Pfizer's Biotherapeutics and Bioinnovation Center. Modeling Protein-Protein Interactions in Weeks, Not Years.

---

The Science

Antibodies have been used as medicine since the late nineteenth century, explains Jacob Glanville, an Informatics Associate at Pfizer’s Rinat biotherapeutics research unit. "Antibody therapeutics probably date back to anti-venom," Glanville says. "Basically, it was found that when a horse was bitten by a cobra, the horse is big enough that it won't die, but the proteins in the snake's venom annoy the horse and cause it to create anti-venom antibodies." Scientists discovered that it was possible to isolate serum from the horse's blood and inject it into human beings to stop the deadly effects of a poisonous snakebite.

Today, making antibody therapeutics involves sophisticated techniques for protein design and modeling. "Say you have a disease, and you think that a certain protein is implicated in the disease," says Andrea Rossi, a Senior Scientist at Rinat. "You take the protein that you want to inhibit, you put it into a mouse, you allow the mouse to have an immune response to this protein, and then you extract the cells that produce the antibody against this protein." Next, says Rossi, the mouse antibodies are "humanized," their amino acid sequences altered so that human bodies won't identify them as foreign matter.

Antibody engineering is the specialty of Rinat, a former biotechnology company that was formed in 2001. Located in South San Francisco, Rinat was acquired by Pfizer in 2006 and now forms part of the Biotherapeutics and Bioinnovation Center (BBC), a new R&D division of Pfizer characterized by its biotech-like entrepreneurial agility. "Our primary objective is to use information technology to drive the science quicker and get the therapeutics out the door faster," explains Giles Day, Senior Director of Informatics at the BBC. One facet of the BBC's approach is a unique computational modeling technology that Glanville and Rossi developed to predict a hard to know but crucial variable for antibody engineering: the location of the epitope, or the spot where an antibody binds to its target protein. To power the computational process, which can generate up to a terabyte of data per project, the Rinat team turned to Mac.

The Challenge

It is possible to create an antibody therapeutic without locating the epitope, but protein engineers who are developing drugs want the information for a couple of reasons. First, Rossi explains, knowing where an antibody binds allows researchers to understand a drug's mechanism of action precisely, which allows for more advanced and targeted engineering of the protein. Second, a description of the binding site is used to substantiate a drug maker's patent application—it's called "epitope protection," and it's crucial to safeguarding a protein engineer's intellectual property or IP.

Classically, scientists used x-ray crystallography to determine the binding site of an antibody. The intensive process involves creating a crystal out of the joined antibody and antigen, and using x-ray radiation to determine the position of every atom in the complex. It's time-consuming, expensive, and not fail-safe.

"Traditional methods of identifying the epitope can take six months to a year," says Giles Day. "So waiting to solve the structure of the antibody target complex is a huge time sink, especially in the BBC where every week counts." Rossi and Glanville's challenge was to develop a successful predictive model of protein-protein interaction that would deliver results in weeks, not months or years.

The Solution

"Proteins are very complex macromolecules with thousands of atoms," says Andrea Rossi, and their surfaces are irregular and extremely intricate. "You have to use software to predict these interactions. We've developed a computational platform that runs on the Xgrid, which takes as its input the sequence of an antibody and the structure of the target, and as output it predicts where the antibody binds to the target."

Rossi and Jacob Glanville worked together to evaluate available protein modeling software, and modify it to meet their ambitious requirements. The end result is something they call "the pipeline": a funnel of software programs with hand-written scripts that modify the data output of one program into a suitable input for the next. The software rotates spatial models of an antibody around a model of the target protein, searching for possible interaction sites, and evaluating tens of millions of them to determine the most likely site for the epitope.