Monoclonal antibodies are increasingly popular therapies for diseases such as cancer, arthritis and multiple sclerosis. They are also very expensive, due in part to the requirement that they are given intravenously at high concentrations to achieve their therapeutic benefits. Attempts to redesign the therapies to allow for easier and cheaper subcutaneous delivery have been stymied by the tendency of the antibodies to clump together, producing an unusably viscous solution. While experimental studies have identified some of the reasons for this viscosity, fully understanding these protein-protein interactions requires zooming in to a scale that's currently beyond the ability of experiments.
Enter computational modeling, which can help scientists determine why some antibodies aggregate and others don't, pointing the way to designing better treatments. While a postdoctoral scholar with the Center for Multiscale Theory and Simulation, Anuj Chaudhri worked with CMTS director Gregory Voth and scientists Dan Zarraga, Steve Shire and Tom Patapoff from the Late & Early Stage Pharmaceutical Development teams at Genentech to construct a model of what exactly happens when you put a lot of these antibodies into close proximity. The work was published by The Journal of Physical Chemistry.
"For high concentration proteins, not many experimental methods are available to get a deeper understanding of the fundamental interactions involved," Chaudhri said. "This is where computation comes in. Using theoretical and computational methods, we can model the problem step by step by putting each piece together."
Computational chemists are capable of modeling the motion of molecules in solution down to individual atoms. But monoclonal antibodies (MAbs) are large, complex molecules, and modeling the behavior of even two antibodies in water would require simulating the motion of well over half a million atoms for millions of iterations, Chaudhri said. Using this scale to simulate a high concentration scenario where hundreds of antibodies interact with each other and the surrounding solution is beyond the ability of even the world's fastest computers.
"Modeling this problem using atomistic techniques would be impossible even with today’s computational power," said Chaudhri, now a Mathematics & Computational Science Postdoctoral Researcher at Lawrence Berkeley National Laboratory. "Hence approaching this problem using coarse-grained models is the right way to go."
In a coarse-grained model, scientists break their target down into a smaller number of subregions that can be simulated more easily. MAbs contain 12 domains, so the first iteration of the computer model broke down each molecule into 12 sites. They then simulated the behavior of 1,000 of these simplified antibodies over 5 microseconds (five millionths of a second), moving it forward 1 picosecond (one trillionth of a second) at a time for 5 million total steps. The model was run for two different antibodies, one that became viscous at high density (MAb1) and one that did not (MAb2), at various concentrations.
The simulation fell in line with the experimental results: at high concentrations, MAb1 formed many more "associations" where antibodies clumped together than MAb2, where such clumps were rarely detected. These associations would be expected to create a more viscous solution -- though simulating viscosity was beyond the scope of this model. The simulations could also identify which regions of these interacting MAb1s tended to stick together, confirming experimental results that the "arms" of the antibody were most likely to connect, and adding in new information about arm-tail and tail-tail interactions. A second study looked at how selectively mutating different amino acids of the antibody could change this clumping behavior, again mirroring the changes in viscosity observed for these mutants in experiments.
What good is a model that confirms the results of already completed experiments? The value comes when you flip the order around, running simulations of redesigned or mutated antibodies before the slower and more expensive process of producing and testing each one in the laboratory. Computer models could help pharmaceutical makers design and test for optimal antibody therapies -- or other drugs -- that are equally effective but less likely to form viscous solutions that make them more difficult to use in the clinic.
"A model such as this can be used a potential screening test for stable/unstable protein sequences, where stable would imply non self-associating solutions," Chaudhri said. "This would save millions of dollars and a lot of tedious work that is spent on performing this experimentally."
===== Chaudhri A., Zarraga I.E., Kamerzell T.J., Brandt J.P., Patapoff T.W., Shire S.J. & Voth G.A. (2012). Coarse-Grained Modeling of the Self-Association of Therapeutic Monoclonal Antibodies, The Journal of Physical Chemistry B, 116 (28) 8045-8057. DOI: 10.1021/jp301140u Chaudhri A., Zarraga I.E., Yadav S., Patapoff T.W., Shire S.J. & Voth G.A. (2013). The Role of Amino Acid Sequence in the Self-Association of Therapeutic Monoclonal Antibodies: Insights from Coarse-Grained Modeling, The Journal of Physical Chemistry B, 130125135828003. DOI: 10.1021/jp3108396