When two inpatients at a Canadian hospital both tested positive for a gene capable of defying carbapenems, one of medicine’s last-resort antibiotics, it looked like an outbreak. It wasn’t.
The gene in question was VIM, a well-known resistance gene that infection control teams dread. Finding it once is concerning. Finding it twice, in the same facility, days apart, raises an immediate question: did one patient infect the other? A simple yes-or-no from a resistance screen wasn’t going to cut it.
This gap between detecting a resistance gene and understanding its context is one that does not get talked about enough. Scientists have become quite good at detecting antimicrobial resistance genes, but detection alone tells us only that they are present. It does not tell us where they came from, how they are moving or whether two cases are connected. For that, we need interpretation, and interpretation is where the real work begins.
Peeling back the layers
For those two Canadian patients, investigations started where most do: with the gene itself. Both patients carried VIM, but a closer look revealed different gene variants: VIM-78 and VIM-4, separated by a single nucleotide. They are easy to mistake for the same gene if you aren’t looking carefully.
But uncovering the specific gene identity was only the first layer. The team reconstructed the plasmids: the mobile DNA elements that carry resistance genes between bacteria. One patient’s gene sat on plasmid inside a Proteus mirabilis bacterium, likely acquired during travel abroad. The other patient’s gene appeared on two entirely different plasmids, in two different bacterial species. This means there were three different plasmids across three different organisms — which is not what a transmission event looks like. Even the integrons, the genetic structures that capture and shuffle groups of resistance genes, had diverged structurally. Every layer of analysis pointed the same way: these were independent, unrelated events. No outbreak. The genetic structures that capture and shuffle groups of resistance genes, had diverged structurally. Every layer of analysis pointed the same way: these were independent, unrelated events. No outbreak.
Each step — gene comparison, plasmid reconstruction, integron analysis — added clarity that the initial detection alone could not provide.
What gets missed when the tools aren’t there
That kind of layered interpretation requires resources, expertise and infrastructure. Not every setting has them. In 2016, a patient with leukemia at a Nigerian teaching hospital developed a bloodstream infection. Conventional laboratory methods identified the bacterium as Klebsiella pneumoniae and found it susceptible to most antibiotics. Colistin resistance (resistance to a true last-resort medicine) wasn’t even tested. It rarely is. Most labs only test for the antibiotics doctors are likely to prescribe first and colistin, being a last- resort antibiotic, doesn’t usually make that list.
Years later, retrospective genomic analyses told a different story. The bacterium wasn’t Klebsiella pneumoniae at all, it was a closely related species, Klebsiella quasipneumoniae, which conventional identification tools struggle to distinguish because these bacteria share similar genotype profiles to Klebsiella pneumoniae. And it carried mcr-10, a plasmid-borne colistin resistance gene that hadn’t been formally described until 2020, four years after the sample was collected. The full picture only emerged through the kind of genomic analysis that wasn’t available at the point of care
Two very different cases. But the same underlying problem: the resistance gene was found, or in the Nigerian case, the resistance genes that were found told an incomplete story. What happened after detection is what mattered, and in both instances, only deeper genomic analysis could provide answers.
This is where computation comes in
The interpretive steps that unravelled the VIM case and the retrospective analysis that recharacterized the Nigerian isolate are not inherently trivial. However, these follow a logic that can be encoded, automated and applied across large datasets. The challenge is building systems that move beyond simple detection and into the kind of layered, contextual analysis that distinguishes a real outbreak from a coincidence, or a misidentified pathogen from an accurate diagnosis.
That’s the space I work in. Not replacing the expertise behind studies like these, but making their kind of reasoning applicable at scale: across thousands of genomes, not just two isolates at a time. In my upcoming OpenThink blog posts, I’ll dive into what that looks like in practice.
Detection is the starting line. Interpretation is the race. And computation may be the only way to run it at the pace the data requires.
