliminated by real experiments is not really a problem."

Many interesting drug targets, including kinases, occur in large protein families. Within these protein families, different family members often have similar docking sites and drugs bind to those sites in similar ways. The docking program uses what is known about the structure of a drug bound to one receptor as a model for how that drug will interact with similar receptors.

Thus, an essential prerequisite of this approach is the availability of at least one x-ray crystal structure of a drug bound to a real receptor protein. A crystal structure is obtained by firing X-rays at a crystallized protein. This produces a diffraction pattern and analysis of that pattern reveals the three-dimensional shape of the protein.

The receptor-drug structure identifies a correct binding orientation for the drug and the location of the docking site -- the region on the receptor protein where the drug binds. These are critical pieces of information for the docking program.

The shape of a drug-bound receptor also provides a template to construct models of similar receptors. This approach, known as homology modeling, allows researchers to build models of receptor proteins with known protein sequences even when no x-ray crystal structure is available for those proteins. The success of the docking program demonstrated by the UI study suggests that homology modeling may be broadly useful in screening many proteins for potential side effects.

With the initial study proving the feasibility of using computational methods to predict unintended receptor-drug interactions, Elcock hopes to extend the research on two fronts. Working with known drugs for which there are crystal structures of the drug bound to a receptor, the UI team will screen those drugs against all other receptors to identify potential side effects.

The researchers also will use the docking program to investigate normal metabolic pathways. Substances called metabolites are formed naturally during metabolic processes. Determining which proteins interact with metabolites may help researchers understand metabolic pathways and also identify important regulatory proteins.