When Miami-based physicist Neil Johnson heard about the Orlando nightclub massacre he felt a particular chill, and not because it had happened in his home state. “The first thing I thought was, ‘Oh, no, I wonder if this person is in our database.’”

It’s a painful coincidence for Johnson that his new paper about the ecosystem of online pro-ISIS groups is being published Thursday, just five days after professed ISIS follower Omar Mateen murdered at least 49 people and wounded another 53 in the worst mass shooting in U.S. history. Johnson leads an interdisciplinary team at the University of Miami that studies complexity in real-world systems. In the first eight months of 2015 Johnson and his colleagues uncovered 196 self-organized online groups of apparent pro-ISIS followers—a total of 108,086 individuals—consuming and sharing the most extreme ISIS propaganda, including beheading videos on social media. Was Mateen, who died in the Orlando attack, one of them? “Absolutely it’s possible,” Johnson says. “It’s possible that someone he knows is in there, it’s possible that the San Bernardino people are in there. All of this is possible.”

Johnson says the paper’s key point is that it would be more effective for agencies combating ISIS to shift their focus from millions of individual followers worldwide to these self-organizing groups, “of which there will typically be only a few hundred.” Breaking up such clusters of hardcore followers while they are relatively small, he says, can prevent the development of larger, potentially more dangerous ones. (Johnson says the groups his team studied, which the paper refers to as aggregates, averaged around 500 followers.)

As described in the June 17 Science, Johnson and his team have come up with a tool for analyzing the “online ecology of adversarial aggregates,” looking at things such as how the groups grow and shrink over time and how individual followers navigate among them. The tool, Johnson explains, is a precise mathematical model that correctly reproduces patterns observed in online pro-ISIS activity and can be used to assess future risks and possible interventions, as well as exploring “what if” scenarios.

An aggregate can behave like an ever-changing digital organism with a life of its own, complete with a kind of survival instinct. As the paper puts it, “…the data reveal that pro-ISIS aggregates exhibit the ability to collectively adapt in a way that can extend their lifetime and increase their maximal size.” Their online survival strategies include name changes, flipping from visibility (open access) to invisibility (current followers only), and reincarnation, in which the aggregate disappears completely from social media then “reemerges at a later time with another identity but with most (greater than 60 percent) of the same followers.”

For their analysis, Johnson and his fellow researchers chose the Russian social network VKontake, or VK, whose 350 million users make it Europe’s largest social network. The paper says one reason for this choice was that pro-ISIS pages are not immediately shut down on VK the way they are on Facebook.

In their search Johnson and colleagues used hashtags such as #ISIS and #Caliphate in multiple languages. They ignored pages with casual chatter about ISIS, such as mentions of news reports, and zeroed in on those that explicitly expressed support for ISIS, published ISIS-related news or propaganda, called for violent jihad in the name of ISIS and so on. As the paper points out, the ad hoc groups who gather at such pages, “who likely have never met, do not know each other, and do not live in the same city or country,” have to be adaptable because they are under constant predatory pressure from antiterror agencies, Web site monitors and individual hackers seeking to shut them down.

How scary is it that researchers found 196 extreme pro-ISIS groups with more than 100,000 followers in just six months of searching? Considering the radical nature of the content circulating among the pro-ISIS aggregates identified by Johnson’s team, the vast range of people engaging within them is really surprising, says Paul Gill, a terrorism researcher at University College London who has collaborated with Johnson on other studies. He adds that those numbers become less startling when put into context, however. The 108,086 figure, for example, represents just 0.03 percent of total VK users. It is also quick and relatively easy for an individual to engage with one of the groups. “By no means is it a pure indicator of how many people are actively radicalized,” he says. “The fact that those engaged in ISIS's violent activities are far fewer in number is testament to this.”

As for the power of his mathematical model to predict violent pro-ISIS activity, Johnson says, “It’s never good as a scientist to start predicting things.” Nevertheless, the paper does suggest that tracking the behavior and proliferation of pro-ISIS aggregates “can indicate an alignment of favorable conditions” for an act of terror without relying on online chatter about likely dates for the action.

J. M. Berger, co-author most recently of ISIS: The State of Terror, thinks Johnson’s ideas about degrading ISIS’s online networks show promise. “But the paper only looked at how users interacted with the aggregates and didn't factor how users interacted with each other,” says Berger, who is a Fellow in The George Washington University’s Program on Extremism and was not involved in Johnson’s research. “I’m confident there is value to this approach but I don't think it replaces analysis and countertactics on follower-to-follower relationships.” Still, he adds, “I’d like to see more research in this general area.”

For Johnson, the unexpected issue at hand is Omar Mateen. Johnson wants to look for him in his database. To do so, he and his team would have to collect as much information as possible about Mateen from the media, encode it and use it to search through all 108,086 apparent pro-ISIS followers in the database. They would look for subsets of followers from the Orlando and Port Saint Lucie, Fla., areas that might fit the bill. They would then have to narrow it down further—and then, just maybe, be able to assign a likelihood that user X was Mateen.