I ended my last post on a limit. A resistance gene can pick up so many changes over generations that direct comparison of its DNA letters to the catalogue of known resistance genes stops recognizing it, even though the protein it builds still works and still defeats the antibiotic. It is a gene we know. We just cannot find it by matching letter to letter anymore.
So, for genes like these, we can turn to the family. A family of genes is a set of genes descended from a common ancestor. Over generations their letters may drift apart, but they go on building proteins that work the same way. These proteins hold the resemblance long after the DNA has lost it, which is why this comparison is made on proteins rather than on DNA letters.
The question, then, is not how closely one gene resembles a single catalogued sequence. It instead is whether its protein product carries the pattern that runs through the whole family.
A face in a crowd
Think about how you recognize your cousin. You do not hold up a photograph and check whether her face lines up point by point. If you did, you would fail constantly: new haircut, bad light three years older than the picture.
Yet you would pick her out of a crowded room in a second, because you are not matching an image. You are recognizing a family resemblance. You know which features run in the family and stay put, which ones change, and you weigh them differently without deciding to.
Notice what makes that possible. You can recognize your cousin because you know the family. Meet a stranger and the skill is useless; however good you are at faces.
Building a gene family portrait
This level of familiarity can be built for gene families. Last time, I described alignment as lining up two sequences, letter by letter. A gene family portrait starts the same way, except you do not line up two. You take the protein built by every confirmed member of the family and stack them all at once. Proteins are spelled out in letters too, twenty of them instead of four, and stacking lets you see which positions hold steady down the whole column and which vary. Then you turn that stack into a model. Scientists call it a hidden Markov model.
The model is a set of expectations, one for each position. Where every known member carries the same letter, that letter is near certain, and a sequence carrying anything else is heavily penalized. Where the members already vary, whatever a sequence carries barely counts. No single position decides whether a sequence belongs. Each one adds or subtracts evidence, and the sum of all of them gives a score.
Positions hold steady for a reason. Take the family blaNDM-1 belongs to, the gene behind the Nevada case. In every confirmed member, a handful of positions in the protein hold two atoms of zinc. Without that zinc, the protein cannot break down antibiotics. So, no working member has ever changed them.
Putting the portrait to work
To identify a new sequence, you run it through the model, which fits it to the portrait position by position and returns a score. That score is a measure of odds: how much better this sequence fits the family than it would by chance. A cut-off still applies to that score rather than to a count of matched letters.
A group of scientists in Sweden demonstrated this in 2017. They built a model from 20 confirmed members of the blaNDM-1 family and pointed it at public data: thousands of finished bacterial genomes and trillions of letters of DNA from river sediment, sewage and soil, all of it translated into protein first.
It returned 76 resistance genes nobody had catalogued, all of them members of that same family. Thirteen of these were assembled from the environmental samples, and every one of these genes shared under 70 percent of its letters with any resistance gene already known. The team synthesized 21 of these and put them into harmless laboratory E. coli. Eighteen destroyed imipenem, an antibiotic held in reserve for infections nothing else will treat. It is the same class of antibiotic that failed in the Nevada case.
Those genes had been sitting in public data the whole time. What was missing was a way to recognize them. A gene can drift until almost nothing about it looks familiar. It cannot drift out of the family it came from.
So far, we have looked at two of the ways scientists identify resistance genes: by matching them against a catalogue, and by recognizing the family they belong to. Some resistance genes, however, pose greater risks than others. In my next OpenThink blog post, I will look at what sets the riskier genes apart.
Photo by 21zere
