Finding new drugs is hard. Sometimes we don’t even know how a disease works, and drug tests in animals don’t always go the same as in humans. Drugs can even behave very differently from person to person. And because companies don’t tend to share data with one another about failures, we can’t learn from each other and the larger data pool. The whole process is extremely expensive, and the cost is ultimately borne by us, the consumers.
But what if drug discovery could go from finding a needle in a haystack to making small piles of needles?
That’s what Abe Heifets wants to do. Abe is the CEO and co-founder of Atomwise, a 50-person biotech startup based in San Francisco. And he thinks he’ll find the next blockbuster drug using a technology you carry in your own pocket: neural networks.
What’s a neural network anyway?
“If you’ve ever used Siri or Alexa, or uploaded a photo to Facebook, then you’ve used neural networks,” says Abe. A neural network is a set of computer instructions (algorithms) that resemble human brain function where it comes to recognizing patterns and clusters in data. The data can be images, sound, text, or other information — like molecules at the atomic level.
Consider three kinds of data. Speech is one-dimensional data: a single audio signal varying over time. Images are two-dimensional data because the pixel color depends on both the x coordinate and the y coordinate. Atoms are three-dimensional because they have x, y, and z coordinates: height, width, and depth.
Speech-to-text software uses 1D neural networks. The vision systems of self-driving cars use 2D neural networks. Atomwise’s insight was to develop a 3D neural network that could “see” and understand molecules in space in the same way a self-driving car sees the world. “Instead of red, green, and blue color channels at every grid point, we have carbon, oxygen, sulfur, and nitrogen channels,” he says.
How does it work?
“Let’s say you’re a professor at UC San Francisco,” says Abe, “and you think that if you can just block protein XYZ, you can cure Alzheimer’s… That’s a great paper you can publish in Nature, but you can’t help a patient with that. You actually need the drug.”
That’s where Atomwise comes in. “You say, ‘Give me a molecule for XYZ.’ And it can be on Alzheimer’s, cancer, malaria, whatever you want…” Atomwise’s AI system searches for the best small molecules among millions and millions.
By default, Atomwise starts with a chemical library of 10 million small molecules. From this pool, Atomwise’s algorithms sift through and identify the most promising molecules — 7% of 1% of 1%, just a tiny sliver. They then order them inexpensively from a third-party manufacturer and ship them to their customer on a 96-well plate. From there, Atomwise’s customers can test the molecules and see how they work in their systems.
How big is big? Ultra-economies of scale
For context, big pharma companies typically have 3 to 5 million small molecules in their entire collections. So Atomwise can double that.
“A decade or maybe 15 years ago, you and I could buy a million molecules off-the-shelf. Last year, we could buy 300 million. This year it’s 11 billion molecules that you and I can order for 100 bucks a pop and get shipped to us in six weeks,” Abe told me. He thinks next year it’ll be 100 billion.
Atomwise’s business model is akin to Dell in the 90’s: You custom-design your computer from any possible combination of peripherals and memory, enter your credit card info, and press submit. Dell goes out and buys the peripherals and builds only the computers it needs, and assembles the parts on-demand.
“Chemistry has undergone the same transformation in the last decade,” says Abe, where chemical manufacturers are storing all the building blocks and making chemicals on-demand. “What they’re selling you is the Cartesian product of how to put those together.”
With an important difference in Atomwise’s case: They are also selling a highly intelligent selection of chemical products, based on customers’ needs.
“This is virtual chemistry, on-demand chemistry, right?” Abe says. “We’ve shifted from a world of scarcity in chemistry, to a world of abundance.”
Abe likens the space to other neural network we use all the time: “Netflix has way more movies than you could ever watch, and YouTube has way more cat videos than you can ever see, right? But how do we get a new cat video, one that you feel like watching right now? These are questions of filtering, matching, searching. These are AI questions.”
The origins of a good idea
Abe studied computer science at Cornell, where he worked on the AI system for soccer-playing robots (his team won the RoboCup World Champion in 2001). From there, he worked at an IBM research center in Boston. “I worked there on what today we would probably call Big Data,” recalls Abe, “but at the time, we didn’t have that phrase, so we called it high performance data processing.”
The work was rewarding, but Abe wanted to do more. And that time, he got interested in medicine (“Everyone needs a hobby,” he says sheepishly). He started taking chemistry classes at Harvard, where the mixing of chemicals “felt very grainy” to him compared to computer science.
Abe decided to go back for his PhD and landed in a computational biology group at the University of Toronto. Abe’s lab shared a coffee pot with the machine learning group of Geoffrey Hinton — inventor of deep neural networks. That’s also where he met his Atomwise co-founder and CTO, Izhar “Izzy” Wallach. Izzy had been writing structural biology algorithms for a small pharma company. Combined with Abe’s work on big data and the influence of deep neural networks being created in the lab next door, and Atomwise was a natural fusion of it all.
Anything but academic
Applying this thinking is not a mere academic exercise, and investors know it. Atomwise was first selected to join Y Combinator’s Winter 2015 class. By the end of Y Combinator, several well-known venture capitalists were ready to invest in the promise of applying neural nets to drug discovery, including DCVC (where I am an operating partner), Khosla Ventures, Threshold, and Tim Draper. By March 2018, Atomwise closed its $45 million Series A round.
And the technology is maturing nicely, Atomwise just reported the results of a collaboration with Stanford University and the Mayo Clinic that used Atomwise’s technology as a kind of AI virtual drug screen to identify a potential treatment for Parkinson’s disease. It’s also a proof-of-concept for making personalized medicine for this disease quickly and cheaply.
“We’ve been running the world’s largest application of machine learning to drug discovery in history,” says Abe. He recently presented those project results to the American Chemical Society. “This is a project that we’ve been running where we have over 250 projects with hundreds of universities in 36 countries,” he says. “We work on every major disease, we work on every protein class.”
Today, Atomwise is working with a number of big and small pharma companies, particularly around cancer treatments. One partnership, with Hansoh Pharma, marks the largest China-US collaboration for AI drug discovery and could amount to $1.5 billion if all milestones are achieved.
What 21st-century pharma companies will look like
As Old Pharma outsources AI drug discovery and more, Abe thinks it will change the face of pharma companies. “It probably doesn’t look like four brick walls with everything happening inside. It probably looks more like a series of alliances that come together.”
If you’re a small biotech with some deep insight into biology, are you going to spin up your own mouse testing, sales force, and chemical manufacturing? No, says Abe. “You want to partner with Big Pharma, who has those kinds of relationships already in place. And so it’s a question of teamwork.”
Companies like Atomwise are a great example of how the convergence of tech and bio is creating valuable and important new consumer possibilities that were previously off limits, while also disrupting existing value chains in huge industries like pharma.
If your company could biomanufacture any chemical imaginable, what would it be?
Acknowledgement: Thank you to Kevin Costa for additional research and reporting in this post.
Originally published on Forbes: https://www.forbes.com/sites/johncumbers/2019/10/26/how-neural-nets-will-personalize-medicine/1