When it comes to a take-home from Tom Cheatham’s keynote speech on Thursday morning two weeks ago at the TeraGrid ’09 conference in Arlington, Va., the subtitle says it all: “Chronicling the growth of a student to tenured professor in the NSF supercomputing center microcosm.”
Cheatham has been involved in biomolecular simulations at the NSF supercomputing centers since the beginning of his career, which as a grad student was in the early years of the supercomputing centers. Nearly 20 years later, Cheatham (as of July 1) is an associate professor of medicinal chemistry and of pharmaceutics and pharmaceutical chemistry in the University of Utah’s College of Pharmacy, one of the highest ranked pharmaceutical schools in the country (receiving $23 million a year in funding for pharmaceutical research). In his talk, he acknowledged how his career — he’s an international leader in the art and science of simulating proteins and nucleic acids, particularly as it relates to the development of drugs — has tracked the evolution of the NSF centers, now TeraGrid, and would not have been possible without it.
Cheatham highlighted how the TeraGrid and its progenitors enabled the field of biomolecular simulation. “In my case,” he said, “this started with exposure to education and training workshops at the various centers as a graduate student, successful collaboration with vendors and the centers to improve our codes and methods, exposing and fixing lots of problems, to now my continued extensive use and development of, and strong advocating for, HPC resources.”
He briefly sketched the complexities involved in understanding how drugs interact with proteins and nucleic acids, such as RNA, which has been a focus of some of Cheatham’s recent work. As a drug molecule associates with its target protein or RNA, it can alter structure, and the goal is to use biomolecular simulations, codes and methods developed over the past 40 years by various groups and centers, to track the atom-to-atom interactions and thereby to help improve understanding of the structural changes.
In his recent work, Cheatham has learned more about RNA and its roles in regulating gene expression and synthesizing of peptides. While there are some drugs that bind to RNA in the ribosome, he notes, there are few that bind to the many other RNAs molecules present, which are a potentially effective target for therapeutics. Cheatham has begun a research effort with potential for treatment of Hepatitis C. This viral infection is the major cause of liver transplants, and Cheatham’s work aims at direct therapeutic intervention with the viral RNA.
Cheatham sketched a brief history of biomolecular simulations, from the first protein simulation in 1975 to the present. More powerful resources and software advances have made it possible for simulations to extend over longer (more biologically realistic) time scales — from nanoseconds to microseconds (first achieved by Cheatham’s colleague Peter Kollman at PSC in 1998) and now milliseconds.
Inevitably, however, longer simulations have revealed flaws that weren’t apparent in shorter periods. “I’m known in our field as the guy who breaks things,” said Cheatham. “The longer we run, the more problems we find, and without pushing the simulations, we wouldn’t find the problems.” These breakdowns in methods lead to back-to-the-drawing-board efforts — often involving improvement to the force-fields that describe the atom-to-atom interactions — and result in new, better simulations and deeper understanding.
Cheatham also reinforced themes from other TeraGrid ’09 sessions. One of them, from a Wednesday (June 24) session on federal agencies, is the disproportion between the reliance of biomolecular simulation on TeraGrid resources and NIH support for HPC. “If we didn’t have TeraGrid,” said Cheatham, “my field would stagnate, and perhaps die.” Cheatham advocated for collaboration among federal agencies to support HPC.
Cheatham also fortified the message from Ed Seidel, director of the NSF Office of Cyberinfrastructure, Wednesday’s keynote speaker (June 24), about the need for HPC education at the undergrad and grad student level. New grad students in his research group, Cheatham said, have little if any HPC experience and training. “Most students who come into my lab don’t know Unix.” As a result, they experience a steep learning curve that discourages them from using TeraGrid. Cheatham often sends these students to TeraGrid tutorials, and stressed the need for more of such training resources.
The petascale era presents major challenges for biomolecular simulations, says Cheatham. The programs scale well enough up to 256 processors, but getting to 10,000 processors, he says, is “going to be really, really hard.” An approach that to some extent compensates is ensemble simulations — running many smaller, parallel simulations at the same time. While gaining efficiency in the use of hardware, ensemble runs create serious challenges in management of workflow and data.
In raising concerns about the deluge of data resulting from petascale computations, Cheatham echoed one of the prevalent themes of the TeraGrid ’09 gathering. Whereas previously simulations on the nanosecond scale, he said, took months to years, this now happens in hours to days, which presents daunting barriers in data manipulation, analysis and interpretation. “We now run up to 1,000 processors. With storing of data and analyzing it locally, the data is too large, takes too long to transfer, and the analysis has become rate limiting.”
One solution, he offered, is to create ways to interact with the simulations — to “steer” simulations and examine data on the fly. “How can we find hidden correlations in the data?” asked Cheatham. “How can we find things we do not already know?” The goal, he emphasized, is to do less computational science and accomplish more science.