Science1 publisher2 min readPublished
A Wisconsin lab's AI engineers phages to infect bacteria up to a million times more effectively
Vatsan Raman's lab at UW-Madison trained an AI that engineered phages to infect bacteria up to a million times more effectively in lab tests. The Cell Systems work aims at phage therapies that could replace failing antibiotics.
The Scientist · Science desk
What happened
- Vatsan Raman's lab at the University of Wisconsin-Madison built an AI model that learns how phages evolve and proposes mutations to make them better at infecting target bacteria, published in Cell Systems.
- The team also trained the model to find mutations that hit one bacterial species while sparing others, a step toward therapies that would not wipe out beneficial gut microbes.
- The model was trained on the lab's own scans of tens of thousands of mutations in a phage protein that governs how a phage recognizes and infects a host.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- capability Engineering can push a phage past the restraint evolution built in, because a designed phage does not need to keep its host population alive.
- constraint The million-fold figure is lab infection efficiency for the best mutations found; it does not show that an engineered phage cures an infection in an animal or a person.
- precedent Species-selective mutations point toward narrow-spectrum phage therapy that spares the gut microbes broad antibiotics kill.
A phage that kills its host too thoroughly loses the cells it needs to reproduce, so in the wild it holds back. Raman's account is that natural phages evolved for mediocrity, not maximum lethality. [4] Bacteria, for their part, have spent millions of years building defenses to evade the phages that hunt them. [5]
To get past that limit, the model learns the patterns by which phage proteins gain or lose the ability to recognize and infect a host, then proposes mutations a natural phage is unlikely to reach on its own. [6] The training data came from the lab's own high-throughput experiments, tens of thousands of mutations scanned in a phage protein that controls host recognition. [9] Related work on a second protein, one the team describes as a molecular drill that punches through the sugar coating around certain bacteria, mapped how its amino acids let a phage break in. [10]
The lab then built phages carrying the AI's suggested mutations and ran them against their bacterial hosts. The strongest infected up to six orders of magnitude more effectively than the natural versions. [7] That is a factor of a million. [8]
The million-fold figure is the best result among the engineered mutations, and it measures infection efficiency in the laboratory. [7][15] The reported evaluation stayed in the lab. The paper does not describe tests in animals or people. [15]
The team also trained the model to find mutations that attack one bacterial species while leaving others alone. [11] A phage that selective would spare the beneficial gut bacteria that broad antibiotics wipe out along with the target. [11]
Antibiotics are still the frontline treatment for infections such as strep throat and urinary tract infections, but bacteria develop resistance faster than new drugs arrive. [14] Phage therapy is one of the alternatives under study. [3]
"The model can learn the rules by which phages evolve to be successful and can use those rules to engineer phages that are highly effective against pathogens," Raman said. [12] He framed the goal as practical, and said the work has begun: the lab wants to design phages to address real problems, and it has already started moving in that direction. [13]
What to watch
- Whether the engineered phages clear infections in animal models, which the Cell Systems paper does not report.
- Whether target bacteria evolve resistance to the AI-designed phages, given their existing anti-phage defenses.
- Which real pathogen the lab picks as its first named target now that it has begun applying the model.