The Cloud’s the Limit: Rentable Supercomputers for Improving Drug Discovery

July 11, 2013

By Matt Luchette  
July 11, 2013 | Creating a computer program that accurately tells pharmaceutical companies which candidate drugs they should spend millions of dollars developing may seem like a daunting task, but Schrodinger, a software company that specializes in life science applications, hopes to do just that.
“Our mission is to advance computational drug design to the point of becoming a true enabling technology,” said Alessandro Monge, Schrodinger’s VP of Strategic Business.
Schrodinger won the Bio-IT World Best Practice Award for IT Infrastructure at the Bio-IT World Expo this past April for a drug discovery project they ran in collaboration with Cycle Computing that harnessed the power of cloud-based computing, a tool that allows companies to rent high performance computing hardware.
Since the mid-1900s, the power of the cloud, or infrastructure that provides remote access to digital information, was restricted mainly to scientists and academics, but by the 1990s, with the birth of the internet and email clients like Hotmail, the cloud entered the public realm, providing users access to their files from anywhere they had an internet connection. Users didn’t own the storage space; the company housed the hardware, but allocated a certain amount of storage for each customer.
In 2006, Amazon opened up its Amazon Web Services (AWS) to businesses by providing remote computing through the cloud, as opposed to just remote storage. While Amazon provided the infrastructure, other companies such as Cycle Computing helped clients tailor AWS hardware to their computational needs.
A few years ago, Schrodinger began a project that they hoped would show the power AWS’s supercomputing could have in drug discovery. One of their programs, Glide, could simulate the interaction between a small chemical compound and its target on the molecular level (see, “Going Up: Cycle Launches 50,000-Core Utility Supercomputer in the Cloud”). 
These so-called “Docking Algorithms” have been the Holy Grail for many pharmaceutical companies; an efficient, reliable program that could mimic the interaction between a drug and its target, and quickly scan thousands of small molecules for the drugs that provide the strongest fit, could mean enormous savings for a process that can take over a billion dollars and nearly a decade to complete. 
Yet the computational requirements for algorithms like Glide are extensive; for each of the thousands of small molecules these algorithms screen, the program must simulate each drug’s many possible conformations, as well as the multiple ways for it to bind with its target. The hardware that runs the program needs to be efficient and high-performing; any time or computational constraints on the program would decrease its accuracy and lead to false positives or negatives. 
“To run simulations quickly, it comes at the cost of accuracy,” said Monge.
Schrodinger turned to Cycle Computing for help.  In collaboration with Nimbus Discovery, a computational drug discovery company, Schrodinger wanted to test Glide’s capabilities by screening a staggering 21 million small molecule ligands against a protein target. By building a 50,000 core cloud-based supercomputer in AWS, Cycle Computing provided Schrodinger with the computational power their program required, without the upfront capital needed to purchase new hardware. 
“There are a lot of questions in the cloud” in terms of its reliability and security, Monge explained, “but Cycle was able to work with us and build our infrastructure with AWS.”
Using the 50,000 core supercomputer, Cycle was able to screen the 21 million compounds in three hours, a process that would have taken Schrodinger engineers an estimated 12 years to run on their own. Furthermore, while Schrodinger would have needed to invest several millions of dollars to build a similar supercomputer in-house, “the project cost was less than $4,900 per hour at peak,” according to Cycle. The software even identified a number of promising candidate compounds that the program would have rejected without the increased accuracy AWS provided.
As Monge explained in a presentation at the “Big Data and High Performance Computing in the Cloud” conference in Boston last year, a “50,000 core Glide run represents a proof of concept that we can start attacking a scientific problem without being constrained by computational resources.” 
As evidenced by the high efficiency and fidelity Schrodinger was able to achieve by running Glide on the cloud, Monge remarked that “the cloud is the next level of Moore’s Law.”
While Monge was not able to comment on updates to the program or new projects the company is undertaking, he said that winning the award has generated even more momentum within Schrodinger to pursue cloud-based computing. “Our customers know we have a serious effort in the cloud,” he said.
In the nomination application for the Best Practice award, Cycle Computing summarized the possible implications of the project, stating that rentable supercomputing hardware can make drug-testing algorithms possible that would otherwise be “too compute intensive to justify the cost.”