C&EN talks with Rommie Amaro, computational chemist

Determining the structures of enzymes and other biomolecules with X-ray crystallography has deepened biologists’ understanding of the inner workings of cells and led to the design of many important drugs. Increasingly, researchers are using computer modeling to attain a more realistic picture of the movement of these biomolecules in their natural environment. Rommie E. Amaro, a professor of chemistry and biochemistry at the University of California, San Diego (UCSD), uses computational techniques to predict how enzymes regularly shift their configurations, revealing potential vulnerable areas for drug targeting. C&EN talked with Amaro about how she’s using dynamic models of enzymes to develop a new class of cancer drugs, work that is being commercialized by Actavalon, a San Diego start-up Amaro cofounded.

How did you get into working on computational drug design?

I fell in love with physical chemistry as a chemical engineering undergraduate at the University of Illinois, Urbana-Champaign (UIUC). I ended up doing a biophysics research project with Zan Luthey-Schulten predicting the structure of a protein involved in histidine biosynthesis, and I just loved it. I could not wait for Saturday, and not because of tailgating but because I could go to the lab. I didn’t have to worry about classes, and I could just do research. I was there for hours. That was when I knew something was wrong with me (laughs)!

From there I went to work for Kraft Foods for a couple of years doing very traditional chemical engineering and product development for Philadelphia Cream Cheese, which was a ton of fun. But I missed something—I went back to graduate school to find out what it was. And I fell back in love with computational chemistry, and with molecular dynamics simulations. The work of Klaus Schulten at UIUC was tremendously influential, and later, as a postdoc with Andy McCammon at UCSD, I learned about pharmacology and drug design.

What can computational techniques give you that X-ray crystallography alone cannot?

These days, X-ray crystallography is constantly driving to higher and higher resolution. In rational drug discovery, usually people start with a crystal structure, which is a high-resolution image of the relative locations of the atoms in the drug receptor, then try to screen and discover small molecules that bind with very high affinity and stay for a long time. That’s the standard protocol—to rely on static images of proteins. Of course, we know that, as important as X-ray crystallography data are, it’s like taking a Polaroid snapshot of a dynamic, moving machine.

If a drug receptor is flopping around, then the technique can’t resolve the relative positions of the amino acid residues. So researchers do all these tricks to try to immobilize these proteins. It’s different than the protein’s actual environment in vivo.

Where my lab has been trying to push the envelope is in trying to use molecular dynamics simulations to predict how these proteins or drug targets move over time. We then select out structures from what we call the trajectory of the drug target and use those alternative structures in drug discovery.

How are these simulations enabling you to make new cancer drugs?

It turns out that in more than 50% of human cancers, the gene for a protein called p53 has one single mutation that basically renders it inactive. This protein binds to DNA to activate gene expression that triggers cell death when cells are damaged, causing them to die instead of grow into tumors. When p53 is inactive, it can’t prevent tumor growth. We try to develop small molecules that would reactivate p53 because this would be a broad-spectrum cancer therapy.

More than 100 static images of p53 are available in the Protein Data Bank, and they all have fairly similar structural features. We might start with one of these and do an analysis on it. We place the protein in what we call a bath of water—all these simulated water molecules—and ions to mimic the solution environment. Then we might see a movement of a side chain, or we might see an entirely new pocket open up—a pocket that wasn’t evident in the crystal structure but reveals itself in these predictive simulations. Oftentimes, those pockets are actually druggable. We find these pockets and then we try to design small molecules that bind to them.

In p53, a pocket basically revealed itself after we simulated just a few tens of nanoseconds of the enzyme dynamics (PLOS Comput. Biol. 2011, DOI: 10.1371/journal.pcbi.1002238). Then we used this pocket to design and discover reactivation compounds. The p53 protein turned out to be a nice example of the power computation can play in drug discovery.

How are computing advances changing your work?

One thing that’s exciting about our field right now is the meteoric rise in the ability of commodity graphical processing units, or GPUs, to carry out molecular dynamics simulations. GPUs have become very high performance because they’re used to render realistic graphics in real time for video games. It turns out that those same processors are highly amenable to the kinds of scientific calculations that we use for drug discovery.

This has been a game changer. Coming back to p53, when we originally did that work we ran the simulations on supercomputers, and it took a couple of months to generate the dynamics data. Today we can do those exact same simulations in probably three hours on a single GPU.

Are there still things that have to be done on supercomputers?

As the system size scales, we still need to go to supercomputers. We have simulations now that contain multiple hundreds of millions of atoms. Those have to run on some of the biggest computers in the world.

More recently, we’ve been interested in continuing to develop models that are more accurate and more realistic. We have built a model of p53 on a supercomputer that’s gone from 50,000 atoms to 1.5 million (Oncogene 2016, DOI: 10.1038/onc.2016.321), and we’re looking at time scales of the enzyme dynamics that are not just tens of nanoseconds but more on the order of microseconds. We can ask totally different questions now that we have these much larger, more realistic simulations of p53. For example, how does p53 bind to DNA, and how does the sequence within the DNA itself change the binding dynamics?

You can’t run that newer work on a GPU—or you could, but it would be ­super slow. I think five years from now, with improvements to GPUs, maybe we will be routinely simulating 2 million ­atoms on a single GPU chip. And 10 years from now, we’ll be simulating whole cells with more than 1 billion atoms on ­future supercomputers now being designed.