This biotech-med tech crossover is working on AI approaches to cancer immunotherapies.
President and CEO: Andrew Allen, M.D., Ph.D.
Based: San Francisco Bay Area and Cambridge, Massachusetts
Clinical focus: The biotech’s plan is to find out which tumor-specific neoantigens (TSNA) a patient has and then use an algorithm to figure out which drug candidate is most likely to activate tumor-specific T-cells, an deliver an immunotherapy to target them.
The scoop: This biotech got off a deeply impressive series A back in 2015 and is managed by some big names out of Clovis Oncology, Pfizer and most recently Genentech. Its aim? To boost the odds of success in treating a number of solid tumors with new artificial intelligence (AI) predictor tech alongside a cancer vaccine.
What makes Gritstone fierce: You want a big money raise? How does a $102 million series A grab you, with a series B this September worth $92.7 million?
You want big names? How about a chief that cofounded Clovis, Andrew Allen? Or a CTO and tech side manager formerly serving as VP of Foundation Medicine, Roman Yelensky?
Furthermore, its CSO, Karin Jooss, is an ex-Pfizer cancer immunotherapy chief, and the company recently recruited a CMO from Genentech, Raphaël Rousseau. Put all this together, and you get an experienced dream team backed up by some serious cash.
But what's the goal? Allen, a highly driven Brit who said he doesn’t take on projects that won’t work, said Gritstone crystallized in 2015, and the company drew money sources early on.
“I’ve actually avoided immunotherapy in cancer for a long time, because it had just been such a graveyard," Allen says. "I’ve been at teams in the past working on this sort of platform, and I discovered that, in the main, we didn’t really know how it works. We saw that some patients responded, and that some had some pretty good responses, but across the board, we weren’t seeing that.”
Allen spent many frustrating years trying to understand which patients were responding and why, but could only see a 4% to 6% response rate, and only around half of those were actually cured. What he sorely needed was technology that could help answer his questions, but in his former role, this tech simply didn’t exist.
“It was grim,” he recalled, and so he stepped away from work in the field, turning instead to targeted therapies for more than a decade.
But then, immunotherapies came back with Yervoy (ipilimumab) and then the class of PD-1 antibodies from Bristol-Myers Squibb and Merck. At this point, Allen was thinking of moving on from Clovis, which he cofounded, to create a startup in immunotherapy, because “I want to cure cancer,” he said matter-of-factly.
“But the problem I had was that I couldn’t quite see a mechanism; there are lots of other people developing checkpoint [inhibitors] and to me, it looked a little like throwing darts and just seeing which ones would stick. That’s not a satisfying basis for a company," he says.
Arriving at T-cell therapy
Allen wanted a fundamental piece of science to exploit and develop, and at the time, he couldn’t see what that was. At the end of 2014, however, "the scales fell from my eyes and a paper was published, by what later became our cofounders, showing that patients responding to PD-1 appeared to do so because they were developing T-cell responses to tumor neoantigens.”
That was the “Aha” moment for Allen. If you believe in cancer immunotherapy, then today, it’s basically T-cell therapy—and T-cells recognize T-cell antigens, which Allen says aren't well understood. “That to me was the big idea,” he said. “If we can really ... build a comprehensive, best-in-class understanding of the nature of tumor antigens, then that is a company that is going to be highly relevant, and generate real insights and real revenue.”
That was the story venture capitalists bought into, but Allen said they realized very quickly that their goal was inherently a complex and tough problem to crack. Because tumors present short peptide fragments in the context of highly polymorphic molecules, it’s a major challenge to actually work out which mutation is relevant for an individual patient.
Then there's the second part of the equation. Of every one hundred mutations in a tumor, only one of them will actually make it as a neoantigen. To work, they must possess the right properties that allow a particularly mutated protein to be chopped up by the proteasome into peptides. Those peptides must then be transported into the right cell compartment, loaded up and later presented on the cell surface to passing T-cells. “It’s a pretty stringent set of biological processes and only 1% of mutations actually make it through,” Allen said.
“That was the first major insight. The second was that basically everyone in the field was using this public domain tool to try and do the same predictions, and that is a narrow tool that ignores most of the biology," he added. The medical literature shows that this approach and its predictive power comes in at just 3%, he said.
Using machine learning to predict mutations
“So, I’ve got these two problems,” says Allen: “A typical lung cancer patient has around 300 mutations, so then I’m trying to identify the three that are the right ones, and I’m trying to put them into a therapeutic. So, I needed to predict, with pretty high accuracy, but I was getting a 3% hit, which is woefully short.”
These were the challenges from day one for the startup. Their answer was to bring in Yelensky, the “perfect partner” for Allen on the tech side, back up with a host of “crazy smart machine-learning folks.” Allen said Yelensky and the team solved the problem by collecting large numbers of human tumors and doing the sequencing work to get the data, then perform a mass spectrometry analysis.
“So now, for a given specimen, I know I have the info I need," Allen said. "And we can create a deep model in between those datasets, and use these to help better predict what we need in the future.”
Mass spectrometry has been used for years, but Allen said the tech is now much better, and the company can produce huge amounts of data very quickly. “We now have 200 tumor samples from across the world and more than one million peptides in our database, and this is at a scale no one has ever done before. We have a lot of high-quality data to train our model on, and so our machine learning guys have been using this to train the prediction model.”
The result? With just the sequencing alone, the startup is predicting, not with a 3% positive value, but 50%. “If we put 10 antigens into a vaccine, then roughly every other one should be a good antigen that actually is found on the tumor cell surface. This is best-in-class, and an elegant solution, and brings in a machine learning approach to a complex biological problem.”
Investors: Versant Ventures, The Column Group, Clarus Ventures, Frazier Healthcare Partners, Redmile Group, Casdin Capital, Transformational Healthcare Opportunity, Lilly Asia Ventures, Google’s venture arm GV, Trinitas Capital (Beijing) and Alexandria Venture Investments.