Atomwise inks China deal as list of AI collaborations lengthens
The long list of major AI biopharma collaborations has gotten longer as one of the first artificial intelligence startups has inked its first deal in China.
San Francisco-based startup Atomwise has signed an agreement to develop targeted drugs with Hansoh Pharmaceuticals, a deal that could ultimately be worth up to $1.5 billion. Hansoh is flush with cash after a $1 billion IPO on the Hong Kong exchange in June.
The promise of machine learning to speed up preclinical work and save developers millions of dollars has led to a string of new collaborations between a handful of software startups and some of the biggest drug developers and research institutions including Merck, AstraZeneca, J&J, Bristol-Myers Squibb, Pfizer and Duke University School of Medicine. They’ve agreed to work on research ranging from oncology to chronic disease.
Part of the swarm likely comes from the hype that periodically surrounds a new technology — and few words are buzzier right now in both tech and popular culture than “artificial intellgience” and “machine learning” — and that has concerned some key figures in pharmaceutical development. But although it’s too early for the AI platforms to have brought a drug to market, early studies have indicated there could be something beneath the buzz. That includes last week’s landmark study from Insilico in Nature Biotechnology, in which over 21 days the company found six molecules that could be potential treatments for fibrosis.
At its most basic, artificial intelligence works like this: You feed an AI system a vast number of, say, images of a cow and images not of a cow, and you tell it which is which. With each image of a cow and not-cow, the AI develops a more and more refined set of criteria for what constitutes a cow (even if that criteria is far different from what a human might give). Pretty soon it can very accurately recognize whether a new picture has a cow or not. You can also do this with, say, an image of your mom. It’s how your iPhone’s facial recognition works.
And you can do this with a molecule.
Atomwise works by what’s called “virtual screening,” meaning it uses its AI system to rapidly search databases for molecules that resemble what its partners are looking for. Its June partnership with Ukraine-based Etamine, the world’s largest chemical supplier, gives it access to a database of billions of compounds to scan. Atomwise can scan 10-20 million per day, up from conventional computer methods that cap out at about 100,000. This latest deal with Hansoh will see the company design and discover drugs for 11 undisclosed target proteins.
However, the Insilico study that grabbed headlines was for a slightly different form of AI.
This newer AI, only put forth in 2014, goes further. Rather than recognizing a face, it can imagine a face (or, say, art). The idea Insilico is betting on and getting close to proving is that if it can imagine a face, it can imagine a drug. Accordingly, these are called “generative” networks, as opposed to the “convolutional” ones Atomwise uses.
We’ll use cows again for the model. These new AIs actually consist of two systems. Loaded with data, the “generative” one attempts to come up with an image of a cow. Then a second one, which is called the “discriminator” and works likes the tech described above, tells the generative one if it got a cow or not. The generator learns from the discriminator, which learns from the vast store of uploaded information. You have a learning feedback loop that should eventually gets you a brand new pretty picture of a cow.
In the Insilico study, they were searching for a new tyrosine kinase inhibitor for discoidin domain receptor 1 (DDR1). The system was taught all DDRI literature, a larger set of kinase inhibitors, databases of medicinally active structures and a database of structures that have already been patented. The result? 30,000 candidate structures, which the company then whittled down to 40. They produced 6 of them in the lab, tested 2 of them on cells and one on mice.
Prominent science writer and noted skeptic of biotech AI hype Derek Lowe dampened the exuberant headlines, noting the DDRI is already well researched (creating an ideal sample size to train the neural networks), the discoveries weren’t drugs but possible drug targets, and generalizing these techniques to other drug areas will take years and lots of cash. This accords with a consensus view of the trial as a proof-of-concept. Still, he found it one of the most interesting papers he had read on virtual screening.
“The good news, though, is that there is no reason that virtual screening can’t do great things, eventually,” he wrote in his blog, In the Pipeline. “We just have to get a lot better at it than we are now, and that’s as true as it was when I first heard about it in the mid-1980s.”