Artificial intelligence has made it possible to design viruses capable of infecting bacteria and multiplying. Of nearly 300 genomes created and then tested, 16 produced functional viruses. Several even proved more effective than their natural model in certain experiments on Escherichia coli bacteria.
These viruses are bacteriophages, often simply called phages. They do not attack human cells: their target is a specific bacterium. When a phage infects its target, it uses the bacterium’s machinery to produce new viral particles. The bacterium generally ends up destroyed, which explains the medical interest in these viruses for treating certain bacterial infections.
Representation of a phage attached to a bacterium.
© Adrien Bernheim
For this experiment, Samuel King and his colleagues used Evo 1 and Evo 2, artificial intelligence models trained on genetic sequences. Their operation is reminiscent of models that predict the continuation of a text. Here, however, the elements being manipulated are the letters of DNA. The objective was to produce complete genomes with an organization compatible with that of a functional virus.
The starting point was ΦX174, a small, well-known natural phage that infects E. coli. The models generated numerous variants of its genome. The scientists then selected candidates to manufacture and experimentally verify whether they could actually form viruses capable of infecting bacteria.
Sixteen candidates passed this stage.
Their interest does not come solely from the fact that they work. Several possess genetic sequences that are markedly different from those of the natural phage used as a reference. Observation of one of them under an electron microscope also revealed the use of a DNA-packaging protein very different from the one found in ΦX174. The AI therefore did not simply reproduce the original virus.
The scientists then compared some of the generated phages with their natural model. Several multiplied more efficiently in competition experiments or destroyed bacteria more quickly. A mixture of AI-designed phages also successfully attacked three strains of E. coli that had become resistant to ΦX174. This resistance is a common problem: a bacterium can evolve and no longer be vulnerable to the virus that previously infected it.
The envisaged interest concerns phage therapy in particular, which involves using phages to fight pathogenic bacteria. This approach is being studied as a possible complement to antibiotics, especially when bacteria become resistant to drugs.
The next question will be whether this method can be extended to other phage families and to bacterial targets of medical relevance. The authors primarily present their work as a demonstration: a generative model can now propose an entire viral genome and then produce, among its creations, genuinely functional viruses after laboratory validation.