This tantalizing possibility comes from a study reported recently in the journal Cell. In this study, researchers took blood samples from people with mild to severe COVID-19 and analyzed them for nearly 2,000 proteins and metabolites [1]. Their detailed analyses turned up hundreds of molecular changes in blood that differentiated milder COVID-19 symptoms from more severe illness. What’s more, they found that they could train a computer to use the most informative of the proteins and predict the disease severity with a high degree of accuracy.

The findings come from the lab of Tiannan Guo, Westlake University, Zhejiang Province, China. His team recognized that, while we’ve learned a lot about the clinical symptoms of COVID-19 and the spread of the illness around the world, much less is known about the condition’s underlying molecular features. It also remains mysterious what distinguishes the 80 percent of symptomatic infected people who recover with little to no need for medical care from the other 20 percent, who suffer from much more serious illness, including respiratory distress requiring oxygen or even more significant medical interventions.

In search of clues, Guo and colleagues analyzed hundreds of molecular changes in blood samples collected from 53 healthy people and 46 people with COVID-19, including 21 with severe disease involving respiratory distress and decreased blood-oxygen levels. Their studies turned up more than 470 proteins and metabolites that differed in people with COVID-19 compared to healthy people. Of those, levels of about 300 were associated with disease severity.

Further analysis revealed that the majority of proteins and metabolites on the list are associated with the suppression or dysregulation of one of three biological processes. Two processes are related to the immune system, including early immune responses and the function of particular scavenging immune cells called macrophages. The third relates to the function of platelets, which are sticky, disc-shaped cell fragments that play an essential role in blood clotting. Such biological insights might help pave the way for potentially effective new ways to treat COVID-19 down the road.

Next, the researchers turned to “machine learning” to explore the possibility that such molecular changes also might be used to predict mild versus severe COVID-19. Machine learning involves the use of computers to discern patterns, or molecular signatures, in large data sets that a human being couldn’t readily pick out. In this case, the question was whether the computer could “learn” to tell the difference between mild and severe COVID-19 based on molecular data alone.

Their analyses showed that a computer, once trained, could differentiate mild and severe COVID-19 based on just 22 proteins and 7 metabolites. Their model correctly classified all but one person in the original training set, for an accuracy of about 94 percent. And importantly, in further prospective validation tests, they confirmed that this model accurately identified mild versus severe COVID-19 in most cases.

While these findings are certainly encouraging, there’s much more work to do. It will be important to explore these molecular signatures in many more people. It also will be critical to find out how early in the course of the disease such telltale signatures arise. While we await those answers, I find encouragement in all that we’re learning—and will continue to learn—about COVID-19 each day.

Reference:

[1] Proteomic and metabolomic characterization of COVID-19 patient sera . Shen B et al. Cell. 28 May 2020. [Epub ahead of publication]