SAN FRANCISCO — It took only a single, one-hour meeting in a Starbucks to convince venture capitalist Robert Nelsen that he was going to invest in the company being built by machine-learning whiz Daphne Koller. “She’s just an amazing force of nature,” said Nelsen, who invests for ARCH Venture Partners.

Koller’s company, Insitro, in a matter of months raised over $100 million from big-name investors including ARCH, Foresite Capital, Andreessen Horowitz, and the firm that manages Jeff Bezos’ personal VC investments. (Koller refers to Bezos by his first name.) Now, it’s time for the next step. On Tuesday, the giant drug maker Gilead Sciences said it would pay Insitro $15 million — and up to $1 billion more down the line if it meets certain goals — for help in developing drugs to treat a common liver disease. It is the South San Francisco-based company’s first deal with a drug maker, and offers a test of whether artificial intelligence can remake drug development.

The very nature of the financials emphasize the exciting but risky reality: The technology has great promise, but nothing is guaranteed.

advertisement

“I’ve been working at the boundary of these disciplines — machine learning and the life sciences — for about 20 years,” Koller said. “But I’ve constantly been stymied by the fact that there hasn’t been enough data to apply machine learning in the way that you’d like it to be applied.”

She’s betting that her approach, focused on generating datasets aimed at solving intractable problems, will get around this problem.

In spite of all the excitement around Koller’s approach, the biology of disease has thwarted many past efforts to tame it with new technology. It remains to be seen whether investments in artificial intelligence are actually saving money. And there are cultural barriers, too, in biopharma, where the old-school drug developers who control the levers of power haven’t quite figured out to do with computational biologists and AI whizzes.

Koller, 50, earned a master’s degree from Hebrew University when she was 18 and a MacArthur “genius grant” at 36.

She was always interested in biology. In the early 2000s she tried to use data sets from the first DNA microarrays, which can measure how genes are working inside the body, and from the first sequence of the human genome, the collection of human DNA.

More recently, Koller founded the online education company Coursera, and built out computing at Calico, the Alphabet startup focused on researching the biology of aging.

But Koller, who also had taught computer science at Stanford University, soon decided she wanted to focus more broadly on what machine learning could do in drug discovery. She started raising money for Insitro at the J.P. Morgan Healthcare Conference in January 2018.

What is machine learning? For Insitro’s purposes, Koller uses this definition: “It’s the task of building machine algorithms that, given access to a very large dataset, are able to provide prediction tasks with very high performance.” The prediction being made in this case is what will happen when a chemical, a potential medicine, is added to the human body. What will it do?

There are many new tools, Koller said, that make this a good time to try to use computers to make such predictions. Insitro can take data from clinical trials of thousands of patients, but also from experiments on individual cells, which can be altered with CRISPR, the gene editing technology, to answer specific questions about what happens when one gene is turned on or off.

With Gilead, the goal will be to use this approach to understand non-alcoholic steatohepatitis, or NASH, the liver disease many companies believe could be a big market. The goal is to identify five proteins that could be targets for new drugs, and then try to move them forward. In addition to the $15 million, Gilead will pay Insitro $35 million more in the short term if the collaboration meets certain near-term goals. If drugs based on those five targets move through the clinic and to the market, Insitro could be paid $200 million for each of them.

Gilead has been working on NASH since 2011. It’s bet big on the condition in the hopes of creating a new source of earnings growth as sales of its other blockbusters wane. But earlier this year, Gilead’s lead NASH drug, called selonsertib, failed in the first late-stage clinical trial testing a medicine for the condition. Results are expected later this year for a second Phase 3 study of that same drug, this time in patients with earlier stage NASH.

“We need to come up with unique ways of increasing our confidence in targets” for NASH, said Mani Subramanian, Gilead’s senior vice president for liver diseases. Subramania said he’s “quite clear” what the deal brings. “It’s mainly Daphne, right?” Subramanian said.

Insitro’s deal with Gilead ”really bears out our investment thesis that pharma is excited and ready to embrace machine learning,” said Vijay Pande, an Andreessen Horowitz VC and one of Insitro’s investors. “And this deal is the first financial demonstration of that.”

One indicator of how well the approach is working, he predicted, is whether this is the first of many deals that Insitro and its competitors ink with big drug developers.

Insitro, which made its public debut last May in a post on the blogging site Medium, now employs 30 people. The goal, according to Koller, is to “figure out what the really important problems are — and then let’s figure out what data sets need to be generated in order to make those problems solvable.” Sure, that will cost money, Koller acknowledged. “But think about the cost of a single failed clinical trial, and how expensive that is, compared to generating the right data to begin with.” Maybe, she said, the right machines can help.

Leave a Comment

Please enter your name.
Please enter a comment.

Sign up for our Daily Recap newsletter

A roundup of STAT’s top stories of the day in science and medicine

Privacy Policy