Accelerating Drug Development

New tools in the form of ground-breaking techniques and the latest computer hardware are enabling drug developers to design compounds more quickly and accurately than ever before.

Introduction
As in most fields of scientific endeavour, the drug development community has adopted computer-based simulation alongside theory and experiment as the third pillar of its research. These in-silico methods running on departmental clusters based on commodity components present tremendous advances in the computational power now available to scientists on a day-to-day basis.

However there are limits to this approach: departments are often constrained by the infrastructure and location of their cluster. These limits might be:

• The amount of space that is available, or on the available power supply (one fully configured “rack” in a cluster can consume up to 25 kW of electrical energy);
• The ability to cool the cluster (much of the 25 kW per rack is turned into waste heat which then has to be efficiently removed to keep the cluster operating properly);
• The strength of the floor on which the cluster will stand (a full-size rack when fully loaded can approach 1000kg/2,200 pounds in weight).

Even where physical infrastructure constraints aren’t the limiting factor, other issues may affect how much compute power a scientist has access to. The software they are using may not scale well across a bigger cluster. The bigger a cluster, the more difficult it becomes to manage, and bigger clusters often require a greater proportion of the available budget to be spent on more expensive interconnects. In addition, larger clusters use more power, which means that more CO2 is emitted into the atmosphere. As corporations become increasingly conscious of their carbon footprints, IT infrastructure capability will increasingly be capped by corporate carbon emission goals.

These constraints can restrict how much computational power is available to a scientist, limiting their ability to run more, or longer, simulations or to adopt more accurate (and usually more computationally expensive) methods. Fortunately, there are now solutions available that directly address these issues.

Application accelerators
Application accelerators are systems designed to be installed as part of a commodity cluster. Much as a graphics card added to a home PC can make a game run faster, application accelerators can be added to the servers in a cluster to make a wide range of life science applications run much faster than before, while at much lower power consumption. These accelerators significantly increase the computational power available for a given space, weight, power supply or cooling limit. Because of the improvements in power consumption, they also reduce the CO2 generated by the cluster, helping to meet corporate carbon emission goals.

ClearSpeed Technology has developed a range of professional accelerators to complement commodity cluster solutions in the life science space. ClearSpeed’s accelerators use a “many core” approach, exploiting more of the parallelism available in an application than a conventional dual- or quad-core x86 CPU is able to. Most applications that contain lots of parallelism can be accelerated to some degree, with examples of applications that can be significantly accelerated including:

• Molecular Mechanics (MM) based simulations
• Protein-Ligand docking
• Density Functional Theory (DFT) methods from Quantum Mechanics (QM)
• Mixed QM/MM methods

Scientists are already putting ClearSpeed accelerators to good use in these areas.

Molecular mechanics acceleration
At the Tokyo Institute of Technology, Professor Matsuoka has constructed the world’s largest accelerated cluster – TSUBAME. This cluster, consisting of 648 Sun Microsystems X4600 servers, each with its own ClearSpeed accelerator delivers 56 trillion floating point operations per second (TFLOPS), 47% of which comes from the ClearSpeed accelerators. Because the ClearSpeed accelerators are installed inside each node, the 47% performance increase came with no increase in space requirements, and with a tiny increase in power consumption (under 2%). TSUBAME primarily runs life science applications, including the AMBER molecular mechanics package which has been modified to run on ClearSpeed accelerators.

Protein-ligand docking acceleration
Dr Richard Sessions at the department of biochemistry at Bristol University has been developing next-generation, highly accurate protein-ligand docking methods. Dr Session’s group has been focusing on protease inhibitors using a specific type of peptide against human elastase, an enzyme which is involved in the extensive scarring of lung tissue in emphysema. As part of this work, new methods for evaluating the “fitness” of a protein-ligand docking have been developed which include a novel atom-atom based empirical free energy force field. Implemented in a computer program called BUDE, this method promises to more accurately determine the best possible candidates from a large virtual library of potential drugs. The downside is that these more accurate methods are significantly more computationally expensive than the previous, less accurate methods.

Dr Sessions estimated that calculating the 8 billion poses for the elastase peptide library would take over 100 days on an x86 quad core server, an unacceptably long time. Fortunately, the calculations BUDE performs are very parallel, with poses being able to be processed simultaneously, ideal for ClearSpeed acceleration. Dr Sessions ported just 500 lines of code from BUDE to run on a 1 TFLOP ClearSpeed CATS system and reduced the execution time to just 7.5 days, a time span much more acceptable to most projects. Further developments enabled an entire peptide library calculation to be performed in just 18 hours on a cluster of ten CATS nodes, a system that would easily fit in just half a rack within any department’s machine room.

Quantum mechanical simulation acceleration
Quantum mechanical simulations are playing an increasingly important role within many areas of the pharmaceutical industry. However, the greater accuracy enabled by methods such as Density Functional Theory are often out of the reach of scientists because of the prohibitive computational cost associated with these methods, especially when applied to non-trivially sized molecular systems.

Researchers at Bristol University’s Chemistry department, lead by Dr Fred Manby and Dr Adrian Mulholland have been developing mixed QM/MM methods to take advantage of the extra accuracy enabled by DFT. In collaboration with Dr Christopher Woods and PhD student Mr Philip Brown, a QM/MM software package called Sire is being developed to efficiently calculate free energies. Sire itself uses the Molpro quantum mechanics software package to provide the DFT methods.

While still early in its development, Sire has been used to study the binding free energies of potential neuraminidase inhibitors in the fight against the influenza virus. Neuraminidase is an enzyme found on the surface of the influenza virus, and neuraminidase inhibitors work by blocking the function of this protein. This then prevents cell reproduction and eventually eliminates the virus. Sire uses Molpro to perform the QM DFT calculations on the entire ligand under consideration – typically 50-60 atoms in the studies so far. The neuraminidase molecule itself is simulated using MM methods, hence the hybrid QM/MM approach.

This method is computationally extremely expensive so, even though it should offer many advantages, it may be out of reach for most potential users. To solve this problem, the group has collaborated with ClearSpeed to accelerate the QM part of the computation, which represents the vast majority of the computation required.

To achieve this acceleration, the DFT part of Molpro has been ported to ClearSpeed accelerators. Running on a single CATS system, a speedup of 7.3X has already been achieved for Molpro, with speedups of around 15X believed possible. Using Sire accelerated on ClearSpeed hardware, the neuraminidase study took 5 days to complete; without acceleration the same computation would have taken over 5 weeks.

Summary
In-silico computational chemistry methods are already a valuable tool within the pharmaceutical industry, and this trend is set to continue to grow. Commodity clusters have helped fuel the success of these methods so far but commodity clusters are rapidly reaching their natural limits in terms of size, scaling, power consumption, heat dissipation and carbon emissions. New, more accurate and more automated computational methods will demand even more compute power, further exacerbating these issues.

Professional application accelerators from ClearSpeed Technology are designed to solve this problem, building on the existing successful infrastructure of commodity clusters and delivering significant increases in computational power available within given space, power and cooling budgets, while dramatically reducing carbon emissions for a given simulation.
For more information on ClearSpeed’s products and how they can give your business an edge, please visit www.clearspeed.com.

Simon McIntosh-Smith is Vice President for Customer Applications at ClearSpeed Technology plc. Simon joined ClearSpeed in 2002 and has helped lead the development of ClearSpeed’s application acceleration products, both in terms of the hardware design and the supporting software.