In its quest to advance AI, Facebook is leaning into medical research

Facebook’s AI Research lab (FAIR) is trying to teach machines how to think like humans. That means being able to pick up a skill or information and use it to create something entirely new. Consider the act of cooking.

“When we learn how to cook, we first learn a few simple recipes, and then we can recombine them into more sophisticated dishes,” says David Lopez-Paz, a research scientist at Facebook’s AI Research lab in Paris. You may not know how to make gravy, but if you know how to make a roux, you can probably figure it out. “Machines are not there yet.”

This concept is known as “compositional learning.” To teach machines how to learn more like people, Facebook’s AI team has increasingly put its resources to use in the field of medicine, which has an unlimited array of complex problems that need solving. Its most recent work includes a collaboration with a lab in Germany called Helmholtz Zentrum München, which is researching how it can make medicine more personalized. Together, the two groups have come up with an artificial intelligence model that can predict, to varying degrees of efficacy, how combinations of treatments, such as drugs and gene therapy, can impact an individual cell. The hope is that this experimental open-source model will help researchers learn how to tailor treatments to patients based on how illness manifests on a cellular level.

For Facebook, the exercise offers yet another opportunity to refine its artificial intelligence. Lopez-Paz says what attracted him to this project was the rich data set and a need for combinatorial analysis that would challenge FAIR’s machines to learn in a compositional way.

“We’re interested in advancing research in artificial intelligence, and in order to do that, we’re always looking for challenging high-impact problems,” says Lopez-Paz.

Lopez-Paz began working with Helmholtz Zentrum München and researcher Fabian Thies two years ago, after being introduced through a mutual connection. Thies studies individual cells, a field known as single-cell genomics, which seeks to advance human health through decoding single cells.

“What we are doing is essentially trying to understand how cells make decisions,” says Thies.

It may seem counterintuitive, but one cell can tell scientists a lot about a whole person’s health. Cancer, for instance, can start with a mutated gene in a single cell that multiplies. Scientists believe that having deeper knowledge about an individual cell or cells that go rogue will help them devise more appropriate treatment regimens. Traditionally, researchers have analyzed cells in large groups to understand how they work. But with recent technological advancements, it’s become easier to look at the makeup of an individual cell.

Since 2015, scientists have been amassing data on individual cells through an effort called The Cell Atlas. Thies, Lopez-Paz, and a multidisciplinary team of researchers designed their AI model to sit on top of this data set and others like it. The goal is to help put these large data sets to work.

The model attempts to tell researchers how combinations of therapeutics at specific doses will impact a cell. It does this for drugs and various other therapies, including newer CRISPR-based gene editing. For this reason, not all of its predictions are equally accurate.

For instance, CRISPR is just going through clinical trials as a therapy for sickle-cell anemia, where it edits out the offending piece of genetic code that causes the disease. The Facebook AI model might attempt to calculate the impact of using both a CRISPR-based therapy and a secondary drug on a patient. But because CRISPR is such a new technology, there just isn’t enough data (yet) to understand how a CRISPR edit might impact a cell, especially in combination with another therapy.

Still, Thies says, even though these predictions are based on limited data, they still give scientists ample starting points for further research.

“You need to have models to guide where you go,” says Thies. “I think potentially quite a bunch of new research directions could be taken from this, which is super exciting.”


Leave a Reply

Your email address will not be published. Required fields are marked *