Few natural phenomena are trickier to understand than the interactions between molecules in the body. They govern everything from immunity to motion to memory.
Researchers can spend decades just trying to understand how and why one particular drug or protein interacts with another.
A young company called Atomwise is employing high-powered computing to help answer some of these questions, in the hope of completing the task exponentially faster than clinical research ever might.
“Initially we didn’t think about it as a business,” says Atomwise’s chief executive officer Abraham Heifets. “We were thinking about it in terms of the science.”
In March, the company graduated from Silicon Valley business incubator Y Combinator, yet it’s also working with partners like IBM and Dalhousie University. Heifets hopes to leverage algorithms that learn as they operate, to predict which drugs interact with biological molecules and from that, to develop new treatments.
Molecules, especially proteins, create complicated structures that interact with one another like giant, 3-D puzzle pieces. The teeth of any piece can change depending on temperature, salinity, nearby molecules and a host of other factors.
Heifets makes the analogy that understanding molecular interactions can be like trying to tune your radio to the right station, except you have millions of knobs instead of one. For decades, scientists have hoped that computers could someday tackle this sea of variables. Only recently have the advances in computer processing shown that dream could come true.
Biopharma companies screen millions of molecules all the time to see which ones will inhibit a given disease target, and only a small minority of them do, says Professor Michael Goldberg of Harvard Medical School’s Dana Farber Cancer Institute.
“The ability to increase the efficiency with which drug screens are performed would be greatly desirable,” Goldberg says.
Recently, Atomwise used an IBM supercomputer to screen 7,000 drugs that might be effective in treating Ebola. During a pandemic, anything that can save time will inevitably save lives. The company found a short list of potential drugs — not previously thought useful for the disease — and is now moving to further tests.
In a recent hunt for drugs that would inhibit two proteins believed responsible for spreading multiple sclerosis, Atomwise’s software chewed through 8.2 million molecules and ranked them for which was most likely to affect the targeted proteins. Lab researchers then tested the top 50 molecules and found that nine of them bound to the protein, making them great suspects for further experimentation. (Note: Just because it binds to the protein doesn’t mean it’s useful in treating the disease.)
That quick triage through millions of molecules is where Heifets hopes Atomwise can prove most useful. But Heifets acknowledges that the four-person company has a long way to go, to prove its approach works.
“There’s a lot of skepticism,” he says. “People have been trying to solve this problem for years.”
Heifets had been working on high-performance data processing at IBM before co-founding Atomwise in 2012. With increased investment dollars available for leveraging computers to recognize language and imagery — and process that data in useful ways — Heifets is making the bet that computing has finally hit that tipping point where it can predict molecular interactions.
The company is now working with its third software iteration. To align interests with potential partners, the company does the analysis work for free. Atomwise doesn’t operate labs, only software, and will take a royalty if their partners find a drug that generates revenue.
“The kind of work that we’re doing takes a very long time to build out,” Heifets says.