FDA authorizes new test, built with machine learning, to detect past Covid-19 infections

The Food and Drug Administration on Friday issued an emergency authorization for a new test to detect Covid-19 infections — one that stands apart from the hundreds already authorized.

Unlike tests that detect bits of SARS-CoV-2 or antibodies to it, the new test, called T-Detect COVID, looks for signals of past infections in the body’s adaptive immune system — in particular, the T cells that help the body remember what its viral enemies look like. Developed by Seattle-based Adaptive Biotechnologies, it is the first test of its kind.

Adaptive’s approach involves mapping antigens to their matching receptors on the surface of T cells. They and other researchers had already shown that the cast of T cells floating around in an individual’s blood reflects the diseases they’ve encountered, in many cases years later. The next step is trying to unlock that information to help diagnose those past infections.

That challenge is extremely data-heavy. “When you think about a patient level, we’re looking at 300,000 to 400,000 T cells, on average,” said Lance Baldo, Adaptive’s chief medical officer. “When you look at a population level we’re looking at hundreds of millions, and then ultimately billions of T cells. So it ends up being a web scale problem.”

Enter Microsoft. In 2018, Adaptive developed a partnership with its tech giant neighbor to build up the cloud infrastructure and machine learning models necessary to deal with those reams of data — in particular, to build a complete map of which T cells bind to which antigens.

“Microsoft wants and wanted to get into healthcare,” said Baldo. “Adaptive needed expertise in cloud computing, machine learning, and AI. So it was a pretty ideal fit.” Teams from both companies worked together one day a week, in Adaptive’s Seattle office or Microsoft’s in Redmond.

When the virus started gaining speed, they quickly pivoted a large part of that team to work on Covid-19. By June, they were able to access blood samples from people who had been infected with the coronavirus, and sequence the genomes of the T cell receptors therein. Then they could compare that dataset to their control group — the database of T cell receptor sequences they had been working on for years — and within two months, they had collected enough data to publish their first results.

The machine learning models necessary to develop the T-Detect test, in the end, were relatively straightforward. “To me that’s actually a big plus,” said Jonathan Carlson, senior director of immunomics at Microsoft and leader of the partnership with Adaptive. “It’s a viral infection and drives a raging T cell response, and it turns out you can find the exact same T cell receptors in many people. And that allows you to use a pretty simple statistical approach.” The test reported a sensitivity of 97.1% and 100% specificity.

The EUA issued by the FDA is a reflection of that first approach — but it’s not the end of the test’s evolution. “When we file with the FDA we do something called ‘locking the classifier,’” said Baldo, the algorithm that determines whether a blood sample’s T cell receptors say “Yes Covid” or “No Covid.”

Those T cell responses, though, can vary depending on the version of a virus you’re exposed to.

“We’ve already discussed this with the FDA,” said Baldo. “You’ve got mutations and other variants coming.” So Adaptive and Microsoft are continuing to improve the classifier. “The models get better frequently,” said Carlson. “Weekly, monthly.” The question that remains, then, is, “when is better enough? That’s where Adaptive is really spending a lot of time thinking.”

At some point, when the test reaches a new threshold of sensitivity and specificity, they plan to file a second version of the test for the FDA’s review.

That’s just one of Adaptive’s three areas of focus in the coming months, said Baldo. “One pillar is to improve the current algorithm and make sure we continue to have a great test as the virus continues to mutate,” he said. The second is to turn the company’s T cell expertise toward other questions surrounding Covid-19, including the impacts of long Covid and the efficacy and durability of the immune response elicited by different vaccines.

The third is to continue its work on other diagnostics, for conditions like celiac disease and multiple sclerosis. Before the pandemic, the company was focused on developing a proof of concept diagnostic for Lyme disease, which it announced in November 2019.

That distributed focus will force the company to continue building not just its biological capabilities, but its machine learning approaches. While its approach to Covid-19 screening is  relatively straightforward, said Carlson, “I don’t expect that will work for every disease.”