[Updated 4/20/17 10:27 am. See below.] Artificial intelligence company Atomwise is offering a shortcut for as many as 100 university scientists who are searching for new drugs to fight disease. The San Francisco-based startup, which uses deep learning algorithms to ferret out drug candidates by sifting through masses of data, wants to send each researcher 72 compounds that might work.
Atomwise has two motives for launching its Artificial Intelligence Molecular Screen (AIMS) awards program. The first goal is to help scientists leap over a gap in government funding, which covers basic research on the molecular mechanisms behind illnesses, as well as clinical trials on experimental drugs, but doesn’t support the hunt for drugs to test, says former UC Berkeley researcher Han Lim, who manages academic partnerships for Atomwise.
The second objective for Atomwise is to further field-test its AI-enhanced drug screening system, AtomNet, on a broader class of research projects on numerous diseases.
“That could produce case studies that allow us to demonstrate what we can do,” Atomwise co-founder and chief operating officer Alexander Levy says.
Levy co-founded the company in 2012 with former colleagues at the University of Toronto, a hotbed of AI innovation. Their idea was to apply the techniques of AI—already proven powerful in the analysis of images and speech—to the complex mysteries of chemistry and biology.
Atomwise was a 2015 participant in Y Combinator’s accelerator program and raised $6 million in seed funding that year. It has collaborated with 27 research institutions and companies, including Stanford University, IBM, and Merck. About 14 of those projects are still ongoing, and the company is earning revenues, Levy says.
Atomwise’s virtual drug screening system is part of a long-running drive to find quicker ways to identify compounds that could block off-kilter processes on the molecular level that lead to illness.
Researchers looking for the causes behind a disease often have strong suspects in mind. The culprits are usually biologically active proteins in the body that are knocking essential functions out of whack.
While discovering a disease’s molecular cause does reveal it as a possible target for therapeutic drugs, that’s only part of the path toward a medical treatment. Scientists still need to identify compounds that might counteract the malign effects of the suspect protein. Finding those drug candidates is one of the most challenging tasks in medical research.
The compounds most likely to succeed are those that will bind with the target molecule. In the first attempts at automated mass drug screening, companies such as Exelixis did physical binding tests by exposing millions of compounds to the target.
Such empirical studies are now part of the mass of data that AI companies such as Atomwise and Numerate—along with big players like IBM and Google—can rifle through with deep learning software. Levy says the inputs for Atomwise’s analysis include chemical descriptions of the small molecule compounds it evaluates as drugs, and X-ray crystallography images of the target proteins—which can reveal their structure and possible binding sites. Atomwise also scans data on the known interactions between small molecules and targets.
Atomwise is inviting university researchers to send in a short application for the AIMS program by June 12, stating the disease they’re working on, their target molecule, and any particular requirements for their drug candidate. For example, if the drug must reach a target molecule in the brain, it would need to be able to cross the blood-brain barrier.
Atomwise will study the feasibility of each applicant’s project, and notify those chosen for the program in September.
Levy says Atomwise might screen as many as 10 million molecules to find the six dozen best compounds for a single scientist’s needs.
Lim and Levy acknowledge that some scientists might be reluctant to reveal the target molecule they think is causing a disease, if that causal link is an original discovery that might become extremely valuable if it leads to lucrative drugs. It’s up to the scientists to decide whether to identify their targets, Lim and Levy say, but the collaboration agreement that participants would sign includes confidentiality terms calling for non-disclosure of proprietary information on both sides.
Once on board, the researchers will each receive 72 compounds to test. The number 72 was chosen because researchers often test compounds by placing them in the wells, or depressions, in 96-well plates. The plates can be used on the lab benchtop or in automated testing. By providing 72 compounds, Atomwise leaves open the option for scientists to add other compounds in the extra wells. These might include control compounds whose responses in a certain test are already known, Levy says. The participating scientists are free to choose the tests they’ll use to evaluate the compounds sent by Atomwise.
“It’s possible some programs could lead to a successful invention,” Levy says. The collaboration agreement includes terms that govern any jointly developed intellectual property. If the scientist applies for a patent, Atomwise might have standing as a co-inventor, Levy says.
[This paragraph was amended to include further detail from Atomwise about its spending on AIMS.] Atomwise isn’t providing any cash to support the scientists’ work, but it might offer further consultation and in-kind support as the experimental results emerge. The company’s largest contribution is to narrow down the number of compounds each lab will test—a drug screening project beyond the means of a university researcher, Levy says. None of the costs of the AtomNet screening computations—including the labor of staff scientists—is counted in Levy’s estimate of the amount Atomwise will spend solely to buy, prepare, and ship thousands of compounds for testing by AIMS participants. On those costs alone, Levy says, “We expect to spend a million dollars on AIMS.”
Illustration courtesy of Atomwise.