Folding@home exascale supercomputer finds potential targets for COVID-19 cure

The Folding@home project has shared new results of its efforts to simulate proteins from the SARS-CoV-2 virus to better understand how they function and how to stop them.

Folding@home is a distributed computing effort that uses small clients to run simulations for biomedical research when users’ PCs are idle. The clients operate independently of each other to perform their own unique simulation and send in the results to the F@h servers. (Read more about where the Folding@home network is administered and how it broke the exaFLOPS barrier.)

In its SARS-CoV-2 simulations, F@h first targeted the spike, the cone-shaped appendages on the surface of the virus consisting of three proteins. The spike must open to attach itself to a human cell to infiltrate and replicate. F@h’s mission was to simulate this opening process to gain unique insight into what the open state looks like and find a way to inhibit the connection between the spike and human cells.

And it did so. In a newly published paper, the Folding@home team said it was able to simulate an “unprecedented” 0.1 seconds of the viral proteome. They captured dramatic opening of the spike complex, as well as shape-shifting in other proteins that revealed more than 50 “cryptic” pockets that expand targeting options for the design of antivirals. 

Dr. Greg Bowman, associate professor of biochemistry at Washington University and leader of Folding@home, said the spikes hide from the immune system by folding up on themselves to protect their receptor-binding sites, kind of like how a turtle pulls into its shell. Eventually, though, they would have to open up to find a potential host. 

That was what F@h targeted. “We knew it was happening but not what it looked like. In the simulations, we see far more extensive opening than had been seen experimentally,” Bowman told me.

The model derived from the F@h simulations shows that the spike opens up and exposes buried surfaces. These surfaces are necessary for infecting a human cell and can also be targeted with antibodies or antivirals that bind to the surface to neutralize the virus and prevent it from infecting someone.

“By generating over 100-fold more data than anyone else has access to, we were able to capture events like a dramatic opening of the spike that exposes surfaces one wouldn’t otherwise have expected were viable targets. Likewise, we also found opening motions that create novel pockets in many other viral proteins. All these new structural features could be useful drug targets. We’re sharing all the data online so that others can use it to understand the virus and develop antivirals in parallel with our own efforts,” Bowman said.

To that end, Folding@home is taking part in a project called Covid Moonshot, an open science collaboration to design small molecules to block viral replication. The Covid Moonshot team is using Folding@home to screen through large libraries of chemicals for the ones that are most likely to be useful antiviral drugs, and those chemicals are being synthesized and tested by experimental collaborators.

Tech titans step up

Before COVID-19, Folding@home was a rather modest project with 30,000 users and a few servers at Washington University and other schools. When the pandemic hit, and Nvidia put out a call to arms to its users, the organization suddenly gained nearly a million users and its servers were massively overloaded. 

From there, the tech industry really stepped up to help. Bowman said Microsoft, Nvidia, AMD, Intel, AWS, Oracle, and Cisco all helped with hardware and cloud services. Pure Storage donated a one petabyte all-flash storage array. Linus Tech Tips, a hobbyist YouTube channel for home system builders with 12 million followers, set up a 100TB server to take the load off.

“We’re extremely grateful for all the help,” Bowman said. “The massive outpouring of support from citizen scientists and the help from the tech industry opened up scientific opportunities we wouldn’t have had otherwise.”

At its peak, F@h had more than 2.5 exaFLOPS of computation to throw at the problem, or 2,500 petaFLOPS. As of late October, that has dropped to 247 petaFLOPS.

However, Bowman sees an upside to it: That’s still more horsepower than Summit, the fastest supercomputer in the United States.